# Paying Attention: The State of Attention Measurement in 2026

*Attention measurement now has a final standard, one accredited attention product and a documented trade, mostly as a bidding input, whose scale is not disclosed. What it lacks, in the public record screened, is evidence that its numbers predict outcomes better than time in view, and any independent audit of whether its guarantees are met.*

Evgeny Popov · No Fluff Advisory · Research paper · October 2026 · Research cutoff 1 October 2026

## Abstract


*What attention measurement measures in 2026, who checks it, whether it predicts anything that time in view does not, and how it is traded. Research current to the 1 October 2026 cutoff, with status-sensitive facts as read on 28 to 30 September and the United States as the main market.*

> **In short.**
> Attention measurement now has a final industry guideline, one accredited product and a documented trade whose scale is not disclosed. Almost every campaign number it sells is a model estimate. As of 28 September 2026, MRC listed one accredited attention product, in three environments and not in CTV; the standards the accreditor's letter lists do not include the attention guidelines, and whether they informed the audit is not public. In the sources screened, no public study shows that attention predicts memory, brand or sales outcomes better than time in view on held-out data, and no test shows that buying on it pays back once costs are counted. Attention is marketed as a bidding input and, in a few offers, used to guarantee or price media. Where a guarantee says who measures delivery, it is the metric's own vendor, and no independent audit of delivery was found. No metric found meets the conditions of a trading currency. The AdCP protocol can represent a named vendor's attention score; beyond that schema, only test-mode handling with a placeholder vendor is documented, and no production use of a named vendor's attention signal in agent-driven buying was found. On the Focus Quotient, no product reaches the Most documented band.

> **What a buyer can do with this.**
> Use attention as a diagnostic of media quality, not as a settlement currency, until four things are in writing: the accreditor’s record for the environment you buy, a held-out test against time in view for the outcome you buy for, the guarantee’s threshold and remedy, and an audit of delivery by someone other than the vendor. [Chapter 12](#next) turns this into a checklist. This paper audits what vendors document in public by the cutoff; it does not test whether buying on attention works.

### The question

Attention metrics claim to measure whether a person looked at or listened to an ad, and they are sold as the step beyond viewability. The author's own playbook argues the category has to be created and then defaulted by brand committees; his 2026 essays argue it is a diagnostic with no settlement infrastructure, and the first metric to lose meaning when agents buy media. This paper tests both positions on the public record, without taking either side in advance. It asks what the products measure, who has checked them, whether they predict outcomes beyond time in view, how attention is traded and bought, where it works, how privacy law treats its signals, and whether agent-driven buying carries it.

### How it was done

Six hypotheses and four coded questions were registered with their decision rules before any research began, alongside a product register and scoring rules with fixed anchors. AI research agents then ran 9 workstreams, 9 vendor-universe and product sweeps and 8 company profiles from one brief. A separate AI verifier, told to refute, checked each file; eight workstreams were verified twice. In all, 985 sources were opened and logged, and the ledger holds 584 claims, each traced to a source. Four outside reviews by another AI model (of the briefs, of the evidence, of the full draft and of the built paper before publication) and two internal review passes of the full draft tested the work, and a separate AI consistency audit compared every score against the anchors. The corrections ledger holds 260 entries, 23 of them scoring corrections. 20 products are scored against 12 dimensions, less creative diagnostics for media attention measurement and activation for creative attention testing. This is AI-assisted research with a published process. It has not been audited or peer reviewed, and no vendor saw it before publication. The author works at Samba, whose tune-in data Adelaide, a scored and profiled vendor, names among the inputs that train its CTV model [106]. He co-leads AgenticAdvertising.org's Signals & Measurement working group. His own firm, No Fluff Advisory, runs two agents in that organisation's registry, and one of them passed the test-mode storyboard reported in [chapter 10](#agents) [33, 572]. The full statement of interests, including past employers, is in [chapter 13](#method).

### What it found

1. **Almost every number sold is a model estimate.** The campaign figures behind Adelaide AU, Lumen's platform and IAS Quality Attention are model estimates [10, 63, 73, 74, 9]. DoubleVerify reports indices built from device signals and, since April 2026, also accredited metrics from a modelled methodology [43, 44, 564, 45, 73, 433] [47, 209]. In CTV, TVision is the only scored media measurement product whose data come from observing people, in opt-in panel homes projected to the US population [45, 106, 220, 110]. Vendor units differ in construct and basis and cannot be compared [43, 44, 564, 45, 73, 433]. [Chapter 1](#defining).
2. **The standard is final; the audit trail is thin.** The IAB and MRC Attention Measurement Guidelines reached Version 1.0 in November 2025 [41, 430, 425, 563]. As of 28 September 2026, MRC listed one accredited attention product, DoubleVerify's Authentic Attention, for desktop, mobile web and mobile in-app, not CTV [2] [42, 205]. MRC first accredited it in December 2022 and continued and extended it on 13 April 2026 [215, 433] [42, 205]. H3, the registered hypothesis that accreditation is granted against the guidelines, is contradicted at 28 September 2026 on a narrow basis: the standards the April letter lists do not include them [42, 2, 425, 563]. The letter's stated basis also includes DoubleVerify's answers to MRC's questionnaire, so whether the guidelines informed the audit is not public [42, 2, 425, 563]. MRC removed IAS's attention measurement from in-process status in January 2026 [49, 339]; on 28 September 2026 Adelaide's AU had a pre-audit review, with the audit not begun [48, 339]. [Chapter 2](#standards).
3. **Predictive value beyond time in view is not shown.** In the sources screened, no public study compares an attention metric with time in view on held-out data; the closest comparisons are in-sample and not directly comparable with one another [46, 51, 52]. In a creative test run by the ARF, in which twelve companies scored ads from a common set, pairs of eye-tracking-only companies barely agreed, and no method group's scores tracked advertisers' own success ratings above R-squared 0.12 [4]. No randomised or holdout test of buying on attention with costs was found, in independent, academic or vendor sources [51, 4, 5] [69, 65, 11, 10]. The evidence is missing, not negative. [Chapter 3](#validation).
4. **Most outcome evidence comes from the sellers.** No commissioned or vendor study screened makes a held-out comparison against time in view [8, 10, 71, 64]. PwC's review, commissioned by Lumen, found attention correlated better than viewability with click-through and conversion. That comparison is in-sample, on a Lumen-supplied subset whose data cleaning PwC could not verify [8] [8, 10, 71, 64]. In a brand-lift pilot for one advertiser, attention was not significantly better than viewability on awareness or recall [8]. [Chapter 4](#vendor-evidence).
5. **Both verification companies left or agreed to leave public markets.** Novacap took IAS private on 23 December 2025 [83, 437], and on 6 August 2026 DoubleVerify agreed to be acquired by the parent of Nielsen for $13.60 a share, a deal still pending on 29 September 2026 [465, 84, 436, 221] [181, 465, 565]. Viant, a buying platform, completed its purchase of TVision on 1 May 2026 [82]. No source screened by 28 September 2026 links verification companies' attention bundling to specialists' demand, pricing or revenue, so H4 is unresolved [47, 231, 285, 232]. [Chapter 5](#market).
6. **Attention is traded at an undocumented scale, and no independent audit of delivery was found.** It is marketed as a live bidding input in named buying platforms, on vendor and seller statements [91, 92, 334, 93, 25]. Where a guarantee says who measures delivery, it is the metric's own vendor, and this review found no independent audit of delivery [94, 93, 91, 66]. Two offers price media on attention, and neither documents US sales [392, 96, 566]. No offer discloses US spend or impressions against a market total, so H2 is supported with its scale undocumented [91, 92, 334, 93, 25]. No metric-environment pair found meets all four currency conditions [63, 222]. [Chapter 6](#trading).
7. **Buyers use it to plan.** Most buyer-use records found involve planning [484, 476]. In IAB's US buy-side surveys, the share expecting to focus more on attention rose from 36% to 47% and then held at 47% in the next wave; that is intent, not spend [103, 102]. [Chapter 7](#buyers).
8. **Coverage follows the screen.** Under the registered rule, display, online video, CTV and social are covered, and social has platform integrations on Snap and TikTok [105, 119, 42, 45, 115, 110, 113] [116, 118, 117]. In the sources screened, no attention product is in Meta's or YouTube's partner programmes or has a certified integration with them; xpln.ai states that it measures Facebook and Instagram through the platforms' reporting interfaces or APIs [42, 568, 569, 570, 109] [571, 121, 496]. Audio, DOOH and retail media are thin, and in-game is study only [105, 119, 42, 45, 115, 110, 113] [113, 29]. [Chapter 8](#channels).
9. **The law turns on identification and emotion.** Emotion recognition for advertising is high-risk under the EU AI Act, subject to the Article 6(3) derogation [30, 124]. The obligations for such systems were moved to 2 December 2027, and they reach products already on the market only after a significant design change [30, 123]. [Chapter 9](#privacy).
10. **Agents can hold the number; no production use was found.** AdCP v3.1.24, the stable release of 23 September 2026, can represent a vendor-bound attention metric in forecasts, optimisation goals, delivery reports and performance floors [567] [144, 138, 140, 142]. This review found no documented production use of a named vendor's attention signal in agent-driven buying; beyond schema representation, the only handling documented is in test mode with a placeholder vendor, so H5, the agentic hypothesis, is unresolved [142, 141, 151, 33]. [Chapter 10](#agents).
11. **On the Focus Quotient, the field is close and no product reaches the Most documented band.** The field FQ is 51 in cohort A (media attention measurement) and 51 in cohort B (creative attention testing). Outcome validation and independent assurance are the two lowest-scoring dimensions in cohort A and, with production scale, the three lowest in cohort B. [Chapter 11](#fq).

### What it means

On this paper's reading, attention is a diagnostic of media quality that is being sold, in places, as a trading unit before the evidence and the plumbing of a trading unit exist. A buyer who treats it as a billing unit or guarantee should ask for the accreditor's record for the environment bought, a held-out test against time in view for the outcome that matters, the threshold and remedy in writing, and an audit by someone other than the vendor. [Chapter 12](#next) turns that into a checklist.



## 1. What attention measurement measures

*Attention measurement estimates whether and how much a person looked at or listened to an ad. In 2026 almost every attention number sold for a live campaign is a model estimate or an index of device signals, and people are observed only in panels and creative tests. Vendor units measure different things and do not convert into one another.*

> **In short.**
> Attention sits between the opportunity to see or hear an ad and what the ad does to memory, brands and sales. The Interactive Advertising Bureau (IAB) and Media Rating Council (MRC) guidelines define it as whether an ad "was seen or heard by a person and to what depth", and say it should not be used as a measure of outcomes [41] [41, 425]. They define no common unit [41]. Vendor units differ in construct and basis, and none is comparable across vendors [43, 44, 564, 45, 73, 433]. In connected TV (CTV), of the media products this paper scores, only TVision's attention data come from direct observation of people, in its in-home panel [45, 106, 220, 110]. Lumen's and Adelaide's campaign numbers are model estimates [282, 281] [43, 106, 1]. DoubleVerify's indices, which MRC's April 2026 letter lists as accredited, are built from device signals [42] [47, 209]. The people observed are panellists and creative-test participants.

### The definitions

The audio and gaming papers in this series cite these terms as defined here.

- Attention measurement: estimating whether and how much a person looked at or listened to an ad.
- Four layers, kept apart. Layer 1 is the opportunity to see or hear: the ad was served and was viewable or audible. Layer 2 is attention. Layer 3 is memory or brand response, such as recall or brand lift. Layer 4 is behaviour or sales. A number from one layer says nothing about the next until a study links them ([chapter 3](#validation)).
- Construct: visual attention (looking at the ad) or auditory attention (listening to it), each judged against its own methods.
- Candidate proxy: dwell, time in view, audibility and screen coverage. They stay proxies until validated for the stated construct.
- Method families: reference panels (eye tracking or webcam for visual attention, and the audio equivalent), device or on-screen signals, modelled attention, and neuro or biometric studies such as EEG (electroencephalography) and facial coding.
- Units: attention seconds; attentive seconds per 1,000 impressions (APM); vendor-defined attention units, such as Adelaide's AU on a 0–100 scale; and percentage attended. They are not interchangeable.
- Measurement basis: direct observation of people, or a model estimate, recorded with what the model predicts and over what horizon.
- Application stage: creative pretest, media planning or pre-bid, in-flight, or post-campaign.

![Almost every attention number sold is a model estimate or device-signal index built on the layer below it.](figures/fig-four-layers.svg)

*Figure fig-four-layers. **Almost every attention number sold is a model estimate or device-signal index built on the layer below it.** The four layers this paper keeps apart, and where the numbers in the market sit. Campaign-level attention figures from the scored media products are model estimates or device-signal indices built from opportunity signals; people are observed in reference panels used to train those models and in one in-home TV panel. No study in the sources screened tests an attention metric against time in view on held-out data. Units: none (diagram). Scope: products in cohort A of the product register, at the cutoff. Sources [41], [42], [43], [44], [45], [46], [47]. Caveat: A schematic of the paper’s definitions, not a measurement. Data: data/product-register.json*

Cohort A products measure, plan or optimise attention on media exposures. Cohort B products test creative before launch. No model output is called observed attention here, whatever the vendor calls it.

### Where the layers separate

The IAB and MRC Attention Measurement Guidelines, Version 1.0 of November 2025 [41, 430, 425, 563], draw the first boundary the same way. Viewability is the opportunity to see the ad, or to hear it for audibility. Attention is "whether the ad was seen or heard by a person and to what depth". MRC calls attention a measure of audience, meaning persons, because persons generate attention [41]. The guidelines add that attention "should not be considered or used as a measure of outcomes" [41, 425].

The guidelines use the word more widely, for "the exposure, engagement, focus, and cognitive impact measured through these methods" [41]. This paper keeps viewability, time in view and audibility in layer 1, because the data show the layers coming apart. In a German lab study of a mock Instagram feed, viewport time correlated at .89 with eye-tracked gaze duration [46]. Under the MRC one-second rule, 94% of viewable post exposures received at least a second of gaze, but 41% of the small group of non-viewable exposures also received more than a second [46]. TVision's 2016–17 panel showed viewers absent from the room during 30% of ads played in active viewing sessions [52].

Two scores cross the boundary between layer 2 and layers 3 and 4 by design. Adelaide defines AU as "a placement's probability of attention and subsequent impact", produced by a model trained on media-quality signals and full-funnel outcome data [43, 106, 1]. IAS says its Quality Attention combination "can predict if an impression is likely to lead to a business result" [73, 238], and its report-level definition describes the score as click or conversion propensity [237, 236, 115]. Both are outcome predictions with attention inputs. This paper's inference follows: a comparison showing that high scores go with better outcomes can partly restate what the model was trained to predict.

### Seen or heard

Even visual attention is inferred. Eye trackers report eye movements and gaze, and attention is always inferred from them, as Orquin and Holmqvist put it [419, 55]. TVision defines attention as the time a viewer is looking at the TV screen. Independent academics who used its 2016–17 panel data describe that measure as the co-occurrence of eyes-open and eyes-on-screen inferences, and call it "only one possible measure of attention" [45, 52].

A construct can drift inside a single product. Kantar's passive attention is the time viewers' eyes are on screen while the ad is visible. Its active attention is the time their facial expressions indicate emotional engagement [249]. Kantar's own publications report that the two barely relate at ad level [259, 258, 249]. This paper codes active attention as an emotion measure and a candidate proxy for attention.

Auditory attention is the thinner construct. In the sources screened, the three vendors that attach attention metrics to audio use three different constructs. None is documented as observing a listener's auditory attention to an ad [107, 105, 120, 112]. Dentsu's 2023 audio research with Lumen (funding not stated) reported an average of 10,126 attentive seconds per 1,000 impressions, a modelled visual-equivalent score [112]. DoubleVerify is not one of the three. Its accredited statistics, for display and video, include quartile audible rates and volume-change rates [42]. They record what the device did, so this paper codes them as candidate proxies for listening. [Chapter 8](#channels) covers audio.

### Four method families

The guidelines list a similar set (data signals, visual tracking, physiological and neurological observation, panel and survey methods), plus predictive models that combine them to predict visual attention from logged signals. They do not say which is best [41].

Reference panels observe people. Per its privacy policy, Amplified recruits a paid panel through Cint and collects facial images from device cameras and an in-home TV micro-camera [184]. TVision reports "an opt-in panel of 5,000 homes across the United States" [404, 298, 405, 24, 52, 292, 415]. Webcam accuracy is limited. An early open-source webcam system showed about 3.94 degrees of offset in the lab; a 2023 lab comparison co-authored by developers of Labvanced, a commercial webcam eye-tracking system, reported 1.4 degrees against an EyeLink 1000. Neither tested a commercial ad-attention panel [58, 59].

Device and on-screen signals record what the screen and the device did. DoubleVerify's Authentic Attention is built from signals such as viewable time, share of screen, audibility, touches and playback, and its filing names no reference method [47, 209].

Modelled attention turns those signals into expected attention with a model trained on panel data. Lumen's tag collects domain, format, page geometry, scroll speed, % viewable and time in view, and the model estimates attentive seconds per impression [282, 281]. A Lumen-commissioned PwC review found that the 2022 view and dwell models were a logistic regression and a quasi-Poisson model on ad size, log viewable time and domain terms [8]. It follows that the estimate depends on the same layer 1 signal it is often compared with.

Neuro and biometric methods measure the body. Brainsights, which describes itself as a neuroanalytics company using EEG, names no attention construct on the page for its creative service [420, 421, 422].

The families do not agree by default. In the Advertising Research Foundation's (ARF) Phase 2 test of creative measurement (June 2024), 12 companies scored ads from a common set of 32. Two pairs of eye-tracking-only companies agreed at R-squared 0.08 and 0.005 (0.0005 in the report's appendix). Two eye-tracking-plus-facial-coding companies agreed at 0.78, two survey companies at 0.66 (0.61 in the figure label) and two neuro-based companies at 0.10 [4]. The ARF is funded by its members, who include attention vendors [423, 5].

### Units that do not convert

The guidelines define no common unit. A text search of Version 1.0 found no definition of attention seconds, APM, a 0–100 unit or percentage attended. The guidelines list per-measure units and encourage reporting attentive time relative to viewable time [41].

- Attention seconds. Amplified reports active and passive attention seconds [43, 44, 564, 45, 73, 433].
- APM. Lumen defines it as % viewed × average eyes-on dwell time × 1,000, but its February 2023 definitions page omits the × 1,000 [392, 44].
- Vendor units. Adelaide's AU runs from 0 to 100 [43, 106, 1]. DoubleVerify's indices are normalised to 100 as the average across its own impressions over a 28-day rolling window, so 125 means 25% above that moving benchmark [71, 211, 370]. In the sources screened, IAS publishes no scale definition; one IAS case splits scores into 1–57 (low) and 58+ (high) [11, 228, 235].
- Percentage attended. TVision's Ad Attention is the share of ad impressions in which the viewer looked at the screen for two or more seconds [74].

Adelaide reports that connected TV (CTV) placements average 67 AU, a vendor model output with no stated denominator or period [106]. None of these units is comparable across vendors [43, 44, 564, 45, 73, 433]. The problem reaches inside single studies. The News UK and Lumen release gives its baseline as "25 minutes per 1,000 impressions" in one sentence and "25 seconds" in another, so only its percentage can be cited [402, 62].

### Basis and stage

Every attention number in this paper carries its measurement basis and application stage. The guidelines ask for the same split. Observed measurement "should be distinguished" from modelled measurement, models "must have empirical support" that they fit their purpose, and model validation should run at least annually on fresh deterministic data [41].

Vendor wording does not always hold the line. Amplified contrasts its "measured behaviour" with competitors' "modelled probability". Yet it told the ARF in 2023 that tag data are combined with panel ground truth to predict in-flight attention [187, 186, 424].

The stage can change the basis even when the source is observation. TVision's scores in demand-side platforms (DSPs) are "based on historic viewer attention" at daypart and channel level. OpenX's targeting uses "predictive attention models" built from TVision panel data [407, 315, 296, 45]. Adelaide's pre-bid segments in Amazon DSP are built from historical placement-level AU [162]. A pre-bid number is a forecast about an impression that has not yet happened. The target matters too. IAS's is click or conversion propensity [237, 236, 115], which is layer 4 behaviour.

### What each product observes

Q2 asks what each product measures. Its registered rule codes a number as direct observation only when the source documents that it was observed on people for the impression or asset reported; anything else is a model estimate. Fields the public record does not state are coded "not disclosed", which is itself reported under method disclosure. The table applies the rule to the profiled products, from their verified metric cards.

| Product | Construct | Observed in production | Reference labels | Basis | Stages | Evidence |
|---|---|---|---|---|---|---|
| Adelaide AU (cohort A) | 0–100 score of a placement's probability of attention and subsequent impact | Tag, pixel and page signals about the placement; no gaze or presence of the exposed audience | Gaze data licensed from five suppliers (Adelaide's September 2026 FAQ says Lumen's were used in early models); brand-lift, tune-in and sales outcomes | Model estimate at every stage | Planning, pre-bid, in-flight, post-campaign | [43, 106, 1] [169, 327] [168] [106] [1] |
| Lumen LAMP (cohort A) | Visual attention: APM, % viewed, view time | Tag signals: domain, format, page geometry, scroll speed, % viewable, time in view | Webcam eye tracking of opt-in panellists; TVision data for US CTV (2023–24 releases) | Model estimate for every campaign number | Planning, pre-bid, in-flight, post-campaign | [282, 281] [271, 392, 115, 287] [392, 427, 20, 428] [290, 289, 295] |
| TVision (cohort A) | Time a viewer looks at the TV screen; presence and co-viewing | Optical sensors and audio matching in opt-in US panel homes | Not disclosed | Direct observation of panellists, classified by a model; market figures are weighted projections | Planning, pre-bid (historical aggregates), in-flight, post-campaign | [45, 52] [407, 315, 296, 45] [52] |
| Amplified Campaign Performance (cohort A) | Active and passive attention seconds per ad view | Tag or server-to-server signals per exposure; no gaze of the exposed audience | Paid camera panel: device cameras, in-home TV micro-camera | Model estimate: in-flight predictions trained on panel data | In-flight, post-campaign; planning in Campaign Strategy | [186, 350] [184] [187, 186, 424] [119] |
| DoubleVerify Authentic Attention (cohort A) | Index of exposure and engagement | Tag device and on-screen events: viewable time, share of screen, audibility, touches, playback, volume | Not disclosed for the open web; Lumen on Snapchat; TVision for CTV (announced 2023) | Device-signal index (candidate proxy); model estimate for Authentic Modeled metrics and pre-bid segments | Pre-bid, in-flight, post-campaign, planning | [47, 209] [42, 205] [116, 117] [220, 426, 364, 219, 82] [222, 213] |
| IAS Quality Attention (cohort A) | Composite score framed as predicting a business result | Tag signals: visibility, situation, interaction | Lumen eye-tracking data; click and conversion events | Model estimate at every stage; target: click or conversion propensity | Post-bid reporting, pre-bid optimisation (beta), sell-side packaging | [73, 238] [237, 236, 115] [115, 232, 226, 231] |
| Kantar LINK+ (cohort B) | Passive attention (eyes on screen); active attention (facial expression of emotion) | Webcam facial coding with gaze monitoring (an optional LINK+ module) of survey participants; the survey base is N=150 per study, and the base for the attention measures is not stated | Not disclosed for the gaze component | Direct observation of participants; model estimate for AI predictive eye tracking and LINK AI | Creative pretest | [249] [263, 246] [247, 391] |
| Vevo Attention Guaranteed (trading offer) | Adelaide AU, by clear implication | Not explained for Vevo's inventory | As for Adelaide AU | Model estimate | Campaign guarantee; reporting | [78] [78, 94, 48] |

Adelaide's tune-in outcome data come from providers that include Samba, the author's employer [106].

For live media exposures, one product in the table reports numbers observed on people: TVision, for the exposures in its panel homes. Its figures for the wider market are projections of those observations, and partner scores built on its data are models [407, 315, 296, 45] [52]. In CTV, Adelaide, DoubleVerify and Lumen score impressions with models, and all three document TVision panel data as an input: Adelaide in its July 2026 method blog, DoubleVerify only in its June 2023 partnership announcement, and Lumen in 2023–24 partner releases [45, 106, 220, 110]. In the sources screened, whether that supply continues after Viant completed its acquisition of TVision on 1 May 2026 is not documented [45, 106, 220, 110] [82]. Adelaide's July 2026 blog says in the present tense that CTV AU "incorporates TVision data", which could describe historical training data [297, 311, 106].

Every other campaign number in the table is a model estimate or a device-signal index, and the pattern holds beyond it. The Snapchat attention products from Lumen and DoubleVerify, and the one IAS announced in June 2025 (live status unconfirmed), all rest on Lumen's eye-tracking model, so they are not independent measurements [116, 118, 117]. xpln.ai sells metrics built from eye-tracking panel data combined with machine learning to give real-time estimates [122, 41]. On MRC's list as read on 28 September 2026, it is the only attention product found in audit, recorded as in process as of November 2025 [48].

People are observed at two other points. Reference panels observe panellists whose data train the models. Lumen's platform model was trained on 650,000+ eye-tracking sessions [392, 427, 20, 428]. Creative tests observe recruited participants before launch: Kantar LINK+ with webcam facial coding and gaze monitoring [249], Lumen SPOTLIGHT with webcam eye tracking [279]. Their model-based siblings observe no one. Kantar's LINK AI attention outputs are model estimates [247, 391], and Amplified's AttentionAI simulates 20,000 viewing personas per ad [192, 429].

Among the fields coded "not disclosed" are the ones a buyer needs to judge a model: the size and recruitment of the reference panel, and held-out agreement with it. In the Adelaide documentation screened, neither a reference-panel size nor a held-out agreement statistic against gaze data is published [66, 168]. The IAS sources opened state neither for the Lumen data behind Quality Attention [237, 236, 115]. Amplified claims "99% data accuracy" without naming a validator, metric or reference set [186, 191, 187]. These gaps are scored under method disclosure in [chapter 11](#fq).



## 2. The standard and the audit


*Attention measurement now has a written standard from the Interactive Advertising Bureau (IAB) and the Media Rating Council (MRC), and one accredited product. MRC's accreditation of DoubleVerify's Authentic Attention predates the standard, and the standards its April 2026 renewal lists do not include it. By this paper's registered rule, that narrow basis is enough to code H3 contradicted on the records read in late September 2026. Whether the guidelines informed the audit is not public. One other product is in audit and another has had only a pre-audit review.*

> **In short.**
> The IAB and MRC Attention Measurement Guidelines became final as Version 1.0 in November 2025, after a public-comment draft in May [41, 430, 425, 563]. They build attention on viewable impressions, full filtering of sophisticated invalid traffic and a present user, and they define no common unit [41] [41]. MRC lists one accredited attention product, DoubleVerify's, for desktop, mobile web and mobile in-app [2]. Its CTV attention is not accredited [42, 2, 48]. The standards the April 2026 letter lists do not include the attention guidelines [42, 41]. Under the registered rule, H3 is therefore contradicted on the records read on 28-29 September 2026. The letter's stated basis also includes the procedures in DoubleVerify's answers to MRC's questionnaire, so whether the guidelines informed the audit is not public [42, 2, 425, 563]. XPLN is in audit, Adelaide has had a pre-audit review, and IAS left the process in January 2026 [48] [48, 339] [49, 339]. In this paper's reading, accreditation certifies how a number is produced and disclosed. Whether the number predicts outcomes is a separate question.

### What the guidelines require

The IAB and MRC published a public-comment draft of the Attention Measurement Guidelines on 12 May 2025, with comments due by 12 July. The final Version 1.0 followed in November 2025. IAB's web page for it is dated 11 November 2025, and MRC lists it among its guidelines as issued that month [41, 430, 425, 563]. IAB's page says the guidelines "serve as the basis for MRC accreditation audits". That sentence is on the web page and does not appear in the PDF's text layer. The PDF still frames several provisions in accreditation terms: base inputs must be audited for accreditation consideration, MRC audits will verify upstream processes, and users should encourage providers to be independently audited against the guidelines [425, 41].

The guidelines define viewability as the opportunity to see or hear an ad, and attention as "whether the ad was seen or heard by a person and to what depth". MRC treats attention as a measure of audience, meaning persons rather than devices [41]. The standard base is compliant viewable or audible impressions (one second for display, two continuous seconds for video), full filtration of sophisticated invalid traffic (SIVT), and the presence of a user. Panel or survey methods in which respondents affirm exposure may substitute for these checks, but only with "strong empirical support". The base inputs must themselves be audited for accreditation consideration [41].

Attention measured without or before viewability may also be reported, under four conditions. Among them, the standard viewable version must be reported beside it, and the non-viewable version is labelled non-standard and diagnostic. Both presentations, and aggregates that combine them, are eligible for accreditation with proper disclosure [41].

The guidelines' definitions draw the first boundary where this paper does, but their measurement categories cross it. Their Exposure category includes viewability, time in view and audibility, which [chapter 1](#defining) places in the first of four layers, the opportunity to see or hear [41].

The guidelines recognise four measurement approaches: data signals, visual tracking, physiological and neurological observation, and panel and survey methods. Predictive models, which combine them to predict visual attention from logged data signals, form a fifth. The text does not say which approach is best [41].

On models the text is specific. Observed measurement "should be distinguished" from modelled measurement. Models "must have empirical support" that they fit their intended purpose. That includes a link to validated attention and, where a model is used to predict outcomes, a link to campaign objectives and outcomes. Validation "should" run at least annually on fresh deterministic data, and training data "should" be collected deterministically. Error margins and standard errors around sample-based projections must be disclosed. Randomised or A/B validation is strongly encouraged and is not required [41].

The final text is firmer than the May draft. It adds rules on synthetic training data and the sentence "These guidelines strongly require validation for models and measurement approaches" [41, 430].

### What the guidelines leave open

The guidelines define no common unit. A text search of Version 1.0 found no definition of attention seconds, attentive seconds per 1,000 impressions (APM), a 0-100 attention unit or percentage attended. The text lists per-measure units and strongly encourages reporting attentive time relative to viewable time [41]. Each vendor's unit therefore stays its own. Adelaide's AU, Lumen's APM and DoubleVerify's indices differ in construct and basis, and none is comparable across vendors [43, 44, 564, 45, 73, 433].

On outcomes and trading the guidelines are cautious. They state that attention "should not be considered or used as a measure of outcomes". They only strongly encourage independent third-party auditing of attention methods used in buying and selling [41, 425]. The Attention Measurement Playbook for Marketers, published in November 2025 by IAB and CIMM, a measurement coalition, summarises IAB and MRC's position as "never as a stand-alone currency" [438].

The vendors whose products this chapter tracks helped write the text. Staff of DoubleVerify, Integral Ad Science (IAS), Adelaide, GumGum, Amplified Intelligence, Lumen Research and TVision are named key contributors to Version 1.0, alongside MRC's senior vice president for digital research and standards [41]. The list says nothing about influence on any clause.

The Advertising Research Foundation (ARF) runs an attention validation initiative as research. IAB Europe publishes guidance, and France's CESP publishes a grid of vendor answers and MRC statuses. None of them accredits [423, 5, 41] [431, 432, 425, 3].

### The accreditation record

MRC, which accredits measurement services [2], records each accreditation by service, metrics and environments [42, 205] [2]. The table reads its records for every attention product found. Status is as of 28-29 September 2026, when MRC's pages were read for this paper.

| Product | Metrics | Environments | MRC status | Standards MRC lists | Evidence |
|---|---|---|---|---|---|
| DoubleVerify Authentic Attention | Display and video Authentic Attention metrics | Desktop, mobile web, mobile in-app | Accredited December 2022; continued March 2024 and 13 April 2026 | Not listed in the 2022 report (the 2022 letter was not opened). April 2026: viewability, invalid-traffic, impression, audience-measurement, outcomes and data-quality, ad-verification; not the attention guidelines | [215, 433] [210] [42, 41] |
| DoubleVerify Authentic Attention V4 | Indexed Impressions, Direct-Indexed-Rate; metrics from the Authentic Non-viewable and Authentic Modeled methodologies | Desktop, mobile web, mobile in-app | Initial accreditation 13 April 2026 | As above | [42, 205] |
| DoubleVerify pre-bid Authentic Attention | Pre-bid attention | Desktop, mobile web, mobile in-app | Denied March 2024; no attention offering in the April 2026 pre-bid table | n/a | [210] [42, 2] |
| DoubleVerify CTV attention | Attention | CTV | Not accredited (CTV accredited only for impressions, video viewability, completion, quartiles and verification); no entry under review | n/a | [42, 2, 48] |
| IAS attention measurement | Attention Score and associated metrics that contribute to it | Not stated | On the pre-audit list by 16 September 2025; removed from in-process status January 2026 | n/a | [49] [233, 48] |
| Adelaide AU | AU | Digital display and video, audio, linear TV, cinema, digital out-of-home | Pre-audit reviewed 5/2026; audit not begun | Not stated | [48] |
| XPLN (xpln.ai) | Impressions, viewable impressions, viewability and attention metrics, SIVT-filtered | Mobile web, mobile in-app | Pre-audit reviewed 9/2025; audit in process as of 11/2025 | Not named | [48] |
| Lumen | None found | n/a | No entry on the MRC lists screened | n/a | [2, 48, 3] |
| TVision (owned by Viant since 1 May 2026) | None found | n/a | No entry on the MRC lists screened | n/a | [2, 302, 50, 48] [435, 82] |
| Amplified | None found | n/a | No entry on the MRC lists screened | n/a | [2, 48] |

![One attention product is accredited, in three environments and not in CTV.](figures/fig-accreditation.svg)

*Figure fig-accreditation. **One attention product is accredited, in three environments and not in CTV.** MRC status of attention metrics by product and environment on 28 September 2026, from MRC’s accredited and under-review lists, its quarterly decisions and its 13 April 2026 letter to DoubleVerify. An audit in process counts only when MRC confirms it; a vendor announcing one is a vendor assertion. Accreditation certifies process and disclosure, not that a metric predicts outcomes. Units: status category. Scope: MRC records for attention metrics, US. Sources [42], [2], [48], [49], [50]. Caveat: Status-sensitive: read on 28 September 2026 and not rechecked before publication (chapter 13). "No record found" is bounded to MRC’s digital, TV and multimedia lists screened that day. Data: data/figure-data.json#accreditation*

### DoubleVerify: accredited, within limits

DoubleVerify's Authentic Attention is the only accredited attention product in the MRC records read for this paper. On 28 September 2026 the "Attention Metrics" category of MRC's digital list has one entry, DoubleVerify Quality Analytics Platform, for desktop, mobile web and mobile in-app [2]. MRC first accredited the Authentic Attention metrics suite in those environments in December 2022, nearly three years before the guidelines were final [215, 433]. In March 2024 it continued that accreditation and denied accreditation for pre-bid Authentic Attention, giving no reason [210].

MRC's letter of 13 April 2026 continued the display and video metrics. It also granted initial accreditation to new reporting: Indexed Impressions and Direct-Indexed-Rate, and metrics built on DoubleVerify's Authentic Non-viewable and Authentic Modeled methodologies. MRC's July 2026 report labels this "Authentic Attention V4" [42, 205].

The accredited statistics are indices and device-event rates. They include Attention, Exposure, Engagement, Intensity, Prominence, User Presence and Ad Interaction indices; rates for ad touches, scrolling, tabbing into background and volume changes; and quartile audible and viewable rates [42]. DoubleVerify's 10-K describes the inputs as exposure and engagement signals and names no reference method [47, 209].

The letter's scope has limits. CTV is accredited for impressions, video viewability, completion and quartile metrics and property-level verification. CTV attention is not accredited, and on 29 September 2026 MRC's digital and television under-review lists carried no DoubleVerify entry [42, 2, 48]. The letter's pre-bid table lists brand safety, fraud, viewability and other offerings, with no attention offering [42, 2]. The standards it lists cover the whole Pinnacle service: viewability, invalid-traffic, impression, audience-measurement, outcomes and data-quality, and ad-verification standards. The Attention Measurement Guidelines are not among them. A DoubleVerify vice president of attention product management is a named key contributor to the guidelines, as are staff of IAS, Adelaide and other vendors [42, 41] [41].

DoubleVerify describes its own scope differently from MRC. Its product page says "only direct indexed impressions", a November 2024 release dates first accreditation to 2023, and the 10-K calls the product "MRC-accredited" without naming environments [209, 214, 47, 215]. DoubleVerify has carried that 10-K sentence since the report filed on 1 March 2023 [47, 207, 436]. For scope, the letter is the record.

### The rest of the field

IAS never reached audit. Digiday reported in October 2024 that IAS was in pre-assessment for attention [49, 339]. IAS Attention entered MRC's public pre-audit list between 20 June and 16 September 2025 and was gone by 20 January 2026 [233, 48]. In January 2026 MRC removed "IAS Attention Measurement" from in-process status, stating the reason as IAS' lack of intent to engage in third-party audit [49]. That is a withdrawal from the process. The service was never accredited, so nothing was denied or revoked. MRC names services, not products, so reading that service as IAS's Quality Attention is this paper's inference [49]. Novacap had completed its take-private of IAS on 23 December 2025 [83, 437, 244]. No source screened links the two events [49, 339].

IAS does hold MRC accreditation for viewability and SIVT and, since June 2026, for pre-bid viewability and invalid traffic. Quality Attention uses such signals as inputs, and no IAS entry carries an attention metric [2, 205].

Adelaide is in the MRC process and not in audit. MRC lists its AU (attention unit) with "Pre-audit reviewed 5/2026, Audit not begun", across digital display and video, audio, linear TV, cinema and digital out-of-home [48]. MRC told Digiday in October 2024 that Adelaide was in pre-assessment. The pre-audit review came about 19 months later [339, 48]. Under this paper's assurance anchors, a pre-audit review does not count as an audit in progress [48, 339]. Adelaide says it asked MediaSense, a media advisory firm, to audit AU in 2025. That review is not an accreditation [66, 22]. Vevo's Attention Guaranteed does not state the unit of its minimum. Reading it as AU is this paper's inference from the partnership, and on that reading the guarantee rests on an unaccredited metric [78, 94, 48] [48, 78]. [Chapter 6](#trading) covers its terms.

XPLN (xpln.ai) is the only attention product MRC lists as in audit. The record reads "Pre-audit reviewed 9/2025, Audit in process as of 11/2025". It covers display and video impressions, viewable impressions, viewability metrics and attention metrics, all with SIVT filtration, in mobile web and mobile in-app. It names no standards. Desktop, CTV and social, where xpln.ai says it also sells its metric, sit outside the audit [48]. XPLN was not on MRC's accredited digital list on 29 September 2026. In CESP's grid of 10 June 2026, it answered no to whether its attention solution had been audited by an independent third party [2, 3]. Adelaide, DoubleVerify and IAS answered yes, and of these only DoubleVerify holds an MRC attention accreditation [3] [2] [3].

For Lumen, TVision (owned by Viant since 1 May 2026 [435, 82]) and Amplified, this review found no entry on the MRC accredited or under-review lists screened by 29 September 2026 [2, 48, 3] [2, 302, 50, 48] [2, 48]. MRC's outside-US page showed no readable entries, so assurance outside the US is not established here [2].

### H3, coded by the rule

H3 states that by the cutoff MRC accredits attention metrics against the Attention Measurement Guidelines. The registered rule codes it contradicted when every attention accreditation found in MRC records lists standards that exclude the guidelines. Audits in progress confirmed by MRC are reported separately and do not change the verdict, which the rule dates to the cutoff. MRC's records, as read on 28-29 September 2026, hold one accredited attention product. Its continued accreditation and its initial V4 accreditation both rest on the April 2026 letter, and the standards that letter lists do not include the guidelines. On that narrow basis, H3 is contradicted on the records read on those dates [42, 2, 425, 563] [42, 205] [42, 41]. The letter's stated basis also includes the procedures described in DoubleVerify's answers to MRC's questionnaire. Whether the guidelines informed the audit is therefore not public, and only MRC could settle it [42, 2, 425, 563]. The status was not rechecked before the 1 October 2026 cutoff, so an MRC decision published after that reading is not reflected ([chapter 13](#method)).

The verdict rests on one product, and it can change. XPLN's audit, and Adelaide's if it begins, may apply the guidelines [42, 2, 425, 563]. Why the DoubleVerify letter does not list them is not stated in the records screened. One possibility is that the DoubleVerify audit cycle began before November 2025. That is inference and unverified [42, 2, 425, 563]. In one respect the V4 scope is consistent with the guideline text. The guidelines make attention measured without viewability eligible for accreditation under four conditions and with proper disclosure, and V4 accredits metrics built on DoubleVerify's Non-viewable methodology [41] [42, 205]. Neither the letter nor MRC's July 2026 report describes that methodology, so whether it meets the four conditions cannot be told from the record [42, 205]. The consistency is this paper's inference. The letter does not draw it.

### What accreditation certifies

An MRC accreditation covers named metrics in named environments, against the standards the letter lists [42, 205]. The guidelines point the audit at process. Base inputs are audited, upstream processes verified, models validated against deterministic attention data, and errors disclosed [425, 41] [41]. Outcome validation is required only for a model used to predict outcomes [41]. The DoubleVerify letter does not define the accredited metrics or describe the Modeled and Non-viewable methodologies [42] [42, 205].

In this paper's reading, accreditation therefore certifies how a number is produced, filtered and disclosed. It does not show that the number predicts memory, brand or sales outcomes. That is the question for [chapter 3](#validation). The ARF's Phase 3 release (June 2026), covering four live national campaigns, reports "no simple one-to-one relationship between attention and upper-funnel brand lift". It states no funding, so its independence is unknown [5, 434].

The distinction matters most for scores sold as outcome predictors. Adelaide describes AU as "a placement's probability of attention and subsequent impact", and IAS says Quality Attention predicts whether an impression is likely to lead to a business result [43, 44, 564, 45, 73, 433]. The guidelines say attention should not be used as a measure of outcomes, while allowing its use "in conjunction with measures of outcomes" [41, 73]. They also require outcome validation for any model used to predict outcomes [41]. MRC has not judged IAS's score against those clauses, because IAS left the process [41, 73]. Adelaide's audit has not begun [48].

Accreditation is also narrow. An accredited desktop and mobile index says nothing about the same vendor's CTV attention, which DoubleVerify's pages describe without naming a data source [42, 2, 48] [220, 426, 364, 219, 82]. Accredited viewability and SIVT inputs do not make the score built on them accredited [2, 205]. A vendor's own claim of audit is not accreditation either. Lumen markets its technology as "the only solution validated and audited by an independent third-party in the industry". It cites a 2023 PwC report that Lumen requested, in which PwC says it "has not independently verified, validated, or audited the data" [20, 8]. Amplified's pages call the guidelines "Validation from the MRC and IAB", although the guidelines validate no vendor [41, 355, 193].

A buyer who wants assurance can ask for the MRC letter, check that its environments match the buy and read which standards it lists. [Chapter 11](#fq) scores independent assurance on that scope, and [chapter 12](#next) turns it into questions.



## 3. Does attention predict outcomes?


*The public record cannot yet say. In the sources screened, no study tests an attention metric against time in view on held-out data, and no test of buying on attention reports its cost per incremental memory, brand or sales outcome. Both hypotheses are unresolved for memory, brand and sales, because no screened source reports the tests that would decide them.*

> **In short.**
> In the sources screened by the cutoff, this review found no study that compares an attention metric with a time-in-view baseline on held-out data [46, 51, 52]. The closest comparisons are in-sample and not directly comparable with one another. In a German lab study, viewport time and eye-tracked gaze had the same reported in-sample R-squared for brand recall, .027 for each single-predictor model [46]. In a US and UK experiment, eye-tracked dwell showed the larger association with recall per standard deviation, but the dwell and visibility results come from different sub-samples and are not directly comparable [51]. On TVision panel data, attention explained slightly more brand-search lift than tuning or presence [52]. In a creative test run by the Advertising Research Foundation (ARF), pairs of eye-tracking-only vendors scoring the same ads agreed at R-squared 0.08 and 0.005 (0.0005 in the report's appendix) [4]. In the sources screened, no independent, vendor, agency or media-owner test of buying on attention reports cost per incremental outcome [51, 4, 5] [69, 65, 11, 10]. H1a and H1b are unresolved for memory, brand and sales.

Two hypotheses carry the commercial case for attention. H1a says attention metrics add predictive value for memory, brand or sales outcomes beyond viewability and time in view, on data the model has not seen. H1b says buying media on an attention metric improves those outcomes at comparable cost. Both carry registered decision rules ([chapter 13](#method)). The Interactive Advertising Bureau (IAB) and Media Rating Council (MRC) guidelines build attention on a base of viewable or audible impressions [41], so what a buyer pays to learn is what attention adds to time in view.

This chapter's screen for independent and academic work covered the Journal of Advertising and the Journal of Advertising Research through Crossref title filters, ARF reports and logged web searches, but not the International Journal of Advertising [46, 51, 52]. This review also found two 2025 Journal of Advertising Research articles on attention in CTV and linear TV that draw on TVision panel data, but did not read them [300, 301]. Whether either compares attention with time in view on held-out data is not known, so every H1a code below is bounded by that gap. Vendor and commissioned studies are set out in [chapter 4](#vendor-evidence); where they bear on a verdict, this chapter says so.

![Most outcome evidence is published by the vendor that sells the metric.](figures/fig-evidence-ladder.svg)

*Figure fig-evidence-ladder. **Most outcome evidence is published by the vendor that sells the metric.** The 48 study records coded in workstreams B (independent and academic) and C (commissioned and vendor), by the outcome layer each measures and by who paid. A record that measures more than one layer is counted in each, so the bars sum to more than 48. Validity is the only layer where independent or academic records are the largest group; brand and sales evidence is mostly vendor-published. ARF Phase 3 is counted as funding unknown because its release states no funding (round 2). Units: count of study records. Scope: studies screened by the cutoff, any geography. Sources [46], [51], [52], [6], [4], [5], [53], [54], [55], [56], [57], [58], [59], [60], [61], [62], [21], [63], [13], [14], [15], [64], [65], [66], [67], [10], [8], [68], [69], [70], [71], [72], [11], [73], [9], [74], [17], [75], [76], [77], [78]. Caveat: Counts of records, not of independent studies: one study can yield several records, and syndicated copies were merged. Every count is recomputed in data/derived-checks.json. Data: data/figure-data.json#evidence_ladder*

### The closest comparisons are in-sample

The most direct test of a time-in-view proxy against observed gaze comes from a lab at the University of Wuppertal. Graduate students scrolled a mock Instagram feed on their own smartphones while wearing mobile eye trackers (30 assessed, four excluded; 806 post exposures to ads and content posts). Viewport time, a log of how long each post sat on screen, correlated .89 with gaze duration across all posts. In separate single-predictor models on ad exposures, viewport time predicted unaided brand recall (odds ratio 1.12 per second, p = .045) and so did gaze duration (odds ratio 1.14, p = .047), with the same reported in-sample R-squared, .027 for each single-predictor model; the number of ad exposures and the type of R-squared are not reported [46]. The MRC 1-second viewability rule caught gaze in 94% of viewable cases, but 41% of non-viewable cases also received more than a second of gaze [46]. For recall in a feed, the time-in-view proxy matched eye tracking on this one in-sample measure. That is not a test of equivalence: the models were fitted separately, with no formal increment test, and the explained variance was small for both [46].

A US and UK online experiment by Simonov, Valletti and Veiga recruited 1,013 participants in July and August 2020, paired articles with brands at random, and used webcam eye tracking supplied by Lumen [51]. Imperial College and the NET Institute funded it, and it kept only high-quality eye-tracking data, about 70% of observations [51]. One extra second of eye-tracked dwell went with a 3.43 percentage-point higher probability of aided recall; one extra second of MRC-standard visibility went with 0.32 points. The per-second gap misleads, because dwell has a standard deviation of 3.16 seconds and visibility 17.37. Per standard deviation, the effects are 10.83 and 5.55 points [51]. But the dwell and visibility results come from separate models on different sub-samples, 3,925 eye-tracked observations and 5,707, so they are not directly comparable [46, 51, 52]. Their R-squared values (.143 and .091) are full-model values with extensive fixed effects, so they are not an increment test either. The authors conclude that visibility still gives reliable results [51]. For a purchase proxy, a voucher choice made right after exposure, the per-second figures were 0.73 points for dwell and 0.13 for visibility; per standard deviation the effects are almost identical, 2.31 and 2.26 points, again from different sub-samples [51].

On TV, McGranaghan, Liaukonyte and Wilbur used TVision panel data from 1,155 paid, consenting US households between July 2016 and June 2017. Across 180 national ad slots from two NFL playoff games on 22 January 2017, eyes-on-screen attention explained slightly more brand-search lift (R-squared .116) than exposure tuning (.082), presence (.070) or opportunity-to-see tuning (.035), each in its own regression, with no hold-out. With quarter-hour fixed effects the gap nearly closes: .180 for attention against .176 for exposure tuning. Search lift is a behavioural proxy for brand interest; it measures neither memory nor sales [52]. Attention here is a classifier's reading of camera images, an eyes-open detection plus an eyes-on-screen inference from head orientation, with no reported accuracy against human-coded gaze [52].

All three studies are in-sample [46, 51, 52]. The two memory studies are not directly comparable: one is a lab study of unaided recall, the other an online experiment of aided recall with dwell and visibility modelled on different sub-samples [46] [51]. Under the registered rule, in-sample associations alone leave H1a unresolved.

### The industry's own validation work

The ARF runs its attention validation initiative as research; it sets no standard and grants no accreditation. Its first phase, in October 2023, found vendors' definitions of attention shaped by their own tools [423, 5, 41]. In Phase 2 (June 2024), 12 attention-measurement companies scored ads from a common set of 32 with their usual methods. Two pairs of eye-tracking-only companies agreed at R-squared 0.08 and 0.005 (0.0005 in the report's appendix). Two eye-tracking-plus-facial-coding companies agreed at 0.78, two survey companies at 0.66 (0.61 in the figure label) and two neuro-based companies at 0.10. No method group's scores correlated with advertisers' own success ratings above R-squared 0.12 [4]. The ARF says those ratings cannot be treated as ground truth, and only three of the 12 companies agreed to share cost information [4]. The test covered creative and anonymised the vendors, so no result attaches to a named product [4]. Inference: where agreement is this low, validating one vendor's score says little about another's.

Phase 3, released on 9 June 2026, measured four live national campaigns across TV, online video, social and display with multiple providers. The release reports "no simple one-to-one relationship between attention and upper-funnel brand lift", with methods agreeing mainly at channel level and diverging at platform and placement level [5, 434]. It adds that lower-attention channels sometimes delivered recall or awareness gains when they received enough media weight, reach and frequency [5]. The full report is member-only. The release gives no effect sizes, does not say whether time in view was a comparator, and states no funding or participation terms, so this paper codes it as the ARF's synthesis, with independence unknown [5, 434].

### Sales, and the studies that do not test it

Venkatraman and colleagues (2015) is the only screened study linking lab measures to TV advertising elasticities: 37 thirty-second TV ads, market-response data for 26 of them from five firms, and eye tracking on 29 participants. On their own, eye tracking and EEG (electroencephalography) were moderate predictors. Once traditional self-report measures were controlled, eye tracking added no significant explanatory power; only fMRI (functional brain imaging) did. The comparator was self-report and the exposure was forced in a lab, so the study does not test time in view [6]. The ARF and the corporate sponsors that supplied ads and sales data funded it [6]. Zhang, Wedel and Pieters (2009), read at abstract level, link gaze duration on grocery feature ads to sales beyond the mere presence of the ad; the abstract reports no comparison with time in view and no hold-out [55, 56]. A 2019 Journal of Advertising Research study of attention measures for TV creative was read only as an abstract and is not used [7].

Two often-cited items do not bear on H1a. The Kwon et al. 2021 meta-analysis, carried from the author's earlier papers, covers media-context effects on attitudes and purchase intention and, on its abstract, includes no attention metric; its full text was not obtainable [60]. A frequently cited neuro and eye-tracking study reports 82.9% test accuracy in classifying ads' YouTube view counts, but its random 70/30 split of cases across 8 Super Bowl ads puts the same ads in training and test sets, so on this review's reading the accuracy is not prediction for new ads [54].

Held-out validation can be done. Martinovici and colleagues (2023) validated attention trajectories out of sample and ahead of period on 325 US consumers, with an 85% brand-choice hit rate against 20% chance, and report that only one of 21 earlier attention-and-choice studies used a classic two-group validation. None of the advertising-attention outcome studies screened here validated on held-out ads, people or time periods, except the one whose split leaks ad identity [57, 46, 54].

### H1a, coded by outcome

Under the registered rule, H1a is supported only by a published design showing a positive increment over a time-in-view baseline on held-out data, outcome by outcome. It is unresolved when no such study is found, when only in-sample associations are published, or when comparable studies conflict. Each outcome is unresolved.

| Outcome | Closest academic and industry-body evidence | Baseline tested | Hold-out | Result | H1a |
|---|---|---|---|---|---|
| Memory | Bruns et al. 2025, lab, Germany [46]; Simonov et al. 2025, online, US and UK [51] | Viewport time; MRC-standard visibility | None | Same reported in-sample R-squared for viewport time and gaze (.027 for each single-predictor model); dwell 10.83 and visibility 5.55 points per standard deviation, from different sub-samples and not directly comparable | Unresolved: in-sample only |
| Brand | McGranaghan et al. 2021, TVision panel, US [52]; ARF Phase 3 release, independence unknown [5, 434] | Exposure tuning and presence; Phase 3 not stated | None (TVision); not stated (Phase 3) | Search-lift R-squared .116 against .082 (.180 against .176 with fixed effects); no simple one-to-one link with upper-funnel brand lift | Unresolved: in-sample only |
| Sales | Venkatraman et al. 2015, US [6]; Zhang, Wedel and Pieters 2009 [55, 56] | Self-report; ad presence | None | Eye tracking adds no significant explanatory power beyond self-report; gaze adds to presence for print feature ads | Unresolved: no time-in-view comparison |

The TVision and elasticity studies use US data, and the webcam experiment covers US and UK participants; none is held out [46, 51, 52].

The commissioned and vendor record does not change the codes. None of the commissioned or vendor studies screened makes a held-out comparison against time in view; several vendor models take viewable time as an input, and the one commissioned review with published statistics, PwC's for Lumen, found no significant memory advantage over viewability [8, 10, 71, 64]. On all the evidence screened, H1a is unresolved for memory, brand and sales. It is not contradicted either: no study made the held-out comparison and found no increment.

### No test of buying on attention with costs

In the sources screened by the cutoff, this review found no randomised or hold-out test, by independent or academic researchers, of buying media on an attention metric against a standard-metric control with costs reported [51, 4, 5]. The vendor, agency and media-owner sources screened contain no such test with a memory, brand or sales gain, cost per incremental outcome and uncertainty [69, 65, 11, 10]. The only cost figures in the independent studies are Lumen's own attentive-CPM (cost per 1,000 attentive views) estimates in the Simonov experiment, about 21.88 GBP on desktop and 13.54 GBP on mobile, which are not a tested buying outcome [51].

The nearest tests are vendor and agency material, and each lacks part of what the rule requires.

| Outcome | Nearest test found | What it reports | What it lacks for H1b |
|---|---|---|---|
| Memory | Amplified and Chalice trials, US and Australia (vendor release) [195] | Lucid measured a 70% ad-recall increase against the standard non-attention line item | Design, sample, uncertainty and cost |
| Brand | MAGNA Media Trials and Adlook, US (joint release, funding unknown) [342] [69, 13, 67, 15] | A +131% awareness "delta lift" for ads optimised pre-bid on AU and delivered contextually; lower display CPM and 42% lower cost per completed video for the AU pre-bid arm | An awareness result separated from contextual delivery, uncertainty and cost per incremental outcome |
| Brand | Adelaide and Lucid "structured A/B tests" (vendor page) [65] | Brand impact increases of up to 22% against viewability, from in-flight optimisation toward AU | Design, sample, cost and uncertainty; "up to" is a maximum |
| Sales | IAS and Kimberly-Clark, NCSolutions tests (vendor case) [11] | Test 1: no significant difference in sales lift; Test 2: 313% higher sales lift for above-average-attention households | A bought arm against a control; segments split after the fact on a modelled score |
| Attention only | Lumen consumer-goods attention bidding (vendor case) [69] | Against a control campaign on the same budget, +104% viewed ads at 40% lower CPM | Any memory, brand or sales outcome; the same vendor's model bought and measured |

Under the registered rule, H1b is unresolved for memory, brand and sales: the published tests lack a holdout or randomisation, omit cost or cost per incremental outcome, report no uncertainty, or are case studies without a design. No study screened discloses measurement fees, and none reports a campaign's media premium for attention-bought inventory [69, 13, 67, 15].

Inference: in the 2016-17 TVision panel, attention fell across the first slots of a break, so the earliest slots drew the most, and it fell slightly with ad duration. Optimising to a per-second attention rate could therefore favour early slots and shorter ads without an outcome gain, while a total-attentive-seconds metric would favour longer ads. No screened study tests whether attention optimisation improves exposure quality or only the score [52, 5].

### Drift, updates and transfer: nothing published

In the sources screened, none of the independent studies reports a commercial attention model's performance after model updates, calibration drift, calibration by channel or population, or transfer to new channels. Their data vintages are 2016-17 for the TVision panel, 2020 for the Simonov webcam experiment, which used Lumen's technology, and a recent lab feed study, so they validate those vintages only. A reported model-update date is not evidence that performance held [52, 51, 46]. The guidelines ask for validation at least annually on fresh deterministic data, and strongly encourage randomised or A/B validation without requiring it [41]. The accuracy evidence for webcam eye tracking is thin and partly vendor-authored, and none of it tests a commercial ad-attention panel [58, 59].

### What study would settle it

H1a needs one design, run by a team with no stake in the metric and repeated by another. Take one vendor's attention metric and the viewable seconds for the same impressions; the guidelines already encourage reporting attentive time relative to viewable time [41]. Fit each outcome on time in view alone, then add attention. Score both models on data neither saw: campaigns held out at random, and a later period that falls after the vendor's next model update. Report memory, brand and sales separately, with intervals, the funder and the model version, and register the comparison before the data arrive.

H1b needs a randomised buying test. Split one budget at random between an arm bought on the attention metric and an arm bought on standard metrics such as viewability, CPM or completion. Measure the outcome against a holdout, through a party other than the vendor whose score steered the buying. Report cost per incremental outcome with the media premium and measurement fee included, and repeat in a second channel before claiming transfer.

The study table lists the 48 study records coded from independent, academic, commissioned and vendor work, one row per study and outcome, with design, sample, sponsor and uncertainty. Some studies discussed here and in chapter 4, such as the Amplified and Chalice trials, are not among them.

**Study evidence table**

| Study | Year | Design | Sample | Setting | Format / channel | Outcome measure | Headline result | Uncertainty | Sponsor | Generalisability | Src |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Bruns, Kopka, Borgmann, Prior, Langner (J Advertising) | 2025 | Lab validation: viewport logging vs mobile eye tracking (reference) | 30 students assessed, 4 excluded; 806 post exposures | Germany, lab, mock Instagram feed on own smartphones | Static social feed posts (10 ads, 23 content) | Concurrent validity vs gaze duration | r = .89 (gaze duration), .87 (fixation duration), .81 (fixation count) | all p < .001; no CIs | European Advertising Academy grant; no COI | Low-moderate: students, static ads, simulated feed | [46] |
| Bruns et al. (J Advertising) | 2025 | Same; single-predictor logistic regressions, in-sample | as above | as above | as above | Unaided brand recall (memory) | Viewport time OR 1.12/s, R2 .027; gaze duration OR 1.14/s, R2 .027 | p = .045 and p = .047; no hold-out | European Advertising Academy grant; no COI | Low-moderate | [46] |
| Simonov, Valletti, Veiga (J Marketing Research) | 2025 | Randomised online experiment (random article-ad pairing); OLS | 1,013 participants; 5,707 obs (visibility), 3,925 obs (eye tracking) | US and UK online panel, desktop and mobile | Display ads beside news articles | Aided brand recall (memory) | +3.43 pp per s ad dwell (R2 .143) vs +0.32 pp per s ad visible (R2 .091) | SE .004 and .001, clustered by individual; R2 from different samples; no hold-out | Imperial College COVID-19 Response Fund, NET Institute; Lumen webcam tech in kind | Moderate: two countries, realistic pages, immediate recall | [51] |
| Simonov, Valletti, Veiga (J Marketing Research) | 2025 | Same; OLS and IV | same | same | same | Incentivised voucher-vs-cash choice (purchase proxy) | +0.73 pp per s ad dwell (R2 .136) vs +0.13 pp per s ad visible (R2 .130) | SE .003 and .001; different samples; no hold-out | same | Low for sales: immediate choice, not in-market | [51] |
| McGranaghan, Liaukonyte, Wilbur (MSI WP; Marketing Science 2022) | 2021 | Observational panel; OLS, one viewing measure per model | 180 ad slots; panel 1,155 households, 3,659 viewers | US TV homes, TVision panel, two NFL playoff games | Linear TV ads | Google brand-search lift (behavioural proxy) | R2 attention .116 vs exposure tuning .082, presence .070, OTS tuning .035 (no QH FE); .180 vs .176, .165, .158 (QH FE) | Attention coef .064 (SE .029) and .106 (SE .052); no hold-out | No funding; TVision data and feedback | Low: one evening, 2017 data | [52] |
| Venkatraman et al. (J Marketing Research) | 2015 | Lab multi-method + market response models; in-sample F-tests | 37 ads (26 with sales data, 5 firms); eye tracking 29 participants | US lab; weekly sales/GRP data | 30-s TV ads | Advertising elasticities (sales) | Eye tracking not significant beyond self-report; only fMRI added (p < .011) | Company fixed effects; small df; no hold-out | ARF and six corporate sponsors (Neuro 2.0) | Low-moderate: forced exposure, 26 ads | [6] |
| ARF Attention Measurement Validation Initiative Phase 2 | 2024 | Multi-vendor comparison on a common ad set | 32 ads, 12 companies | Vendors' own samples; forced and natural exposure | Creative (TV, social, digital video, other) | Cross-vendor agreement | R2 eye-tracking-only pair 0.08 and 0.005; eye+facial pair 0.78; survey pair 0.66; neuro pair 0.10 | Decile regressions; no CIs | ARF; vendors in kind, anonymised | Low-moderate: large brands, creative testing only | [4] |
| ARF Phase 2 | 2024 | Same | same | same | same | Advertiser-rated campaign success | R2 0.003 (facial coding) to 0.12 (eye tracking + facial coding) | No CIs; advertiser ratings not ground truth | ARF | Low | [4] |
| ARF Phase 2 | 2024 | Same | two companies' sales proxies vs a survey company's | same | same | Agreement of vendors' sales-proxy scores | R2 0.0007 and 0.006 | No CIs | ARF | Low | [4] |
| ARF Phase 3 (press release) | 2026 | Live-campaign comparison of multiple providers | 4 national campaigns | not stated | TV, online video, social, display | Upper-funnel brand lift | No simple one-to-one relationship between attention and brand lift (no figures public) | not disclosed | ARF; funding not stated, so independence is unknown (round 2, R2-02) | unknown | [5] |
| Büyükakgül, Yüce, Katırcı (J Eye Movement Research) | 2025 | Lab eye tracking; GLMM | 79 participants | Turkey lab, mobile tennis game | In-game static ads | Visual-aided brand recall (memory) | Total fixation duration B = -0.003 (n.s.) after placement; placement (brand dummy) OR 22.4 | SE 0.006, p = .672; brand-placement confound | No external funding | Low: one game, young convenience sample | [53] |
| Guixeres et al. (Frontiers in Psychology) | 2017 | Lab EEG, HRV, eye tracking; ANN with random 70/30 case split x10 | 35 participants, 8 ads | Spain lab | Super Bowl TV ads | YouTube view class over one year | 82.9% test accuracy (4 classes) | Same ads in train and test (leakage) | Heineken Endowed Chair in Neuromarketing | Very low: 8 ads | [54] |
| Wedel and Pieters (Marketing Science) | 2000 | Infrared eye tracking, hierarchical Bayesian model | 88 consumers, 65 print ads | Natural magazine exposure | Print ads | Indirect brand memory | Pictorial and brand fixations promote memory; text fixations do not | not extracted (abstract only) | unknown | Low for digital media | [55] |
| Zhang, Wedel, Pieters (J Marketing Research) | 2009 | Lab eye tracking matched to sales; Bayesian mediation | not extracted (abstract only) | not stated | Grocery feature ads (print) | Sales of featured products | Gaze duration affects sales beyond mere presence of the ad | not extracted | unknown | Low for digital media | [56] |
| Martinovici, Pieters, Erdem (J Marketing Research) | 2023 | Eye tracking in choice task; Bayesian K-fold cross-validation | 325 consumers | not stated | Brand choice task (not ads) | Brand choice | 85% of choices predicted; 45% from first quarter vs 20% chance | K-fold CV on out-of-sample participants | unknown | Not advertising | [57] |
| Semmelmann and Weigelt (Behavior Research Methods) | 2018 | Webcam (WebGazer) eye tracking, lab vs online | 30 in-lab, 32 complete online (84 started) | Germany lab and crowd sample | Fixation, pursuit, free-viewing tasks | Spatial accuracy | Offset 172 px (3.94 degrees) lab; 207 px online | Higher online variance | unknown | 2016-era open-source tracker | [58] |
| Wisiecka et al. (ETRA) | 2022 | Webcam vs remote vs integrated eye tracking, between-subjects | 83 participants | not stated | Point detection and visual search | Measurement error and replication of known effects | Higher webcam error; expected effects replicated | not extracted (abstract only) | unknown | Low-moderate | [59] |
| Kwon, Nyilasy, King, Reid (J Advertising) | 2021 | Meta-analysis of media-context effects | 96 studies, 597 effect sizes, 139,233 participants | Published literature, 31 years | Mixed media | Attitudes and purchase intention (no attention metric) | Congruence very weak; involvement, transportation, trust stronger | not extracted (abstract only) | Grady College (University of Georgia) internal funds | Not about attention metrics | [60] |
| Han et al. (J Advertising) | 2025 | Biometric coupling (facial EMG, HR, SCL) | not in abstract | not stated | Video ads | Attitudes, purchase intention, recognition memory | Emotional pattern stability predicted attitudes and PI; engagement stability predicted recognition | not extracted (abstract only) | unknown | Not an attention metric | [61] |
| News UK / Lumen: context and relevance (REG-002) | 2022 | comparison of context conditions; model-estimated attention | about 980,000 impressions | UK, The Sun website | digital display | attention per 1,000 impressions (no outcome) | +31% relevant vs non-relevant context; viewing time +28% (1.8 s to 2.3 s) | not reported | News UK with Lumen (commissioning not stated on page) | one publisher; unit inconsistency in release | [62] |
| Outbrain / Lumen: native vs social (REG-001) | 2019 | consumer responses; method not stated; no attention measure | 900+ consumers | UK, FR, DE | native vs social vs display | stated trust, click, purchase | +31% trust, +16% click, +18% future purchase vs social | not reported | Outbrain | not an attention study; never merge with REG-002 | [21] |
| Nativo / Adelaide inventory audit (REG-003) | 2025 | AU audit vs Adelaide format benchmarks | about 25 million impressions | Nativo inventory | native display, article, story | AU (model estimate; no outcome) | 28.21 AU; +40% vs native benchmark; +21% vs display benchmark | not reported | Nativo-issued release; audit funding not stated | one seller; benchmarks unpublished | [63] |
| Adelaide 2024 Outcomes Guide | 2024 | compilation of selected case studies (high vs low AU); designs not disclosed | 45 case studies, 18 verticals | not stated | omnichannel | upper- and lower-funnel KPIs; cost | average +40% upper-funnel, +53% lower-funnel, 37% cost savings | not reported | Adelaide | selected cases; guide gated | [13] |
| Adelaide 2025 Outcomes Guide | 2025 | compilation of selected case studies; designs not disclosed | 52 case studies, 18 industries | not stated | omnichannel incl. audio, display | brand lift; lower-funnel KPIs | average +41% brand lift, +55% lower-funnel | not reported | Adelaide | selected cases; guide gated | [14] |
| Adelaide 2026 Outcomes Guide | 2026 | compilation of selected case studies; designs not disclosed | 60 case studies, 16 industries | not stated | omnichannel incl. CTV, audio, display | upper- and lower-funnel KPIs | average +33% upper-funnel, +53% lower-funnel; bank case 2.8x conversions at lower cost | not reported | Adelaide | selected cases; guide gated | [15] |
| Adelaide: yogurt brand AU vs Nielsen ROI | 2020 | in-sample correlation, AU vs viewability | not stated | not stated | not stated | Nielsen ROI (sales) | AU 'over +180% more correlated' with ROI than viewability | not reported | Adelaide | single brand; AU model since updated | [64] |
| Adelaide: insurance brand AU vs conversion | 2020 | in-sample correlation | not stated | not stated | not stated | conversion rate; incrementality | AU per impression correlated with conversion rate; 'strong trend' with incrementality | not reported | Adelaide | single brand | [64] |
| Lucid / Adelaide brand lift | 2022 | brand lift split by AU level; A/B tests vs viewability optimisation (design not described) | 7 brands, 15 campaigns | not stated | not stated | ad recall, familiarity, unaided awareness | +40% ad recall, +28% familiarity, +27% unaided awareness (high vs low AU); up to +22% vs viewability optimisation | not reported | Adelaide page; Lucid integration partner | partner-reported; no costs | [65] |
| MediaSense review of Adelaide AU | 2025 | commissioned methodology review; reconstruction of AU curves from vendor-supplied aggregated data | not stated | not stated | omnichannel | aided awareness, familiarity, recall, intent | higher AU 'reliably correlate[s]' with KPIs (Adelaide summary) | not in accessible text | Adelaide (asked MediaSense); payment not stated | report gated; not accreditation | [66] |
| iFIT / Adelaide / Upwave (The Trade Desk) | 2025 | campaign case study; comparator not stated | not stated | The Trade Desk | CTV, OTT, OLV, audio, DOOH | aided awareness, consideration, favourability | +61% aided awareness, +17% consideration, +10% favourability | 'statistically significant scale' after four weeks; no intervals | published by The Trade Desk-owned outlet; Adelaide, Upwave vendors | single advertiser | [67] |
| Havas / Lumen / Brand Metrics: 9,000 brand lift studies | 2024 | descriptive correlation; model-estimated attention; no control group | 9,089 campaigns; 1,879,764 respondents; 5,618,088 viewable impressions | mostly UK; some US, Singapore | display | awareness, consideration, preference, action intent | more aggregate attention associated with higher scores; about 2.5 s aggregate attention threshold | 'significant' stated; no intervals | Havas Media Network, Lumen, Brand Metrics (funding not stated) | display only; no CPMs; campaign count given as 9089 and 9087 | [10] |
| PwC review of Lumen attention analysis | 2022 | commissioned review; re-performed in-sample correlations and brand-lift pilot regressions | ~1 billion impressions (3 advertisers); model test 40,000 impressions; pilot 2,400+ panelists, 6 studies | three anonymised advertisers | banner, outstream, skins, instream | CTR, conversion; 12 brand metrics | attention correlated better than viewability with CTR and conversion; not significantly better on first-mention awareness, any-mention awareness, digital ad recall | p-values and McFadden R-squared reported | Lumen (requested PwC review) | data not verified by PwC; model uses viewable seconds | [8] |
| Teads / Lumen brand lift meta-analysis | 2023 | Dynata exposed vs control brand lift segmented post hoc by predicted dwell time | 16 campaigns, 14 advertisers | Teads inventory; North America, APAC, EMEA | Teads formats (not specified) | 7 brand lift metrics | online ad recall lift 40% higher at 5+ s vs 1+ s; purchase intent needs about 8 s+ for significance | significance referenced; no intervals | Teads with Lumen (funding not stated) | selected sample; single exposure; one seller | [68] |
| Lumen CPG attention bidding | 2024 | two-arm campaign test vs traditional programmatic control, same budget (not described as randomised) | not stated | not stated | programmatic (formats not stated) | Lumen-estimated viewed ads and view time; CPM | +104% viewed ads; +110% view time; -40% CPM; -44% aCPM | not reported | Lumen | outcome is the vendor's own score | [69] |
| Newsworks / Lumen / Peter Field | 2025 | IPA Databank analysis by media attention class plus Brand Metrics and Lumen data | not stated in text read | UK | cross-media | market share; brand effects | +12% market share; +17% brand effects; 58% more attentive seconds per pound | not reported | Newsworks (UK newsbrands' body) with Lumen | media-class comparison; award-database cases | [70] |
| DV Authentic Attention / Mondelez / Kantar | 2021 | Kantar control/exposed brand lift; exposed split post hoc by DV attention quadrant | not stated | US, category buyers | static display (35% mobile web, 50% app, 15% desktop) | unaided awareness, favourability, consideration, purchase intent | high-Exposure +5 pp unaided awareness vs +3 pp (80% sig.) for all authentic ads; high-Engagement +9 pp | 90% confidence; 80% where asterisked | DoubleVerify (published); Kantar measured | one campaign; Exposure index includes viewable time | [71] |
| DV Authentic Attention / Club Med / Kantar Millward Brown | 2023 | campaign case study; comparator not stated | not stated | not stated | not stated | top-of-mind and online ad awareness | +19% top-of-mind awareness; +41% online ad awareness | not reported | DoubleVerify | summary only; case gated | [72] |
| IAS Quality Attention / Kimberly-Clark / NCS, Test 1 | 2024 | NCS sales effect vs modelled baseline; post hoc high vs low attention segments | not stated | not stated | browser display | sales lift | no significant difference between high and low attention segments | significance referenced; no intervals | IAS (published); NCS measured | one advertiser | [11] |
| IAS Quality Attention / Kimberly-Clark / NCS, Test 2 | 2024 | NCS sales effect vs modelled baseline; post hoc high vs low attention segments | not stated | not stated | display, video, mobile in-app | sales lift; ROAS | +313% sales lift above- vs below-average attention; ROAS +78% (comparator unspecified); attention score 72 vs 53 in Test 1 | not reported | IAS (published); NCS measured | format and season changed from Test 1 | [11] |
| IAS attention page aggregate claims | 2026 | unspecified high vs low attention comparison | not stated | not stated | incl. in-app | unaided awareness; 'success rate'; cost per success | +52% unaided awareness; +277% success rate; -63% cost per success | not reported | IAS | no base or period | [73] |
| IAS / Stellantis / Publicis | not stated | optimisation case study | 13 billion impressions | global | not stated | engagement; ad recall | 165% engagement boost; 33% ad recall | not reported | IAS | tile summary only | [73] |
| dentsu Attention Economy Phase 2 | 2021 | eye-tracking panels and controlled experiments with control groups; planning simulation at fixed budget | panels of 3,400 (Amplified) and 6,000 (Lumen); TVision data from 5,500 US households | US, UK, Australia | cross-media | prompted recall; brand choice; modelled attention | significant recall and choice uplifts vs control rising with dwell time; +3.7 years attention in one month at same budget (modelled) | 'statistically significant' stated; no intervals | dentsu | agency programme; summary document | [9] |
| TVision / Upwave: attention and brand lift | 2022 | cross-campaign linear regression (correlational) | 40+ campaigns | US, TVision panel | TV | aided awareness; ad recall | close to 1% aided awareness lift per percentage point of attention (high-frequency ads) | not reported | TVision and Upwave (funding not disclosed) | correlational; outcome partly modelled | [74] |
| Wurl / TVision: emotion alignment | 2025 | observational aligned vs mismatched placements | 50+ advertiser campaigns | streaming TV (TVision panel) | CTV | share of ad duration attended (no outcome) | 2.1x (33% to 71% of ad, as reported); 2.4x for 30-second ads (27% to 66%, release) | not reported | Wurl with TVision (funding not stated) | vendor sells the targeting tested; report gated | [17] |
| Amplified / VCCP Media: distinctive assets | 2025 | creative experiment, branded vs de-branded ads in real social environments | 72 ads, 8 brands, 20,000+ views | UK | social video | memory; 'attention-adjusted ROI' | 3.5x attention-adjusted ROI for best brand code vs unbranded; £66bn waste extrapolated | not reported | VCCP Media with Amplified (funding not stated) | creative test; extrapolation is a scenario | [75] |
| Amplified / Audience Group | not stated | sequential before/after of bidding tactics; no control described | not stated | not stated | programmatic | ad and brand recall | +14% ad recall (algorithm alone); +74% brand recall after attention score, same budget | not reported | Amplified | undated; not isolable in business results | [76] |
| Amplified / Spotify India | not stated | panel attention study, audio vs video formats | 7,900+ ad views; 2,500 panel users | India, Spotify | audio and video ads | attention seconds; STAS | video held attention 17.3 s average; +23 pp STAS vs audio | not reported | Spotify with Amplified (funding not stated) | non-US; audio attention method not stated | [77] |
| Vevo / Adelaide inventory benchmark (REG-004) | 2026 | inventory measurement vs Adelaide benchmarks | not stated | Vevo YouTube CTV | CTV | AU (no outcome) | +28% vs average benchmarks | not reported | Vevo release | benchmark undefined | [78] |



## 4. What vendors publish


*The public case that attention pays is made almost entirely by the companies that sell it. None of the study records coded here is independent measurement, and on this evidence H1a and H1b are both unresolved. The results that cut against the attention case come from the same interested parties.*

> **In short.**
> None of the study records coded for this chapter is independent measurement. The commissioned ones are PwC's review for Lumen, dentsu's Attention Economy programme and a Havas-commissioned study with Lumen and Brand Metrics; the rest are vendor, partner, agency or media-owner publications [8, 71, 66]. In PwC's brand-lift pilot, attention was not significantly better than viewability on awareness or ad recall [8]. IAS reports no significant sales-lift difference between above- and below-average-attention impressions on browser display; its second test, which added formats in another quarter, reports a 313% higher sales lift for above-average households [11]. No commissioned or vendor study screened compares attention with time in view on held-out data [8, 10, 71, 64], and none tests attention buying against a randomised or holdout control bought on standard metrics and reports a memory, brand or sales gain with cost per incremental outcome and uncertainty [69, 65, 11, 10]. No study screened discloses a measurement fee or a campaign's media premium [69, 13, 67, 15].

Chapter 3 weighed the independent studies. This chapter takes the rest: case studies, white papers and reviews paid for or published by a vendor, partner, agency or media owner. Independent measurement is run by a party with no financial interest. Commissioned measurement is paid for by a named interested party. A result resting only on a release is a vendor assertion, and where funding is not stated, independence is unknown. H1a concerns predictive value beyond time in view; H1b concerns outcome gains at comparable cost.

### Who stands behind the numbers

Many of the study records coded here used third parties such as Kantar, NCSolutions, Dynata, Upwave, Lucid, Brand Metrics or Nielsen to measure the outcome. In every record the result still reaches the public through an interested party, and this review found no independent replication of any vendor claim in the sources screened [8, 71, 66]. Attention figures from Adelaide's attention unit (AU), Lumen's LAMP and IAS Quality Attention are model estimates. DoubleVerify's 2021 Authentic Attention index, a mix of device and on-screen signals with no reference labels described, is coded as a candidate proxy. Direct observation of people appears only in panel work by TVision, Amplified and dentsu, and TVision's campaign-level figures mix observation and projection [10, 63, 73, 74, 9].

The main records, by class (CTV is connected TV, CPM cost per thousand impressions, ROAS return on ad spend):

| Record | Paid or published by | Class | Comparison | Outcome |
|---|---|---|---|---|
| PwC review, April 2023 [8] | Lumen | Commissioned | Attention against viewability, in-sample, Lumen-supplied data | Click-through, conversion; awareness and recall |
| Attention Economy Phase 2, 2021 [9] | dentsu | Commissioned | Exposed against control, by dwell time | Recall, brand choice; modelled attention |
| Havas, Lumen, Brand Metrics, September 2024 [10] | Havas Media Network | Commissioned | Association with survey responses; no control | Awareness to action intent |
| Outcomes Guides, 2024 to 2026 [13, 14, 15] | Adelaide | Vendor assertion | Selected cases; designs not disclosed | Upper- and lower-funnel lift |
| MAGNA Media Trials and Adlook, July 2023 [342] [69, 13, 67, 15] | Joint release; funding unknown | Vendor assertion | AU pre-bid optimisation, mixed with contextual delivery | Awareness; CPM, cost per completed video |
| Consumer-goods attention bidding, March 2024 [69] | Lumen | Vendor assertion | Attention algorithm against a traditional control, same budget | Lumen attention estimates; CPM |
| Mondelez, 2021 [71] | DoubleVerify | Vendor assertion | Kantar exposed-control lift split by attention quadrant | Unaided awareness |
| Kimberly-Clark, September 2025 [11] | IAS | Vendor assertion | NCSolutions sales lift split after the fact at average attention | Sales lift; ROAS |
| Emotion alignment, July 2025 [16, 17, 18, 440] | Wurl and TVision | Vendor assertion | Aligned against mismatched CTV placements | Attention only |

### The commissioned studies

Lumen commissioned PwC's review, which Lumen released on 27 April 2023. PwC re-performed Lumen's campaign analysis for three advertisers on a subset of data Lumen supplied, and could not verify Lumen's data cleaning or recreate its exact results. Attention correlated better than viewability with click-through and conversion. In a brand-lift pilot for one advertiser (2,400+ panelists, six studies, 12 metrics), both attention and viewability were significantly related to first-mention awareness, any-mention awareness and digital ad recall, with higher z-values for attention [8, 283]. Attention was not significantly better than viewability on any of the three, and PwC found no significant result for either measure on the eight other metrics it lists [8] [8]. Lumen's model takes viewable time as an input [8], so the pilot never set attention against an independent time-in-view baseline.

dentsu's Attention Economy Phase 2 report (2021) combined eye-tracking panels built with Amplified Intelligence and Lumen with purchased TVision data. It reports significant recall and brand-choice uplifts against control groups, rising with dwell time. Its planning example gained 3.7 years of attention in a month at the same budget, a modelled figure with no measured outcome [9]. dentsu sells attention planning, and the report states its aim "to advocate attention as a media currency" [9].

Havas Media Network commissioned the third, published on 17 September 2024 with Lumen and Brand Metrics. It matched Lumen attention, predicted from viewability data, to survey responses across 9,089 display campaigns, mostly in the UK. More attention went with higher awareness, consideration, preference and action intent; there was no control group and no campaign CPM data [10]. The paper itself says "Attention time can be accurately estimated from typical viewability metrics" [10].

### Results that cut against the attention case

PwC's pilot, from the one commissioned review with published statistics, found no significant memory advantage over viewability [8, 10, 71, 64]. Lumen's own release said the brand-lift relationship was "in line with viewable time metrics" [8, 283], and the Teads and Lumen white paper notes that PwC did not strongly support the attention-over-viewability hypothesis [68].

IAS's Kimberly-Clark case reports two NCSolutions sales-lift tests with impressions split after the fact into above- and below-average Quality Attention. Test 1 (Q2 2024, browser display only) found no significant difference in sales lift between the segments, although IAS says higher attention correlated with stronger ROAS in both tests without giving a Test 1 figure. Test 2 (Q4 2024) added video and mobile in-app and reports a 313% higher sales lift for above-average households. Its 78% ROAS rise came "with the addition of high-attention formats", which reads as Test 2 against Test 1 rather than high against low [11] [11]. Test 2 changed formats and season at once, confounding any attention effect with both (inference).

At channel level, Kantar reports from 873 campaigns in its LIFT+ database that channel attention levels per impression do not seem to be directly related to channel-level cost-effectiveness [258].

### Adelaide's case-study compilations

Adelaide's annual Outcomes Guide releases report headline averages over selected case studies: 45 case studies with 40% upper-funnel lift, 53% lower-funnel lift and 37% cost savings (January 2024); 52 case studies with 41% higher brand lift and 55% stronger lower-funnel impact for 2024 campaigns (January 2025); and 60 case studies with 33% upper-funnel and 53% lower-funnel lift for 2025 campaigns (January 2026). In August 2023 Adelaide cited averages of 31% upper-funnel and 56% lower-funnel lift [12]. The averages cover different campaign years and undisclosed case sets, so no trend can be read from them. The releases disclose no designs, holdouts, uncertainty or averaging method, and the guides are gated [13, 14, 15] [13, 14, 15]. They show which cases Adelaide chooses to publish; they cannot show what AU buying adds over a control bought on standard metrics. The lift comparators are not stated and "37% cost savings" is undefined. The releases do not say how cases were chosen, so selection toward favourable cases cannot be ruled out (inference), as for any vendor's published cases.

Adelaide also says pre-bid beta partners in The Trade Desk saw a 58% increase in cumulative ROAS against control groups, with no design, sample, control definition or cost [328, 161]. In the iFIT case, Upwave measured a 61% increase in aided awareness; Adelaide describes the lifts as "compared with the control group" without saying whether that group was unexposed or bought on standard metrics [67]. An unexposed control would make it a measure of the campaign as a whole.

AU is itself trained and calibrated on historical outcome data, including brand lift, conversions and sales [168]. Inference: part of any AU-to-outcome correlation in an Adelaide case may be built into the score, and the public record does not allow a check.

The only outside review of AU found in the sources screened is MediaSense's, which Adelaide requested [94, 93, 91, 66]. MediaSense rebuilt AU-to-outcome curves from sampled data Adelaide supplied, did not independently verify that data and did not review AU source code [66, 170]. That tests whether the curves can be reproduced; it does not test them against independent data [66, 22]. Because Adelaide asked for it, the review is commissioned work; this chapter counts it with the vendor publications only because its report is gated and was not read [66, 22].

### What the Wurl and TVision ratios compare

The Wurl/TVision "2.1x" carried from the author's earlier papers compares shares of an ad's duration. As reported by The Measure, attentive time was 71% of the ad for CTV placements emotionally aligned with surrounding content against 33% for mismatched ones, an observational comparison across more than 50 advertiser campaigns. The release's own key finding is a 2.4x gap, 66% against 27% of the ad, for 30-second ads. TVision's chief executive later gave roughly 30% and 70% [16, 17, 18, 19] [16, 17, 18, 440]. The denominator is share of ad duration attended, neither attention per impression nor per dollar, and the outcome is attention only [17, 18]. Wurl sells the contextual targeting tested, and AppLovin owned Wurl when the report was released on 31 July 2025 [16, 17, 18, 440].

### Three figures that travel between papers

The News UK and Lumen release of 22 March 2022 reports that, across 980,000 impressions on The Sun's website, ads in a relevant editorial context received 31% more attention per 1,000 impressions, as calculated by Lumen's model, than ads in a non-relevant context. It reports no memory, brand or sales outcome [402, 62]. The release gives its baseline in minutes per 1,000 impressions in one sentence and in seconds in another, so only the 31% can be cited [402, 62].

The Outbrain and Lumen study of 25 November 2019 is a different +31%. It found native ads on premium news sites more likely to be trusted (+31%) than ads on social platforms, from over 900 consumers in the UK, France and Germany. It is a trust survey with no attention measure [21], and its headline claims "44% more trusted" without reconciling the two figures [21].

Nativo's 28.21 AU is a vendor assertion from Nativo's own release of 26 August 2025: an Adelaide audit of about 25 million impressions gave its native formats an average 28.21 AU, 40% above Adelaide's native-format benchmark, with no outcome measured [63]. AU is vendor-defined and cannot be compared with any other vendor's unit.

### What "independent" means in vendor copy

Under this paper's rules, a study is independent only when whoever ran it has no financial interest. Sellers apply the word to commissioned Lumen work. Lumen markets its technology as "the only solution validated and audited by an independent third-party in the industry" on the strength of the PwC report, which says Lumen requested it and that PwC "has not independently verified, validated, or audited the data" [20, 8] [20, 8]. dentsu's report on a programme it funded, with Lumen panels among its sources, calls the vendor studies independent [9]. Outbrain's 2019 page calls Lumen an "independent research agency" and does not say whether Lumen was paid [21]. Nativo calls Adelaide's audit "independent", although Adelaide sells the metric and the funding is not stated [63], and Adelaide titled its release of the MediaSense review an "Independent Review" [22]. This paper does not repeat the label for any of them.

### Cost: what no study discloses

No study screened discloses measurement fees, and none reports a campaign's media premium for attention-bought inventory. Cost appears instead as CPM or cost-per-completion changes (Lumen's consumer-goods case, 40% lower CPM; MAGNA and Adlook, 42% lower cost per completed video for AU pre-bid), cost-per-acquisition or cost-per-conversion claims, undefined "cost savings", IAS's "cost per success", ROAS, or a fixed budget. The only premium figure is an estimate: the Newsworks and Lumen report puts UK news-brand display at a 24% CPM premium over general display [23] [69, 13, 67, 15]. Buyers mention premiums without pricing them: Haleon warned that the highest-attention CTV carries very high CPMs [441]. Fees exist outside the studies. By 29 September 2026, Google's Display & Video 360 documentation listed an additional media fee for its "Maximize Adelaide AU" algorithm [330, 159]. Adelaide said in October 2025 that publishers such as Condé Nast and Hearst cover the cost of AU measurement [167, 176]. [Chapter 6](#trading) covers pricing.

### How H1a and H1b code on this evidence

H1a needs added predictive value over time in view on held-out data. None screened makes that comparison. Published comparisons are in-sample, mostly against viewability or completion rate, and several vendor models take viewable time as an input [8, 10, 71, 64]. Adelaide's 2020 correlations set AU against viewability with no coefficients or held-out data [64], and TVision and Upwave report a cross-campaign regression with no uncertainty or held-out test [74, 439]. H1a is unresolved on this evidence. PwC's null does not contradict it, because the contradiction test needs a held-out comparison with time in view.

H1b needs a randomised or holdout test with cost reported. In the vendor, agency and media-owner sources screened, this review found no public study that compares media bought on an attention metric with a randomised or holdout control bought on standard metrics and reports a memory, brand or sales gain with cost per incremental outcome and uncertainty [69, 65, 11, 10]. The nearest candidates each miss a condition. Lumen's consumer-goods test had a control and CPMs, but its outcomes were Lumen's own attention estimates, +104% viewed ads and +110% view time [69]. MAGNA and Adlook's +131% awareness "delta lift" combines AU optimisation with contextual delivery [342]. An Amplified and Chalice release reports a 70% ad-recall increase against a standard non-attention line item, with no sample, uncertainty or cost [195]. Amplified's Audience Group case reports brand recall up 74% with no control group [76]. Adelaide's page announcing its Lucid integration (16 March 2022) says "structured A/B tests" of in-flight optimisation toward AU gave brand impact increases of up to 22% against viewability, with no design, sample, cost or uncertainty [65]. Viant's chief executive said on 10 August 2026 that more than 80% of 42 pilots on TVision high-attention segments beat a benchmark conversion rate, with an average lift of 1.4x; Viant has owned TVision since 1 May 2026, and no design, benchmark definition, control or cost was given [304] [435, 82]. H1b is unresolved on commissioned and vendor evidence. IAS's Test 1 does not contradict it: a post hoc split with no cost is not the registered test.

### What a vendor study would need to publish

A vendor or commissioned study would count toward H1a or H1b if it published:

- who paid, who ran the study and who measured the outcome, with any financial link between them;
- for H1a, a time-in-view baseline on the same data and the increment on held-out data, ideally held out in time, with the model version tested;
- for H1b, a randomised or holdout arm bought on standard metrics (an unexposed control or a post hoc split does not qualify);
- memory, brand and sales results reported separately, with sample sizes and uncertainty;
- cost per incremental outcome in both arms, including media premium and measurement fee;
- the full set of campaigns tested, and an independent replication.

The IAB and MRC attention guidelines ([chapter 2](#standards)) already ask for part of this: modelled measurement distinguished from observed, annual validation on fresh deterministic data, disclosed error margins, and outcome validation for any model used to predict outcomes [41]. [Chapter 12](#next) turns the list into questions a buyer can ask.



## 5. Who sells attention


*Attention measurement is sold by two large verification companies, a small group of specialists and a set of creative testers. Since 2021 Integral Ad Science has gone private, DoubleVerify has agreed to a sale to Nielsen's parent, and the media-attention specialists that were sold went to buyers outside verification. No public source screened shows whether verification bundles changed the specialists' demand, pricing or revenue, so H4 is unresolved. No source found ties any event to a metric failing.*

> **In short.**
> Oracle ended Moat, with all its advertising products, on 30 September 2024 [87, 88]. Novacap funds took Integral Ad Science (IAS) private on 23 December 2025 [83, 437]. DoubleVerify (DV) agreed on 6 August 2026 to be acquired by the parent of Nielsen [84, 465, 466]. At 29 September 2026 the deal was pending, with closing targeted by the first quarter of 2027 [181, 465, 565]. Viant, which runs a demand-side platform (DSP), completed its purchase of TVision on 1 May 2026 [435, 82]. By 28 September 2026 GumGum's website no longer presented GumGum Attention as a standalone product [443]. In the filings read, neither DV nor IAS discloses attention revenue [47, 207] [231]. Lumen's unaudited accounts show two loan covenant breaches at 31 December 2025, and the directors state that no material uncertainty exists on going concern [288]. Its loss widened in the year to September 2025, almost entirely through a non-cash share-based charge by this paper's inference [285] [285, 288]. H4, the bundling hypothesis, is unresolved [47, 231, 285, 232].

This chapter answers Q4, how the market has consolidated since 2021, and tests H4. Each event gets one class under shared rule 11: acquisition, rebrand, restructuring, insolvency, product retirement or pivot. A class records what happened to a company or product. It says nothing about whether a metric worked unless a source establishes the cause.

### Three kinds of seller

The verification companies sell attention inside wider suites. DV's annual report (10-K) for 2025 calls its Authentic Attention product "MRC-accredited" [47, 207, 436]. IAS's last 10-K describes Quality Attention as built from time in view, share of screen and ad interaction [231]. On 29 September 2026 the Media Rating Council (MRC) listed accredited attention metrics for one company, DV [2, 205]. [Chapter 2](#standards) sets out their scope.

The specialists sell attention as the product. Adelaide sells AU, its own attention unit, through Adelaide Metrics Inc., a Delaware corporation based in New York [105, 334]. Lumen Research Ltd, a UK company, reports no ultimate controlling party [427, 285]. Amplified trades through an Australian private company whose owners were not established [462, 456]. TVision has been a Viant subsidiary since 1 May 2026 [435, 82]. xpln.ai sells attention estimates built from eye-tracking panels and machine learning [122, 41], and on 28 September 2026 was the one vendor on MRC's digital under-review page with an attention audit in process [48].

The creative testers sell tests run before launch. Kantar sells LINK+, LINK AI and Context Lab [246, 247, 248], and its audited 2025 accounts put its ownership at 60% Bain Capital and 40% WPP affiliates [468, 269]. Smart Eye AB, listed on Nasdaq First North, owns iMotions and Affectiva outright [452]. Adverteyes Inc., whose owner no source read discloses, now sells the creative prediction Realeyes used to sell [86, 449, 450] [460, 461].

### The events, classified

The table lists the events found from 2017 to the cutoff. Dates come from a filing, a registry or the parties' own release where one exists. The list is not a census; some vendors were not screened for events [87, 80, 82, 83, 465, 86].

| Date | Event | Class (rule 11) | Claim |
|---|---|---|---|
| 17 May 2017 (announced) | Tobii buys substantially all the assets of Sticky | Acquisition | [79, 455] |
| June 2021 | Smart Eye buys Affectiva | Acquisition | [87, 80, 82, 83, 465, 86] |
| November 2021 | Smart Eye buys iMotions | Acquisition | [87, 80, 82, 83, 465, 86] |
| 16 December 2021 (announced) | GumGum buys Playground xyz; price and completion date not disclosed | Acquisition | [80] |
| November 2022 (reported) | Lumen combines with Avocet, a DSP | Acquisition | [398, 285] |
| 13 December 2022 (announced) | Amplified buys AttentvAds; terms not disclosed | Acquisition | [464, 193] |
| 23 April 2024 | GumGum merges Playground xyz's Asia-Pacific media business and keeps the attention platform standalone | Restructuring | [442, 85] |
| 19 August 2024 (announced) | Adelaide buys Rita, a consumer-data app, for stock; price not disclosed, completion date not stated | Acquisition | [341] |
| 30 September 2024 | Oracle ends all Oracle Advertising products, including Moat | Product retirement | [87, 88] |
| 3 October 2024 | Playground xyz's platform renamed GumGum Attention | Rebrand | [442, 85] |
| January 2025 | iMotions integrates Affectiva's Media Analytics business | Restructuring | [452] |
| February 2025 (completed) | Outbrain buys Teads | Acquisition (trading table) | [87, 80, 82, 83, 465, 86] |
| 29 April 2025 | Amplified Intelligence becomes Amplified | Rebrand | [456, 193, 463] |
| 6 June 2025 | Outbrain Inc. becomes Teads Holding Co. | Rebrand | [448] |
| 1 August 2025 | Kantar sells Kantar Media to H.I.G. Capital for $924.1 million | Acquisition (not an attention business) | [265] |
| 18 November 2025 | Realeyes launches Adverteyes, called a spin-out in the release | Restructuring | [86, 449, 450] |
| 23 December 2025 | Novacap funds complete the take-private of IAS | Acquisition | [83, 437] |
| 2 January 2026 | Life360 closes its purchase of Nativo | Acquisition (trading table) | [446, 447] |
| 1 May 2026 | Viant completes its purchase of TVision | Acquisition | [435, 82] |
| 6 August 2026 (agreed) | DV agrees to be acquired by Neptune BidCo US Inc., parent of Nielsen Company (US); pending at 29 September 2026 | Acquisition, announced | [84, 465, 466] [181, 465, 565] |
| Date unknown (observed 28 September 2026) | GumGum Attention no longer presented as a standalone product | Product retirement (inferred) | [443] |
| Date unknown (observed 28 September 2026) | SPARK Neuro's home page sells dementia care, with no advertising offer | Pivot (inferred) | [451] |

![Both verification suites left or agreed to leave public markets within eight months; the panel company went to a DSP.](figures/fig-consolidation.svg)

*Figure fig-consolidation. **Both verification suites left or agreed to leave public markets within eight months; the panel company went to a DSP.** Transactions, rebrands, spin-outs and retirements in the attention market, classified per shared rule 11, with the date each source settles (announcement or completion). No event found has a documented cause of metric failure. Units: dates. Scope: companies and products in the vendor universe. Sources [79], [80], [81], [82], [83], [84], [85], [86], [87], [88]. Caveat: SPARK Neuro’s pivot and the end of GumGum Attention as a standalone product have no documented date and are not plotted. The Nielsen–DoubleVerify merger was pending at the cutoff. Data: data/figure-data.json#events*

None of the events found is an insolvency. The media-attention specialists bought by other companies went to Viant, a DSP, and to GumGum, which sells its own targeting [458, 459]. DV's own acquisitions from 2023 to 2025 were Scibids, for AI campaign optimisation, and Rockerbox, for attribution; its filings describe neither as an attention purchase [47, 207]. No source found attributes any event in the table to a failure of attention metrics [87, 80, 82, 83, 465, 86].

### The verification companies change owners

Novacap announced its take-private of IAS on 24 September 2025 at $10.30 a share in cash, in a deal the parties said valued IAS at about $1.9 billion, basis not stated. On completion, IAS became a wholly owned subsidiary of Igloo Group Parent, Inc., an affiliate of Novacap funds. The price paid for the outstanding shares came to about $1.6 billion, and IAS filed to deregister on 2 January 2026 [83, 437, 244]. In January 2026 MRC removed IAS Attention Measurement from in-process status, citing IAS's lack of intent to engage in a third-party audit. No source links that timing to the take-private [49].

DV agreed on 6 August 2026 to be acquired for $13.60 a share in cash, an enterprise value of about $2.15 billion. The buyer is Neptune BidCo US Inc., the parent of Nielsen Company (US), LLC and an affiliate of funds managed by Elliott and Brookfield. Termination fees are $60 million payable by DV and $144 million by the buyer [84, 465, 466]. At 29 September 2026 the deal was pending. The antitrust waiting period under the Hart-Scott-Rodino Act expired on 21 September, and DV filed its definitive proxy on 29 September for a stockholder meeting on 29 October. DV targets closing by the first quarter of 2027, a forecast set in a corrected release of 7 August 2026 [181, 465, 565]. If the deal closes, neither verification company is expected to file periodic reports with the Securities and Exchange Commission (SEC). That would remove the main public financial series for testing H4, an inference about future disclosure [83, 84].

Nielsen already touches the specialists. Adelaide's AU has been available in Nielsen's Outcomes Marketplace since 7 October 2025 [331], and Adverteyes is a partner there [460, 461]. Viant took on Nielsen's pending patent litigation against TVision when it bought the company [435, 82]. No source read links any of this to the DV merger. If the merger closes and MRC's list is unchanged from 29 September 2026, the only accredited attention metrics on it would belong to a company owned by Nielsen's parent [2, 205] [84, 465, 466]. That is this paper's inference from the two records; in the sources read, neither party has said it.

### The specialists change hands

Viant agreed on 14 April 2026 to buy TVision for $22.5 million in cash plus 1,656,701 Class A shares, set by an agreed equity value of $17.5 million [435, 308]. Its quarterly report (10-Q) estimates the consideration at about $37.5 million, with $13.4 million of goodwill, and cites the "integration of TVision's attention signals" [435, 82]. Viant's releases call that integration exclusive. Its filings do not say whether TVision data stay available to others [82, 470, 308, 314, 309]. Vendor sources suggest they do: TVision's site still markets data licensing, and Adelaide wrote in July 2026 that AU for connected TV (CTV) "incorporates TVision data" [297, 311, 106]. Before the deal closed, TVision panel data fed products at Adelaide, Lumen and xpln.ai, and DV had announced a planned fusion with TVision signals [294, 295, 290, 289, 220, 296, 293].

GumGum kept Playground xyz's attention platform as a standalone data solution in April 2024 and renamed it GumGum Attention in October [442, 85]. By 28 September 2026 gumgum.com/attention redirected to a results page that presents attention through Attentive Lift. The retirement date cannot be set, and the product may still be sold under contract or another name [443]. Attentive Lift joins GumGum's attention data to On Device's brand-lift data, with no US availability documented in the sources opened [458, 459]. Oracle's end-of-life notice ties Moat's retirement to Oracle's exit from advertising as a whole, and does not attribute it to Moat or to attention metrics [87, 88].

Realeyes' chief executive became executive chairman of Adverteyes, and the launch release discloses no ownership, stake or funding [86, 449, 450]. Realeyes' own site now sells human verification and training-data licensing [460, 461]. Three specialists bought companies of their own. Amplified bought AttentvAds, which it described as tracking and verifying impression-level data, on 13 December 2022 [464, 193]. Adelaide bought Rita, an app with about 200,000 members that uses EU data-access rights to collect behavioural data, to feed AU models [341]. Lumen combined with Avocet, a DSP [398, 285]. The date of SPARK Neuro's pivot is unknown [451].

Outbrain's purchase of Teads and two other deals sit outside attention measurement [87, 80, 82, 83, 465, 86]. Life360 recorded $104.0 million of consideration for Nativo, a media seller on this paper's trading table, against about $120 million announced; its acquisition disclosures do not mention attention [446, 447]. Kantar Media, the audience measurement, data analytics and media planning business Kantar sold, is separate from creative testing [265]. [Chapter 6](#trading) covers the trading-table sellers.

### Who discloses what

H4 needs revenue or pricing evidence, and almost none of it is public.

| Company | Status at the cutoff | What is public | Attention revenue | Claim |
|---|---|---|---|---|
| DoubleVerify | Listed; merger pending at 29 September 2026 | Revenue by customer type: $748.3 million in 2025, of which measurement $249.7 million | None disclosed | [47, 206, 207] [47, 207] [181, 465, 565] |
| IAS | Private since 23 December 2025 | Last periodic report for Q3 2025; revenue $530.1 million in 2024 | None disclosed | [383, 467] [231] |
| Lumen Research | Private, UK | Unaudited group accounts without an income statement | No turnover disclosed | [285] |
| Adelaide | Private, US | Funding announcements; Form D notices | No revenue disclosed | [174, 81] |
| TVision | Viant subsidiary | A 2022 Form D; Viant calls its results not material | None disclosed | [445, 436] [82] |
| Smart Eye (iMotions, Affectiva) | Listed | Behavioral Research sales of SEK 208.8 million in 2025 (SEK 221.2 million in 2024), including research outside advertising | No media-analytics figure | [452] |
| DAIVID | Private, UK | Micro-company accounts | No revenue disclosed | [454] |
| Amplified, Neurons | Private | Vendor funding statements (Amplified); a reported 2022 seed round (Neurons) | No revenue disclosed | [456, 457] [204, 201, 193, 191] |

Lumen's filed accounts show losses and loans. For the year to 30 September 2025 its group loss was £3,761,807 as first filed, against £1,134,773 a year earlier, with share-based payments of £2,720,974 credited to reserves [285]. If that whole credit was charged in the year, the loss before the non-cash charge was about £1.04 million, by this paper's inference [285, 288]. After Lumen shortened its accounting year [427, 453], amended accounts for the three months to 31 December 2025, filed on 27 June 2026, state that its Salica venture-debt loan breached a cash covenant and its Innovate UK loan breached a debt service cover ratio, with balances of £2,150,050 and £245,793. The accounts reclassify both loans as due within one year, show group net current liabilities of £63,945 and restate the group loss for the year to September 2025 to £3,813,533. The directors state that no material uncertainty exists on going concern [288]. Group cash fell from £1,920,236 to £1,147,844 over the same three months, while average staff rose from 68 to 77 [288, 453, 469]. The accounts are unaudited and show no turnover, so none of this can be read as a change in demand.

Adelaide announced a $7 million seed round in June 2022, saying it brought total capital raised to $9.5 million, and a $1.4 million extension in August 2024: about $10.9 million announced, by this paper's sum. Its Form D notices, filed with the SEC for exempt offerings, report equity offerings of $9.48 million in 2022 and $3.0 million in 2024, with $4.59 million sold as of the filing dates [174, 81].

### Bundling and the specialists (H4)

H4's registered rule needs a source with a comparison that links bundling to specialists' demand, realised pricing or revenue. Without one, bundling and the specialists' commercial changes are reported separately, and the causal claim stays open.

The bundling is documented. DV has held MRC accreditation for display and video Authentic Attention metrics since 2022 [47, 207, 436]. It charges a Measured Transaction Fee per thousand impressions, mostly fixed, and its filings give no attention revenue, price or adoption figure [47, 207]. IAS charges a cost per thousand impressions (CPM) on the volume of ads analysed and launched Quality Attention with eye tracking in 2024 [231]. It names attention in its measurement-revenue definition and, separately, lists attention among its optimisation solutions, so attention revenue cannot be separated [231]. DV's average fee per thousand fell 3% in 2025 and 7% in the second quarter of 2026, a series that is not attention-specific [444, 206, 207, 47].

The specialists' commercial changes point in no single direction. Lumen's group loss widened, almost entirely through a non-cash share-based payment charge by this paper's inference [285] [285, 288]. Its accounts to 31 December 2025 also record two loan covenant breaches and a rise in average staff from 68 to 77 [288] [288, 453, 469]. Adelaide added pre-bid segments in Amazon DSP on 9 June 2026 [162], and Google's Display & Video 360 help pages documented a packaged Adelaide bidding algorithm as live by 29 September [330, 159]. By Adelaide's own count, its AU Ecosystem grew from 54 partners in October 2025 to more than 70 sell-side members and over 50 buy-side partners by July 2026 [167, 176]. Viant paid an estimated $37.5 million for TVision while calling its revenue "pretty de minimis" [82] [304, 82]. None of these is a revenue or pricing series for an attention product.

The evidence also cuts across the hypothesis. The verification companies buy from specialists as well as competing with them. IAS reports a Lumen model for social attention, and its Quality Attention uses passive eye tracking "powered by Lumen data" [115, 237, 381, 379, 380] [73, 238]. DV's Snapchat product combines its exposure data with Lumen eyes-on-screen signals [116, 117]. Adelaide says it licenses gaze reference data from Amplified, Lumen, Tobii, TVision and Viomba [169, 327]. By this paper's inference, a few reference panels sit under many products, and a verification bundle can carry a specialist's data. No buyer of a specialist in the table was a verification company [87, 80, 82, 83, 465, 86].

By the registered rule, H4 is unresolved. Reads of the DV and IAS filings, including DV's preliminary merger proxy, UK accounts for Lumen and DAIVID, and releases from Adelaide, Amplified and Neurons found no filing, revenue or pricing series, or buyer-level evidence, linking bundling to a change in specialists' demand, pricing or revenue. That holds for the sources screened by 28 September 2026 [47, 231, 285, 232]. Acquisitions do not settle the question, and Lumen's accounts do not tie its losses to demand. What would settle it is a specialist's revenue or pricing series over time, or buyers documented switching because of a bundle. The closest public series are the verification companies' own filings: IAS no longer files, and DV is expected to stop if its merger closes [83, 84].



## 6. How attention is traded


*In the documented record, attention is traded in the US mainly as a targeting and bidding input inside buying platforms; guarantees and priced offers are few, and neither priced offer documents US sales. Where a guarantee names who measures delivery, it is the company that sells the metric, and no offer discloses its volume against a market total. On the registered rules, H2 is supported with its scale undocumented, and no metric-environment pair qualifies as a currency.*

> **In short.**
> Attention is marketed as a live optimisation input in named buying platforms: Adelaide segments in The Trade Desk and Amazon DSP, and an Adelaide bidding algorithm in DV360 [91, 92, 334, 93, 25] [330, 159]. DoubleVerify offers attention segments and Scibids optimisation without naming a platform [222, 213]. Guarantees are rarer. The earliest found is A+E's for Budweiser in 2020, on TVision panel data [475, 89, 476, 24]. Among the guarantees, only the Kargo case publishes a numeric threshold, a minimum average of 25 AU [93, 91, 94, 97, 95] [475, 89, 476, 24]; Impact Media's billing suite publishes seconds seen instead [101]. Where a guarantee names who measures delivery, it is the metric's own vendor, and this review found no independent audit of delivery [94, 93, 91, 66]. Two models bill on attention, and neither documents US sales [392, 96, 566]. No offer discloses US spend, impressions, campaigns or buyers against a denominator and period, so H2 is supported with its scale undocumented: the size of the trade is unknown, not shown to be small [91, 92, 334, 93, 25]. On MRC's list of 28 September 2026, the only accredited attention metric, DoubleVerify's, has no guarantee or billing use found, and none of the guaranteed or priced metrics is documented as accredited. No metric-environment pair codes as a currency [2] [63, 222].

### Five uses and two attributes

This chapter settles two registered tests: H2, whether attention is used in trading and not only in planning, and Q1, whether it is a currency or a diagnostic. The H2 rule sorts each documented use into one of five classes. Planning uses attention to choose where to buy. Reporting describes a campaign after it runs. Optimisation feeds the metric into targeting or bidding. A billing unit sets the price per unit of attention. A guarantee promises a minimum level. Two attributes are recorded separately: whether anyone outside the sale audits delivery, and what the buyer receives if the minimum is missed. The IAB and MRC guidelines encourage independent third-party auditing of attention methods used in buying and selling, without requiring it [41, 425].

Digiday reported in March 2024 that media groups at five holding companies had signed on to Adelaide's Flight Control, a planning tool that sets a target AU and flags media below an AU floor [333]. AU, Adelaide's attention unit, is a 0 to 100 score from a machine-learning model that blends media-quality signals with outcome data [43, 44, 564, 45, 73, 433]. [Chapter 7](#buyers) covers the buyers' side.

### Optimisation carries most of the trading use

AU has the most named buying routes. Yahoo launched AU pre-bid segments in its DSP (demand-side platform, the software buyers bid through) on 13 June 2023 [334]; the latest status evidence is Adelaide's own June 2026 statement [334]. The Trade Desk has offered low, medium and high attention tiers since 8 October 2024, and Amazon DSP AU tiers and an AU Quality Floor since 9 June 2026. Neither release states geography or price [161, 162]. The Trade Desk's tiers are relative ranks of its inventory by Adelaide's score [161]. Google announced AU custom bidding for DV360 (Display & Video 360) at the May 2025 NewFronts. By 29 September 2026 its Help Center documented live steps for the Maximize Adelaide AU algorithm, with an additional media fee; the go-live date is not documented [330, 159].

Sellers package attention into curated deals. Adelaide's High Attention Markets launched on 7 March 2023 with Magnite, PubMatic, Xandr, DV360, Teads, Sovrn, TRUSTX and Sharethrough, in 20+ countries including the US. Comscore-Certified Deal IDs with AU followed on PubMatic in June 2025 and AU reached Index Marketplaces in January 2026. No curated deal found discloses pricing or uptake [183, 165, 166].

DoubleVerify launched pre-bid attention segments on 4 May 2023 and, with Scibids, the DV Algorithmic Optimizer for attention bidding on 19 June 2023 [222, 213]. DV acquired Scibids in August 2023 [47, 207]. Several DSPs named then no longer operate as named, and DV's 2026 product page names no platform [222, 213]. On 29 September 2026 MRC accredited no DoubleVerify pre-bid attention offering [42, 2]. Lumen Attention Bidding is a pre-bid custom algorithm that plugs into the buyer's DSP [392, 427, 20, 428], with no DSP or price named [392, 96, 566]. Havas Media Network and Viant announced an Attention Marketplace on 24 June 2026 that scores inventory in Viant's DSP by attention decile [310].

Every AU route bids on a model estimate: Adelaide's segments are built from historical placement-level AU [161, 162]. DoubleVerify's attention indices are built from device signals, which this paper codes as a candidate proxy, extending the Q2 rule's clause for dwell, time in view, audibility and screen coverage [43, 44, 564, 45, 73, 433] [10, 63, 73, 74, 9]. The outcome figures attached to the launches are vendor-reported, with no cost per incremental outcome ([chapter 3](#validation)) [161, 213].

### Guarantees

A guarantee promises the buyer a minimum attention level. The earliest documented case is A+E's guarantee to dentsu for AB InBev's Budweiser in Q4 2020, repeated in Q1 2021. It rested on TVision panel data counting eyes on screen for at least two seconds, and no guarantee level, remedy or independent audit is disclosed [475, 89, 476, 24]. The advertiser described it as based on people being present in the room, so whether it guaranteed presence or attention is ambiguous ([chapter 7](#buyers)) [24]. The account comes from a trade article published in an A+E-branded channel [24].

The earliest case with published terms is an Adelaide case study of a 2022 campaign run by MediaCom (now EssenceMediacom) with Kargo on DV360 private marketplaces (PMPs). The parties agreed a minimum average AU of 25, and Kargo served added-value PMP impressions when the average dipped below it. The sources call it an AU guarantee and do not call it contractual [93]. Among the guarantees found, it is the only one with a public numeric threshold [93, 91, 94, 97, 95] [475, 89, 476, 24].

The rest publish less. Adlook's AU-CPM buying model (12 April 2023) was described as the first DSP offer of guaranteed media quality on attention metrics; its status at the cutoff was not checked [97, 98, 78, 95]. TripleLift announced Guaranteed AU deals for CTV on 13 September 2023, certified by Adelaide, with no threshold, remedy or price, and no later TripleLift source was opened to confirm they are still sold [94]. The Wall Street Journal offered an Adelaide AU guarantee for qualified campaigns with impressions measured between 15 October and 15 November 2024, timed to the US election, with a makegood of added-value impressions for any miss. All five participating clients, running six campaigns, exceeded the benchmark, by the publisher's account. Adweek called it a promotion, and no numeric threshold is readable [90, 91, 92, 167].

Teads presented Guaranteed Attention at the May 2025 NewFronts. It describes Attention-to-Outcome Guarantees as a pilot and Effective Attention Guarantees as a beta, both combining AU with Teads' own benchmarks, engagement and purchase data, with no threshold, remedy or live date [95]. Its FY2025 10-K and Q2 2026 10-Q do not mention an attention guarantee [96, 471].

Vevo announced on 27 January 2026 that, with Adelaide, it had launched Attention Guaranteed, which "secures a minimum attention score" across CTV, mobile and desktop. The release states no threshold, remedy, audit or price, and does not say who measures compliance [78]. It does not state the unit of the minimum; reading it as AU is an inference from the partnership [78, 94, 48]. By April 2026 Vevo described the offer as a way to understand engagement [324], and on 29 September 2026 its page for advertisers did not mention it. No retirement announcement was found [318, 473, 417].

SeenThis launched SeenThis Attention on 30 June 2026: a video offer measured by Lumen Research and bought on attentive CPM (aCPM) or cost per attentive second (CPAS). An attentive-outcome threshold is agreed before launch, and the product page promises extra delivery at no cost if it is missed. The release says it is available globally. US sales, the threshold values and any independent audit of settlement are not documented [99, 100].

Adelaide announced on 15 September 2026 that Duolingo is launching AU Guarantees, described as a contractual commitment to a minimum level of media quality across its in-app inventory, with no threshold or remedy stated [97]. Adelaide says several members of its AU Ecosystem offer guaranteed minimum AU scores but names only the Journal, so the rest cannot be classified [167].

### Attention as a price

Two models bill on attention. SeenThis prices on aCPM or CPAS [99, 100]. Impact Media announced a Guaranteed Attention Suite from New York on 18 April 2024 that bills clients only when a display ad is seen for one second or a video ad for three seconds. The release does not say how "seen" is measured, and this review found no later source on the suite's availability. If "seen" means visible, the suite bills on time in view, the opportunity-to-see layer, and the attention label does not fit [101].

Viant announced an attention-adjusted CPM on 15 April 2026 and said in May that its platform would be able to deliver a metric pricing CTV and linear inventory on TVision attention. Viant has owned TVision since 1 May 2026, so the platform that would set the price also owns the metric [82]. An EDGAR full-text search finds no later filing that mentions the attention-adjusted CPM, to 29 September 2026. On 10 August 2026 Viant said it had begun testing TVision pre-bid attention intelligence. The metric has been announced but not launched, and its billing role is undocumented [308, 219, 472, 309, 304] [308, 219, 309, 474].

Beyond these, this review found no attention-priced billing unit in the sources screened by the cutoff. Teads' filings list CPC, CPM, CPV, CPA and other models, none defined on attention [392, 96, 566].

The table lists every guarantee and attention-priced offer found. Viant's attention-adjusted CPM, announced but not launched and with its billing role undocumented, is left out and coded under Q1 below [308, 219, 472, 309, 304]. "None found" is bounded to the sources screened by the cutoff.

| Offer | Seller | Metric vendor | Threshold public? | Remedy public? | Independent audit? | Status | Date |
|---|---|---|---|---|---|---|---|
| Attention guarantee on a Budweiser buy | A+E, for dentsu and AB InBev | TVision (presence or attention ambiguous) | No | No | None found (measurer not stated) | Completed; repeated in a second wave | Q4 2020 and Q1 2021 [475, 89, 476, 24] [24] |
| Guaranteed AU campaign | Kargo, for MediaCom (now EssenceMediacom) | Adelaide | Yes: minimum average 25 AU | Yes: added-value PMP impressions | None found (Adelaide dashboards) | Completed, one campaign | 2022 [93] |
| AU-CPM buying model | Adlook (DSP) | Adelaide | No | No | None found | Status at cutoff not checked | 12 April 2023 [97, 98, 78, 95] |
| Guaranteed AU deals, CTV | TripleLift | Adelaide | No | No | None found (certified by Adelaide) | Announced; later status not verified | 13 September 2023 [94] |
| Guaranteed Attention Suite | Impact Media | Not stated | Yes: 1 second display, 3 seconds video; method not stated | Not applicable: bills only when seen | None found | Later availability not documented | 18 April 2024 [101] |
| Election AU guarantee | The Wall Street Journal | Adelaide | No | Yes: makegood of added-value impressions | None found (Adelaide scores campaigns) | Completed; five clients, six campaigns | 15 October to 15 November 2024 [90, 91, 92, 167] |
| Attention-to-Outcome (pilot) and Effective Attention (beta) Guarantees | Teads | Adelaide, with Teads benchmarks and purchase data | No | No | None found | Pilot and beta; no later status found | May 2025 [95] |
| Attention Guaranteed | Vevo | Adelaide (unit of the minimum not stated) | No | No | None found | Launched per Vevo; last marketed April 2026 | 27 January 2026 [78] [324] |
| SeenThis Attention (aCPM, CPAS) | SeenThis | Lumen | No: agreed per campaign, values not published | Yes: extra delivery at no cost | None found (verified by Lumen) | Launched, offered globally; US sales not documented | 30 June 2026 [99, 100] |
| AU Guarantees, in-app | Duolingo | Adelaide | No | No | None found | Announced as launching | 15 September 2026 [97] |

![10 offers guarantee or price attention; none documents an independent audit of delivery.](figures/fig-guarantees.svg)

*Figure fig-guarantees. **10 offers guarantee or price attention; none documents an independent audit of delivery.** Every attention guarantee or attention-priced offer found, by what its public terms state. Two publish a numeric threshold (Kargo’s campaign, at an average of 25 AU, and Impact Media’s seconds seen); three state a remedy, each a form of extra delivery. SeenThis rests on Lumen’s measure, A+E’s guarantee on TVision panel data and Impact Media’s suite on a measure its release does not name; the rest use Adelaide’s AU, Vevo’s by inference. Units: categories. Scope: offers documented by the cutoff; US availability varies (see chapter 6). Sources [24], [89], [90], [91], [92], [93], [94], [95], [96], [78], [97], [98], [99], [100], [101]. Caveat: Contract terms are not public; "not stated" means not stated in the documents found, bounded to the sources screened. Data: data/figure-data.json#guarantees*

### Who checks delivery, and who pays

Where a guarantee discloses who measures delivery, it is the vendor whose metric the offer uses: Adelaide for the Journal, Kargo and TripleLift, and Lumen for SeenThis. For Vevo, Duolingo and Teads, Adelaide's role is implied but not stated, and for A+E's guarantee on TVision data the measurer is not stated. Most guarantees found are defined on Adelaide's AU; the exceptions are SeenThis and A+E. This review found no independent audit of guarantee delivery in the sources screened [94, 93, 91, 66]. The third-party reviews found were requested by the vendors. MediaSense's 2025 methodology review of AU, which Adelaide requested, worked from aggregated data Adelaide supplied and is not an accreditation [66, 22]. PwC's review of Lumen's campaign analysis, commissioned by Lumen, worked on a Lumen-supplied subset of data [8]. On 29 September 2026 MRC listed AU as pre-audit reviewed in May 2026, with the audit not begun [48, 78].

The few public statements on who pays point to sellers. Adelaide said in October 2025 that Condé Nast and Hearst cover the cost of AU measurement on their inventory, and in April 2024 it launched a free audit tier for publishers; no fee level is disclosed [167]. No buyer-side source screened states who pays, what it costs, or a guarantee level, remedy or audit clause for an attention-based buy [477, 333]. Inference: where the seller pays for measurement and the measurer arbitrates delivery, no party outside the transaction checks the result. The sources contain no evidence of biased grading.

### Coding H2 and Q1

On the registered rule, H2 is supported, with its scale undocumented. The rule needs one live US offer in any of the three trading classes, and optimisation in named buying platforms meets it. Yahoo launched AU pre-bid segments in its DSP with Verizon as brand launch partner [334], and Adelaide said in June 2026 that the targeting is available there [334]. iFIT bought Adelaide segments in The Trade Desk in 2025 [67]. JLR North America described custom bidding on AU in DV360 in May 2025 [25], and DV360's Help Center now documents a live Adelaide algorithm with no go-live date [330, 159]. The Trade Desk and Amazon DSP releases themselves state no geography [161, 162]. Guarantees do not carry the verdict on their own. The Journal's ran from October to November 2024, but the sources do not say its advertisers were in the US, and Adweek called it a promotion [90, 91, 92, 167]. A+E's ran in the US in 2020 and 2021 on terms not disclosed, and may have guaranteed presence rather than attention [475, 89, 476, 24] [24]. DoubleVerify's 2026 page offers attention segments and Scibids optimisation without naming a platform [222, 213], and Viant, whose DSP hosts the Havas marketplace, called its TVision attention targeting "testing" in August 2026 [308, 219, 309, 474]. Apart from JLR's account to a trade publication, all of it rests on vendor or seller statements [91, 92, 334, 93, 25]. "Supported at scale" also needs disclosed US spend, impressions, campaigns or buyers with a denominator and period. No offer discloses them, so the size of the trade is unknown, not shown to be small [91, 92, 334, 93, 25]. The counts that are disclosed have no denominator: the Journal's five clients running six campaigns, Kargo's one campaign [78, 95, 94, 92, 93], and 42 pilot campaigns reported by Viant's CEO with no design, control or cost [304]. No guarantee states threshold, remedy and independent audit together, so none is currency-grade [91, 92, 334, 93, 25].

Q1's rule codes each metric-environment pair on four conditions: accreditation for the environment traded, a billing or guarantee use with stated threshold and remedy, independent audit, and disclosed adoption. A pair is "currency" with all four and "partial currency" with accreditation plus a guarantee or billing use; otherwise it is "diagnostic". Unreadable accreditation or trading terms make it "unresolved" [63, 222]. The registered rule does not say how to code an offer that publishes only part of its terms. This paper codes such offers on what can be read, which places the Journal and SeenThis pairs in "diagnostic" rather than "unresolved"; neither reading produces a "currency" or "partial currency" pair. The table covers every pair that is accredited or has a documented guarantee, billing or announced pricing use, whatever its terms. Pairs with neither, such as AU segments in The Trade Desk, are diagnostic and left out. The coding is this paper's inference from the claims in each row.

| Metric and environment | Accredited there? | Guarantee or billing use | Threshold and remedy public | Independent audit | Adoption with denominator | Code |
|---|---|---|---|---|---|---|
| DoubleVerify Authentic Attention: desktop, mobile web, mobile in-app (three pairs) | Yes (MRC, April 2026); pre-bid not accredited | None found; optimisation only | Not applicable | Not applicable | No | Diagnostic [42, 205] [42, 2] |
| Adelaide AU: Journal inventory, screens not stated | No: pre-audit reviewed 5/2026, audit not begun | Guarantee, 2024 | Remedy only | None found | No | Diagnostic [48] [90, 91, 92, 167] |
| Adelaide AU: Kargo PMPs, screens not stated | No | Guarantee, 2022 | Both | None found | No | Diagnostic [48] [93] |
| Lumen: SeenThis video, screens not stated | No MRC entry | Guarantee and price | Remedy only | None found | No | Diagnostic [2, 48, 3] [99, 100] |
| TVision data: A+E inventory for Budweiser, screens not stated | No | Guarantee, 2020 and 2021; presence or attention ambiguous | Neither | None found | No | Unresolved [475, 89, 476, 24] [24] [2, 302, 50, 48] |
| Adelaide AU: Vevo CTV, mobile and desktop; TripleLift CTV; Duolingo mobile in-app; Teads, screens not named; Adlook, environment not stated (seven pairs) | No | Guarantees or AU-CPM model; terms unpublished, pilot, beta or announced | Neither | None found | No | Unresolved [78, 48] [94] [97] [95] [97, 98, 78, 95] |
| Impact Media's unnamed measure: display, video (two pairs) | Cannot be read | Billing unit | Threshold only | None found | No | Unresolved [101] |
| TVision attention: Viant CTV, linear (two pairs) | No | Announced price; billing role undocumented | Neither | None found | No | Unresolved [2, 302, 50, 48] [308, 219, 472, 309, 304] |

No pair codes "currency" or "partial currency". Of the pairs in the table, six code "diagnostic" and twelve "unresolved"; the pairs left out are also diagnostic and are not counted here. The registered rule codes an unaccredited metric "diagnostic" even when it prices media, as SeenThis's aCPM and CPAS do [99, 100], so the code records missing assurance, not the absence of a billing function. On 28 September 2026 MRC's list had one entry under Attention Metrics, DoubleVerify's, and no accredited attention product from Adelaide, Lumen, TVision or Amplified [2]. No guarantee or billing use of DoubleVerify's accredited pairs was found [63, 222]. Neither AU nor TVision is accredited, so their unresolved pairs could not reach either higher code whatever their terms turn out to be [78, 48] [2, 302, 50, 48]. On MRC's lists of 28 and 29 September 2026, the accredited metric and the guaranteed or priced metrics are different metrics.

Three changes would move pairs up. MRC accreditation of AU for the environments traded would lift the Journal and Kargo pairs to "partial currency", and accreditation of Lumen's metric would do the same for the SeenThis pair, since each carries a guarantee with readable terms [48] [2, 48, 3]. A guarantee or billing use on DoubleVerify's accredited metrics would do the same from the other side. None was found by the cutoff, and the CIMM and IAB playbook advises buyers against treating attention as a currency [438]. AdCP v3.1.24, published on 23 September 2026, can already express an attention floor paired with a makegood, as schema only ([chapter 10](#agents)) [567] [144, 138, 140, 142]. [Chapter 12](#next) turns the missing terms into questions a buyer can put to a seller.



## 7. What buyers do with it


*In the buyer records found, planning is the most common use of attention. Every major holding group has a documented attention partnership and several announced optimisation. Buyer-side sources record one executed guarantee, from 2020 and 2021, and no billing-unit use, contract term or statement of who pays; the one guarantee an agency received with a public threshold and remedy, Kargo's for MediaCom in 2022, comes from the metric vendor's case study. IAB's surveys of US buyers record intent that rose once and then held flat, and Haleon describes attention as a check on media quality.*

> **In short.**
> Digiday reported in March 2024 that media groups at five holding companies, not including dentsu, had signed on to Adelaide's Flight Control, a planning tool [333], and planning is the most common use in the buyer records found [484, 476]. Optimisation was announced by dentsu in the UK in 2022 [486], Havas in the US in 2023 [484, 476] and IPG Mediabrands in 2024 [477, 476]. Xaxis Switzerland was reported in February 2023 to optimise media purchases on Adelaide data [476, 490]. Nespresso and Mars tested Adelaide's attention deals on Xandr Curate [26] [492, 26]. The only executed guarantee found in buyer-side sources, A+E's for Budweiser in 2020 and 2021, states no level, remedy or audit [475, 89, 476, 24]; Kargo's 2022 guarantee to MediaCom, a minimum average of 25 AU with added-value impressions as the remedy, comes from Adelaide's case study [93]. No buyer-side source screened states who pays or a contract term [477, 333]. Haleon's effectiveness lead wants proof that attention drives incremental outcomes [441]. In IAB's US buy-side surveys, the share expecting to focus more on attention was 36% for 2023 (n=223), 47% for 2024 (n=203) and 47% for 2025 (n=200): intent rose once and then held flat [103, 102]. H2 stays supported with its scale undocumented, and no buyer record codes as a currency [91, 92, 334, 93, 25] [63, 222].

### Five groups at the cutoff

WPP renamed its media company from GroupM to WPP Media on 28 May 2025 [478]. Omnicom completed its merger with IPG on 26 November 2025, and its FY2025 10-K does not name IPG Mediabrands or KINESSO [479, 480]. The status of IPG Mediabrands' Adelaide deal after the merger was not found [477, 476]. Havas N.V. listed on Euronext Amsterdam on 16 December 2024 after its spin-off from Vivendi [481]. In February 2026 dentsu said it was no longer considering a sale of its international business [482].

Each group has at least one documented attention partnership, mostly with Adelaide or Lumen [333] [487, 476]. The table lists what public sources document; private practice is invisible to this review.

| Group at the cutoff | Attention vendors named | Uses documented | Latest record found |
|---|---|---|---|
| WPP Media (GroupM until May 2025) | Adelaide, Lumen | Planning (Flight Control, 2024; Open Intelligence, 2025); optimisation (Xaxis Switzerland, reported 2023); a guarantee received by MediaCom from Kargo (2022) | Open Intelligence, 3 June 2025 [333] [476, 490] [93] |
| Omnicom (with IPG's agencies since 26 November 2025) | Amplified, Lumen, Adelaide | Planning (OMD since 2021; Lumen benchmarks in Omni from May 2024); optimisation announced (IPG Mediabrands, April 2024) | Lumen benchmarks in Omni, May 2024; the next-generation Omni (January 2026) is described without attention [343, 476, 278, 489] [477, 476] [490, 488, 479] |
| Publicis | Adelaide | Planning (Flight Control, 2024); reporting (Haleon, US, mid-2023; Starcom for Visa, Germany, 2023) | Flight Control sign-on, 2024; no Publicis-branded attention product found [491, 476, 333] |
| dentsu | Lumen, TVision, Amplified | Planning; optimisation (Google custom algorithm for Co-op, 2019; Avocet DSP, reported 2022; UK, 2022); a guarantee (A+E for Budweiser, 2020 to 2021) and guarantee testing (2023) | Global Media Planner, 18 March 2026 [486] [487, 476] [488] [475, 89, 476, 24] [278, 9] |
| Havas | Lumen, Adelaide; Viant (TVision named as a component) | Planning (2022); optimisation (Avocet DSP, reported 2022); optimisation announced (US, 2023; Viant marketplace, 2026); commissioned research (2024) | Havas and Viant Attention Marketplace, 24 June 2026 [483] [278, 9] [484, 476] [485] [310] |

### Planning is the common use

Digiday reported on 12 March 2024 that Havas Media Network, GroupM, IPG Mediabrands, Omnicom Media Group and Publicis Media had signed on to Flight Control, which outputs a target in AU (Adelaide's attention unit) and flags media below an AU floor for reduction. No terms, fees or payer were reported, and signing on shows neither scope nor continuing use [333]. AU is a 0 to 100 score from a machine-learning model that blends media-quality signals with outcome data [43, 44, 564, 45, 73, 433], so an AU planning target is a model estimate.

OMD has planned with attention since 2021, first with Amplified Intelligence. In February 2023 OMD's CEO said OMD had applied attention in 11 markets, 13 categories and 25 brands, and called it a planning metric; the count has no denominator [343, 476, 278, 489]. Havas's Converged platform added a planner in November 2022 that uses Lumen eye-tracking-panel data to score over 9,000 domains by format for the average attention they are likely to deliver [483]. On 18 March 2026 dentsu launched Global Media Planner across 15 markets, powered by an attention dataset built with Lumen; US inclusion is not stated [488]. WPP Media's Open Intelligence launch on 3 June 2025 lists Adelaide and Lumen Research as partners, and Adelaide's CEO says AU is embedded in WPP's Large Marketing Model; how, and whether in the US, is not documented [476, 490].

Inference: in the sources screened, holding-company attention announcements moved from standalone vendor partnerships in 2021 to 2024 to inputs inside broader planning platforms in 2025 and 2026 [490, 488, 479], with the Havas and Viant marketplace of June 2026 as an exception [310]. Announcement volume follows PR choices and does not measure use.

Of the groups' annual filings, only Omnicom's and Publicis's were searched, and both are silent. A full-text search of Omnicom's FY2025 Form 10-K found no reference to attention metrics, attention vendors or viewability [479], and Publicis's 2025 Universal Registration Document, filed 30 April 2026, contains no attention-metric reference [491, 476, 333]. WPP's, dentsu's and Havas's annual filings were not searched [477, 333]. Filings rarely name measurement vendors, so the silence says nothing about operational use [479].

### Optimisation is announced more often than shown live

dentsu's optimisation record is the oldest found. Dentsu Aegis Network bought for Co-op through a Google custom algorithm using Lumen data in 2019, and dentsu and Havas were reported in 2022 to use Lumen signals in the Avocet DSP (demand-side platform, the software buyers bid through) [278, 9]. On 7 April 2022 dentsu UK&I launched Effective Attention Forecasting with Lumen, a predictive model that optimises programmatic campaigns in the UK only [486].

Havas Media North America and Adelaide announced the Meaningful Ad Unit for US clients on 19 June 2023, for planning and for activation through custom algorithms, pre-bid segments and custom private marketplaces (PMPs, invitation-only auctions), with no buying platform named [484, 476]. Xaxis Switzerland, then part of GroupM, was reported in February 2023 to use Adelaide data in an algorithm that optimised media purchases [476, 490]. IPG Mediabrands' April 2024 Adelaide partnership covers planning, measurement, in-platform custom bidding and auction packages delivered through supply-side platforms, without naming the bidding platform [477, 476].

Brand records name platforms more often. Nespresso began optimising to AU as a KPI (key performance indicator) in Q2 2023 and was testing Adelaide's high-attention, low-emissions PMPs built on Xandr Curate; the geography is not stated [26]. Mars was testing the same PMPs [492, 26]. Xandr's DSP was retired on 28 February 2026, and any AU route through Microsoft Curate is undocumented at the cutoff [183, 164, 165, 166, 1]. JLR (Jaguar Land Rover) North America told AdExchanger in May 2025 that it uses Adelaide's AU for planning, optimisation and negotiation, and described custom bidding on AU in DV360 (Display & Video 360, Google's DSP) [25]. The newest agency record names one too: the Havas and Viant Attention Marketplace, announced on 24 June 2026, scores inventory in Viant's DSP by attention decile for pre-bid targeting [310]. Viant described its TVision-based pre-bid targeting as testing in its 10 August 2026 release and as integrating in its 10-Q and a September prospectus supplement [308, 219, 309, 474].

The agency announcements from 2022 to 2024, by dentsu UK, Havas, Xaxis Switzerland and IPG Mediabrands, name no buying platform in the text read, and none is confirmed live at the cutoff [484, 476]. Every AU-based route bids on a model estimate of placement quality [43, 106, 1]. The performance claims attached to agency announcements, such as Havas's, have no published comparator or holdout [484, 476], so under the registered rule they leave H1b unresolved ([chapter 3](#validation)).

### Guarantees stayed at the pilot stage

Buyer-side sources record one executed guarantee, which TVision, its data supplier, calls an attention guarantee [475, 89, 476, 24]. A+E guaranteed dentsu's Q4 2020 Budweiser buy for AB InBev and repeated it in Q1 2021. It rested on TVision panel data counting eyes on screen for at least two seconds, was optimised in flight in A+E's serving stack, and was priced on the principle that attentive inventory costs more [475, 89, 476, 24]. The advertiser described it as based on people being present in the room, so whether the guaranteed metric was presence or attention is ambiguous [24]. Presence would sit in the opportunity-to-see layer, below attention.

The rest are aims, tests and unnamed claims. In May 2021 AB InBev said it wanted to guarantee attention the way it guarantees in-demo reach [475, 89, 476, 24]. In February 2023 dentsu told Digiday it was testing attention guarantees [487, 476], and Adelaide's CEO said an unnamed agency offered them [475, 89, 476, 24]. In March 2023 Xaxis said it could offer guaranteed attention-based outcomes on its curated Adelaide marketplace, without terms [183, 477]. None of these discloses a guarantee level, remedy or independent audit [475, 89, 476, 24].

One seller guarantee had an agency buyer and a public threshold. In a 2022 campaign on DV360 PMPs, documented in an Adelaide case study, Kargo and MediaCom (now EssenceMediacom) agreed a minimum average of 25 AU, with added-value impressions served when the average dipped below it, monitored in Adelaide's dashboards [93]. As in every guarantee that names who measures delivery, the metric's own vendor did so ([chapter 6](#trading)); A+E's guarantee does not name its measurer [94, 93, 91, 66].

### Terms and payers

No buyer-side source screened states who pays for attention measurement, its fee, or a specific contract term (guarantee level, remedy, audit or make-good) for an attention-based buy. Buyers describe commercial principles instead. AB InBev and A+E held in 2020 and 2021 that attentive inventory should cost more and that the efficiency could be shared, JLR uses attention scores to negotiate media premiums (2025), and Haleon flags CPM (cost per thousand impressions) premiums on the highest-attention CTV (connected TV) inventory (2024) [477, 333]. The absence is bounded to the holding-company releases, filings, trade reports and brand interviews screened by the cutoff [477, 333].

Platform and seller documents show where fees arise without stating their level. By 29 September 2026 Google's DV360 Help Center documented an additional media fee for its Maximize Adelaide AU algorithm [330, 159]. From 25 March 2024 Lumen measurement could be switched on and billed per ad group inside The Trade Desk, at a fee not stated [275]. On the sell side, Adelaide says Condé Nast and Hearst cover the cost of AU measurement on their inventory [167]. Inference: the buyer pays through the platform in the first two routes and the seller pays in the third.

### What brands say

Haleon treats attention as a check on media quality and asks for outcome proof. With its agency Publicis, it evaluated Amazon DSP inventory for Flonase in the US in mid-2023 with Adelaide AU, treating 60 or more as high for CTV and under 20 as low. AdExchanger paraphrased Haleon's position in February 2024 as attention being just another tool for assessing media quality. Haleon's effectiveness lead, Phil Jackson, said "we need to prove [that] attention drives incremental business outcomes", and warned that the highest-attention CTV carries very high CPMs [441].

Mars said in March 2020 that it had not bought media on attention. In May 2023 its senior director of global media told AdExchanger that Mars had seen promising signs of a high correlation between attention scores and sales, across tests with Adelaide, Realeyes, Lumen, IAS (Integral Ad Science) and DoubleVerify. Its layering of attention metrics onto Nielsen Catalina Solutions sales-lift panels tested a link between publishers' sustainability scores and attention scores, and did not test sales. No design is published [492, 26]. Mars's stated aim in 2023 was attention measurement as a pre-buy option in DSPs [26], so its record combines diagnostic testing with a buying goal.

On cost, The Trade Desk's own publication reported in October 2025 that iFIT, buying Adelaide segments in The Trade Desk, achieved a 61% increase in aided awareness with no stated comparator, and iFIT's executive said outcome lifts were needed to justify the higher costs [67].

### Surveys measure intent, not spend

In the US buy-side Outlook surveys of the IAB (Interactive Advertising Bureau), the share expecting to focus significantly or somewhat more on attention metrics was 36% for 2023 (n=223, fielded October to November 2022), 47% for 2024 (n=203, fielded October to November 2023) and 47% for 2025 (n=200, fielded November to December 2024). The share rose once and then held flat, while the N/A share fell from 24% to 19% to 14% [103, 102]. Intended focus cannot be turned into use, spend or a market share. IAB is a trade body whose members include sellers and measurement vendors; it has no direct stake in attention vendors but is not free of interest [103, 102].

![Intent to focus on attention rose once, from 36% to 47% of US buy-side respondents, then held at 47%.](figures/fig-surveys.svg)

*Figure fig-surveys. **Intent to focus on attention rose once, from 36% to 47% of US buy-side respondents, then held at 47%.** Share of IAB Outlook respondents (US buy-side ad investment decision-makers) answering that they expected to focus significantly or somewhat more on attention metrics, fieldwork late 2022, late 2023 and late 2024. The question measures intent, not use or spend. Units: % of respondents. Scope: US buy-side respondents, IAB Outlook. Sources [102], [103], [104]. Caveat: Three samples of about 200; no margin of error is published. Not a market-wide share of spend. Data: data/figure-data.json#surveys*

Surveys run or relayed by vendors report higher shares, or the same number for a different question, and the releases read do not publish their questions. IAS's Industry Pulse preview publishes no question wording or base per statistic [27]. IAB's 47% and DoubleVerify's 47% differ in population, wording and geography [371].

| Survey | Sponsor | Sample | What it measures | Result |
|---|---|---|---|---|
| IAB Outlook, fielded late 2022, 2023 and 2024 | IAB (trade body) | US buy-side decision-makers; n=223, 203, 200 | Intended focus | 36%, 47%, 47% expect to focus more [103, 102] |
| DoubleVerify Global Insights, June 2024 | DoubleVerify, survey by Sapio | 1,000 advertisers, four regions; base of the 47% not given | Planned integration | 47% of media buyers planned to integrate attention metrics in 2024 [371] |
| IAS, The Attention Payoff, November 2023 | IAS, with YouGov | 200+ US digital media experts, buy and sell side | Use in some capacity | 88% use attention measurement; 36% use a third-party vendor [27, 493, 494] |
| IAS, 2026 Industry Pulse, December 2025 | IAS, with YouGov | Nearly 300 US media experts, buy and sell side | Opinion on social media metrics | 77% name attention and 85% viewability as essential [27, 493, 494] |
| ARF (Advertising Research Foundation) Attention Report, relayed by TVision, 2022 | ARF with the Attention Council; funding unknown | Qualitative study of 63 participants; buy-side subsample under 63 | Future consideration | Over 75% of buyers would consider attention for planning and buying [475] |
| ARF figure cited in the playbook of CIMM (Coalition for Innovative Media Measurement) and IAB | ARF | Not stated | Use | About 50% of advertisers and 60% of agencies; no sample, date or question [438] |

### What the buyer record settles

H2 asks whether attention is used in trading and not only in planning and reporting. The registered rule needs one live US use in a named buying platform, a billing unit or a contractual guarantee. Read on its own, the buyer record meets it only narrowly. JLR's AU custom bidding in DV360 rests on a 2025 trade report [25]. The Havas and Viant marketplace was announced in June 2026 and described by Viant as testing in August [310] [308, 219, 309, 474]. The A+E guarantee was live in 2020 and 2021, though its guaranteed metric may have been presence rather than attention [475, 89, 476, 24] [24]. No buyer record is a billing unit, and none discloses US spend, impressions, campaigns or buyers with a denominator [484, 476]. On the seller offers in [chapter 6](#trading), the verdict stands: H2 is supported with its scale undocumented, and no use is currency-grade [91, 92, 334, 93, 25]. The buyer-side reading is this paper's inference from the records above.

Q1 codes each metric-environment pair on the four currency conditions set out in chapter 6. The buyer records add none of the four conditions beyond what chapter 6 coded, and none codes as currency [63, 222]. The only executed guarantee found in buyer-side sources, A+E's on TVision data, states no threshold or remedy, so chapter 6 codes it "unresolved" [475, 89, 476, 24]. In this paper's reading, buyers' own descriptions fit the diagnostic code [343, 476, 278, 489] [441]. The CIMM and IAB playbook of November 2025 warns buyers against treating attention as a currency [438].



## 8. Where it works


*Attention measurement is live wherever the ad sits on a screen: display, online video, connected TV (CTV) and social each have at least three cohort A products documenting it. Audio, digital out-of-home (DOOH) and retail media are thin, and in-game rests on one-off studies. The covered channels differ in kind: among the live products only TVision observes people, only on the TV screen, and several CTV models learn from its panel.*

> **In short.**
> Under the registered Q3 rule, display, online video, CTV and social are covered, audio and DOOH are thin, and in-game is study only [105, 119, 42, 45, 115, 110, 113] [113, 29]. CTV is covered on different bases. TVision observes people in panel homes; Adelaide, DoubleVerify and Lumen score CTV with models, and all three have documented TVision panel data as an input, DoubleVerify only in its June 2023 partnership announcement [45, 106, 220, 110]. Platform integrations for social attention run through Snap and TikTok [116, 118, 117]. In the sources screened by the cutoff, no attention product was in Meta's or YouTube's measurement partner programmes, held a partner badge or had a certified partner integration for attention. xpln.ai states that it measures Facebook and Instagram through the platforms' reporting interfaces or API, without a granular-data partnership (vendor disclosure) [42, 568, 569, 570, 109] [571, 121, 496]. None of the vendors that attach attention metrics to audio documents observing a listener's auditory attention to an ad [107, 105, 120, 112]. Retail media, an unscored attribute, is thin on one vendor's assertion [162, 260]. On 28 September 2026, the only accredited attention metrics covered display and video in desktop, mobile web and mobile in-app [2] [42].

### The rule and the count

The registered Q3 rule codes a channel as covered when at least three cohort A products document live measurement there, thin when one or two do, and study only when only one-off studies exist. Beta and announced coverage is reported separately, and an unreadable channel list counts as unresolved. Retail media is an unscored attribute. Coverage here means that a vendor documents live measurement. It does not mean the measurement has been validated or accredited in that channel [105, 119, 42, 45, 115, 110, 113].

| Channel | Cohort A products documenting live measurement | Measurement basis | Q3 code |
|---|---|---|---|
| Display | Adelaide, Lumen, DoubleVerify, IAS, Amplified (in-flight tags); xpln.ai by its own account | Model estimates and indices of device signals | Covered [105, 119, 42, 45, 115, 110, 113] [571, 121, 496] |
| Online video | As for display | As for display | Covered [105, 119, 42, 45, 115, 110, 113] [571, 121, 496] |
| CTV | TVision; Adelaide, DoubleVerify, Lumen; Amplified (US availability not stated); xpln.ai by its own account | TVision: panel observation. Others: models, four of them using TVision panel data | Covered [105, 119, 42, 45, 115, 110, 113] [45, 106, 220, 110] [405, 121] |
| Social | Adelaide, Lumen, DoubleVerify, Amplified (planning); xpln.ai by its own account; IAS reports Lumen's model (stated live, platforms unnamed; Snap announced) | Models; the three Snap products share Lumen's model | Covered [105, 119, 42, 45, 115, 110, 113] [116, 118, 117] [571, 121, 496] [115, 237, 381, 379, 380] |
| Audio | Adelaide; Amplified (planning only) | Model estimates; no ad-level observation of listening | Thin [107, 108, 119] [107, 105, 120, 112] |
| In-game | None | Eye-tracking studies in test settings | Study only [113, 29] |
| DOOH | Adelaide (listed, no method documented); Amplified (planning) | Not documented; planning benchmarks | Thin [114, 495, 105, 119] |
| Retail media (unscored) | DoubleVerify | Not documented | Thin [162, 260] |

Four of the seven scored channels are covered, two are thin and one is study only. Every live cell outside TVision's is a model estimate or an index of device signals. Lumen, Adelaide, IAS and Amplified run machine-learned models, and DoubleVerify builds indices from exposure and engagement signals [282, 281] [43, 106, 1] [73, 238] [186, 350] [47, 209]. xpln.ai, also a cohort A product, documents display, online video, CTV and social as live in its own sources. Adelaide's social and audio coverage likewise rests on vendor sources [571, 121, 496] [105, 1, 329]. Counting xpln.ai changes no code.

![Display, video, CTV and social are covered; audio, in-game, DOOH and retail media are thin or study-only.](figures/fig-channel-coverage.svg)

*Figure fig-channel-coverage. **Display, video, CTV and social are covered; audio, in-game, DOOH and retail media are thin or study-only.** Documented channel coverage of the cohort A products at the cutoff. Only TVision observes people in the channel it measures; every other live cell is a model estimate or an index of device signals. CTV is covered on different bases: several CTV models use TVision panel data as an input. Units: status category. Scope: cohort A products, documented channels, at the cutoff. Sources [105], [106], [107], [108], [109], [110], [111], [112], [113], [114], [115], [45], [116], [117], [42], [73], [118], [119], [120], [48], [121], [122]. Caveat: A blank cell means none found in the sources screened, not a documented absence. Lumen’s Netflix CTV product is offered only in five European markets; its other CTV routes include the US. Data: data/figure-data.json#channels*

### Display and online video

Display and online video have the widest coverage and the only accreditation. On 28 September 2026, MRC's accredited digital list had one attention entry: DoubleVerify's, for desktop, mobile web and mobile in-app [2]. IAS Quality Attention measurement has been generally available for open-web display and video since January 2024, and IAS extended it to mobile in-app on 30 July 2024 [115, 111, 73]. Amplified's Campaign Performance tags capture per-exposure signals for display and video across programmatic, web and CTV in "20+ markets" [186, 350]. xpln.ai's MRC audit, in process as of November 2025, covers display and video in mobile web and mobile in-app only [48].

### CTV: covered on different bases

TVision is the only cohort A product whose attention data come from direct observation of people: optical sensors in opt-in panel homes, projected to the US population by daily weighting [45, 106, 220, 110]. The sensors identify who is in the room and when each person looks at the TV [45, 52]. TVision covers the TV screen, including walled-garden apps viewed on it; Viant names YouTube and Prime Video [304, 297]. Its outputs for the wider market are projections or models built from the panel [407, 315, 296, 45].

Adelaide, DoubleVerify and Lumen score CTV impressions with models, and all three have documented TVision panel data as an input, DoubleVerify only in its June 2023 partnership announcement [45, 106, 220, 110]. Adelaide says its CTV AU combines an image pixel fired at ad serve, automatic content recognition (ACR) matching to programme and pod, and TVision panel data that sets the variable weights. The model is trained on brand lift, tune-in attribution and advertiser sales data. Samba, the author's employer, is among the named tune-in providers [106].

Lumen has documented US CTV measurement through TVision. For LG Ad Solutions (February 2024), US measurement runs through the TVision pixel and is processed in the Lumen platform, and Nexxen's CTV offer (August 2023) uses measurement "powered by TVision data" [290, 289, 295]. Teads' exclusive Lumen HomeScreen measurement, announced in May 2026, is a separate route: a machine-learning prediction from eye-tracking data, covering the US, EMEA, APAC and LATAM [110, 277] [45, 106, 220, 110]. Lumen's Netflix product measures Netflix's global ad-supported inventory but is offered only to clients in the UK, Germany, France, Italy and Spain, so it adds no US client availability [110]. PPC Land reports that xpln.ai's CTV models use TVision's US panel data as a calibration input [405, 121].

DoubleVerify announced CTV attention with TVision on 13 June 2023, with a closed beta anticipated in the second half of 2023, and by 30 June 2025 described its attention measurement as spanning CTV. This review found no dated general-availability release in DV's newsroom [220, 116]. DV's current CTV pages describe CTV attention without naming a data source [220, 426, 364, 219, 82]. MRC's April 2026 letter accredits DV's CTV reporting for impressions, video viewability, completion and quartile metrics and property-level verification. It lists attention metrics only for desktop, mobile web and mobile in-app [42, 2, 48].

The shared panel is a dependency. Viant completed its acquisition of TVision on 1 May 2026, and whether TVision data remain licensed to Adelaide, DoubleVerify and Lumen is not documented [45, 106, 220, 110]. TVision's site still markets data licensing, and Adelaide's July 2026 method blog says in the present tense that CTV AU "incorporates TVision data". No post-closing statement from DoubleVerify, Lumen or xpln.ai on continued access was found [297, 311, 106]. This paper infers that agreement among CTV models trained on the same panel would not count as independent corroboration.

### Social: through the platforms, and around them

Platform integrations for social attention run through Snap and TikTok. Lumen announced a Snapchat attention measurement integration on 18 March 2024, with a model trained on eye-tracking data gathered with Snap since 2021. DoubleVerify's Authentic Attention for Social became globally available on Snapchat on 30 June 2025, combining DV exposure data with Lumen eyes-on-screen signals. IAS announced Snap Attention Measurement the same day. DoubleVerify extended Authentic Attention to TikTok on 10 December 2025 as a badged TikTok Marketing Partner [116, 118, 117]. In the sources screened, no later IAS source confirms that its Snap product went live [238, 115, 226, 118]. All three Snap products rest on Lumen's eye-tracking model [116, 118, 117], so three vendors' Snap numbers trace to one reference source.

Platform integration brings no attention accreditation. In MRC's 13 April 2026 letter, DoubleVerify's accreditations on Meta, YouTube (via Ads Data Hub) and TikTok cover impressions and viewability only [42]. IAS reports a Lumen social model through its own platform, stated as live from January 2025 with no platforms named [115, 237, 381, 379, 380]. xpln.ai says it reaches Facebook, Instagram, Snapchat, TikTok and X through platform reporting interfaces or API, with no granular-data partnership [571, 121, 496]. Amplified's planning tool lists Meta, TikTok, Snapchat, X and Pinterest [119].

In the sources screened by the cutoff, this review found no attention measurement product in Meta's or YouTube's measurement partner programmes, with a platform partner badge, or with a certified partner reporting integration for attention. DoubleVerify's June 2026 Meta expansion is Authentic AdVantage, with no mention of attention [42, 568, 569, 570, 109]. The absence is that narrow. xpln.ai's stated access to Facebook and Instagram through reporting interfaces or API, without a granular-data partnership, is a vendor disclosure [42, 568, 569, 570, 109]. Attention on these platforms is also measured without platform integration. Adelaide has offered AU ratings for YouTube, Facebook and Instagram placements since September 2020, from licensed Amplified panel data. Lumen measured YouTube campaigns for Silverpush and Epitaph in February 2026. TVision's panel observes attention in YouTube's CTV app [42, 568, 569, 570, 109]. Adelaide's Google-commissioned AU analysis of 395 US campaigns from 2022 to 2024 spans 30 placement types, including competitor streaming platforms [338].

### Audio: thin, and no listening observed

Audio is thin under the Q3 rule. Adelaide announced show-level podcast AU in Magellan AI on 10 January 2024, with availability stated from early 2024; current availability is not confirmed. On 9 May 2024 it extended Adelaide for Publishers to podcasts, streaming audio and terrestrial radio, with NPR, SiriusXM Media and Westwood One as launch partners. Amplified covers audio only with planning-stage prediction: Spotify and radio appear in Campaign Strategy, and audio is not listed for its in-flight measurement [107, 108, 119]. In the sources screened, IAS and xpln.ai document no audio attention measurement [115, 237, 381, 379, 380] [571, 121, 496].

The three vendors that attach attention metrics to audio use three different constructs, and none documents observing a listener's auditory attention to an ad [107, 105, 120, 112]. Adelaide's podcast AU is a model estimate built from placement exposure data such as genre, ad position, ad load and ad type, with eye-tracking and outcome data in the model by Adelaide's account. Amplified classifies active, passive and non-attention from listener tasks and surveys, alongside STAS, a brand-choice measure. Lumen's 2023 audio score for dentsu was modelled to be equivalent to visual attention from listening data, survey results and exposure type [107, 105, 120, 112]. That dentsu research, whose funding is not stated, reported an average of 10,126 attentive seconds per 1,000 impressions across audio formats, against dentsu norms of 6,501. Sample size and geography are not stated [112]. The audio numbers on sale are models of placements, not observations of listening.

The measures closest to listening sit in layer 1. DoubleVerify's accredited statistics include quartile audible rates and volume-change rates on video [42], which this paper codes as candidate proxies for listening ([chapter 1](#defining)). A peer-reviewed EEG study of radio, podcasts and music streaming measured attention to audio formats and excluded responses to ads. The Australian Radio Network, an interested party, funded and co-designed it [28]. Adelaide's AU is under MRC review for a scope that includes audio, with the pre-audit reviewed in May 2026 and the audit not begun [48, 339]. The audio paper in this series cites this paper's definition of auditory attention.

### In-game: study only

No cohort A product documents live in-game attention measurement in the sources screened; the evidence is one-off studies [113, 29]. Dentsu and Lumen's January 2024 report measured attention with eye-tracking while respondents played real games on mobile and PC. It reports gaming APM of 6,736, rewarded video at 10,043 APM against 6,100 for online video, and intrinsic in-game ads at 3,442 against 1,416 for traditional online ads. The last pair is stated as "seconds". Anzu renders it as 3.4 and 1.4 seconds, from which this paper infers a unit of attentive seconds per 1,000 impressions [113, 29]. These are direct observations of participants in a test setting. Lumen and Anzu released a study in June 2023, and Amplified published one with Digital Turbine in April 2023 [113, 29]. All three are vendor or agency studies whose funding is not established here, so their independence is unknown.

Anzu's guide names Adelaide and Amplified as in-game attention vendors [29]. The channel lists of both vendors opened for this review omit gaming [113, 29] [119]. One in-game ad seller, Frameplay, markets a proprietary viewable-time metric as attention, on the basis of one-campaign eye-tracking validation [113, 29]. Under this paper's definitions that metric is a candidate proxy. IAS's in-game service reached MRC's pre-audit list for display and video rendered and viewable impressions in desktop, mobile and gaming-console environments [233, 48]. MRC removed it from in-process status in January 2026, citing IAS's lack of intent to engage in third-party audit [49, 339]. That process concerned impressions and viewability.

Independent evidence is cautionary. In an eye-tracking study of a mobile tennis game with 79 participants in Turkey, total fixation duration on in-game ads had no significant effect on aided brand recall once brand placement was controlled. Brands were not counterbalanced across placements [53]. Creative testing does reach the format: Lumen SPOTLIGHT tests in-game creative with webcam eye-tracking of opt-in participants [279, 260]. Those tests use recruited participants in simulated contexts and measure no live media exposure. The gaming paper in this series cites this paper's definition of in-game attention.

### DOOH: thin, and measured differently in kind

DOOH is thin. Adelaide lists DOOH in its coverage without a documented method, and Amplified lists OOH and DOOH in its planning tool [114, 495, 105, 119]. Adelaide's MRC review scope has included digital OOH since at least November 2024 [48, 339]. xpln.ai told CESP, a French industry body, that it does not cover DOOH [571, 121, 496].

The DOOH evidence rests on methods that differ in kind. Lumen's study with Ocean Outdoor (UK, June 2025) tracked 2,500 online participants' eyes by webcam while they watched filmed walks past OOH sites. Amplified's study with QMS in Australia used in-situ eye, gaze and pose tracking across 1.2 million exposures at twelve sites. Quividi's explainer defines DOOH Attention Time as time spent watching, measurable only with camera-based anonymous video analytics [114, 495, 105, 119]. All three are vendor sources. The OOH audience currencies already apply likelihood-to-see adjustments derived from eye-tracking: Geopath in the US and Route in the UK. Those adjustments correct audience counts for visibility and measure no attention to a given ad [114, 495, 105, 119].

### Retail media: one vendor's assertion

One cohort A product documents retail media. DoubleVerify presents Authentic Attention within its commerce-media offering and reported in May 2024, from 2023 data, that retail media network owned-and-operated inventory had engagement rates 183% above DV's attention baseline. It did not name the networks or placements measured or publish a method for them [162, 260] [366, 367]. The figure is relative to DV's own baseline, and DV normalises its attention indices to the average of its own impressions over a rolling 28-day window [71, 211, 370].

Adelaide's AU pre-bid segments in Amazon DSP (9 June 2026) cover display, online video and streaming TV inventory, and the release does not say whether Amazon-owned retail placements are included [162]. Kantar extended LINK+ creative testing to ads in retail media networks on 6 May 2025, but that release does not mention attention. Lumen's in-store work covers physical stores [162, 260].

### What coverage does not settle

A covered code says nothing about assurance. On 28 September 2026, MRC's digital list had one accredited attention entry, DoubleVerify's, for desktop, mobile web and mobile in-app. No accredited attention product from Adelaide, Lumen, TVision or Amplified appeared on MRC's digital, television or multimedia lists [2]. DoubleVerify's CTV attention is not accredited [42, 2, 48]. So no accredited attention metric found covers CTV, social, audio, in-game or DOOH. On that date two MRC reviews were open. xpln.ai's audit covers mobile web and mobile in-app [48]. Adelaide's scope reaches audio, linear TV, cinema and digital OOH, but its audit has not begun [48, 339]. [Chapter 2](#standards) sets out the accreditation record.

US availability is uneven too. Amplified's in-flight product names no US market [354, 119, 186]. xpln.ai's launch release names US-headquartered advertisers among its clients but places those clients in Europe and APAC, so in the sources screened no US campaign or US production use is documented [496, 121, 405]. GumGum launched Attentive Lift in Australia with no channel named, so its coverage stays unresolved [458].

Per channel, a buyer needs to know which product measures it live, whether its number is observed or modelled, and from which panel ([chapter 12](#next)).



## 9. Privacy and law


*Two questions decide an attention product's legal exposure: can its signal identify a person, and is it used to infer emotion? Gaze and device signals that report only where and how long people look fall mostly under ordinary personal-data and consent rules. On the European Commission's reading, facial coding that infers emotion is emotion recognition under the EU AI Act. In the sources screened by the cutoff, this review found no privacy or biometric-data regulator action or lawsuit naming a seed attention vendor.*

> **In short.**
> In most statutes screened, whether webcam gaze, landmark or facial-coding data counts as biometric turns on identification or identifiability. The EU AI Act and California's profiling definition reach emotion inference without it [130, 134, 132, 128, 30, 131]. The AI Act bans emotion inference only in workplaces and education; advertising uses that infer emotion from biometric data are high-risk, subject to the Article 6(3) derogation [30, 124]. Deployers of emotion recognition systems have had to inform the people exposed since 2 August 2026 [30, 123]. High-risk obligations start on 2 December 2027 and reach products already on the market only after a significant design change [30, 123]. Illinois and Texas define biometric identifiers to include face geometry [128] [129, 505], and the Ninth Circuit held that face signatures which cannot identify fall outside the Illinois law [418]. No privacy or biometric-data action found in the searches run names a seed attention vendor; Nielsen's patent actions against TVision are a separate matter [507, 136] [82]. The certifications found cover information security or research process; none attests privacy compliance [384] [500, 501, 502, 503, 504].

### Two questions sort the products

Attention products capture different things about people. Tag-based products read the device: DoubleVerify's notice lists viewport and screen size, tab and browser focus, time in view, scroll position and ad interactions [340, 372]. Panel products film people. TVision's US panellist policy lists face geometry and periodic snapshot images of the webcam's field of view [306, 45, 291]; Amplified uses device cameras and an in-home TV micro-camera [184]; Lumen turns a webcam into an eye tracker [393]. Creative-testing products add facial coding: iMotions' Media Analytics codes 20 Action Units and seven core emotions [513, 514, 515, 452], and Adverteyes measures "Visual Attention (eyes on screen) and Emotions (facial expressions)" [512, 499].

The law sorts these signals by two questions: can the signal identify a person, and is it used to infer emotion? In the statutes screened, biometric status for webcam gaze, landmark or facial-coding signals turns mostly on identification or identifiability. The EU AI Act and California's profiling definition reach emotion inference without identification. GDPR and UK GDPR can reach such inferences as special-category health data where they support a health inference, and Washington's health-data law reaches physiological measurements linked to a consumer [130, 134, 132, 128, 30, 131]. By this review's reading, a gaze product that reports only where and how long people look, and cannot identify anyone, sits low on both questions. A facial-coding product that reports emotion sits high on the second whether or not it identifies anyone. Other US state laws, such as Virginia's, Connecticut's and Oregon's, were not opened [130, 134, 132, 128, 30, 131].

[Chapter 1](#defining) coded what each product observes under Q2. This chapter reads those observations for what they capture about the person, and supplies the evidence behind the privacy-posture dimension scored in [chapter 11](#fq).

### In the EU, emotion inference is the line

The AI Act defines an emotion recognition system as one that identifies or infers emotions or intentions from biometric data [30]. The Act's definition of biometric data drops GDPR's requirement that the data allow or confirm unique identification. It still covers only personal data, so truly anonymous data fall outside it [30, 125]. The Commission's guidelines treat eye-tracking data as biometric data under the Act, and say a system that tracks only the gaze point does not infer emotions. Inference: a gaze-only product stays outside the definition only while its outputs stay at where and how long people look. Once gaze or facial data is used to infer emotions, interest, arousal or intentions, such as "engagement" or purchase intent, the Commission's reading places it inside [30, 124]. One of the guidelines' prohibited examples is an education institution using emotion recognition to infer the "interest and attention" of students. An attention product that infers engagement from facial expressions may therefore be caught [30, 124]. iMotions' AI Act statement of 13 February 2025 calls its own facial expression analysis an emotion recognition system [510, 514, 516, 517].

The ban itself is narrow. Since 2 February 2025 the Act has banned emotion inference in the workplace and in education, with medical and safety exceptions. The Commission's non-binding guidelines place customer-facing and advertising emotion inference, including "intelligent billboards", outside the ban. Where those systems infer emotions from biometric data, they are high-risk under Annex III 1(c), subject to Article 6(3) [30, 124].

Two duties run on different clocks. From 2 August 2026, anyone deploying an emotion recognition or biometric categorisation system in the EU must inform the people exposed to it (Article 50(3)); the July 2026 Digital Omnibus did not amend that paragraph in the text screened [30, 123]. That omnibus, Regulation (EU) 2026/1744, moved the Chapter III Sections 1-3 obligations for Annex III systems to 2 December 2027. Under Article 111(2) as amended, an emotion recognition system placed on the market or put into service before that date is subject to them only if its design changes significantly afterwards; public-authority uses are excepted, with a 2030 deadline. Every facial-coding attention product on the market at the cutoff is therefore outside the high-risk regime until a significant design change, while the transparency duty applies to deployers now. Whether a creative-pretest facial-coding tool qualifies for the Article 6(3) derogation is unresolved [30, 123].

### GDPR, the UK and the device

Under GDPR, biometric data is a special category, needing an Article 9(2) condition such as explicit consent, only when processed "for the purpose of uniquely identifying a natural person" [125]. The UK ICO takes the same line, adds that transient processing lasting "a fraction of a second" is still processing, and is reviewing its guidance because of the Data (Use and Access) Act [127]. By inference from the definitions, a facial-coding test can process ordinary personal data under GDPR and still be an emotion recognition system under the AI Act [30, 125] [125]. Some vendors ask for the stricter consent anyway: iMotions' 2022 policy makes "the Data Controller" (by this review's inference, the client) responsible for collecting consent under GDPR Articles 6(1)(a) and 9(2)(a) [510, 511] [510, 514, 516, 517].

Device signals meet a different test. The European Data Protection Board's (EDPB) Guidelines 2/2023 say ePrivacy Article 5(3) protects information on a user's device whether or not it is personal data, and that JavaScript instructing a browser to send information "clearly" falls within it. The inference is that tag-based attention signals collected in the EU need consent unless an Article 5(3) exemption applies [126]. Two vendors' entries in IAB Europe's Transparency and Consent Framework declare legitimate interest instead. Amplified registers its attentionPROVE tag with legitimate interest as the default basis and declares no device storage access [184, 190]; xpln.ai declares legitimate interest for measuring advertising performance and no cookie use [190]. Both entries are the vendors' own declarations, and neither is a compliance audit. Whether any authority treats attention measurement as strictly necessary, and so exempt, was not checked [126].

Adelaide's AU Ratings policy lists the tag's device type, IP address, location to city, DMA and zip code, and inferences as personal information [340, 372]. The same policy states that AU Ratings process no personal data under the GDPR and use no cookies [340, 168].

### In the US, identification decides most cases

| Jurisdiction | What brings a signal in | What it requires | From |
|---|---|---|---|
| California (CCPA and CPPA regulations) | Biometric processing to uniquely identify, and neural data, are sensitive; profiling includes facial expression used "to identify or infer emotion"; device identifiers, pixel tags and probabilistic identifiers are personal information | Written risk assessment before selling or sharing personal information, processing sensitive personal information, or training emotion-recognition technology | 1 January 2026 [130] [131, 528] |
| Colorado (HB24-1130) | Biometric identifiers: data that can be processed to uniquely identify a person, including iris scans and facial geometry | Notice before collecting any identifier; consent before collecting biometric data used or intended for identification; retention and deletion policy; any volume, except for the access right | 1 July 2025 [134] |
| Washington (My Health My Data Act) | Biometric data that identifies a consumer | Opt-in consent for a specified purpose, or necessity for a requested service | Compliance from 31 March 2024 [132, 133] |
| Illinois (BIPA) | Retina or iris scan; scan of hand or face geometry | Written notice, purpose and term, and a written release; $1,000 or $5,000 per violation | 2008; amended 2 August 2024 [128] |
| Texas (CUBI as amended by HB 149) | Record of hand or face geometry | Exempts biometric identifiers involved in AI systems unless a system is used to uniquely identify a specific individual; later commercial use falls back under the penalty provisions | Amendment 1 January 2026 [129, 505] |
| Federal (FTC policy statement) | Facial features, iris or retina, "characteristic movements or gestures"; derived data where identification is reasonably possible | Surreptitious and unexpected collection named as potentially unfair | 18 May 2023; still listed by the FTC on 28 September 2026 [135, 506] |

California's profiling definition excludes characteristics that cannot reasonably be linked to a consumer. Processing already under way must be assessed by 31 December 2027, with submissions due by 1 April 2028. Which attention vendors meet the CCPA's business thresholds was not established [131, 528]. By this review's inference, EEG-based ad testing on California consumers produces sensitive personal information, while eye tracking that does not identify produces personal information that is not sensitive [130].

In Zellmer v. Meta Platforms (17 June 2024), the Ninth Circuit held on its record that Facebook "face signatures" "cannot identify and therefore are not biometric identifiers or information as defined by BIPA". A federal court applying Illinois law does not bind Illinois state courts. By inference, the ruling helps vendors whose facial landmarks cannot identify, but offers no safe harbour [418]. Whether attention-panel capture falls inside Texas's AI exemption depends on actual use and deployment, which this review did not establish. The Texas Attorney General's first action under the law ended in a $1.4 billion settlement with Meta on 30 July 2024 [129, 505].

Washington's separate biometric statute excludes photographs, video or audio recordings and data generated from them. By inference, webcam landmarks that do not identify fall outside both Washington definitions [132, 133].

### No privacy action found against a seed vendor

In the sources screened by the cutoff, this review found no public privacy or biometric-data regulator action or lawsuit naming a seed attention-measurement vendor. The searches named Realeyes, Affectiva, TVision, Tobii and iMotions, plus general BIPA searches on eye tracking, facial coding and webcam research. Adelaide, Lumen, Amplified, DoubleVerify, IAS, Neurons, DAIVID and Kantar were not named in any litigation search [507, 136]. A vendor sweep named eye square in one general web search, which found no regulator action, fine or lawsuit [500, 501, 502, 503, 504]. This is not a finding of no litigation. Nielsen has brought patent actions against TVision since 2021, and Viant took them on when it bought the company [82]. The nearest privacy matters involve advertising screens and a call centre.

| Matter | Place and date | What was captured | Outcome |
|---|---|---|---|
| Roberts v. Cooler Screens (putative BIPA class action) | Cook County, Illinois; filed 3 March 2022 | In-store ad screens with a camera and an alleged "iris tracker to detect exactly where the customer is looking", plus facial profiling of age, gender and emotional response | Allegations only; disposition not found [507] |
| Cadillac Fairview (Canadian privacy commissioners) | Canada; 28 October 2020 | "Anonymous Video Analytics" in mall directories estimating age range and gender; about 5 million numerical face representations stored | Sensitive biometric information collected without valid consent; express opt-in consent recommended [136] |
| ViaQuatro (Metro Line 4 advertising doors) | São Paulo; judgment released 7 May 2021 | Passengers' images captured to detect characteristics, emotions and reactions, without notice | Stop capture without prior consent; R$100,000 collective moral damages, reported raised to R$500,000 on appeal in May 2023 [137, 509] |
| Budapest Bank (Hungarian data protection authority) | Hungary; 8 February 2022 | AI voice analysis of customer-service calls detecting "emotional/mood elements" | HUF 250,000,000 fine; ordered to stop analysing emotions [508] |

In the three camera matters, the finding or allegation was capture without valid consent or notice. ViaQuatro had described its system as counting views, "tempo de atenção" (attention time) and emotions [137, 509].

### What vendors document

Vendors describe opt-in panels, on-device or in-home processing and landmarks that do not identify. Every row below is the vendor's own statement.

| Vendor | Consent stated | Processing, retention and identification |
|---|---|---|
| TVision | Panel "100% opt-in", compensated [291] | Face geometry collected as a biometric identifier, destroyed within three years of last interaction; data leaving the home "completely anonymous" after an initial training period [306, 45, 291] |
| Amplified | Explicit prior consent for sensitive data | TV images processed in the home and "promptly deleted"; landmarks used "solely for attention measurement"; no facial recognition [184] |
| Lumen | Panels "fully informed, consented and compensated" (2021 page) [271, 392, 115, 287] | Processing on the user's device, no video or personally identifiable information uploaded; no panel privacy notice found [393] [393, 401] |
| Kantar | Facial-expression and eye-tracking data only with consent | Affectiva stores facial-expression video for seven years; anonymised data may train AI [267, 387] |
| Tobii Sticky | "Legitimate interest, or consent"; emotion analysis opt-in | Webcam framing limited to face and head; recordings deleted after up to 90 days [520, 521] |
| Adverteyes | Explicit consent to facial images and biometric data | No facial recognition; models being trained to determine whether the same individual appears in two recordings [497, 498, 499] |
| eye square | Consent under GDPR Article 6(1)(a) for online eye tracking | Face video recorded; models "trained to uniquely identify one person from another" to remove duplicate participants; video deleted after three months at most; the SEAL page instead says "Only X/Y gaze coordinates leave the device" [500, 501, 502, 503, 504] |

Three statements need reading against the law. The first is TVision's "anonymous": data keyed to household and viewer numbers is pseudonymous under GDPR-style tests and plausibly personal information under the CCPA, by this review's inference [291]. TVision's policy also permits sharing personal information with clients in hashed form that may be re-identified [306, 45, 291]. The second is non-identification. Amplified's TV camera can capture guests who are not the consenting panellist [184]. Adverteyes and eye square describe models that tell one person from another [497, 498, 499] [500, 501, 502, 503, 504]; by inference, such a model bears on the first question whatever the product outputs. The third is jurisdiction. Tobii's white paper names GDPR as its only legal framework [520, 521], and no US state biometric-law documentation was found for Media Analytics [510, 514, 516, 517].

On the tag side, DoubleVerify's Solutions Privacy Notice does not name Authentic Attention [372, 47], and IAS's Technology Solutions policy does not name Quality Attention [385, 386, 384].

The certifications found cover information security or research process. IAS announced ISO/IEC 27001:2022 certification on 22 May 2024, with a scope that names "Attention"; its validity at the cutoff, after IAS's December 2025 change of ownership, was not verified. ISO 27001 certifies information-security management, not privacy compliance, consent or measurement validity [384]. Tobii's Sticky page asserts ISO/IEC 27001 certification. The certificate behind it is Tobii AB's company-wide ISO/IEC 27001:2022 certificate, valid to 8 May 2028, which does not name Sticky [455, 518, 519]. Adverteyes' SOC 2 statement names Realeyes [512, 499]. The certifications xpln.ai cites are attributed to its host, Google Cloud Platform [522, 121, 523]. eye square's ISO 20252 certificate, valid to 13 November 2026, covers its market-research process and does not attest privacy compliance for the product [524, 525, 526, 502, 527] [500, 501, 502, 503, 504].

For a buyer, the documents that matter are the panel privacy notice, the tag's legal basis in each market and, for facial coding in the EU, the Article 50(3) notice to the people tested [30, 123]. [Chapter 12](#next) turns these into questions for vendors.

### The instruments read

The instrument lists the laws and guidance read for this chapter. Each row describes the primary text as read; none is legal advice.

**Rules and governance instruments**

| Jurisdiction | Instrument | What it requires | Effective | Scope | Law or practice | Src |
|---|---|---|---|---|---|---|
| European Union | AI Act Art 5(1)(f) | Prohibits placing on the market, putting into service or using AI systems to infer emotions of a natural person, except for medical or safety reasons | 2025-02-02 | Workplace and education institutions only; advertising use not prohibited (Commission guidelines para 266) | regulation | [30] |
| European Union | AI Act Art 50(3) | Deployers must inform exposed natural persons of the operation of the system and process personal data per GDPR | 2026-08-02 | Emotion recognition and biometric categorisation systems (not amended by Reg 2026/1744 in the text screened) | regulation | [30] |
| European Union | AI Act Annex III 1(c) with Art 6 (as amended by Reg 2026/1744) | High-risk system obligations (Chapter III Sections 1-3); Art 6(3) derogation possible except where the system profiles natural persons | 2027-12-02 | AI systems intended to be used for emotion recognition | regulation | [123] |
| European Union | Commission Guidelines on prohibited AI practices C(2025) 5052 | Interpretation: gaze-point eye tracking does not infer emotions; emotion recognition in advertising is outside Art 5(1)(f) but high-risk | 2025-07-29 (document date) | Article 5 AI Act | regulator guidance (non-binding) | [124] |
| European Union | GDPR Art 4(14), Art 9(1)-(2)(a) | An Art 9(2) condition (e.g. explicit consent) for biometric data processed to uniquely identify a person | 2018-05-25 | Biometric data only where it allows or confirms unique identification; photos not systematically special category (recital 51) | regulation | [125] |
| European Union | ePrivacy Directive Art 5(3) per EDPB Guidelines 2/2023 v2.0 | Consent (unless exempt) for storing or gaining access to information in terminal equipment, including JavaScript-instructed collection, regardless of personal-data status | 2024-10-07 (guidance adopted) | Device signals collected by tags, pixels and SDKs | regulator guidance | [126] |
| United Kingdom | UK GDPR Art 4(14)/Art 9 per ICO biometric data guidance | Special-category condition only when biometric data is used to uniquely identify; transient processing is still processing | undated; under review after Data (Use and Access) Act | Biometric recognition | regulator guidance | [127] |
| Illinois | BIPA 740 ILCS 14/15(b), 14/20 | Written notice, purpose and term, and written release before collection; public retention policy; private right of action ($1,000 negligent / $5,000 intentional per violation; one recovery per person per method since 2024) | 2008-10-03; amendment 2024-08-02 | Retina or iris scan, fingerprint, voiceprint, scan of hand or face geometry | statute | [128] |
| Texas | CUBI Bus. & Com. Code 503.001 as amended by HB 149 (TRAIGA) | Informed consent before capture for a commercial purpose (per (b), referenced in new (b-1)); exemption for AI training unless used to uniquely identify a specific individual | amendment 2026-01-01 | Retina or iris scan, fingerprint, voiceprint, record of hand or face geometry; AG enforcement | statute | [129] |
| California | CCPA Civ. Code 1798.140(c), (ae), (aj) | Sensitive-PI treatment for biometric processing to uniquely identify and for neural data; persistent and probabilistic device identifiers are personal information | as amended 2026-01-01 | Consumers and households; businesses above CCPA thresholds | statute | [130] |
| California | CPPA regulations 11 CCR 7001(ee), 7150, 7155 | Written risk assessment before selling or sharing PI, processing sensitive PI, or training emotion-recognition or physical/biological profiling technology; legacy processing assessed by 2027-12-31 | 2026-01-01 | CCPA-covered businesses | regulation | [131] |
| Washington | My Health My Data Act RCW 19.373.010 | Opt-in consent for consumer health data (incl. identifying biometric data); consent may not be obtained by hovering, muting, pausing or closing content | 2023 c 191 (effective date not verified) | Biometric data that identifies a consumer | statute | [132] |
| Washington | RCW 19.375.010 | Biometric identifier regime; excludes photographs, video or audio recordings and data generated therefrom | 2017 c 299 | Data used to identify a specific individual | statute | [133] |
| Colorado | Colorado Privacy Act as amended by HB24-1130 | Consent before collecting biometric data; written retention and deletion policy; applies to any controller processing any amount of biometric data | 2025-07-01 | Data that can be processed to uniquely identify, incl. eye retina or iris and facial geometry; photos and recordings excluded unless used for identification | statute | [134] |
| United States (federal) | FTC Policy Statement on Biometric Information and Section 5 | Clear disclosure; surreptitious or unexpected collection of biometric information may be unfair | 2023-05-18 | Facial features, iris or retina, characteristic movements; derived data where identification is reasonably possible | regulator policy statement | [135] |
| Canada | PIPEDA as applied in Findings #2020-004 | Meaningful express opt-in consent for camera-based facial analysis (recommendation) | 2020-10-28 | Facial analytics in mall directories for age and gender estimation | regulator finding | [136] |
| Brazil (São Paulo) | Consumer and image-rights law as applied in Idec v. ViaQuatro | Abstain from capturing images, sounds or personal data of users without prior consent | 2021-05-07 (first instance) | Advertising doors measuring views, attention time and emotions of passengers | court judgment | [137] |



## 10. Agents


*Attention can be written into AdCP, the agentic advertising protocol maintained by AgenticAdvertising.org. Version 3.1.24 has fields for a named vendor's attention metric in forecasts, optimisation goals, delivery reports and contractual floors. In the workflows screened, no named vendor's attention signal is documented in a connector, an interoperability test, production use or a transaction. Beyond the schema, the only handling found is a test-mode check with a placeholder vendor. Under the registered rule, H5 is unresolved.*

> **In short.**
> AdCP v3.1.24, the stable release of 23 September 2026, can represent a vendor-bound attention metric in forecasts, optimisation goals and delivery reports, and as a performance-standard floor [567] [144, 138, 140, 142]. The 3.0 attention goals named no vendor and could not be carried end to end, and 3.1 deprecated them [146, 32, 143, 150]. Viewability graduated to standard reporting in 3.1, while attention stayed a vendor-defined metric [143, 556]. The only tests of the attention fields found use a placeholder vendor and simulated delivery [33, 572]. On 28 and 29 September 2026 the AdCP registry listed 34 agents and no measurement agent [141, 551]. No IAB Tech Lab AAMP artifact screened names an attention metric [36, 153, 154, 147]. DoubleVerify runs a live agent server whose attention coverage is undocumented [217, 559, 218, 148, 560]. H5 is unresolved. No production use of a named vendor's attention signal is documented in the workflows screened, and no named vendor's signal is documented beyond schema representation; separately, sellers' handling of the attention fields is shown only in test mode, with a placeholder vendor [142, 141, 151, 33].

### The test

H5 asks whether attention signals are carried and used when software agents buy media. Shared rule 12 records five separate states for each protocol and version: schema representation, an implemented connector, a demonstrated interoperability test, documented production use and a demonstrated transaction. Schema representation proves only that a value can be expressed. H5 is supported when an attention signal reaches documented production use or a demonstrated transaction in an agent-driven buying workflow, with the protocol name and version recorded. It is contradicted when a source documents an attention signal dropped, not carried or failing in an agent-driven buying workflow or interoperability test, while none reaches production use. Everything in between is unresolved. Membership, working-group roles and announcements earn no state ([chapter 13](#method)).

The protocols screened are AdCP (the Ad Context Protocol), maintained by AgenticAdvertising.org (AAO), and IAB Tech Lab's AAMP (Agentic Advertising Management Protocols), at the versions current on 28 and 29 September 2026. The screen also covered vendors' own agent surfaces, such as servers built on MCP (the Model Context Protocol), through which AI assistants call outside tools. AAO is an industry consortium and holds no accreditation as a standards body [567].

> **Interests.**
> The author co-leads the Signals & Measurement working group of AgenticAdvertising.org ([chapter 13](#method)). A sales agent operated by No Fluff Advisory, the author's firm, is one of three registered agents that passed AdCP's vendor-metric accountability storyboard, a test run with a placeholder attention vendor [33, 572]. Two of the 34 agents in AAO's registry belong to the firm [141, 551]. Both are declared interests, and neither earns a state. This review's verification pass found the author's GitHub account among neither the authors nor the commenters on the two issues that led AdCP to redesign its attention fields; a different account cannot be excluded [146, 32, 143, 150].

### What AdCP can represent

AdCP's stable release at the screen was v3.1.24, published 23 September 2026, with 3.2.0 at release candidate 7. The 3.1 line began on 18 June 2026, and 3.0.0 shipped on 22 April 2026 [567].

Version 3.1.24 can carry a named vendor's attention value in four places. A seller can forecast it for planning. A buyer can set it as an optimisation goal of kind vendor_metric, bound to a vendor. Delivery reports can return it in vendor_metric_values, where the value is required and the unit and count of measurable impressions are optional. A performance standard can set an attention_score floor that names a vendor, and a breached standard can be paired with a makegood remedy [144, 138, 140, 142]: additional delivery, credit or invoice adjustment [31]. All of this is schema representation. Units stay vendor-defined, so two vendors' values in the same field cannot be compared ([chapter 1](#defining)) [144, 138, 140, 142].

The design came out of a reconciliation problem. In AdCP 3.0 a buyer could set an attention_seconds or attention_score optimisation goal, but the goal named no vendor and delivery reporting had no field for it. AdCP's issue tracker recorded the gap between 15 and 17 May 2026. Version 3.1 added vendor-bound goals (PR #4668, merged 17 May 2026), reports attention through vendor_metric_values, and deprecated the seller-native attention values, which remain in v3.2.0-rc.7 pending removal at the next major version [146, 32, 143, 150]. A maintainer called the vendor-agnostic goal "a meaningless question" [32]. This was a design review of the specification, and no failure in a live agent buy was observed [146, 32, 143, 150].

Viewability moved the other way. AdCP 3.1 graduated it to standard delivery reporting, adding viewed seconds, while attention stayed a vendor-defined Tier 2 metric. AdCP graduates a metric when a standards body publishes a measurement specification. The specification names the IAB Attention Measurement Guidelines only as an example standard a vendor's metric may implement, and its maintainers state that no MRC-or-equivalent body has graduated a shared attention definition. AdCP therefore does not treat the November 2025 guidelines as grounds to graduate attention [143, 556]. The only MRC attention accreditation found does not cite them either ([chapter 2](#standards)) [42, 2, 425, 563].

Two gaps follow, on this review's reading. The attention floor is under-specified: AdCP's worked example sets an Adelaide attention standard with a threshold of 65, but the schema defines the threshold as a rate between 0 and 1, and attention_score as a rate over measurable impressions with no rule for which impressions count [561, 145, 144]. And accreditation scope has no structured field. A vendor declares accreditations in its own measurement catalog, a surface marked experimental in 3.1, with no field for environment, and AAO treats such entries as vendor assertions [139, 149, 141]. DoubleVerify's MRC accreditation, per the letter of 13 April 2026, does not cover CTV (connected TV) attention [42, 205], and that limit could travel only as free text or behind a link. Both gaps are inference [139, 149, 141] [561, 145, 144].

### Beyond the schema

Past schema representation, the evidence is from tests. AdCP 3.1 defines seller conformance storyboards, scripted tests, for vendor-metric optimisation and accountability. Their fixtures use the placeholder domain attentionvendor.example, and a test controller injects the delivery. They check how a seller handles vendor metrics and involve no real attention vendor. AAO's public compliance records show three registered sales agents passing the accountability storyboard between 15 and 23 September 2026: Addiscovery, Purrsonality Seller and the agent operated by No Fluff Advisory. No agent had been tested on the optimisation-flow or catalog-precondition storyboards, and no pass involving a named attention vendor was found [33, 572]. The optimisation storyboard makes its catalog check a SHOULD in 3.1 "while measurement-vendor adoption of AdCP-conformant capability publication catches up" [33].

Prebid SalesAgent, a community AdCP seller agent that AdCP's docs recommend for self-hosting, records coverage "none" for the vendor-metric optimisation flow and catalog precondition and "partial" for vendor-metric accountability, in a map dated 17 September 2026 and keyed to AdCP 3.1.1. Its end-to-end vendor-metric scenario, written with a placeholder vendor and simulated delivery, is listed as dormant [151, 152]. A coverage map shows what is graded. It does not prove the code path is absent [151, 152].

No attention vendor had registered an agent in the registries screened. On 28 and 29 September 2026, AAO's registry of member-enrolled agents listed 34: 29 sales, three signals, one creative and one buying agent, by self-declared type. None was a measurement agent, and a query for the metric attention_units returned none, although AdCP's registry documentation gives Adelaide-style attention as its first example of measurement-vendor discovery [141, 551]. In those agents' capability records, attention appears only inside tool schemas inherited from the specification, and no record names the domains of Adelaide, Lumen or IAS [141]. IAB Tech Lab's Agent & MCP Server Registry listed 47 agents on 29 September 2026, three in its measurement category and none from an attention vendor [34]. Neither AAO's brand records nor the brand.json discovery files at eight attention vendors' domains listed an agent [35]. A private or partner-only integration would not appear in any of these sources [157, 141].

### AAMP names no attention metric

IAB Tech Lab's AAMP is an umbrella of separately versioned repositories. AAMP 2.3 was released on 30 July 2026, and its buyer-agent and seller-agent v2.4.2 followed on 18 August. IAB Tech Lab announced AAMP 3.0 on 22 September 2026. It adds OpenProposal, which is in public comment until 22 October 2026 and states that it is not a ratified IAB Tech Lab standard [552, 553, 36, 554, 147].

No AAMP artifact screened names an attention metric. OpenProposal's examples of committed metrics are viewable rate, completion rate and an audience-modelled metric for DOOH (digital out-of-home). The buyer and seller agents use the word only as ordinary English and in bundled copies of the Agentic Audiences draft taxonomy, which lists attention embeddings as a future extension. The reference buyer agent reports viewability and video completion rate [36, 153, 154, 147]. By inference, OpenProposal's measurement-source and shortfall-remedy fields could carry an attention guarantee, but no document shows one [36, 153, 154, 147]. The draft also records a problem any attention guarantee would inherit: once one line item guarantees viewable rate and another completion rate, "there is no normalised way for a buying agent to compare or optimise across them" [36].

![Beyond schema representation, attention reaches only test-mode handling with a placeholder vendor.](figures/fig-protocol-states.svg)

*Figure fig-protocol-states. **Beyond schema representation, attention reaches only test-mode handling with a placeholder vendor.** Each attention signal in the agentic protocols current at the cutoff, against the five states: schema representation, implemented connector, demonstrated interoperability test, documented production use and demonstrated transaction. The only connector evidence is a conformance test with a placeholder attention vendor and simulated delivery. Units: state. Scope: AdCP 3.0 and 3.1.24, AAMP 2.3 and the 3.0 draft. Sources [138], [139], [140], [33], [141], [142], [143], [144], [145], [146], [147], [36], [148]. Caveat: The author co-leads AdCP’s Signals & Measurement working group, and a sales agent operated by No Fluff Advisory is among those that passed the test-mode storyboard; neither earns a state. Data: data/figure-data.json#protocols*

### Viewability travels further

H5's descriptive comparator is the state other measurement signals reach in the same protocols. Its other comparator, the rival hypothesis, is taken up below. Viewability reaches more. It is a standard field in AdCP 3.1 delivery reporting [143, 556], a committed-metric example in OpenProposal and a reported metric in IAB Tech Lab's reference buyer agent [36, 153, 154, 147].

Its furthest evidence is a vendor's account. DoubleVerify's blog of 24 July 2026 describes Butler/Till using Claude as a buyer agent through AdCP, connected to DV's Media AdVantage Platform through MCP, with execution on PubMatic's AgenticOS. Brand suitability is the core control, and fraud and invalid-traffic avoidance, viewability thresholds and contextual alignment are optional add-ons. The post does not say which protocol carries which control, and it does not mention attention, although DoubleVerify sells Authentic Attention [148]. That is vendor-reported production use of DV data in an AdCP workflow for verification controls, with no attention in it [217, 559, 218, 148, 560]. PubMatic's AgenticOS page is titled as an early-access programme while claiming live campaigns, and neither it nor PubMatic's Decision Fabric release mentions attention [37, 38].

On this review's reading, the difference tracks definition, since AdCP's maintainers say no MRC-or-equivalent body has graduated a shared attention definition [143, 556]. No source explains why DoubleVerify's account of the Butler/Till workflow does not mention attention, and its absence from a case study does not show it was dropped [148].

### Vendors' own agent surfaces

Of the media-attention vendors whose agent surfaces were screened (DoubleVerify, IAS, Adelaide, Lumen, TVision and Amplified), only DoubleVerify was found with a live agent server [157, 141]. It announced DV Neura on 17 June 2026: clients can use Claude to connect through MCP to the DV Neura Insight Agent, which analyses DV's "media quality and performance data", and DV claims AdCP support. The endpoint is live and gated, and its tool list is not public. The DV Neura Activation Agent, announced for the third quarter, was still coming "in the months ahead" on 17 September 2026. That day DV launched Ask Neura, inside its DV Pinnacle platform, which lets DV customers analyse campaign effectiveness across fraud, brand suitability, viewability, geographic alignment and attention [217, 559, 218, 148, 560]. Whether the MCP server returns Authentic Attention to a buyer's agent is undocumented [142, 141, 151, 33].

IAS announced IAS Agent on 16 December 2025 as an assistant inside the IAS user interface for brand safety and suitability, insights and campaign setup, with a global rollout expected early in the first quarter of 2026. No source found confirms general availability, and the release does not mention attention [155, 555]. It has no agent-facing interface and reaches none of the five states for attention [155, 555].

In the vendor pages screened, this review found MCP access to attention predictions advertised by Viomba, Adverteyes, Amplified and DAIVID. Viomba markets "Viomba MCP" as a human attention intelligence layer for AI agents that predicts, scores and optimises ad attention across display, social and video. Adverteyes has a dedicated MCP product page, and Amplified offers AttentionAI through an API and MCP. None of the three documents interoperability, production use or transactions [562, 344]. DAIVID's site sells its Creative Intelligence via "API, MCP or Self Serve" [544, 545, 546, 547, 548, 39, 549, 550], and its published API reference documents the API but not the MCP route [39]. Adverteyes describes its MCP server as read-only and opened it "first with a limited group" on 21 July 2026; its answers come from a trends knowledge base [533, 534, 535, 536, 537, 538, 539, 540, 541, 542, 543]. The official MCP Registry listed no Amplified server on 29 September 2026, and the API host in Amplified's published sample did not resolve in public DNS [196, 429, 558]. Adverteyes' and Amplified's products predict attention to creative before launch, and their outputs are model estimates, never observed attention [529, 530, 531, 532] [192, 429]. A vendor page claiming an MCP server is at most a claimed connector to a vendor API, and none of Viomba, Adverteyes or Amplified states AdCP or AAMP support [562, 344] [196, 429, 558].

Among trading-table sellers, TripleLift says its TL Direct supports MCP-based access and names no attention metric or vendor [40]. For Adelaide, Lumen and TVision, a Viant subsidiary since 1 May 2026 [435, 82], this review found no MCP server, AdCP measurement agent or other agent-facing API returning attention metrics in the registries, code search and vendor sitemaps screened; Viant's newsroom could not be read [157, 141]. Beyond PubMatic's AgenticOS page and Decision Fabric release, the agent products of trading platforms, such as The Trade Desk, Amazon DSP and Yahoo DSP, were not screened. That is a gap in this review.

### H5, coded by the rule

H5 is unresolved at 28 and 29 September 2026. In the agent-driven buying workflows screened, no production use or transaction involving a named vendor's attention signal is documented, and no named vendor's signal is documented beyond schema representation in AdCP v3.1.24. Separately, the test-mode passes show sellers handling a placeholder metric, with no real attention vendor as connector or test partner [142, 141, 151, 33]. The supported rule needs production use or a transaction, so it is not met. The contradicted rule needs a source documenting an attention signal dropped, not carried or failing in an agent-driven buying workflow or interoperability test, with none reaching production use. The 3.0 carriage gap was found in a design review of the specification, not in a buying workflow or a test, and 3.0 is not a version current at the cutoff, so it does not meet that bar [146, 32, 143, 150] [142, 141, 151, 33].

The rival hypothesis cannot be tested with this evidence. The author's essay of 10 July 2026 asks which metric breaks first "once agents become the primary buyers", and argues that attention "fails hardest and fastest" because the registration event it infers has no referent once an agent, not a person, makes the decision [557]. That mechanism concerns agents as the decider being marketed to. H5 concerns attention measured on human viewers and carried through agent-driven buying. The protocol-state evidence neither supports nor refutes the essay, which is the author's own work and counts as no evidence (inference) [557]. The nearest thing to support is AdCP's refusal to treat a vendor-agnostic attention goal as meaningful [32]. That point concerns definition, and on this review's reading it would stand with no agent involved.

### What would move the verdict

Support needs a named vendor's attention metric in production use or a transaction in an agent-driven buy, with protocol and version recorded. Earlier states would show first: an attention vendor's measurement agent in a registry with a published catalog, a storyboard pass with a named vendor in place of the placeholder, or DoubleVerify documenting whether its MCP tools return Authentic Attention. A documented case of an attention signal dropped, not carried or failing in an agent-driven buy or interoperability test would move H5 towards contradicted. AdCP ships often and OpenProposal is in comment until 22 October 2026, so each state holds as read on 28 and 29 September 2026 [567] [552, 553, 36, 554, 147].

A buyer offered an agent-driven buy on attention can ask which vendor's metric travels in which field, and whether any buy has settled on it. [Chapter 12](#next) turns those into questions for sellers.

The register lists the protocol, registry and platform surfaces screened, with each one's status and the observable behind it. Surfaces covered only in the text above, such as vendors' MCP servers and TripleLift's TL Direct, are not in it.

**AI and agentic capability map**

| Platform / product | Capability | Status | Date | Independently observable? | What the record supports | Src |
|---|---|---|---|---|---|---|
| AdCP (AgenticAdvertising.org) v3.1.24 | Vendor-bound attention metric: reporting (vendor_metric_values), forecast, optimisation goal kind vendor_metric | live | 2026-09-23 | Schema files at tag v3.1.24 | Schema representation only; units vendor-defined; examples use placeholder attentionvendor.example | [138] |
| AdCP v3.1.24 | performance_standards metric attention_score (contractual floor, rate over measurable_impressions) | live | 2026-09-23 | enums/performance-standard-metric.json | Threshold semantics under-specified (doc example 65 vs 0-1 schema) | [144] |
| AdCP v3.1.x | Seller-native attention_seconds / attention_score optimisation values | live | 2026-06-18 | product.json and optimization-goal enums; still present in v3.2.0-rc.7 | Deprecated in 3.1.0 in favour of vendor binding | [143] |
| AdCP v3.1 measurement.core | Measurement-vendor metric catalog (get_adcp_capabilities.measurement.metrics[]) | beta | 2026-06-18 | whats-new-in-3-1.mdx | No attention vendor catalog registered (see registry row) | [149] |
| AdCP v3.2.0-rc.7 | Vendor-bound attention metric (unchanged pattern) | beta | 2026-09-27 | Schema files at tag v3.2.0-rc.7 | No attention-specific change found in files checked | [150] |
| AgenticAdvertising.org agent registry | Measurement agent publishing attention metrics | unconfirmed | 2026-09-28 | Registry API responses | 34 agents listed: 29 sales, 3 signals, 1 creative, 1 buying; two are the author's | [141] |
| Prebid sales agent (prebid/salesagent) | Vendor-metric optimisation flow (attention-style goals) | unconfirmed | 2026-09-17 | storyboard-issue-map.yaml | Reference AdCP seller; mapped to 3.1.1 storyboards | [151] |
| Prebid sales agent (prebid/salesagent) | Vendor-metric reporting accountability (placeholder attention vendor, simulated delivery) | unconfirmed | 2026-09-17 | storyboard-issue-map.yaml; dormant_scenarios.txt | Test coverage only; no real attention vendor | [152] |
| IAB Tech Lab AAMP 2.3 | Attention metric in AAMP repositories | unconfirmed | 2026-07-30 | Code search and release text | Buyer agent reports viewability and VCR instead | [147] |
| IAB Tech Lab AAMP 3.0 / OpenProposal v3.0-draft-1 | committed_metrics tuples and measurement_source vendor (attention not named) | announced | 2026-09-24 | spec/openproposal-3.0.md section 5.7 | Generic tuple could carry any metric (inference) | [36] |
| IAB Tech Lab Agentic Audiences | Attention embeddings | announced | 2026 | embedding-taxonomy.md section 6.1 | Anticipated signal type, not specified | [153] |
| IAB Tech Lab buyer-agent v2.4.2 (comparator) | Viewability and video completion rate reporting | live | 2026-08-18 | stats_retrieval.py | Comparator signal; no attention field | [154] |
| DoubleVerify DV Media AdVantage Platform via MCP (Butler/Till) | Brand suitability, fraud/IVT, viewability thresholds, contextual controls in an AdCP buyer-agent workflow | live | 2026-07-24 | Vendor blog | Attention (DV Authentic Attention) not mentioned | [148] |
| PubMatic AgenticOS | Executes agent buys applying selected DV protections before the auction | live | 2026-07-24 | DV blog; PubMatic product page | No attention mention on either page | [148] |
| PubMatic Decision Fabric on AgenticOS | Partner decisioning models inside the auction (ARTF containers) | beta | 2026-06-01 | Press release | No attention mention | [38] |
| IAS Agent | AI assistant in the IAS UI (brand safety/suitability recommendations, insights, setup) | announced | 2025-12-16 | Press release; product page | User-facing assistant, not an agent-facing API; no attention mention | [155] |
| Amplified planning-data API | Attention planning data ingested by Omnicom planning tools via API | live | 2022 | Case study page | Machine-to-machine API, not an agentic or protocol integration | [156] |
| Official MCP Registry | MCP server from an attention vendor | unconfirmed | 2026-09-28 | Registry API responses | Private or unlisted servers not visible | [157] |
| Adelaide, Lumen, TVision/Viant | Agent-facing API, MCP server or AdCP agent for attention | unconfirmed | 2026-09-28 | AdCP registry; MCP registry; sitemap screens (see searches) | Viant newsroom unreadable; bounded absence | [141] |



## 11. The Focus Quotient


*The Focus Quotient (FQ) summarises what each product documents in public, within its cohort. The field is close together, and no product reaches the Most documented band. Among the weakest dimensions in both cohorts are the two a buyer would need before trading on attention: outcome validation and independent assurance.*

> **In short.**
> 20 products are scored: 8 in cohort A (media attention measurement) and 12 in cohort B (creative attention testing). The field FQ is 51 in cohort A and 51 in cohort B. 10 products sit in the Well documented band, 10 in Partly documented, and none in Most documented. No cell scores 5: every product's record rests on vendor documentation, which caps a cell at 4, or on a narrow independent record. Outcome validation and independent assurance are the two lowest-scoring dimensions in cohort A and, with production scale, the three lowest in cohort B. The FQ measures documented capability, not performance: a product with a thin public record scores lower than its engineering may deserve, and the tests below show how much single cells move it.

### What the quotient does

Each product was scored from 0 to 5 on the dimensions that apply to its cohort, against anchors fixed before scoring ([chapter 13](#method)). The quotient groups those cells into three: Depth (what the measurement is: method validity, outcome validation, granularity and timing, and activation or creative diagnostics), Reach (channels, data access, production scale and US coverage) and Standing (independent assurance, method disclosure and privacy). A product's Raw score is 0.45 × Depth + 0.25 × Reach + 0.30 × Standing, each group the mean of its counted cells, renormalised if a group is missing. Cells that are n/e (no evidence) or n/a (outside the cohort's mask) are skipped, never counted as zero.

Raw is then pulled toward the cohort's field mean in proportion to how thin the record is: Shrunk = (k × Raw + K × F) / (k + K), where k is the number of counted cells, F is the mean Raw of the cohort's admitted rows and K, the prior strength, is 4. The FQ is 100 × Shrunk / 5. K is an author-set heuristic, not a statistical estimate, and the tests below rerun the index without it. Bands come from the unrounded distance to the field mean, 20 × (Shrunk − F): Most documented at +12 or more, Well documented at +1 or more, Partly documented at −8 or more, Least documented below. A band needs at least 6 counted cells with every group represented, and a cohort shows bands only when 6 or more of its rows qualify. Both cohorts do.

The field mean here is the mean of the cohort's Raw scores. That differs from the author's Contextual and Native Quotients, which pooled group scores, so an FQ is not comparable with a CQ or an NQ, and a cohort A score is not comparable with a cohort B score.

### The results

![Read the band, not the rank: no product clears the most documented line.](figures/fig-quotient.svg)

*Figure fig-quotient. **Read the band, not the rank: no product clears the most documented line.** The Focus Quotient for the products with enough evidence to band, one section per cohort. The Focus Quotient uses the band offsets of the author’s Contextual Quotient, but its field mean is the cohort’s mean Raw score rather than a pooled group mean, so the two do not read identically; neither construction validates it. Depth, Reach, Standing group means are combined 45/25/30 and shrunk toward the cohort’s field mean (the mean Raw score of its admitted rows) with a prior strength of 4; bands come from the unrounded distance to that field score, and only verified or qualified cells count. A band needs at least 6 counted cells with one in every applicable group, in a cohort with at least 6 such rows. Across the reweightings, positions move by up to 4 places within a cohort, and 17 of the 20 banded products keep their band under every weighting. FQ measures publicly documented capability within a cohort, not product performance. Units: FQ points on a 0–100 scale, pulled toward the cohort field mean; shown rounded, plotted unrounded. Scope: 20 banded products of 20 in the capability matrix, scores as of 1 October 2026. Caveat: One initial scoring pass; six rows nearest a band line re-scored blind. The index inherits every limit of the dimension scores and is sensitive to the weights; the bands, not the positions, are the finding. Data: data/fq.csv*

In cohort A, the field FQ is 51; a product needs 51.7 or more for Well documented and 62.7 for Most documented. In cohort B, the field FQ is 51, with Well documented from 51.8 and Most documented from 62.8. At these field means, a product scoring 5 on every applicable dimension would reach 87 in cohort A and 87 in cohort B, not 100, because of the shrinkage.

The spread is narrow, and a few cells separate the bands. In cohort A, the Well documented rows document more on privacy posture, outcome evidence, activation in buying platforms and data access. Two features a reader might expect to lift a row do not: the one product with an MRC accreditation for attention sits in the Partly documented band, and so do the rows with the largest stated production volumes. In cohort B, the Well documented rows document finer-grained diagnostics, a stronger privacy posture and more outcome evidence.

**The capability matrix** (0–5 evidence-bounded scores; n/e = insufficient evidence; n/a = outside the cohort’s mask; unverified cells are marked and counted as n/e; availability, evidence grade and verification status per cell are in data/capability-matrix-long.csv; one section per cohort; banded products ordered by FQ, the Focus Quotient, within their cohort, and unbanded rows listed A to Z)

*A · Media attention measurement*

| Vendor | Product | Group | FQ (band) | Validity | Outcomes | Grain | Activation | Creative | Channels | Access | Scale | US | Assurance | Disclosure | Privacy | Counted |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Lumen Research | Lumen Attention Measurement & Planning Platform | specialist | 57 (Well documented) | 3 | 3 | 4 | 3 | n/a | 3 | 3 | 2 | 3 | 1 | 3 | 4 | 11/11 |
| xpln.ai | xpln.ai attention metrics | specialist | 56 (Well documented) | 3 | 1 | 3 | 3 | n/a | 3 | 3 | 3 | 3 | 3 | 3 | 4 | 11/11 |
| TVision | TVision | media_owned | 53 (Well documented) | 3 | 3 | 3 | 3 (beta) | n/a | 1 | 3 | 1 | 4 | 1 | 3 | 4 | 11/11 |
| Adelaide | Adelaide AU | specialist | 53 (Well documented) | 2 | 2 | 3 | 3 | n/a | 3 | 3 | 3 | 3 | 1 | 3 | 4 | 11/11 |
| DoubleVerify | DoubleVerify Authentic Attention | verification | 49 (Partly documented) | 2 | 1 | 3 | 2 | n/a | 3 | 1 | 3 | 3 | 4 | 3 | 2 | 11/11 |
| Amplified | Amplified Campaign Performance | specialist | 48 (Partly documented) | 3 | 1 | 3 | 2 (unconfirmed) | n/a | 3 | 2 (unconfirmed) | n/e | 1 (unconfirmed) | 1 | 3 | 4 | 10/11 |
| Integral Ad Science | IAS Quality Attention | verification | 46 (Partly documented) | 2 | 1 | 3 | 2 (beta) | n/a | 3 | 2 | 3 | 3 | 2 | 2 | 2 | 11/11 |
| Amplified | Amplified Campaign Strategy | specialist | 45 (Partly documented) | 3 | 1 | 2 | 1 | n/a | 3 | 3 | n/e | 3 (unconfirmed) | 1 | 3 | 2 | 10/11 |

*B · Creative attention testing*

| Vendor | Product | Group | FQ (band) | Validity | Outcomes | Grain | Activation | Creative | Channels | Access | Scale | US | Assurance | Disclosure | Privacy | Counted |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Adverteyes | Adverteyes PreView | specialist | 55 (Well documented) | 3 | 2 | 4 | n/a | 4 | 3 | 3 | n/e | 3 | 1 | 3 | 2 | 10/11 |
| Dragonfly AI | Dragonfly AI | specialist | 55 (Well documented) | 3 | 2 | 4 | n/a | 4 | 3 | 3 | n/e | 3 | 1 | 3 | 2 | 10/11 |
| eye square | eye square InContext | research | 54 (Well documented) | 3 | 2 | 3 | n/a | 3 | 3 | 1 | 1 | 3 | 4 | 3 | 3 | 11/11 |
| Lumen Research | Lumen SPOTLIGHT | specialist | 53 (Well documented) | 3 | 4 | 4 | n/a | 3 | 3 | 1 | n/e | 1 (unconfirmed) | 1 | 3 | 3 | 10/11 |
| Neurons | Neurons AI | research | 53 (Well documented) | 3 | 1 | 4 | n/a | 4 | 3 | 3 | 1 | 3 | 1 | 3 | 3 | 11/11 |
| Tobii | Tobii Sticky | research | 53 (Well documented) | 3 | 1 | 4 | n/a | 3 | 3 | 3 | 1 | 3 | 1 | 3 | 4 | 11/11 |
| Kantar | Kantar Context Lab | research | 50 (Partly documented) | 3 | n/e | 3 | n/a | 3 | 3 | 1 | 1 | 3 | 1 | 3 | 2 | 10/11 |
| Amplified | Amplified AttentionAI | specialist | 49 (Partly documented) | 3 | 1 | 4 | n/a | 4 | 3 | 1 | 3 | 1 (unconfirmed) | 1 | 3 | 2 | 11/11 |
| Kantar | Kantar LINK+ | research | 48 (Partly documented) | 3 | 1 | 3 | n/a | 3 | 3 | 3 | 1 | 3 | 1 | 3 | 2 | 11/11 |
| DAIVID | DAIVID Creative Data Feed API | research | 48 (Partly documented, provisional: Partly documented with n/e = 1, Well documented with n/e = 4) | 3 | 1 | 3 | n/a | 3 | 3 | 3 | n/e | 3 | 1 | 2 | n/e | 9/11 |
| Kantar | Kantar LINK AI | research | 47 (Partly documented) | 1 | n/e | 3 | n/a | 3 | 3 | 3 | 1 | 3 | 1 | 3 | 2 | 10/11 |
| iMotions | iMotions Media Analytics | research | 46 (Partly documented) | 3 | 1 | 3 | n/a | 3 | 3 | 1 | 1 | 3 | 1 | 3 | 2 | 11/11 |

### Where the field is weak

The pattern across both cohorts matters more than any row.

Outcome validation is the lowest-scoring dimension in cohort A, and among the three lowest in cohort B. In the sources screened, no product documents independent, replicated validation against brand or sales outcomes, and none documents a held-out comparison against time in view [46, 51, 52] [8, 10, 71, 64] ([chapter 3](#validation)). The highest scores, 3 in cohort A, rest on in-sample evidence without a holdout. For Lumen it is PwC's review, commissioned by Lumen, which found attention not significantly better than viewability on awareness and recall in its pilot [8]. For TVision it is an independent academic study of its panel data, in-sample [52].

Independent assurance is the other weak dimension. In cohort A, on MRC's lists as read on 28 and 29 September 2026, one product holds an MRC accreditation for its attention metrics, covering three environments and not CTV [2]; one is in an MRC-confirmed audit [48]; one has had a pre-audit review only [48, 339]; and IAS left the process in January 2026 [49], keeping accreditation for the viewability and invalid-traffic measurement its attention score builds on ([chapter 2](#standards)). In cohort B, in the sources screened, no product documents an independent audit or validation of its testing method. One, eye square InContext, holds an ISO 20252 certificate for its company's market-research process, which does not cover the attention metric and expires on 13 November 2026 [524, 525, 526, 502, 527]. The cohort B anchors score such a certificate 4, and the row's Well documented band depends on that cell: scored as self-attestation, the row would fall to Partly documented on this review's recomputation. Every other scored assurance cell in cohort B rests on self-attestation.

Production scale is often missing, especially in cohort B, where it scores lowest of all. Several products state no volume for a period, or only a cumulative total with no start date, which the anchors cap low. Where a volume is stated, it comes from the vendor, its parent, or a partner or client, and none is independently audited.

The strongest cells are privacy posture and granularity in cohort A, and granularity, creative diagnostics and channel coverage in cohort B, where vendor documentation describes the product in detail. That is what the caps are for: a detailed help page is not proof, so it earns a 4 at most.

### How stable the bands are

The index was recomputed under four other weightings, three other prior strengths and every single-cell deletion, and the bands compared.

Under the four reweightings (equal thirds; Depth-led 60/20/20; Reach-led 30/45/25; Standing-led 30/20/50), recomputing Raw and the field mean each time, the largest position change is 2 places in cohort A and 4 in cohort B; 17 of 20 banded products keep their band under all four.

Removing the shrinkage (K = 0, which reports Raw) changes 0 bands in cohort A and 0 in cohort B; the largest position change is 0 places in cohort A and 1 in cohort B. At K = 2 and K = 8 no band changes in either cohort.

Deleting one counted cell at a time, with every other row and the field mean held fixed (86 deletions in cohort A, 125 in cohort B) changes a band in 6 and 10 cases respectively, and never removes a product from the matrix. Leaving out a low score raises a quotient: missing evidence is not neutral. The largest single-cell effects come from deleting a score of 1: removing Amplified Campaign Performance's independent assurance cell would raise its FQ from 48 to 51, and removing Amplified AttentionAI's outcome validation cell would raise it from 49 to 54.

The missing-evidence scenarios test the opposite case, on the rows that have n/e cells (2 in cohort A and 6 in cohort B). Each such row's n/e cells are filled with 1, then 4, one row at a time with every other row and the field mean frozen. These are scenarios, not bounds. Products that change band under one of them, marked ± in the matrix as provisional: 1. The six rows nearest a band line were also scored a second time, blind ([chapter 13](#method)); a band that the second scorer’s scores would move is marked †. Rows so marked: 0.

### What the quotient is not

The FQ says how much of a product's capability is documented in public, and how well. It does not say which product predicts outcomes best, which is the question chapter 3 could not answer from public evidence. It does not compare vendor attention units, which are not comparable. It rests on one initial scoring pass under published rules, a consistency audit, and blind rescoring of the six rows nearest a band line by a separate AI scorer; every other row was scored once, and no human second scorer was used.

### The profiles

The paper profiles 8 companies in depth: six whose profile row is in cohort A, one in cohort B, and one media owner from the trading table, whose offer is described but not scored. Each profile has a SWOT in which every line cites its sources and is tagged as record, vendor documentation, vendor claim or inference.

### Adelaide (active, independent)

*Adelaide Metrics Inc.*

The attention unit most used in trading offers; MRC pre-audit reviewed, audit not begun.

**SWOT (2026-09-29).** Adelaide sells a model-estimated 0-100 placement score (AU) with wide buying-platform reach and a growing set of seller guarantees, but it has passed only an MRC pre-audit review and publishes no independent validation.

- **Strengths.** AU is marketed as live in several DSPs; Google's DV360 Help Center documented live steps for the 'Maximize Adelaide AU' algorithm by 29 September 2026; go-live date undocumented. (vendor documentation) [158, 159, 160, 161, 162] Adelaide says AU pre-bid targeting is also available in Yahoo DSP, Viant, Adobe Advertising DSP and Equativ, plus SSP and curation routes via PubMatic, Comscore and Index. (vendor claim) [163, 164, 165, 166] AU has the broadest stated scope of any attention metric on MRC's under-review page: digital display and video, audio, linear TV, cinema and digital OOH. (record) [48] Sellers use AU to package and guarantee inventory: the WSJ ran attention-guaranteed campaigns, Vevo launched 'Attention Guaranteed' and Duolingo launched AU Guarantees in September 2026. (vendor claim) [167, 78, 97] Adelaide names its gaze reference suppliers (Amplified Intelligence, Lumen, Tobii, TVision, Viomba) and publishes channel-level method descriptions for display, video and CTV. (vendor documentation) [106, 168, 169]
- **Weaknesses.** AU is not MRC-accredited: MRC lists it as 'Pre-audit reviewed 5/2026, Audit not begun', and Adelaide states its viewability and SIVT filtration are not accredited either. (record) [48, 2, 3] AU is a model estimate, not observed attention; reference panel sizes, model weights and any held-out agreement statistic are not published. (vendor documentation) [43, 66, 169] Outcome evidence is vendor-selected or commissioned: the MediaSense review used Adelaide-supplied data, without independent data verification or source-code review. (vendor claim) [170, 171, 15, 14] Where an AU guarantee names its measurer (Journal, Kargo, TripleLift), it is Adelaide, the metric's supplier; Vevo, Duolingo and Teads state none. No independent delivery audit is described. (inference) [94, 93, 91, 97, 78] Disclosed funding is about 10.9M USD announced to August 2024; two Form D notices report 4.59M USD sold at filing and mark revenue 'Decline to Disclose'. (record) [172, 173, 81, 174]
- **Opportunities.** The IAB and MRC Attention Measurement Guidelines v1.0 (November 2025) give Adelaide a published standard to audit against as it moves beyond pre-audit. (inference) [41, 48] Google's platform-side DV360 algorithm, with a per-line-item media fee, turns AU from a bespoke script into a packaged buying option with a visible fee mechanism. (inference) [158, 160] Custom AU, launched November 2025 with Uber Advertising, lets platforms calibrate AU with first-party signals, a route into retail and commerce media sellers. (inference) [175] Seller guarantees keep arriving (Vevo January 2026, Duolingo September 2026), and the AU Ecosystem reported 70+ sell-side members by July 2026. (inference) [78, 97, 176]
- **Threats.** xpln.ai's attention metrics have been in MRC audit since November 2025, while Adelaide's audit had not begun at 29 September 2026. (inference) [48, 177, 178] Viant has owned TVision, Adelaide's CTV and linear reference-data supplier, since 1 May 2026; whether the data licence continues is undocumented. (inference) [82, 106] Nielsen's Outcomes Marketplace carries AU; Nielsen's parent agreed in August 2026 to buy DoubleVerify, sole holder of MRC attention accreditation, a deal pending at 29 September 2026. (inference) [179, 180, 84, 181, 2] As buyers optimise toward AU, its correlation with outcomes could weaken, a risk Adelaide itself raises about duration metrics; no test of this was found. (inference) [1] Partner platforms can close routes: Microsoft retired the Xandr DSP, a 2023 High Attention Markets partner, on 28 February 2026. (inference) [182, 183]

- **What is sold.** Measurement, planning and activation built on AU, a vendor-defined 0-100 media-quality score that Adelaide describes as a placement's probability of attention and subsequent impact, produced by a machine-learning model from placement and device signals, licensed eye-tracking data and full-funnel outcome data. Sold as campaign measurement in a dashboard (tag, pixel, log-file or post-log based), a planning tool (Flight Control, target AU and AU floors), pre-bid segments and custom bidding inputs in named DSPs, curated high-AU deals on SSPs and curation platforms, publisher audits and packages (Adelaide for Publishers, AU Ecosystem, AU Guarantees) and platform-calibrated variants (Custom AU).
- **Who buys it.** Advertisers and their agencies (named US users in vendor and partner sources include Hershey, Verizon, iFIT, the NBA; five holding-company media groups were reported as signing on to Flight Control in March 2024) and media sellers (publishers such as The New York Times, The Wall Street Journal, the Guardian, Duolingo and Vevo buy audits or build AU packages and guarantees; Condé Nast and Hearst sponsor AU measurement for advertisers on their inventory).
- **How it earns.** Not disclosed. The sources show measurement for advertisers, publisher tiers (AFP Certified, Light, Pro, Enterprise, unpriced, with a free audit tier) and publisher-sponsored measurement; no price, fee level, data-segment fee or revenue figure was found in the sources screened.
- **Supply and data dependencies.** Reference gaze data licensed from third parties (Amplified Intelligence, Lumen, Tobii, TVision, Viomba) plus Adelaide's own eye-tracking tools; outcome labels from brand-lift providers (Kantar, Lucid, Dynata), tune-in attribution providers (Samba, VideoAmp), advertiser sales and conversion data and the Rita panel; ACR data for CTV; podcast transcription data; distribution through DSPs (The Trade Desk, Amazon DSP, Yahoo DSP, DV360), SSPs and curation platforms (PubMatic, Magnite, Index Exchange and others) and Nielsen's Outcomes Marketplace. Disclosure: Samba, named as a tune-in attribution provider used to calibrate AU, is the author's employer.
- **Differentiation.** Placement-level, outcome-trained quality score on one 0-100 scale across channels, positioned against duration-based metrics (ATT-ADELAIDE-S04, S12). Widest documented activation footprint among the attention vendors screened in this profile: The Trade Desk, Amazon DSP and Yahoo DSP segments, DV360 custom bidding, SSP and curation deals (ATT-ADELAIDE-S16 to S25). AU defines most of the guarantees found (WSJ, Kargo, Duolingo and others via the AU Ecosystem, and Vevo's by inference); the exceptions are SeenThis, which guarantees and prices on Lumen's measure, and A+E's 2020-21 guarantee on TVision data (ATT-ADELAIDE-S26, S27, S38; ATT-E-S08; ATT-NFA-S01; ATT-F-S25). Broadest stated MRC review scope among attention vendors on MRC's under-review page (digital display and video, audio, linear TV, cinema, DOOH), though at pre-audit stage (ATT-ADELAIDE-S01). Reference-data strategy built on licensing gaze data from several providers rather than one proprietary panel, with a stated shift to infrared (ATT-ADELAIDE-S10).
- **Risks.** Accreditation risk: about 19 months from MRC pre-assessment (October 2024) to a pre-audit review (May 2026), with no audit begun at 29 September 2026; guarantee and billing uses rest on an unaccredited metric (ATT-ADELAIDE-S01, S36). Arbiter conflict: where an AU guarantee discloses who measures delivery (the Journal, Kargo and TripleLift offers), it is Adelaide, which also supplies the metric; for Vevo, Duolingo and Teads the measurer is not stated. No independent audit of guarantee delivery was found (ATT-E-S07, S08, S12; ATT-ADELAIDE-S27, S38). Reference-data dependency on third parties: TVision (acquired by Viant on 1 May 2026, per Viant's 10-Q) supplies the CTV and linear reference; whether that licence continues is not documented (ATT-ADELAIDE-S05, S11; ATT-D-S02; inference). Optimisation loop: once buyers target AU, the correlation Adelaide warns about for duration metrics could weaken for AU too, and no test of AU's performance after it became a target was found (ATT-ADELAIDE-S12; inference). Small disclosed funding base (about USD 10.9M announced to August 2024) relative to verification-suite competitors (ATT-ADELAIDE-S29, S30; derived).
- **Best fit.** Buyers who want one relative media-quality signal across channels for planning floors and pre-bid exclusion, and accept a vendor-defined, outcome-trained score (ATT-ADELAIDE-S17, S18, S45). Publishers seeking an attention audit and packaging currency for premium inventory, including sponsored or guaranteed AU offers (ATT-ADELAIDE-S26, S34, S43).
- **Poor fit.** Buyers who need an accredited attention metric or independently audited guarantee delivery (ATT-ADELAIDE-S01, S02). Research uses that need observed attention (gaze or presence of real viewers) or an impression-level attention measure with published error statistics (ATT-ADELAIDE-S06, S12). In-game and DOOH measurement, where no documented method was found (ATT-ADELAIDE-S03).
- **Limits.** AU is a model estimate of placement quality trained on outcomes; it is not observed attention and not a per-person gaze measure (ATT-ADELAIDE-S04, S06, S12). No MRC accreditation; MRC shows a pre-audit review (5/2026) with the audit not begun at 29 September 2026 (ATT-ADELAIDE-S01, S02). No published held-out agreement statistic against gaze reference data, and no panel sizes or recruitment for the reference data (ATT-ADELAIDE-S06, S10, S11). Algorithm weights are intentionally not disclosed (ATT-ADELAIDE-S07). Outcome evidence is vendor case studies plus one review requested by Adelaide with a gated report; none independent (ATT-ADELAIDE-S07, S08, S09, S14, S15). AU takes in-view duration as an input (ATT-ADELAIDE-S06) and is trained on outcomes (ATT-ADELAIDE-S05, S06), so it is not independent of time in view and AU-to-outcome correlations risk leakage (inference). No company-wide production volume, revenue or pricing is public (ATT-ADELAIDE-S29, S30, S33). AU values are vendor-defined and, with Custom AU, platform-calibrated; they cannot be compared with other vendors' units (ATT-ADELAIDE-S31).

| Dimension | Availability | Evidence | Observed |
|---|---|---|---|
| Method validity | live | C | AU is a machine-learning estimate, not an observation of the exposed person. Adelaide names its reference gaze data: eye-tracking licensed from Amplified Intelligence, Lumen, Tobii, TVision and Viomba plus its own tools (S10, S11); TVision panel data for CTV and linear TV (S05, S11); from a November 2024 partnership, Viomba's Tobii infrared panels (operated in Europe since 2014) for open-web display, replacing models that 'previously used only webcam-based eye-tracking data' (S10). Channel-specific models are described for display, OLV, CTV, linear TV and podcasts (S05, S06, S03). No reference sample size or recruitment is stated beyond 'fully opted-in panels' and Viomba's 'millions of in-the-wild empirical samples' a month (S06, S10); no held-out agreement or error statistic between AU and a gaze reference is published in the sources screened; calibration is described as calibration to outcomes (brand lift, conversions, sales), not to gaze, by channel (S05, S06). The construct itself mixes attention with outcome prediction ('probability of attention and subsequent impact', S04), so a gaze-reference validation would cover only part of it. Adelaide's September 2026 FAQ says Lumen data were used 'in early models' (S12), so the current reference mix is not fully stated. [105, 43, 106, 168, 169, 327, 1] |
| Outcome validation | live | D | Memory: MediaSense's review (February-June 2025), requested by Adelaide, rebuilt AU-to-outcome curves for aided awareness, familiarity, recall and intent from sampled, aggregated data supplied by Adelaide and reported positive relationships; MediaSense did not verify Adelaide's data or review AU source code, and its sample had limited data at the extremes of the AU scale (S07, S08); the full report is behind a download form and was not read (S09); no sample size, holdout or statistic is public. Brand: Outcomes Guide releases report per-edition averages over undisclosed case sets (52 cases, 41% higher brand lift, 2024 campaigns; 60 cases, 33% upper-funnel lift, 2025 campaigns), each said to be 'supported by third-party outcome measurement', with no design, holdout or uncertainty in the releases (S15, S14); iFIT (The Trade Desk-owned publication) reports +61% aided awareness with Upwave measurement and no stated comparator (S44). Sales: vendor claims only, e.g. a CPG advertiser for which AU was '180% more correlated with in-store sales than viewability' (2020, S35), 58% higher cumulative ROAS versus control groups in The Trade Desk beta (2024, no design, S16, S17), a bank with 2.8x conversions and 69% lower CPA (S06). No independent validation and no replication was found in the sources screened. AU is trained on outcome data (S05, S06), so AU-to-outcome correlations reported on campaign data are not shown to be out-of-sample (inference). [106, 168, 66, 170, 171, 15, 14, 328, 161, 329, 67] |
| Granularity and timing | live | C | All stages output a model estimate (AU); none is direct observation. Capture is impression-level (a tag, VAST tag or pixel fires at ad render or serve, capturing device, timestamp and placement or contextual signals; S05, S06), but Adelaide describes AU's 'Level of measurement' as 'Placement' (S12). Stages documented: planning (Flight Control target AU and AU floors, S45; benchmarks, S03); pre-bid (segments built 'from historical placement-level AU data' in Amazon DSP, S18; tiers in The Trade Desk, S17; custom bidding, S20); in-flight (dashboard optimisation with AU processing refreshed within the day, S41; Nielsen reach-and-attention efficiency metrics for mid-flight optimisation, S32); post-campaign ('impression-level visibility into attention quality after a campaign runs' in Amazon DSP, S18; linear TV scored from post-logs, S05). Impression-level output is documented only for post-campaign reporting in Amazon DSP; pre-bid outputs are placement-level tiers. Adelaide labels AU as predictive, but no stage is labelled with a measurement-basis field. [105, 106, 168, 1, 161, 162, 330, 331, 332, 333] |
| Activation | live | D | Named buying platforms with documented AU activation: The Trade Desk (pre-bid low, medium and high tiers in the self-serve platform from 8 October 2024 after an August-September beta; reported by AdExchanger from Adelaide's announcement, S16, S17); Amazon DSP (AU Media Quality tiers and AU Quality Floor across display, online video and streaming TV, 9 June 2026, vendor blog with an Amazon DSP quote, S18); Yahoo DSP (pre-bid segments launched 13 June 2023, Yahoo release; current status not verified, S21); DV360 (AU custom bidding 'previously available via bespoke scripts for display and online video', per Adelaide, S20; Google announced an off-the-shelf algorithm for mid-July, S19, and by 29 September 2026 its DV360 Help Center documented live activation steps for the 'Maximize Adelaide AU' algorithm, with an additional media fee and no documented go-live date, ATT-V-ADELAIDE-S03). Sell-side and curation: High Attention Markets (2023) with DV360, Xandr, Magnite, PubMatic, Teads, Sovrn, TRUSTX and Sharethrough (S22); PubMatic deal-level AU targeting (December 2024, S25) and Comscore-Certified Deal IDs with AU (June 2025, S24); Index Marketplaces (January 2026, S23). Adelaide claims 125+ integrations (S12). Three or more named DSPs are documented as live by vendor or partner releases; no platform-side documentation of live status was found for The Trade Desk or Amazon DSP. [1, 328, 161, 162, 159, 330, 334, 183, 166, 165, 164, 158] |
| Creative diagnostics | unconfirmed | E |   |
| Channel coverage | live | C | Against the seven scored channels: display and online video live (tag and VAST measurement, S06; AFP audits, S34); CTV live (pixel, ACR and TVision data, S05; Amazon DSP streaming TV segments, S18); social live by vendor claim (YouTube, Facebook and Instagram ratings since September 2020 from licensed Amplified data, S35; Meta report uploads in the dashboard, S41), with no platform integration documented; audio live by vendor claim (podcasts in the glossary and site, S03, S40; a 2025 audio case in the 2026 guide release, S14); DOOH listed in the glossary's coverage (S03) and in MRC's review scope (S01) with no measurement method documented in the sources screened; in-game: no evidence found. Linear TV and cinema are also claimed (S03, S05) and are attributes outside the seven. Adelaide states AU covers 'approximately 95%' of an advertiser's spend (S03) and '19 channels' (S05, S12), while the same September 2026 post also says '18 channels' (S12). Five of the seven channels are documented as live by vendor sources only. [48, 105, 106, 168, 1, 15, 162, 335, 329, 336, 332] |
| Data access and integrations | live | D | Named output routes documented: Adelaide's dashboard (campaign, channel, partner, domain and placement-level analysis, S40, S41); Nielsen's Outcomes Marketplace in Nielsen ONE (AU available there from 7 October 2025, S32); Amazon DSP post-campaign 'impression-level visibility into attention quality' (S18); Kantar LIFT partnership for AU-to-brand-lift correlation analyses (May 2024, roadmap stated, S42). Brand-lift pixels (Lucid) can be uploaded into Adelaide's campaign manager (S41), which is an input route. No public API reference or log-level export documentation was found: docs.adelaidemetrics.com returned HTTP 403 on 29 September 2026 (an access error, not evidence of absence), and no API documentation appears in the pages screened. All routes rest on vendor or partner releases. [162, 331, 336, 332, 337] |
| Production scale | live | D | The only volume stated with a period is the Google-commissioned YouTube analysis: AU measured across 395 US campaigns, 149 advertisers and 32B impressions between 2022 and 2024 (S33). Derived mean annual rate over three calendar years: about 10.7B impressions and about 132 campaigns a year (derived from a cumulative total; the rule caps such a figure at 4); the page does not say whether all 395 campaigns ran on YouTube or include comparison platforms. Other volume statements are not production counts: 'nearly half of Fortune 50 companies' and 'billions in media spend' (spend, 2024, S29); 'hundreds of campaigns' (S07); 'nearly 5 trillion monthly impressions' of addressable supply in High Attention Markets (2023, S22); 60 case studies in the 2026 guide (S14). No company-wide impressions or campaigns measured per year was found in the sources screened. [66, 15, 183, 81, 338] |
| US availability and coverage | live | D | Generally available in the US: operating entity Adelaide Metrics Inc. in New York (S03), US DSP activation (The Trade Desk, Amazon DSP, Yahoo DSP; S17, S18, S21) and named US production use (Verizon, S21; Hershey, S16; iFIT, S44; the NBA, S40; 395 US campaigns in the YouTube analysis, S33). US calibration data are not explicitly documented: the CTV and linear reference is TVision panel data and tune-in attribution from Samba and VideoAmp (S05), but Adelaide's sources do not state the geography of these datasets; the open-web infrared reference is Viomba's panel, which Adelaide describes as operated in Europe (S10). Stating that TVision's panel, Samba and VideoAmp are US datasets would rest on sources not opened here. [105, 106, 169, 328, 161, 162, 334, 338, 336, 67] |
| Independent assurance | live | A | MRC's Digital Services Under Review page (accessed 29 September 2026) lists 'Adelaide (Pre-audit reviewed 5/2026, Audit not begun)' for 'AU (media placements' potential to deliver an attentive Impression)' across digital display and video, audio (podcasts, streaming, terrestrial), linear TV, cinema and digital OOH (S01). MRC's accredited digital list lists 'Attention Metrics' only for DoubleVerify Quality Analytics Platform, and no Adelaide entry appears on it (S02). No MRC accreditation of viewability or invalid-traffic inputs for Adelaide was found. MRC told Digiday on 29 October 2024 that Adelaide was in pre-assessment, 'not yet audit' (S36). Adelaide's own claim that being 'under MRC review' makes it the first pure-play vendor 'to achieve this status' (S12) is a vendor assertion; the accreditor record shows no audit in progress. MediaSense's 2025 review was requested by Adelaide and is not an accreditation or certification (S07, S08). [48, 2, 66, 170, 1, 339] |
| Method disclosure | live | C | Against the seven elements: (1) construct published (0-100 probability of attention and subsequent impact; placement-level media quality, S04, S12); (2) signals published by channel (coverage, duration in view, position, clutter, device for display; audibility and player type for OLV; app, app type, genre, time of day and pod position for CTV; program, network, pod position and duration for linear; genre, ad position, ad load and ad type for podcasts; S03, S05, S06); (3) reference labels named (gaze-data providers and outcome providers, S05, S10, S11) without their weighting; (4) panel size and recruitment not published ('fully opted-in panels'; Viomba sample volumes only, S06, S10); (5) model type and procedure described in summary (machine-learning model; four-step display process, S06), while Adelaide states it does not publicly disclose algorithm weights (S07); (6) no validation statistic published in the sources screened (MediaSense findings are described in words; the report is gated, S07, S09); (7) no recalibration cadence published; AU version 7 was the version under MRC scrutiny in October 2024 (S36) and dashboard processing frequency is a data-refresh statement (S41). About three elements fully and two partly published; no data. [105, 43, 106, 168, 66, 171, 169, 327, 1, 339, 332] |
| Privacy posture | live | C | Product-specific policy: the AU Ratings Privacy Policy (effective 13 September 2024) names the product, lists the data its tracking technology collects (device type, IP address, location to country, city, DMA and zip code, inferences), says Adelaide typically acts as a service provider or processor, does not sell or share personal information under the CCPA and does not process sensitive personal information, and states that AU Ratings do not process personal data under the GDPR or UK GDPR and do not use cookies or device storage under the ePrivacy Directive (S13). Consent model: eye-tracking data come from 'fully opted-in panels', licensed or in-house (S06, S33), and Adelaide says it uses 'fully opted-in eye tracking and outcome data' (S11); Rita's panel is built on members exercising EU data-access rights (S28). No independent certification, regulator record or audit of AU's privacy compliance was found in the sources screened, and no regulator action naming Adelaide was found (bounded to these sources). Tension: the policy lists IP address as personal information collected (CCPA section) while stating that AU Ratings process no GDPR personal data, and the OLV method page says AU works 'without relying on user-based identifiers, cookies, or personally identifiable information' (S06, S13). [168, 327, 340, 341, 338] |

Performance evidence: MediaSense's review found AU scores to be 'strong proxies' for brand outcomes and rebuilt AU curves showing higher AU associated with better aided awareness, familiarity, recall and intent. (commissioned measurement, sponsor Adelaide (asked/requested MediaSense to conduct the review; payment terms not stated)) [66, 170, 171]; Outcomes Guides: 52 case studies (2024 campaigns) averaging 41% higher brand lift and 55% stronger lower-funnel impact; 60 case studies (2025 campaigns) averaging 33% upper-funnel and 53% lower-funnel lift. (vendor assertion, sponsor Adelaide) [15, 14]; The Trade Desk pre-bid beta (August-September 2024): 58% higher cumulative ROAS, 12% better attentive media quality and 62% higher CTR than control groups. (vendor assertion, sponsor Adelaide) [328, 161]; iFIT, buying with Adelaide segments in The Trade Desk and measuring with Upwave, saw +61% aided awareness, +17% consideration and +10% favourability. (vendor assertion, sponsor The Trade Desk (publisher of The Current) with Adelaide and iFIT) [67, 168]; MAGNA Media Trials and Adlook: pre-bid optimisation raised AU by 65% on video; pre-bid plus contextual delivery gave a +131% delta lift in awareness. (commissioned measurement, sponsor MAGNA (IPG Mediabrands trials arm) with Adlook, an interested DSP; funding split not stated) [342]; Google-commissioned AU analysis of YouTube premium video ad solutions (2022-2024) found higher AU for YouTube's premium placements than comparison platforms. (commissioned measurement, sponsor Google) [338]; Display and CTV case claims: highest-AU display placements generated 3x more conversions than lowest-AU; a bank gained 2.8x conversions and 69% lower CPA; CTV placements above an AU floor delivered 94% more sales conversions; NBA viewers exposed to highest-AU CTV were 38% more likely to tune in. (vendor assertion, sponsor Adelaide) [106, 168]; When implementing minimum AU thresholds, advertisers increased conversion metrics by 59% and brand lift metrics by 50% on average (CEO statement). (vendor assertion, sponsor Adelaide (CEO quoted by Digiday)) [333]; A CPG advertiser found AU '180% more correlated with in-store sales than viewability' (2020). (vendor assertion, sponsor Adelaide) [329]

### Amplified (active, independent)

*Amplified Intelligence Technologies Pty Ltd (ABN 21 636 344 360; ACN 636 344 360), Australian private company; trades as Amplified (formerly Amplified Intelligence)*

Planning, measurement and creative prediction sold as separate modules.

**SWOT (2026-09-29).** Amplified owns a camera-based reference panel and sells planning, in-flight and creative-prediction modules on it, but all three report model estimates with no published validation statistics and no MRC accreditation or audit.

- **Strengths.** Amplified owns its reference data: a paid, Cint-recruited camera panel across mobile, PC and in-home TV, plus in-situ field studies such as QMS outdoor sites. (vendor documentation) [184, 185] One construct, active, passive and non-attention with attention decay, runs across its planning, in-flight and creative-prediction products. (vendor documentation) [119, 186, 187] The planning tool lists channels beyond digital, including OOH/DOOH, cinema, audio and TV/BVOD/CTV, and feeds agency planning via API (Omnicom, 2022) and Smartly (June 2025). (vendor claim) [119, 156, 188] Its 2023 ARF self-report discloses about five of seven method elements at summary level for the in-flight product, including a described 60/20/20 held-out design. (vendor documentation) [189] The in-flight tag is declared under TCF vendor 1195 with legitimate interest for purposes 7, 9 and 10 and no device access, matching IAB Europe's GVL v178. (record) [184, 190]
- **Weaknesses.** MRC lists no Amplified product as accredited or under review on 29 September 2026; its '99% data accuracy' claim names no validator, metric or reference set. (record) [2, 48, 186, 191] All three products report model estimates, yet Amplified contrasts its 'measured behaviour' with competitors' 'modelled probability'. (vendor documentation) [187, 189, 192] No held-out agreement, error or outcome statistic is published; the described 60/20/20 split, 9,000-person ground-truth sample and outcome studies come without results. (vendor documentation) [189] Stated panel and dataset sizes vary by date and page, from 108,000 people in 2023 to 2 million+ across 21 markets in May 2026; per-market sizes are unpublished. (vendor claim) [189, 193, 192, 194] No named self-serve DSP or SSP segment or packaged bidding algorithm was found; buying use is documented only via 2023 Chalice trials and an agency pre-bid score. (vendor documentation) [195, 76, 193]
- **Opportunities.** AttentionAI, launched 20 May 2026 with sample REST API calls, webhooks and a claimed MCP integration, gained exclusive African distribution through Connect in July 2026. (inference) [192, 196, 197] Publishing results for the held-out design and outcome studies it described to the ARF in 2023 would address its largest validation gap directly. (inference) [189] An MRC route for attention metrics now exists: DoubleVerify holds an Attention Metrics accreditation and XPLN's audit is in process, and Amplified contributed to the IAB/MRC guidelines. (inference) [2, 48, 41] Media-owner studies in DOOH (QMS, August 2026), cinema (Screenvision) and audio (Spotify India) give its planning data a channel beyond digital video. (inference) [198, 199, 77]
- **Threats.** Competitors hold or are in MRC processes while Amplified documents no accreditation route, leaving it with self-attestation only. (inference) [2, 48] Buyers may read modelled outputs as observations: a September 2026 post describes AttentionAI output as 'measuring ... attention at impression level'. (inference) [200, 187] Amplified names its CEO inconsistently: after Urquhart's December 2024 appointment, posts since November 2025 call the founder 'Founder and CEO' or only 'Founder'; no change announcement was found. (inference) [201, 202, 203, 198, 192, 200] Camera-based panels carry privacy exposure, including in-home TV capture of guests and EU/UK panel data processed on AWS servers in Australia. (inference) [184] Funding is disclosed only in vendor statements ('$10M+' over five years), and no revenue is published, so its scale against larger competitors is unknown. (inference) [191, 204, 201]

- **What is sold.** A SaaS attention platform with three listed modules plus services. (1) Campaign Strategy: pre-launch planning from panel-derived benchmarks and predictions (strategy builder, media plan optimiser, platform and format benchmarks, audience analysis by age and gender, creative recommendations, Attention-Adjusted Reach Curves). (2) Campaign Performance: in-flight attention measurement via tracking tags placed in the buyer's DSP (or server-to-server 'tagless' integration), reporting active and passive attention and attention decay by placement, creative, format and line item, sold on a CPM rate card; this is the former attentionPROVE product (inference from the product-name page and the privacy policy). (3) AttentionAI (launched 20 May 2026; successor to Creative Testing AI of August 2025): a creative-prediction model that simulates 20,000 viewing personas per ad and returns second-by-second predicted attention, a creative score, hook and hold indices and format benchmarks, via self-serve upload, CSV, a REST API and a claimed MCP integration. Services: bespoke research studies (formerly attentionTRACE), analysis, API integrations and partnerships.
- **Who buys it.** Media agencies and holding companies (Omnicom Media Group via planning-tool API since 2022; OMD; dentsu's 2021 research programme; Connect/The Up&Up Group in South Africa), brands (named in vendor boilerplate and releases: Mars, PepsiCo, Shell, easyJet, P&G and WPP in AttentionAI pilots), and media owners and platforms that commission studies to sell their inventory (Screenvision in the US, QMS, Spotify, Pinterest, Snapchat, Yahoo). All client lists are vendor assertions.
- **How it earns.** Vendor-stated models only, no prices published: SaaS subscriptions 'from PAYG usage to Enterprise accounts' for Campaign Strategy; a CPM rate card for in-flight measurement; tailored platform licences ('Enquire for Pricing'); bespoke research and analysis fees; per-test and volume pricing for creative testing (free AttentionAI test as a lead product); API integrations; and exclusive regional rights (Connect's exclusive African rights to AttentionAI, July 2026). No revenue is disclosed.
- **Supply and data dependencies.** Reference data come from Amplified's own paid panel: the Blinq Rewards App (mobile and PC, device-camera facial images) and the Blinq Rewards media hub (TV, micro-camera, facial landmarks for face orientation, audio fingerprinting), with recruitment only through panel partner Cint; plus bespoke in-situ studies (for example eye, gaze and pose tracking at QMS outdoor sites). EU and UK panel data are processed on AWS servers in Australia under Standard Contractual Clauses. In-flight measurement depends on buyers placing tags through their DSP (Yahoo DSP in Australia has an announced integration); activation in buying has depended on partners (Chalice Custom Algorithms custom bidding; agency-built pre-bid scores). Creative prediction depends on the same panel dataset for training. Amplified historically supplied aggregated attention data to Adelaide (Adelaide's 2020 release; Amplified's comparison page acknowledges historic external datasets).
- **Differentiation.** Owns a camera-based, paid, Cint-recruited reference panel across mobile, PC and in-home TV, plus in-situ field studies (OOH, cinema), rather than licensing reference data (S12, S45). Distinguishes active, passive and non-attention and reports attention decay, used consistently across planning, in-flight and creative products (S05, S06, S50). Covers planning channels beyond digital (OOH/DOOH, cinema, radio, linear TV) in its planning tool, which few cohort A rows list (S05). Early planning-tool API integration with a holding company (Omnicom, 2022) and a Smartly planning integration (2025) (S24, S41). A creative-prediction model with a documented REST API, webhooks and a claimed MCP integration, sold alongside media products (S07, S08, S09).
- **Risks.** Labelling risk: the company's 'measured, not modelled' positioning (S50) sits beside products whose outputs are modelled; buyers may treat estimates as observations. Assurance gap: no accreditation route documented while competitors hold or are in MRC processes (S36, S37). Inconsistent public figures (panel size 100,000 vs 1.35 million vs 2 million; data points 38, 45, 50, 59 or 59.7 billion; 16 vs 21 countries) weaken method credibility (S05, S09, S15, S22, S28). Leadership ambiguity (CEO named December 2024; founder called CEO in December 2025 and February 2026 posts) with no announcement found (S20, S25, S29). Privacy exposure from camera-based panels, including incidental capture of non-consenting household members and cross-border transfer of EU/UK data to Australia (S12). Small, private company funding disclosed only in vendor statements; scale relative to verification-suite competitors is unknown (S16, S17, S20).
- **Best fit.** Cross-channel attention planning that must include OOH, cinema, radio or linear TV alongside digital (Campaign Strategy) (S05). Agencies wanting attention benchmarks inside their own planning tools via API (S41). Media owners commissioning bespoke attention studies to sell inventory (Screenvision, QMS, Spotify, Pinterest) (S43, S44, S45, S46). Creative teams wanting fast, per-second predicted attention on social and online video assets before launch (AttentionAI) (S07, S08).
- **Poor fit.** Buyers needing MRC-accredited attention metrics or an accreditor-confirmed audit (S36, S37). Programmatic buyers wanting self-serve attention segments in a named DSP (none documented) (S22, S23). Buyers needing in-flight attention for audio, DOOH or in-game (not listed for Campaign Performance) (S06). Anyone needing published validation statistics or impression-level log data under a documented API (S06, S48).
- **Limits.** Campaign Performance figures are model estimates (tag signals mapped to panel benchmarks), not observations of the exposed audience; the vendor markets all its outputs as 'measured' human attention (S06, S50). No MRC accreditation or accreditor-confirmed audit for any Amplified product at 29 September 2026 (S36, S37). No published held-out validation statistic linking in-flight estimates or AttentionAI predictions to gaze reference data; '99% accuracy' is undefined (S06, S16). Outcome evidence is vendor, joint-release or media-owner case material without published method; no independent or replicated study found (S23, S42, S47). No impression or campaign volume disclosed for the in-flight product (S06, S52). No named self-serve DSP activation; buying use documented only through partner custom-bidding trials in 2023 (S22, S23, S42). APIs (in-flight and AttentionAI) have no public reference documentation; MCP is announcement-level (S08, S48, S56). Company data-scale claims conflict across pages and dates (S05, S09, S15, S22).

| Dimension | Availability | Evidence | Observed |
|---|---|---|---|
| Method validity | live | C | Campaign Performance does not observe the exposed person: its tags 'capture sub-second attention signals for each ad exposure, all mapped to active and passive attention benchmarks' (S06), so the reported figure is a model estimate calibrated to panel data (inference from that wording; the vendor does not label the basis). The reference method is Amplified's paid, Cint-recruited panel: device-camera facial images on mobile and PC (Blinq Rewards App) and an in-home TV micro-camera using facial landmarks for face orientation (S12), described elsewhere as 'machine-learning-powered gaze tracking' (S53) and, in field studies, 'eye, gaze and pose tracking' (S45). The reference sample is stated inconsistently: 500 to 3,000 users per country in 10 countries (2021, S39); over 100,000 people (October 2023, S22); 1.35 million panelists (2025 figures, S15); over 2 million 'real humans across 21 global markets' (May 2026, S09). No held-out agreement or error statistic between the tag-based estimate and gaze reference data, and no calibration by environment, is published in the sources screened. The page claims 'independently validated, with 99% data accuracy' (S06) and the investor page '99% accuracy vs. competitors at 70%' (S16), naming no validator, metric or reference set. ARF's Phase 2 validation report anonymises its 12 participating companies, so no Amplified-specific result can be read from it (S38). [186, 192, 184, 193, 191, 194, 4, 343, 185, 344] |
| Outcome validation | live | D | Memory: Amplified asserts a 2.5-second active-attention threshold for memory from its 'advertising decay' work and a 'direct relationship' between active attention and mental availability from an OMD study of 600 US people (3,300+ impressions, 12 brands), with no design, statistics or holdout published (S47, June 2022; not product-specific). Brand: the joint Amplified-Chalice release reports that Lucid measured a 70% increase in ad recall and a 14-point brand-lift increase for attention-optimised buys versus 'the standard non attention line item' in US and Australian proof-of-concept trials, with no sample, uncertainty, cost or method published (S23, October 2023; the release says the trials used Amplified's attention dataset, and a second automotive campaign fed tag-verified data into bidding, reporting only +16% active attention). The Audience Group case reports +14% ad recall from a custom algorithm alone and then +74% brand recall after an Amplified pre-bid score was added, a sequential before-and-after with no control (S42, undated, Australia). Sales: none found; Audience Group says the tactic cannot be isolated in business results (S42). All are vendor or joint releases (sponsor Amplified or its partner). No independent, commissioned-with-published-method, or replicated validation of the in-flight metric was found in the sources screened. [195, 345, 76, 346] |
| Granularity and timing | live | C | The solutions page says Campaign Performance tracks 'attention at the impression level' (S03); the measurement page says tags capture signals 'for each ad exposure', data flow 'within 30 minutes' of launch, and reporting is by creative execution, format, line item and placement, with active and passive attention levels and attention decay per line item (S06). Stages: in-flight and post-campaign; planning and pre-bid outputs sit in the separate Campaign Strategy row, except partner custom-bidding trials (S23). Measurement basis is not labelled by the vendor; on the vendor's own description the in-flight figures are tag signals mapped to panel benchmarks, i.e. model estimates (inference). Whether impression-level records (rather than line-item aggregates) are released to buyers is not stated; export is 'CSV export or API access depending on your setup' (S06). [347, 186, 195] |
| Activation | unconfirmed | D | No named self-serve DSP or SSP segment, curated deal or packaged bidding algorithm using Amplified in-flight attention was found. Documented: (a) proof-of-concept custom-bidding trials with Chalice Custom Algorithms in the US and Australia, run through agencies Audience Group and 2045, fuelled by 'Amplified’s vast human attention dataset', plus one automotive campaign in which tag-verified engagement data 'was fed into live bidding' (joint release, 9 October 2023; current status not documented) (S23); (b) an Australian agency case in which Audience Group built a pre-bid score from placement attention with Amplified to guide its Chalice algorithm (undated) (S42). The Yahoo DSP integration (Australia, announced 3 October 2023) makes measurement available inside the DSP, not buying on attention (S22). The company timeline lists a 2025 'strategic partnership' with The Trade Desk for 'activation workflows' with no release, platform documentation or segment name found (S15). Search bound: every /insight/ and /case-studies/ slug in Amplified's sitemap was screened by title (S55). [193, 194, 195, 76, 348] |
| Creative diagnostics | unconfirmed | E |   |
| Channel coverage | live | C | Measurement page: 'display and video formats across programmatic, web, CTV and premium digital environments', with 'Social and platform-specific support is expanding' (S06). The integrations page lists 'programmatic, display, video, social and CTV' for in-flight verification (S48); the platform page says 'digital, programmatic, CTV and more' (S04). On the seven-channel scheme: display, online video and CTV are documented as live; social is listed on one page and described as expanding on another (unconfirmed); audio, DOOH and in-game are not listed for in-flight measurement (they appear only in the Campaign Strategy planning list). Retail media: not listed. [349, 186, 350] |
| Data access and integrations | live | C | Routes named by the vendor: the platform dashboard; 'CSV export or API access depending on your setup' (S06); tag-based or 'server-to-server integrations' ('tagless') for in-flight measurement (S48); and measurement made available to advertisers using Yahoo DSP in Australia (announced 3 October 2023; 'now being rolled out'; current status not confirmed) (S22). No API reference or developer documentation for the in-flight API is published (the integrations page offers an enquiry form), so the API does not meet the dimension's documentation test; no ad server, verification suite, clean room or log-level export is named. [186, 194, 350] |
| Production scale | live | D | No count of impressions or campaigns measured with the in-flight product was found in the sources screened. Company-level figures are in other units: 'Capturing millions of ad impressions in real environments' (technology page, no period or product stated) (S52); in-flight benchmarks 'built from 38B+ data points' (S06), where 'data points' are undefined panel observations, not impressions measured; and US$591M of ad spend 'optimised' on the platform to 5 December 2023 (planning platform, cumulative) (S26). Named in-flight users: THE ICONIC via Love Media (September 2024, attentionPROVE) (S27); Irish Life and AIB (named in the measurement FAQ) (S06). [186, 351, 352, 353] |
| US availability and coverage | unconfirmed | D | Amplified opened a New York office on 7 November 2022, described as its 'US headquarters', and said it had built US attention data for two years with Yahoo, Snap, Twitter, Bacardi and Volkswagen (S18). The in-flight product is 'available globally, across 20+ markets' without naming them (S06). US calibration: company data are said to span '21 countries across North America, Europe, APAC and emerging markets' (S05) and an OMD study sampled 600 US people (S47), neither specific to the in-flight product. US production use of the in-flight tag is not documented: the Chalice trials covered the US through North America-based agency 2045 but the release says they used Amplified's attention dataset, and the market of the tag-fed automotive campaign is not stated (S23). The product register records US availability as unconfirmed. [119, 186, 354, 195, 346] |
| Independent assurance | live | A | On 29 September 2026 MRC's accredited digital services list has one 'Attention Metrics' entry (DoubleVerify Quality Analytics Platform, D, MW, MA) and its digital services-under-review list shows XPLN's attention audit in process; neither lists Amplified (S36, S37). No accreditor-confirmed audit of Amplified's attention metrics was found. Vendor self-attestation only: 'Our methodology is independently validated, with 99% data accuracy' (no auditor named) (S06) and 'Our tools are already aligned with these expectations' about the IAB/MRC guidelines (S28). Amplified is listed as an IAB Attention Task Force company participant and its founder as a key contributor to the November 2025 guidelines (S35); participation earns no assurance credit. The company timeline's reference to 'Validation from the MRC and IAB' describes the guidelines' publication, not any validation of Amplified (S15). [186, 193, 355, 41, 2, 48] |
| Method disclosure | live | C | Counted over the seven elements for the in-flight product, from public pages: (1) construct published (active attention 'consciously processing', passive attention 'looking near, but not directly at your ad', non-attention) (S50); (2) signals: tags capture 'sub-second attention signals' (S06), and the privacy policy says the attentionPROVE tag uses a VAST wrapper that collects playback events and does not access device storage (S12), but the signal list is not published; (3) reference labels: panel camera-based gaze and face-orientation data (S12, S53), summary level; (4) panel recruitment via Cint and totals (inconsistent across dates) published, per-market sizes not (S12, S15, S09); (5) model or scoring procedure: only 'mapped to active and passive attention benchmarks' (S06); (6) validation statistics: none beyond an undefined '99%' (S06); (7) recalibration cadence: 'continuously refreshed' with no cadence (S50). That is about three of seven at summary level, no data. The founder's two Palgrave books (2020, 2024; author affiliation Amplified Intelligence) are author-affiliated trade titles; only their tables of contents were read (for example chapters 'Recipe for Good Media Research' and 'Attention Science Becomes a Science'), so their content is not counted (S32, S33). The research-service page that Amplified links as its data 'collection process' gives no collection protocol beyond 'eye-tracking technology and machine-learning models' (S49). [186, 192, 184, 193, 356, 357, 187, 344, 358] |
| Privacy posture | live | C | Product-specific for the tag: the privacy policy (June 2026) says Amplified Intelligence Technologies Pty Ltd is TCF vendor 1195, that the declaration applies to its attentionPROVE tag-based measurement service, that it relies on legitimate interest by default for TCF Purposes 7, 9 and 10 with documented legitimate-interest assessments, that it complies with TCF v2.3, and that the tag 'does not store or access information on a user’s device' (no cookies, localStorage, mobile ad IDs or fingerprinting) (S12). IAB Europe's Global Vendor List (version 178, updated 24 September 2026) records vendor 1195 with legitimate-interest purposes 7, 9 and 10, special purposes 1 and 2, no cookie or non-cookie device access, data categories IP addresses, device characteristics, browsing and interaction data and non-precise location, and a standard retention of 3,650 days (S13); this is a standards-body registry record of the vendor's own declaration, not an audit. Panel side (reference data): consent-based recruitment only through Cint, explicit prior consent for sensitive data, facial landmarks 'TV only', TV images processed in the home and 'promptly deleted', on-device deletion for apparent under-18s, clients receive only anonymised aggregated data, EU/UK data processed in Australia under SCCs; GDPR, UK GDPR, CCPA/CPRA and other US state laws named (S12). No independent certification, regulator record or audit of compliance was found. Tension: legitimate interest is a GDPR Article 6 basis and does not by itself settle ePrivacy device-access consent (the policy's no-device-access statement is the vendor's answer), and the TV camera can capture non-consenting guests (S12). [184, 190] |

Performance evidence: Attention-optimised buys driven by a Chalice custom algorithm fed with Amplified attention data showed a 70% increase in ad recall and a 14-point brand-lift increase versus a standard non-attention line item (Lucid measurement), in US and Australian proof-of-concept trials. (vendor assertion, sponsor Amplified and Chalice Custom Algorithms (joint release); Lucid named as the outcome measurer; funding of the trials not stated) [195]; Audience Group saw +14% ad recall from a Chalice custom algorithm alone and then +74% brand recall after adding an Amplified attention pre-bid score, 'without spending any extra budget'. (vendor assertion, sponsor Amplified (case study page, quoting the agency)) [76]; An OMD study of 600 US people (3,300+ impressions of 12 brands, three categories, three online video platforms) found 'a direct relationship between MA and active attention'; Amplified states a 2.5-second active-attention threshold for memory. (vendor assertion, sponsor Amplified with client OMD (funding not stated)) [346]; Omnicom Media Group's use of Amplified planning data via API made clients' media spend 3% to 11% more effective and could save one CPG brand US$2.5 million a year. (vendor assertion, sponsor Amplified case page reporting OMG estimates) [156]; Screenvision (US cinema) reports that cinema ads drew the highest active attention of media measured and that brand perception and brand choice effects persisted 30 days after exposure. (vendor assertion, sponsor Screenvision with Amplified (media-owner study; funding not stated)) [199]; dentsu's Attention Economy programme built mobile eye-tracking panels of 3,400 respondents with Amplified Intelligence and reports statistically significant recall and brand-choice uplifts versus control that rise with dwell time (pooled across vendors). (commissioned measurement, sponsor dentsu (agency that sells attention planning and advocated attention as a currency)) [9]; Spotify India: 7,900+ ad views from a panel of 2,500 users over a month; video ads held attention for 17.3 seconds on average and outperformed audio by 23 percentage points on STAS. (vendor assertion, sponsor Spotify with Amplified (funding not stated)) [77]; QMS OOH study: 1.2 million unique ad exposures at twelve Australian sites tracked with eye, gaze and pose tracking; average 12 seconds of attention and nearly 90% of sites above a 2.5-second active-attention threshold. (vendor assertion, sponsor QMS with Amplified (media-owner study; funding not stated)) [185]; Pinterest and Amplified report 7.3x more passive attention on Pinterest than other platforms, +6.7% attentive seconds per dollar and +50% more attention when creative evoked happiness and surprise. (vendor assertion, sponsor Pinterest with Amplified (platform-commissioned; funding not stated)) [359]

### DoubleVerify (active, listed; acquisition by Nielsen's parent pending)

*DoubleVerify Holdings, Inc. (Delaware; SEC registrant, CIK 0001819928). MRC's accreditation letters are addressed to the operating company DoubleVerify, Inc.*

The only MRC-accredited attention metric (desktop, mobile web, in-app; not CTV); merger with Nielsen’s parent pending.

**SWOT (2026-09-29).** DoubleVerify holds the only attention accreditation on MRC's record, sold inside its verification suite, but its metrics are device-signal indices without published gaze validation, and Nielsen's parent is acquiring the company.

- **Strengths.** MRC's digital list at 29 September 2026 shows one attention accreditation: DoubleVerify's display and video Authentic Attention metrics in desktop, mobile web and mobile in-app. (record) [2, 42] MRC continued that accreditation on 13 April 2026 and granted initial accreditation of Authentic Attention V4 reporting, covering indexed, non-viewable and modelled impression metrics. (record) [42, 205] Attention sits in the verification suite advertisers already tag; the 2025 Measured Transaction Fee averaged $0.07 per thousand across all solutions, and DV's filings give no attention price. (record) [47, 206, 207] DV says it indexes 100+ billion impressions a month for its attention benchmarks (Q1 2026), although the same report's glossary still says 65 billion. (vendor claim) [208] DV attention measurement has been integrated on Snapchat, with Lumen eyes-on-screen signals, since June 2025, and on TikTok, as a badged partner, since December 2025. (vendor claim) [116, 117]
- **Weaknesses.** The accredited metrics are indices and rates built from device and on-screen signals; no source screened publishes validation against a gaze reference or an error statistic. (record) [42, 47, 209] Accreditation excludes CTV, social and pre-bid attention; MRC denied pre-bid Authentic Attention in March 2024, and the listed standards omit the IAB/MRC attention guidelines. (record) [42, 210, 41] Index values are set against DV's own 28-day rolling average, so levels move with DV's measured mix and cannot be read as absolute attention. (vendor documentation) [211, 208] All outcome evidence is DV-published case material without holdouts; the Scibids result reports a 63% rise in the optimised attention score, not a business outcome. (vendor claim) [71, 212, 213] DV's own descriptions of its accreditation diverge from MRC's record, citing 'only direct indexed impressions', a 2023 first accreditation, and unscoped 'MRC-accredited' in the 10-K. (vendor claim) [209, 214, 47, 215]
- **Opportunities.** The IAB/MRC Attention Measurement Guidelines of November 2025, which DV staff helped write, give DV a path to extend accreditation to the new standard. (inference) [41, 42] Measurement revenue grew 10% in 2025, which DV attributed partly to social adoption; attention already live on Snapchat and TikTok can ride that demand. (inference) [47, 116, 117] Under Nielsen's parent, with DV keeping its name and brand, accredited attention metrics could sit inside a larger media intelligence platform. (inference) [216] DV data already reaches agentic buying through an OAuth-gated MCP endpoint and an AdCP workflow, for suitability and fraud controls; attention is not yet among them. (inference) [217, 148, 218]
- **Threats.** Viant, a DSP, completed its purchase of TVision, DV's only documented CTV attention data source, on 1 May 2026; its closing release and Q2 10-Q omit third-party access. (inference) [219, 220, 82] The merger would delist and deregister DV, ending SEC reporting; the stockholder vote is 29 October 2026 and closing is targeted by Q1 2027. (inference) [221, 216] Named 2023 attention-activation platforms have thinned: MediaMath went through Chapter 11, and Microsoft stopped buying through Invest DSP on 28 February 2026. (inference) [222, 223, 182] IAS's Snapchat attention score also uses Lumen eye-tracking signals, so DV's Snapchat integration is not exclusive. (inference) [224] Future audits against the attention guidelines could test a device-signal construct that has no published gaze-reference validation. (inference) [41, 42]

- **What is sold.** DV Authentic Attention: an attention product inside DoubleVerify's verification and measurement suite (DV Pinnacle reporting, DV Media AdVantage Platform). It reports, per impression and in aggregate, indices built from more than 50 device and on-screen data points: exposure (viewable time, share of screen, video viewability, audibility, quartile completion, fully on-screen) and engagement (touches, scrolls, hovers, orientation changes, playback and volume interactions, user-presence indicators), rolled into an Attention Index, Exposure Index and Engagement Index benchmarked to 100 against DV's own 28-day rolling average. Around it DV sells activation: Universal Attention Segments (pre-bid, machine-learning, updated daily) and attention optimisation through DV Scibids AI. Channel extensions: Snapchat (with Lumen eyes-on-screen signals, June 2025), TikTok (December 2025), CTV (vendor-described; announced with TVision in June 2023), publisher-side DV Attention for Publishers (November 2024).
- **Who buys it.** Advertisers and agencies (DV reports over 2,500 customers, 90% of 2025 revenue from advertiser customers, 131 customers above $1 million a year) and, for DV Attention for Publishers, publishers. Named attention users in vendor case material: Mondelez (US), Club Med with Mediakeys, Havas Hong Kong, Jack in the Box with Attain, Vodafone Germany, Campari Group Spain. DV says its customer base is predominantly US-based.
- **How it earns.** DV charges a Measured Transaction Fee per thousand Media Transactions Measured (average $0.07 in 2025, on 9.5 trillion MTM), fixed for most solutions, with percent-of-media pricing capped to mirror fixed fees; programmatic activation fees are collected by DSP partners; supply-side customers pay subscriptions with minimum guarantees. Revenue is reported by customer type (activation, measurement, supply-side), not by product. No attention-specific revenue, price, attach rate or customer count appears in the FY2025 10-K, the Q2 2026 10-Q or the merger proxy as read. How attention is priced within bundles is not public.
- **Supply and data dependencies.** Tag-based measurement on the open web and in apps (including traffic measured through IAB Tech Lab's OM SDK for the accredited viewability metrics); platform integrations for social (Snap, TikTok) and walled gardens; Lumen Research eyes-on-screen signals for the Snapchat product; TVision viewer-presence and eyes-on-screen panel signals for CTV per the June 2023 partnership (TVision has been owned by Viant, a DSP, since May 2026, and no source screened says whether the data stay available to DV); DSP integrations for activation (programmatic fees collected by DSPs, with revenue-share costs); Scibids AI (acquired August 2023) for bid optimisation; MCP access via DV Neura for agentic workflows.
- **Differentiation.** The only attention product with MRC accreditation on MRC's record at 29 September 2026 (display and video, D, MW, MA), continued in April 2026 and extended to modelled and non-viewable impressions (V4). Attention is sold inside the verification suite advertisers already tag for fraud, viewability and suitability, and priced within per-thousand measured-impression fees, so buyers get it through existing verification contracts (the bundle terms are not public). Measurement at the scale of DV's tag footprint (100+ billion impressions a month in DV's attention benchmarks, Q1 2026, by DV's account; its glossary still says 65 billion) and in near real time (ATT-V-DOUBLEVERIFY-S03). Platform-integrated social attention on Snapchat (with Lumen signals) and TikTok (badged partner). Activation and optimisation from the same vendor: Universal Attention Segments and Scibids attention optimisation, and an MCP route for agentic workflows (attention exposure unconfirmed).
- **Risks.** Ownership change: the pending acquisition by Nielsen's parent (closing targeted by the first quarter of 2027, a forecast from DV's corrected release of 7 August 2026) would take DV private and end SEC reporting, removing the main public financial series; product strategy under Nielsen is not documented (ATT-V-DOUBLEVERIFY-S01, S02). CTV data dependency: the only documented CTV attention input, TVision panel data (2023), now belongs to Viant, a DSP; continued access for DV is not documented. Accreditation scope risk: DV's own descriptions ('MRC-accredited' without environment scope in the 10-K; 'only direct indexed impressions' on the product page; 'first received accreditation ... in 2023' in a 2024 release) diverge from MRC's record, which is broader than the product-page qualifier and dates first accreditation to December 2022; buyers should read scope from MRC's letter (ATT-D-S08, ATT-U-S38, ATT-DOUBLEVERIFY-S41, ATT-A-S08, ATT-A-S05). Validity challenge: the Attention Measurement Guidelines now exist, and later audits of DV against them could test a device-signal construct that has no published gaze-reference validation (inference). Activation erosion: named 2023 DSP partners have exited, and pre-bid attention is not accredited. Agentic buying: attention is not named in DV's documented agentic workflows (Butler/Till, DV Neura), so agent-driven buys may carry DV's suitability and fraud controls without attention.
- **Best fit.** US advertisers already using DV verification who want an accredited, impression-level, near real-time attention-proxy report on open-web and in-app display and video without adding a new tag or vendor. Buyers who need an MRC-accredited metric for reporting or internal governance in desktop, mobile web and in-app, and accept that it measures ad exposure and device interaction, not gaze. Programmatic teams wanting to steer spend away from low-exposure placements through DV segments or Scibids, treating the index as an optimisation input.
- **Poor fit.** Buyers who need a validated measure of visual attention (observed gaze) or a unit comparable across vendors. CTV-led or social-led buyers who need accredited attention in those environments. Buyers wanting an accredited pre-bid attention signal or a contractual attention guarantee (none found). Researchers who need replicable method documentation or data.
- **Limits.** The accredited metrics are device and on-screen signal indices and modelled metrics; no public validation against a gaze reference or published error statistic exists in the sources screened, so the numbers are candidate proxies or model estimates, not observed attention. Accreditation covers desktop, mobile web and mobile in-app post-bid reporting only: not CTV attention, not social attention, not pre-bid segments (denied March 2024). Index values are relative to DV's own 28-day rolling average across its impressions, so levels shift with DV's population mix and are not comparable across vendors or over time as absolute attention. The standards MRC lists for the accreditation do not include the IAB/MRC Attention Measurement Guidelines (November 2025). Named activation platforms date from 2023; two of them (MediaMath, Microsoft Invest DSP) have since gone bankrupt or stopped buying, and no current named-platform list was found. Outcome evidence is vendor-published case material without holdouts, sample sizes or cost; none tests the increment over time in view. No attention-specific revenue, price or adoption count is disclosed; revenue is reported by customer type. Methodology beyond signal categories and index structure is not public (guide gated, developer docs behind login).

| Dimension | Availability | Evidence | Observed |
|---|---|---|---|
| Method validity | live | A | The measure is an index of device and on-screen signals, and no reference method (eye-tracking or webcam gaze) is named for the open-web and in-app product in the sources screened. The 10-K and releases list exposure signals (viewable time, share of screen, video presentation, audibility) and engagement signals (touches, orientation, playback and audio-control interactions); MRC's list of accredited statistics consists of indices and device-event rates (Attention, Exposure, Engagement, Intensity, Prominence, User Presence and Ad Interaction indices; ad touched, scroll, tab into background, volume turned up, quartile audible and viewable rates). The 2026 product page adds 'eyes-on-screen' to the user-initiated events without describing how it is observed; the 10-K list does not include it. Reference data appear only in partner channels: the Snapchat product combines DV exposure data with Lumen Research eyes-on-screen signals (June 2025), and the CTV product was announced in June 2023 as a fusion with TVision viewer-presence and eyes-on-screen panel signals. For neither partner channel is an agreement or error statistic, reference sample or recruitment published by DV. MRC's April 2026 letter accredits metrics based on an 'Authentic Modeled methodology' whose model is not published. MRC accreditation certifies compliance with the listed measurement standards (viewability, IVT, impression and audience standards); the letter does not describe a validation of the index against a gaze reference (inference from the standards list). Grades: the signal set is documented in a filing and the accreditor's record (A); the Lumen and TVision reference roles rest on vendor releases (D). [42, 47, 209, 220, 116, 212, 211] |
| Outcome validation | live | D | All outcome evidence found reaches the public through DoubleVerify; no independent or replicated study, and no holdout or out-of-sample test of Authentic Attention, was found in the sources screened. Memory: Mondelez (US static display, undated whitepaper, filename 2021): a Kantar Brand Lift Insights control/exposed survey that DV proposed, with results cut by DV attention quadrants; high-Exposure impressions +5 points unaided awareness vs +3 points (80% significance) for all DV Authentic Ads, and +6 points in a separate quadrant analysis; no respondent counts, no cost, significance at 90% or 80%. Club Med (2023): Kantar Millward Brown brand lift, 19% higher top-of-mind and 41% higher online ad awareness, comparator not stated on the accessible page, full case gated. Brand: Mondelez +9 points favourability, +8 consideration, +5 purchase intent (80% significance) for high-Exposure impressions among the primary target. Sales and conversion: Vodafone Germany, Engagement Index of 100 or more linked to 'over 2.5x higher' qualified traffic and sales conversion rates (no method); Jack in the Box with Attain, a 'strong correlation' between engagement index and conversion rate and a '13% conversion rate' (gated, no method); Havas Hong Kong '83%' lower-funnel increase (gated). Because the Exposure Index includes viewable time, none of these is a test of increment over time in view. [71, 72, 360, 361, 212] |
| Granularity and timing | live | A | Impression-level measurement with near real-time reporting in DV Pinnacle (in-flight) and campaign, placement, ad size, creative and device breakdowns, downloadable reports and data feeds (post-campaign). Measurement basis for in-flight and post-campaign: an index of device signals relative to DV's 28-day average (a candidate proxy, not observed attention), plus, since April 2026, accredited metrics for non-viewable and modelled impressions (a model estimate for those impressions). Pre-bid: Universal Attention Segments built with machine learning from DV attention data and updated daily (placement-level avoidance of low-attention inventory; model estimate), and Scibids attention optimisation; MRC denied accreditation for pre-bid Authentic Attention in March 2024 and its April 2026 pre-bid table lists no attention offering. Grades: in-flight and post-campaign reporting rests on the 10-K and MRC's record (A); the pre-bid stage rests on vendor releases and pages (D). [42, 210, 47, 209, 222, 71] |
| Activation | live | D | Named buying platforms are documented only in 2023 releases: Universal Attention Segments in Amobee, MediaMath and Invest DSP (4 May 2023; Adobe Advertising, Quantcast and Viant 'will follow', with no later confirmation found), and DV Algorithmic Optimizer (Scibids) bidding on an attention index across DV360, MediaMath, The Trade Desk and Xandr (19 June 2023). Since then MediaMath entered Chapter 11 (30 June 2023) and Microsoft stopped supporting buying through Invest DSP from 28 February 2026. DV's 2026 product page still offers Universal Attention Segments and Scibids attention optimisation but names no platform, and the Scibids page says only 'available across leading buying platforms'. Havas Hong Kong is a named user of the segments (undated case). The CTV Streaming TV product (January 2026) adds app-level 'attention indicators' to pre-bid controls. MRC denied accreditation of pre-bid Authentic Attention (March 2024). Which named platforms carry DV attention segments at the cutoff is not documented in the sources screened. [210, 209, 222, 213, 360, 362, 363, 182, 223] |
| Creative diagnostics | unconfirmed | E |   |
| Channel coverage | live | D | Four of the seven channels are documented as live. Display and online video in desktop, mobile web and in-app are live and accredited (MRC record, A). Social is live on Snapchat (globally available from 30 June 2025) and TikTok (10 December 2025), per vendor releases (D). CTV is described as live on DV's CTV page and fact sheet (D), with no dated general-availability release found; it was announced on 13 June 2023 with a closed beta anticipated in H2 2023. Audio (DV's audibility metrics are device signals on video, not an audio channel), in-game and DOOH: no DV attention product found in the sources screened. Retail media (an unscored attribute): Authentic Attention is listed on DV's commerce-media page, and DV reported aggregate engagement for retail media network owned-and-operated inventory in 2023 data, without naming networks. No attention product integrated with Meta or YouTube was found; DV's accredited Meta and YouTube reporting covers impressions and viewability. [42, 220, 116, 117, 364, 365, 366, 367] |
| Data access and integrations | live | D | Routes named: DV Pinnacle dashboards (DV cites 8 dashboard views and a dedicated TikTok area); downloadable reports and data feeds (Mondelez whitepaper, undated, filename 2021); measurement of Amazon DSP custom-audience buys via a server-to-server integration (October 2023 release); in-platform integrations with Snap and TikTok feeding DV reporting; DV Attention for Publishers with breakdowns by ad unit, order and line item (November 2024; no ad server named). Agentic route: a live, OAuth-gated MCP endpoint (mcp.doubleverify.com/mcp, authorisation server in DV's 'pinnacle' identity realm; 401 without a token on 29 September 2026) exposing DV's 'media quality and performance data'; whether attention metrics are among its tools is not documented. DV's developer documentation portal requires login and access approval, so no public API reference for attention data was found. No clean-room or log-level export route for attention was documented in the sources screened. [209, 217, 218, 71, 214, 368, 369] |
| Production scale | live | D | DV's Q1 2025 benchmark report states 'DV measures 65+ billion impressions each month for the attention benchmarks' (data for January to February 2025; tag-based, open web; vendor figure). Scaled to a year, that is at least 780 billion impressions (derived: 65 billion x 12). DV's Q1 2026 report states 'DV indexes 100+ billion impressions each month for global attention benchmarks' (tag-based, open web; vendor figure), although the same report's glossary still says 65 billion. Earlier vendor figures: 'tens of billions of impressions' a month (June and December 2023) and 'nearly 80 advertisers' activating Authentic Attention in 2023. The 2024 claim that adoption 'tripled' has no base. No campaign count is stated. The figures are DV's own and unaudited. [213, 220, 370, 211, 208, 371] |
| US availability and coverage | live | A | US availability is documented at company level: DV is headquartered in New York, says its customer base is 'predominately U.S.-based', and holds the US accreditor's (MRC) accreditation for the product; the product page does not state geography. US production use is documented in a vendor case: Mondelez, 'US only' targeting, and Jack in the Box, a US brand (gated case). US calibration: the indices are benchmarked against DV-wide 28-day averages, not US norms, although DV publishes a North America index (99 in January to February 2025); no US-specific calibration or reference data are documented. Grades: company-level US base from the filing (A); US campaigns from vendor case material (D). [2, 47, 71, 361, 211] |
| Independent assurance | live | A | MRC accreditation covers display and video Authentic Attention metrics in desktop, mobile web and mobile in-app. It was granted in December 2022, continued in March 2024 and continued on 13 April 2026, when MRC also granted initial accreditation of Authentic Attention reporting (Indexed Impressions, Direct-Indexed-Rate, and metrics based on the Authentic Non-viewable and Authentic Modeled methodologies), labelled 'V4' in MRC's July 2026 report. On 29 September 2026 MRC's digital list shows 'Attention Metrics' for DoubleVerify Quality Analytics Platform in D, MW and MA, without the 'Inclusive of pre-bid data' annotation its viewability lines carry. Not covered: pre-bid Authentic Attention (accreditation denied, March 2024; no attention row in the April 2026 pre-bid table); CTV attention (CTV is accredited for impressions, viewability, completion and verification only, and no DV entry appears on MRC's digital or television under-review lists); social attention (the TikTok accreditation covers video viewability; Meta and YouTube reporting covers impressions and viewability). Standards listed in the April 2026 letter: MRC Minimum Standards, Viewable Impression Measurement Guidelines, IVT 2.0, Digital Audience-Based Measurement Standards, Outcomes and Data Quality Standards; IAB Display Impression, Digital Video and Ad Verification Guidelines; MRC/IAB/MMA Mobile Web and In-App Impression Guidelines. The list does not include the IAB/MRC Attention Measurement Guidelines (November 2025). The product is sold in CTV and social environments that the attention accreditation does not cover. [42, 215, 210, 205, 2, 48, 50, 41] |
| Method disclosure | live | C | Against the seven elements: (1) construct, published (attention as ad exposure plus user engagement; also 'user attention that correlates to business outcomes'); (2) signals, published at category level with over 50 data points, and the accredited metric names are public in MRC's enclosure; (3) reference labels, not published for the core product (Lumen and TVision are named as signal partners for Snap and CTV); (4) reference panel or sample, none published (the benchmark base, 65+ billion impressions a month, is a population, not a reference sample); (5) scoring procedure, partly published (index hierarchy, base 100 against DV's 28-day rolling average, sub-indices; weights and aggregation not published; segments described only as 'machine learning'; Authentic Modeled methodology not published); (6) validation statistics, none published; (7) update cadence, partly published (28-day rolling benchmark; segments updated daily; no model recalibration cadence). Count: 2 elements fully and 2 partly published (at most 4 of 7). No data are released for replication. The detailed guide is gated and the developer documentation needs login. [42, 47, 222, 71, 212, 370, 211] |
| Privacy posture | live | C | DV's Solutions Privacy Notice (effective 9 October 2025) lists the exposure and engagement attributes it collects (whether the ad was clicked, swiped, tapped or touched, resized, skipped, muted, paused, rotated, hovered or had its volume changed) and viewability attributes (viewport and screen size, tab and browser focus, time in view, scroll position). It says DV acts as a GDPR 'processor' for Page Level Engagement Metrics and does not combine IP addresses with data that would identify the individual. The notice does not name Authentic Attention; matching its attributes to the product's signals is inference. No product-specific consent model is described, since DV collects device signals through customer tags and relies on customer contracts. At company level: the 10-K calls the product 'privacy-friendly' and says the core technology does not rely on persistent identifiers or cookies; DV reports 'several external privacy-related audits each calendar year' (auditors not named); DV self-certifies to the EU-U.S. Data Privacy Framework. In the sources screened, no regulator action or privacy lawsuit involving the product was found. A securities class action, not a privacy matter, was voluntarily dismissed on 22 December 2025. [47, 372] |

Performance evidence: High-Exposure impressions (Exposure Index of 100 or more) lifted unaided brand awareness by 5 points vs 3 points (80% significance) for all DV Authentic Ads; brand favourability +9, consideration +8, purchase intent +5 (80%) among the primary target. (vendor assertion, sponsor DoubleVerify (DV proposed the Kantar study and asked for results by attention quadrant; payment terms not stated; published by DV)) [71]; Club Med respondents had 19% higher top-of-mind awareness and 41% higher online ad awareness of the new brand identity. (vendor assertion, sponsor DoubleVerify publication; Kantar Millward Brown study commissioned by Club Med/Mediakeys or DV (not stated)) [72]; Vodafone Germany: high-engagement ads (Engagement Index of 100 or more) drove over 2.5x higher qualified traffic and sales conversion rates than low-engagement ads. (vendor assertion, sponsor DoubleVerify) [212]; Jack in the Box found a strong correlation between average engagement index and conversion rate and achieved a 13% conversion rate. (vendor assertion, sponsor DoubleVerify with Attain (joint commercial solution)) [361, 373]; Havas Hong Kong increased lower-funnel outcomes by 83% with DV attention measurement and Universal Attention Segments. (vendor assertion, sponsor DoubleVerify) [360]; DV Algorithmic Optimizer (with Scibids) produced a 63% increase in attention levels and a 95% increase in impressions won on average across several campaigns. (vendor assertion, sponsor DoubleVerify) [213]; Campari Group Spain used DV Authentic Attention to drive a 14% lift in view-through rate. (vendor assertion, sponsor DoubleVerify) [209]

### Integral Ad Science (active, private (Novacap-owned))

*Integral Ad Science Holding Corp. (SEC registrant, CIK 1842718); operates as Integral Ad Science, Inc. and subsidiaries*

Attention inside a verification suite; left MRC’s attention audit process in January 2026.

**SWOT (2026-09-29).** IAS sells Quality Attention, a model-estimated score built on its verification signals and pitched as a predictor of business results, but it left MRC's attention accreditation process in January 2026 and publishes no validation.

- **Strengths.** IAS holds MRC accreditation for impressions, viewability, SIVT and, since June 2026, pre-bid viewability and IVT segments; none of these entries covers attention. (record) [2, 205] Quality Attention is computed from signals IAS already collects for verification, such as viewability, time in view, ad density and interaction, and is reported as three sub-scores. (vendor documentation) [73, 225] Since October 2024 publishers and SSPs can report Quality Attention metrics and package inventory into custom segments and PMPs, with Tripadvisor quoted at launch. (vendor claim) [226] IAS names brand users including Kimberly-Clark, Stellantis (19 countries, over 13 billion measured impressions), Toyota Spain, Sanofi Consumer Healthcare and Pernod Ricard Japan. (vendor claim) [11, 227, 228, 229, 230] IAS reported FY2024 revenue of $530.1 million and 237 large advertising customers, giving attention a large installed verification base to sell into. (record) [231, 232]
- **Weaknesses.** In January 2026 MRC removed IAS Attention Measurement from in-process status, citing IAS' lack of intent to engage in third-party audit; it never reached audit. (record) [49, 233, 234] The report-level Attention Score is a model estimate of click or conversion propensity, computed only for impressions measured for viewability; no ordinary impression is eye-tracked. (vendor documentation) [225, 235] IAS names Lumen Research as its eye-tracking partner but publishes no panel size, recruitment, geography, or held-out agreement statistic against gaze. (vendor documentation) [73, 236, 237, 229] All outcome evidence is IAS-published without sample or holdout, and headline lifts shifted from 198% to up to 130%, 277% and 246% without stated bases. (vendor claim) [238, 229, 73, 239] IAS's own Threat Lab found Papyrus fraud traffic showed about 13% higher attention scores than non-Papyrus traffic before it was filtered as invalid. (vendor claim) [240]
- **Opportunities.** Mastercard-enriched IAS pre-bid segments, announced March 2026 with attention among their signals, could tie the score to sales if they reach general availability. (inference) [239, 241] Quality Attention Optimization has been in beta since December 2024; general availability with named DSPs would turn measurement into activation. (inference) [115, 118] Funnel's connector already pulls the score and sub-scores into third-party platforms; a documented IAS API or log-level export would widen buyer data access. (inference) [225, 242] The June 2026 pre-bid accreditation shows IAS can complete MRC audits; re-entering the attention process would address its main assurance gap. (inference) [205, 49]
- **Threats.** On MRC's accredited list, only DoubleVerify holds attention-metric accreditation among verification suites, so buyers requiring accredited attention have an alternative to IAS. (inference) [2] The IAB and MRC guidelines of November 2025 say attention should not be used as an outcome measure and exclude unvalidated predictive models from scope. (inference) [41] Lumen supplies IAS's eye-tracking reference data and also DoubleVerify's Snap attention signals, so the two suites' social attention outputs are not independent. (inference) [236, 118, 243] CESP's June 2026 grid appears, by footnote layout, to mark IAS's MRC answer 'Under review', conflicting with MRC's January 2026 removal of IAS Attention Measurement from in-process status. (inference) [3, 49] Novacap's take-private closed on 23 December 2025, ending periodic SEC reports, so changes in attention investment will no longer be publicly disclosed. (inference) [83, 244, 245]

- **What is sold.** IAS Quality Attention is a machine-learned attention score sold inside IAS's verification and optimisation suite. IAS says it weighs signals from its core tags (visibility: viewability, time in view; situation: ad density, share of screen; interaction: scrolling, video pause, resume, skip, start, volume change) plus eye-tracking data from Lumen Research into 'a single attention score' with three sub-scores, and that the score 'can predict if an impression is likely to lead to a business result'. It is sold as post-bid measurement (beta June 2023; general availability announced January 2024; mobile in-app from July 2024), as a publisher and SSP reporting product (October 2024), as Quality Attention Optimization (pre-bid, announced in beta 17 December 2024), and as Social Attention reporting with Lumen (stated live from January 2025) and Snap Attention Measurement (announced 30 June 2025), reported in IAS Signal.
- **Who buys it.** Advertisers and agencies already using IAS verification (named cases: Kimberly-Clark with NCSolutions, Stellantis with Publicis Media, Toyota Spain with WPP Media, HP), plus publishers and SSPs for inventory reporting (Tripadvisor quoted at launch). IAS reported 237 large advertising customers (at least $200,000 a year) in 2024, company-wide, not attention buyers.
- **How it earns.** IAS charges a CPM on the volume of ads it analyses (FY2024 10-K). Attention revenue is not disclosed separately: the 10-K says measurement revenue covers post-bid solutions across 'viewability, ad fraud prevention, brand safety and suitability, attention, and more'. No attention price, fee or customer count was found in the filings or pages opened. IAS has revenue-share agreements with DSPs (Google, The Trade Desk) for pre-bid products generally.
- **Supply and data dependencies.** Depends on IAS's own tags and integrations for the visibility, situation and interaction signals, and on Lumen Research for eye-tracking reference data and predictive attention models (open web since 2023; social and Snap measurement 'powered by Lumen's attention model'). Social measurement depends on platform-supplied data (Snap). Pre-bid use depends on DSP integrations, none named for attention in the sources screened. The Lumen dependency is shared: DoubleVerify's Snap attention product also uses Lumen signals (ATT-G-S18, S19), so the two verification suites' social attention outputs are not independent of each other.
- **Differentiation.** Attention is computed from signals IAS already collects for verification (viewability, time in view, ad density, interaction), so buyers using IAS tags can add it without a separate tag (inference from IAS's description of 'signals obtained as part of IAS’s core products'). Explicitly outcome-trained: IAS trains on conversion events and markets the score as predicting business results, unlike metrics framed as attention alone. Sell-side availability since October 2024 lets publishers and SSPs report and package the same score buyers see. Social attention via Lumen and a bespoke Snapchat score (announced) alongside open-web measurement in one reporting platform.
- **Risks.** Assurance gap: MRC records IAS' lack of intent to engage in third-party audit of its attention measurement, while its main verification rival holds MRC attention accreditation for desktop, mobile web and in-app (MRC's accredited digital list, ATT-A-S09); IAS's own Snap release calls its attention measurement 'trusted, independent third-party'. Outcome framing: the IAB and MRC guidelines (to which IAS staff contributed) say attention 'should not be considered or used as a measure of outcomes' and place unvalidated predictive models outside their scope, while IAS markets Quality Attention as predicting business results (ATT-A-S01, ATT-A-S22). MRC has not assessed the score against those clauses because IAS left the process (ATT-A-S07). Adelaide's AU carries a similar outcome framing (ATT-ADELAIDE-S12). Reference-data dependency on Lumen, which also supplies DoubleVerify's Snap attention signals and sells its own attention products. Interaction signals can be synthetic: IAS's own Threat Lab documents a fraud scheme that scrolls pages 'inflating attention signals'; IAS says IVT avoidance filters it (vendor assertion). Private ownership since December 2025 removes filing-based disclosure; any change in attention investment would not be publicly reported (inference). Self-referential chain: IAS supplies the inputs, trains the score, sells the optimisation and reports the outcome, with no independent link documented (PPC Land makes the same criticism). DoubleVerify sells a similar chain (its own signals and indices, Scibids optimisation, outcome cases published by DV), though MRC accreditation covers its post-bid display and video reporting (ATT-E-S28, ATT-C-S18, ATT-A-S05).
- **Best fit.** Advertisers already on IAS verification who want a relative quality ranking of open-web display and video placements, including in-app, for post-campaign optimisation. Publishers and SSPs wanting to report an IAS-branded attention score to IAS-using buyers. Diagnostic use (which supply paths or placements score low) where the buyer treats the score as a model estimate, not observed attention.
- **Poor fit.** Buyers who need an accredited attention metric for trading, guarantees or billing. CTV, audio, in-game or DOOH attention measurement (none documented). Cross-vendor comparisons or buyers who need published validation statistics, panel details or raw data. Buyers needing attention signals in a named DSP pre-bid workflow at general availability, or through an API or agent protocol.
- **Limits.** Every attention number IAS publishes is a model estimate; IAS sources do not document gaze observation of measured impressions. No accreditor has audited the attention metric; IAS left MRC's process at pre-audit stage in January 2026. No published validation statistic against a gaze reference and no independent outcome validation; the only sales test found with a reported null (Kimberly-Clark Test 1, a post hoc split of NCSolutions results) is vendor-published. Channel coverage documented only for open-web display and video (including in-app) and, by vendor statement, social; no CTV, audio, in-game or DOOH attention product found. No named DSP documented for attention pre-bid segments; the optimisation product's last documented status is beta (December 2024). No attention-specific revenue, volume, price or customer count; IAS stopped periodic SEC reporting after November 2025. Score scale and thresholds are not defined in IAS primary sources opened.

| Dimension | Availability | Evidence | Observed |
|---|---|---|---|
| Method validity | live | D | Quality Attention is a machine-learning estimate, not an observation of the exposed person. IAS names a reference method from a partner: eye-tracking data and predictive attention models from Lumen Research ('passive eye-tracking powered by Lumen data'; 'actionable data from Lumen Research’s eye-tracking technology'), integrated since 2023-2024 (S15, S17, S18, S19). IAS says the model was 'trained based on a pool of data consisting of billions of impressions and millions of conversion events' (S19; S43), so its labels are partly conversions, and it frames the score as predicting 'if an impression is likely to lead to a business result' (S15). No reference sample size, panel recruitment or geography is stated in IAS sources; the only panel description is Lumen's own claim, quoted in IAS releases, of 'the largest, opted-in eye-tracking dataset in the world, covering 30+ countries' (S21) and 'the world’s largest opt-in dataset' (S22). No held-out agreement or error statistic between the score and a gaze reference was found, and no calibration by environment is documented; IAS says only that the July 2024 update 'improved accuracy of the correlation between attention scores and outcomes' (S19). Against the anchors: validity rests on a named partner's reference data with the integration described in vendor releases, and no agreement statistic of the product's own. [73, 236, 374, 237, 115, 118, 375] |
| Outcome validation | live | D | All outcome evidence found is published by IAS, with no study method, sample size or holdout disclosed on the pages opened, and no independent validation. Sales: Kimberly-Clark with NCSolutions (published 29 September 2025): two tests, impressions split post hoc by above- or below-average score against an NCS modelled baseline; Test 1 (Q2 2024, browser display) found 'no significant difference in sales lift'; Test 2 (Q4 2024, adding video and in-app) reported 313% higher sales lift for above-average households and ROAS up 78% 'with the addition of high-attention formats' (S23). A second CPG sales case with NCSolutions (6 May 2024) states sales lift and ROAS were higher for Quality Attention placements, with figures only in a gated download (S28). Brand: Upwave beverage case (gated, no figures on page) (S27); '91% higher brand consideration and 166% higher purchase intent' for greater attention scores, no design (S19). Memory: Stellantis 'A 33% increase in ad recall for key audience, as measured by Dynata’s study' (S24); '52%' unaided-awareness boost for above-average in-app impressions, no base (S15). Conversions: 198% average conversion-rate lift high vs low (June 2023, S16), 'up to a 130% lift' (2024, S19, S21), 277% higher success rate and 63% lower cost per success (S15), HP 3x success events (S26). Every comparison is high versus low score within delivered media or before versus after; none compares against time in view or viewability as a baseline. Against the anchors: vendor case studies reporting outcome associations, with outcome stated and sample not stated. [73, 238, 237, 115, 11, 227, 376, 377, 378] |
| Granularity and timing | live | D | Scores exist at impression level: IAS segments 'impressions by below-average and above-average Quality Attention™ scores' (S23), and the optimisation product steers 'towards high-attention scoring impressions' (S21). Output is 'a single attention score and three sub-scores for visibility, situation, and interaction' (S16), reported by placement and ad size (S15) in IAS Signal (S22). Stages documented: post-campaign and in-flight reporting (post-bid measurement, generally available from January 2024; S16, S21); pre-bid optimisation (beta from 17 December 2024; S21); publisher and SSP inventory reporting (October 2024; S20). No stage is labelled by measurement basis in IAS sources; by the brief's coding every stage is a model estimate. Against the anchors: impression-level output at post-campaign (live) and pre-bid (beta) stages, not labelled by measurement basis. [73, 238, 226, 115, 118, 11] |
| Activation | beta | D | Quality Attention Optimization was announced 'in beta' on 17 December 2024 for open-web display and video, as 'a new, simplified and automated way to optimize their campaigns towards high-attention scoring impressions'; Lumen's CEO describes it as actionability 'via pre-bid segments' (S21). IAS's FY2024 results release states it 'now provides advertisers complete coverage for attention metrics across post-bid, pre-bid, and social media' (S12). No source opened names a DSP, SSP or curation platform in which attention segments are live; the FY2024 10-K names Amazon, Google's Display & Video 360, The Trade Desk and Yahoo as DSPs through which IAS serves advertisers, for its pre-bid products generally, not attention (S11). No general-availability record for the optimisation product was found. Publishers and SSPs can package Quality Attention data into 'custom segments and packages' and 'unique PMP offerings' (S20), with no named SSP or deal. MRC's June 2026 pre-bid accreditation covers viewability and IVT segments only (S05). Against the anchors: beta, with no named buying platform documented. [115, 232, 231, 226, 205] |
| Creative diagnostics | unconfirmed | E |   |
| Channel coverage | live | D | Display and online video on the open web: live (general availability January 2024; mobile in-app added 30 July 2024) (S19, S21). Social: 'Social Attention Measurement with Lumen will be live for IAS customers from January 2025', no platforms named (S21); Snap Attention Measurement announced 30 June 2025 as 'will be available in IAS Signal' (S22); live status after June 2025 not confirmed. CTV: IAS's CTV & Video page (modified May 2026) describes completion and invalid-traffic validation and mentions attention only as consumer attention shifting to CTV (S38); IAS's CTV research with Neuro-Insight is a one-off neurostudy (S39). In-game: MRC's 'IAS In-Game (Pre-audit)' scope was rendered and viewable impressions, not attention, and was removed from in-process in January 2026 (S02, S01); the Anzu partnership covers IVT and viewability (S37). No IAS source found documents Quality Attention in audio, DOOH or retail media. Against the anchors: two channels live (display, online video), a third (social) vendor-stated live without named platforms. [237, 115, 118, 379, 380, 233, 49, 381] |
| Data access and integrations | live | D | The documented route is IAS's own reporting platform, IAS Signal, where Quality Attention and Snap Attention Measurement are reported (S22), building on the 2022 'Attention Metrics Report' in Signal (S41). Publishers and SSPs can see 'IAS Quality Attention metrics and scores' for their inventory (S20). No attention API, log-level export, ad-server integration or clean-room route was documented in the sources screened (attention page, releases 2023-2025, 10-K, Technology privacy policy). Against the anchors: reports and dashboards, plus sell-side access to the same reports. [118, 382, 226, 73, 231] |
| Production scale | live | D | No company-wide volume of impressions or campaigns measured with Quality Attention is disclosed; the FY2024 10-K and Q3 2025 10-Q give none (full-text search for 'attention') (S11, S13). The 10-K's '280 billion digital interactions daily' is all IAS processing, not attention (S11). The largest single stated figure is one advertiser: Stellantis, 'Over 13 billion measured impressions' across 19 countries in a program that 'initially launched in 2024, running for more than six months globally — and up to 18 months in key markets such as France' (S24). That is a cumulative total with a start year but no exact period; if read against the anchors' derivation rule it is a floor for one customer, not the product's volume. [227, 231, 383] |
| US availability and coverage | live | D | IAS is headquartered in New York (S10, S11) and its releases carry New York datelines; the attention page says 'advertisers worldwide' without a US-specific statement (S15). US production use is documented in one vendor case: Stellantis's program analysis reports that 'In the United States, visibility emerged as a primary driver of attention' (S24). The Kimberly-Clark case used NCSolutions household sales data but does not state its geography (S23). No US calibration data (US panel, participants or norms) is documented; Lumen's reference data is described only as covering '30+ countries' (S21). Against the anchors: available in the US (inferred from global availability and US HQ), with US production use documented in one vendor case and no US calibration documented. [245, 231, 73, 115, 11, 227] |
| Independent assurance | live | A | MRC record: an archived Services Under Review page (16 September 2025) lists 'IAS Attention (Pre-Audit)' for 'Attention Score and associated metrics that contribute to Attention Score' (S02). MRC's decisions report for January to March 2026 records 'IAS Attention Measurement' (and 'IAS In-Game Measurement') 'Removed from in-process status due to IAS' lack of intent to engage in third-party audit and the resulting withdrawal from the accreditation process' (S01). On 29 September 2026 MRC's under-review page lists no IAS attention entry (S03), and its accredited digital list shows attention metrics only for DoubleVerify; IAS's entries cover tracked ads, rendered and viewable impressions (display; video including CTV for certified traffic), SIVT, property-level verification, SSAI and pre-bid viewability and IVT (S04, S05). Quality Attention uses viewability as an input (S15), and IAS's viewability and SIVT measurement is MRC-accredited, but no MRC record ties that accreditation to the attention product. The attention service was never accredited, so this is a withdrawal from the process, not a denial or withdrawal of accreditation; that MRC's 'IAS Attention Measurement' is the Quality Attention product is an inference (MRC names no product). Self-reported: on the CESP grid (10 June 2026) IAS answered 'Yes' to an independent third-party audit of its attention solution, with no auditor or date named, and 'No**' to MRC attention accreditation (S34). ISO/IEC 27001 scope includes 'Attention' as a service (information security only) (S31). [49, 233, 48, 2, 205, 73, 3, 384] |
| Method disclosure | live | D | Of the seven elements: (1) construct: stated ('the amount of cognitive focus a viewer gives to an ad', with the score predicting business results) (S15); (2) signals: listed (viewability, time in view, ad density, share of screen, scroll, video pause, resume, skip, start, volume change, 'Eye Tracking/Focus powered by Lumen') (S15, S19); (3) reference labels: named at summary level (Lumen eye-tracking data; 'millions of conversion events') (S19); (4) panel size and recruitment: not published; (5) model type or scoring procedure: 'advanced machine learning' weighing signals into one score and three sub-scores, no model type or weights (S16, S19); (6) validation statistics: none, only high-versus-low lift claims; (7) update cadence: not stated (one model update described in July 2024, S19). IAS's research reports (Taking Action on Attention, Volumes I and II) sit behind download forms and were not read (S43). Against the anchors: two or three elements published in substance, the rest at marketing depth or absent. [73, 238, 237, 375] |
| Privacy posture | live | C | IAS's Technology Solutions and Reporting Platform Privacy Policy (policy IASLG-17, most recent revision dated 01.12.2023) says IAS uses pixels and SDKs to 'collect information about advertising impression opportunities and displayed ads or website/mobile app traffic' and collects IP addresses for fraud and non-precise geo-verification; it does not name Quality Attention or the interaction signals (scroll, pause, volume) the score uses (S32). The website privacy policy (updated 9 March 2026) excludes data processed through IAS's Technology Solutions (S33). The eye-tracking consent model belongs to Lumen and is described in IAS sources only as 'opt-in' or 'opted-in' (S21, S22). ISO/IEC 27001:2022 certification announced 22 May 2024 names 'Attention' among in-scope services, which covers information security, not consent or data minimisation (S31). IAS claims TrustArc Responsible AI, ISO 42001 and AAM Ethical AI certifications, company-level, in the IAS Agent release (S29). No regulator action or breach involving the product was found in this paper's bounded searches. Against the anchors: a company-level policy that describes the data types IAS tags collect, without naming the attention product. [385, 386, 115, 118, 384, 155] |

Performance evidence: Kimberly-Clark: in Test 1 (Q2 2024, browser display) no significant difference in sales lift between low- and high-attention segments; in Test 2 (Q4 2024, adding video and mobile in-app) households exposed to above-average attention media saw 313% higher sales lift than below-average, and ROAS rose 78% 'with the addition of high-attention formats'; average score rose from 53 to 72. (vendor assertion, sponsor IAS (case study on its site); NCSolutions measured sales; who paid for the NCS study is not stated (independence unknown)) [11]; Stellantis with Publicis Media: Quality Attention deployed across open-web campaigns in 19 countries, over 13 billion measured impressions, from 2024; 'A 165% increase in engagement on high-attention inventory' and 'A 33% increase in ad recall for key audience, as measured by Dynata’s study'. (vendor assertion, sponsor IAS (case study); Dynata brand study, funder not stated) [227]; Toyota Spain with WPP Media: over a five-month, three-phase test, attention scores 81 vs a Spanish auto benchmark of 61; in-app 90 vs browser 73; 'A 16% increase in average monthly site visits, rising from 400 in July to 811 in November (calculated as an Compound Annual Growth Rate)'. (vendor assertion, sponsor IAS (case study)) [228]; HP with Lumen: 11.5x higher attention scores for in-view vs out-of-view ads; 3x more success events and 70% lower cost per success event with higher attention scores. (vendor assertion, sponsor IAS (case study, with Lumen, IAS's data partner)) [376]; Headline conversion claims: '198% average lift in the conversion rate when comparing high attention impressions with low attention impressions' (June 2023); 'Up to a 130% lift in conversion rates' with '91% higher brand consideration and 166% higher purchase intent' (July 2024, repeated December 2024); 277% higher success rate and 63% lower cost per success, and a 52% unaided-awareness boost for above-average in-app impressions (attention page, modified August 2025). (vendor assertion, sponsor IAS) [238, 237, 115, 73]; Two 2024 CPG cases: with NCSolutions, 'both sales lift and return on ad spend were higher for Quality Attention™ placements'; with Upwave, IAS and Upwave 'investigated the impacts of varying attention scores on ad placements' for a beverage brand. (vendor assertion, sponsor IAS (case studies); NCSolutions and Upwave measured outcomes, funding not stated) [377, 378]

### Kantar (active, private)

*Kantar Global Holdings S.à r.l. (group holding company named in Kantar's own releases). The entity that sells Kantar Marketplace products in the US was not established from the sources read.*

The largest creative-testing business, selling three separately priced attention-bearing products.

**SWOT (2026-09-29).** Kantar sells the broadest priced creative-pretesting line with attention options, but its attention measures carry no independent certification and rest on an Affectiva partnership.

- **Strengths.** Kantar Marketplace lists three separately priced products with public US list prices: LINK+ from $4,500 self-serve, LINK AI from $500, Context Lab from $13,950 serviced. (vendor documentation) [246, 247, 248] Kantar states a 260,000-ad LINK+ survey database and 50,000 attention ad tests as norm bases; these are vendor figures, not measured against other vendors' bases. (vendor claim) [249, 246] The attention framework launched 14 November 2023 reports passive attention (eyes on screen) and active attention (facial-expression engagement) separately, with second-by-second traces and norms. (vendor claim) [249, 250] Self-serve LINK+ and LINK AI results can be retrieved through the Kantar API, and CreativeX says LINK AI has been available through its Datalink since December 2025. (vendor claim) [251, 252] Kantar reported audited FY2025 revenue of $2,494.4 million from continuing operations, with gross revenue of $1,811.2 million in the Insights division, home of creative testing. (vendor documentation) [253]
- **Weaknesses.** No audit, certification or accreditor record covers LINK+ or its attention measures; MASB's February 2017 certification covers the predecessor Link Copy Test, and Kantar is a MASB member. (record) [254, 255, 256, 257] Active attention is a facial-expression emotion measure, not a gaze measure, and Kantar's own publications report it barely correlates with passive attention at ad level. (vendor documentation) [258, 259] Observed attention is documented for digital and TV video only; attention for print and outdoor comes from AI predictive eye tracking and a survey module. (vendor documentation) [246, 260, 250] Published sales validations cover Link survey metrics such as Short-term Sales Likelihood; Kantar's statements linking attention scores to sales give no sample or design. (vendor claim) [261, 262] Facial coding is an optional, separately priced LINK+ module, so a LINK+ test does not necessarily include any observed attention measure. (vendor documentation) [246, 263]
- **Opportunities.** The June 2026 release added creator-content testing to LINK AI, benchmarked on 21,000+ creator pieces, giving Kantar a priced entry into creator budgets. (inference) [264] Selling Kantar Media to H.I.G. Capital, completed 1 August 2025 for $924.1 million, leaves the group focused on brand building and marketing effectiveness, where creative testing sits. (inference) [265, 253] The May 2025 extension brought LINK+ survey testing to audio, retail media and in-game ads; documenting attention there would widen the attention offer. (inference) [260] Kantar says serviced and batch API retrieval is available from 2026, which could embed LINK results in client and CreativeX creative workflows. (inference) [251, 252]
- **Threats.** Facial coding and gaze depend on Affectiva; the three-year renewal announced 3 October 2024 points to expiry around late 2027 unless renewed. (inference) [266, 267] The ARF's June 2024 Phase 2 test found little to no correlation between twelve companies' attention scores and advertisers' success ratings; the best R-squared was 0.12. (inference) [4, 268] Kantar's May 2026 privacy notice has Affectiva storing facial-expression video for seven years, a retention period that may draw biometric privacy scrutiny; no action was found. (inference) [267] The group recorded a FY2025 statutory loss before tax of $438.6 million, and WPP carries its 40% stake at nil, leaving room for future ownership change. (inference) [253, 269]

- **What is sold.** Three separately listed creative-effectiveness products on Kantar Marketplace, each with its own US list price: (1) LINK+, survey-based ad pretesting of digital, TV, audio, print and outdoor ads, self-serve (US from $4,500) or serviced (US from $8,080), with an optional facial-coding module and Kantar's attention framework (passive and active attention, second-by-second traces, norms), plus webcam eye tracking and AI predictive eye tracking as neuro options; (2) Context Lab, serviced-only in-context digital ad testing with a control versus exposed design (US from $13,950), to which Kantar says the attention framework also applies; (3) LINK AI, a machine-learning prediction of survey-based creative-effectiveness metrics, including attention and predictive eye tracking outputs, self-serve (US from $500) or serviced (US from $1,780). A subscription bundling LINK+ and LINK AI is also offered. All prices are list 'starting from' figures, not realised prices.
- **Who buys it.** Advertisers, insights teams and agencies; Kantar Marketplace describes itself as a platform 'designed for insights professionals, marketers and agencies'. Named users in the sources read include Coca-Cola (early adopter of the 2025 LINK+ format extensions), Genomma Lab (LINK AI testimonial in the North America playbook) and L'Oréal (LINK AI examples). No LINK+ client list or count is published.
- **How it earns.** Per-project fees on Kantar Marketplace (pay-as-you-go) or subscriptions bundling LINK+ and LINK AI; serviced options cost more than self-serve. Price drivers named on the pages: the optional facial coding module and custom questions (LINK+), add-on features such as the content optimiser (LINK AI), and the number of media platforms tested (Context Lab). No product-level or group revenue was found in a primary source.
- **Supply and data dependencies.** Facial coding and the attention measure depend on Affectiva (a Smart Eye company): Kantar calls Affectiva its long-standing attention measurement partner, Affectiva announced a three-year partnership renewal on 3 October 2024 and the inclusion of its Emotion AI in LINK+, and Kantar's survey privacy notice names Affectiva Inc. and eye square GmbH as analytic partners for facial-expression and eye-tracking data. A 2023 Kantar article also names Realeyes as one of its attention measurement partners. Participants are online survey respondents (Kantar's own panels are implied by 'verified audiences' but not described per product). LINK AI depends on Kantar's LINK survey database (300k+ ad tests) as training labels. Distribution routes include the Kantar API (self-serve LINK+ and LINK AI results) and CreativeX's Datalink platform (LINK AI).
- **Differentiation.** Largest stated creative norm base in the sources screened: 260,000 ads for LINK+ survey metrics and 50,000 attention ad tests (vendor-stated; not compared with other vendors' bases on a common definition). Combines survey diagnostics, facial-coding attention and emotion, webcam eye tracking and AI prediction under one brand, with public US list prices for three products. Separates passive (eyes on screen) from active (emotional engagement) attention and publishes second-by-second traces; Kantar's own data show the two barely correlate. Documented API for self-serve LINK+ and LINK AI results and a named third-party integration (CreativeX) for LINK AI.
- **Risks.** Dependence on Affectiva (Smart Eye) for facial coding and gaze; the renewal announced in October 2024 is for three years (inference: expiry around late 2027 unless renewed). Ownership: Kantar's audited 2025 accounts put ownership at 60% Bain Capital and 40% WPP affiliates, with no change reported in H1 2026; WPP carries its stake at nil (ATT-V-KANTAR-S01, ATT-V-KANTAR-S02, ATT-KANTAR-S08). Evidence of outcome value for the attention measures rests on vendor statements; the ARF Phase 2 method-level results show weak links between attention scores and advertiser success ratings. Facial-expression video is retained by Affectiva for seven years under the survey privacy notice, a biometric-adjacent retention period that may draw scrutiny under state privacy laws (inference; no action found).
- **Best fit.** Advertisers who want survey-based pretesting with norms and an optional observed attention and emotion read per asset, at second-by-second resolution. Teams comparing an ad's performance across social and video placements before or during a campaign (Context Lab), with control versus exposed brand metrics. High-volume screening of many assets by prediction (LINK AI), accepting that outputs are model estimates of survey scores.
- **Poor fit.** Buyers who need attention measured on live media exposures or a media-trading metric (not what these products do). Buyers who need an accredited or independently validated attention metric. Audio-only or in-game attention measurement, where attention measures are not documented. Comparisons with other vendors' attention units.
- **Limits.** No published agreement or error statistic between Kantar's webcam passive-attention measure and a reference eye-tracking method, and no calibration by environment, in the sources screened. Active attention is a facial-expression emotion measure; it is not gaze-based and should not be compared with gaze-based attention scores from other vendors. Facial coding is an optional, separately priced module in LINK+, so a LINK+ test does not necessarily include the attention measures. The 6 May 2025 format extension documents survey testing in audio, in-game and retail media but does not state that attention is measured there. No LINK+-, Context Lab- or LINK AI-specific annual test volume is published; the 13,000+ creatives tested in 2025 is company-wide. LINK AI's attention outputs and predictive eye tracking are model estimates validated against survey scores, not observed attention. Outcome validation published by Kantar concerns survey metrics (STSL, Power) or LINK AI overall, not the attention measures. The only third-party certification on record (MASB, February 2017) covers the predecessor Link copy test and is members-only in detail. MASB states that certification is currently included with its membership, and Kantar is a member.

| Dimension | Availability | Evidence | Observed |
|---|---|---|---|
| Method validity | live | D | LINK+ attention measures are direct observations of recruited online test participants by webcam, not model estimates: passive attention is the time or percentage of ad time 'where viewers' eyes are on-screen when the ad was visible', active attention the time where 'facial expressions indicate emotional engagement' (S01). The method is facial coding with 'gaze monitoring components' (S01); Kantar uses Affectiva's emotion AI (S23) and names Affectiva as its long-standing attention measurement partner and Realeyes as another partner (S17). Affectiva's September 2023 attention metric combines facial expression analysis with Smart Eye gaze and head-position tracking, calibration-less, built on '14.5 million face videos from 90 countries' (S29); treating that metric as the one inside LINK+ is an inference from matching wording, not a documented statement. Neuro options listed for LINK+ are facial coding, webcam eye tracking and AI predictive eye tracking (a model estimate) (S10). Sample: 'N=150 respondents per study', 10-20 minute monadic online survey (S10). No agreement or error statistic between Kantar's webcam passive-attention measure and a reference eye-tracking method, no reference sample, and no calibration by environment were found in the sources screened; Affectiva's release claims 'unparalleled accuracy' without a statistic (S29). Active attention is an emotion measure, a candidate proxy for visual attention under the brief's definitions. ARF Phase 2 reported that its eye-tracking-plus-facial-coding companies agreed with each other (R-squared 0.78 between two of them) but anonymised the vendors, so no result can be attributed to Kantar (S20). [249, 263, 250, 4, 387, 388] |
| Outcome validation | live | D | Sales: Kantar's 2022 volumetric validation of Link's survey metric Short-term Sales Likelihood aggregated client sales validations into 970 ad-level cases in 27 markets (high STSL = index 133 vs average; 76% vs 28% of ads followed by a short-term sales increase in 8 weeks; 'holds true within the US') (S16). This validates Link survey metrics, not the attention measures; it is vendor-published, client models and holdout are not described. Attention-specific: Kantar states that 'by mapping creative attention scores to short-term sales shifts' active attention proved more important than passive (S24) and that facial-coding Expressiveness 'drives shifts in sales volume share' (S23); neither gives sample, design or statistics. Brand: Kantar states a Power Contribution metric is validated against long-term brand equity shifts (S16), a survey metric. Memory: none found for the attention measures. Third-party: MASB certified 'Link Copy Test' in February 2017 (predecessor product, before the 2023 attention framework; detail members-only) (S18). The ARF Phase 2 study found that no method group's scores correlated with advertisers' own success ratings above R-squared 0.12 (the eye-tracking-plus-facial-coding group), with vendors anonymised (S20). No independent or replicated validation of LINK+ attention measures against memory, brand or sales outcomes was found in the sources screened. [262, 254, 4, 387, 258] |
| Granularity and timing | live | D | Per asset, the attention framework gives 'a summary of an ad's ability to capture viewers' attention, comparison to norms, and a second-by-second trace' of passive and active attention (S01, repeated with worked US examples in S17 and S25). Element level: AI predictive eye tracking and a point-and-click module pinpoint 'the parts of your TV, digital, print and outdoor ads that attract the most consumer attention' (S02); AI-generated heatmaps and fog maps (S10). Asset-level survey metrics are compared with norms (for example percentile ranks among US LINK+ digital tests, S17). Whether norms exist at second-by-second resolution, rather than for the summary score, is not stated; the heatmaps are model estimates, the traces direct observation. [249, 246, 263, 250, 259] |
| Activation | unconfirmed | E |   |
| Creative diagnostics | live | D | Pre-launch testing from ideas and storyboards to finished ads with a described method: N=150 respondents per study, 10-20 minute online mobile-friendly monadic survey, in-context digital testing on named placements, core metrics (awareness, brand associations, short- and long-term sales) plus optional diagnostics and neuro modules (S10). Norms: 'the world's largest ad testing database of 260,000 ads' (S02; 260K on S15; 250K+ in 2023 and 2025 documents, S01, S10), 'Category & country norms' (S10), and an attention norm base of '50,000 attention ad tests' (S01; 'observed human attention data for over 50,000 ads', S17). No breakdown of the norm base by category or market (counts per cell) is published in the sources screened. [249, 246, 263, 261, 250] |
| Channel coverage | live | D | Testing: the LINK+ page offers 'digital, TV, audio, print and outdoor ads' (S02); the 6 May 2025 release adds social apps (Reels, Shorts), ads and native content on websites, retail media, audio (radio, podcasts, streaming), in-context digital VOD and console in-game ads (S05). Against the seven channels: display, online video, TV spots, social, audio and in-game are documented as live for survey testing; outdoor is documented but not as DOOH. Attention specifically: Kantar documents the attention framework 'in LINK+ for digital and TV ads' (S17) and the attention module for 'TV, digital, print and outdoor ads' (S02); the 2025 format release does not mention attention (a verifier's qualification), so attention measurement in audio, in-game and retail media is not documented. [246, 260, 250] |
| Data access and integrations | live | C | Reports: dashboard and PowerPoint export (S02). Documented API: the Kantar API lets clients 'Retrieve results from Kantar studies (LINK+ & LINK AI)', currently for self-serve projects only, with serviced, Batch and CreativeX retrieval 'available from 2026'; a Kantar API Developer Portal exists but access is through Kantar support, not public, and the client builds the integration (S12). The June 2026 release says the LINK score can be embedded in brands' martech platforms, 'Powered by a new Kantar API' (S14). No named creative-tool or asset-management integration specific to LINK+ was found (the CreativeX integration covers LINK AI, S13). [246, 251, 252, 264] |
| Production scale | live | D | No LINK+-specific count of assets or tests per year was found. Stated totals without a period: a norm database of 260,000 ads (S02), 50,000 attention ad tests (S01), and over 49,000 ads researched with facial coding and 'well over two million face videos' (S23, November 2022). Company level: Kantar cites its 'testing of more than 13,000 creatives globally in 2025' (S27), which is not attributed to LINK+, LINK AI or Context Lab and so cannot be credited to this row. [249, 246, 387, 389] |
| US availability and coverage | live | D | Sold in the US: US list prices on the LINK+ page, 'Starting from $4,500' self-serve and 'Starting from $8,080' serviced, US selected in the market list (S02); a July 2025 Kantar Marketplace North America playbook describes LINK+ (S10). US calibration: Kantar reports percentile ranks 'among the LINK+ digital tests conducted in the US market' (S17), which shows US norms for survey metrics; no US-specific attention norm base is stated. US production use: example tests of US ads (Apple Pay, S17; Netflix 'among the very top performers from the USA', S24), but no named US client of LINK+ was found. Kantar's Link sales validation 'holds true within the US' (S16), for survey metrics. [246, 263, 262, 250, 258] |
| Independent assurance | live | A | MASB's own certification catalog lists 'Kantar Millward Brown / Link Copy Test / Feb 2017' (S18); Kantar describes the MMAP review as an evaluation 'against 10 criteria for ideal metrics' including predictive validity (S19). Scope limits: it covers the predecessor Link copy test, not LINK+ by name and not the attention framework launched in November 2023; the catalog detail is members-only; and MASB states that certification and catalog access 'are currently included with MASB membership' (S18), so the certified provider is a MASB member and independence is qualified. MASB's December 2024 certification of Kantar's MDS framework is a brand-equity framework, not ad testing (S18). No audit, certification or accreditor record covering Kantar's attention measures was found in the sources screened (MRC's accredited lists contain no Kantar entry, per this paper's verified accreditation research; not re-read here). Affectiva's statement that the Kantar work was 'recognized as part of the ARF Attention Validation Initiative' (S11) is a vendor assertion; the ARF report anonymises vendors and is not an audit (S20). [266, 254, 390, 4] |
| Method disclosure | live | D | Of the seven elements: (1) construct, published (passive and active attention definitions, S01); (2) observations, published at summary level (webcam facial coding with gaze monitoring; low-attention signals such as gazing away, speaking and drowsiness, S01, S17); (3) reference labels, not published for the gaze component (Affectiva's facial-coding training corpus is described only in Affectiva's release, S29); (4) sample, partly published (N=150 per LINK+ study, S10; recruitment described only as Kantar's 'verified audiences'); (5) scoring procedure, summary (seconds or percentage of visible ad time, S01); (6) validation statistics, none published for the attention measures; (7) update cadence, none. About four of seven at summary level, no data. [249, 263, 250, 388] |
| Privacy posture | live | C | Company-level, not product-specific: Kantar's Global Surveys Privacy Notice (May 2026) says facial expression and eye-tracking data are collected only with the respondent's consent ('We will always ask your permission to record this data'), names eye square GmbH and Affectiva Inc. as analytic partners that may analyse, anonymise and summarise these data, states that Affectiva stores facial-expression video 'on our behalf for seven years', sets a standard survey retention of 12 to 18 months, and says anonymised data may be used to train AI algorithms (S22). A 2022 Kantar article states facial expressions are captured 'via webcams with explicit consent' and that the data are anonymised and aggregated (S23). The notice names the data types but not LINK+; no product-specific consent flow, data-minimisation statement for LINK+ or independent privacy certification was found. This paper's privacy search found no regulator action or lawsuit naming Kantar, but Kantar was not searched by name there. [267, 387] |

Performance evidence: Link's Short-term Sales Likelihood metric predicts sales volume: across 970 homogenised client validation cases in 27 markets, high-STSL ads generated 33% more sales than average (index 133) and low-STSL ads 27% fewer; 76% of top-third ads vs 28% of bottom-third ads were followed by a short-term sales increase within 8 weeks. (vendor assertion, sponsor Kantar (aggregation of client 'volumetric validation' projects; client funding and models not described)) [262]; Mapping creative attention scores to short-term sales shifts showed active attention matters more than passive attention. (vendor assertion, sponsor Kantar) [258]; LINK AI predictions match LINK survey results: 89-95% correlation on over 15,000 hold-out ads (2025); more than 7 in 10 video and TV ads get comparable ratings; about 3 in 4 TV ads in the US, Brazil and Germany. (vendor assertion, sponsor Kantar) [391, 263]; An FMCG client's MMM validation of 8,500 digital ads found the top third of ads on LINK AI delivered 37% more incremental platform-driven sales than the bottom third. (vendor assertion, sponsor Unnamed client; published by Kantar; funding unknown) [391]; Kantar's own in-market validation on over 400 ads and 35 LIFT+ campaigns found strong LINK AI creative drives 50% incremental brand growth; Whalar's creator-content validation was within +/- 2 percentage points of brand-lift studies. (vendor assertion, sponsor Kantar; Whalar (agency partner)) [391]; In the ARF's Phase 2 test (32 ads, 12 companies, vendors anonymised), attention scores showed little to no correlation with advertisers' own success ratings; the eye-tracking-plus-facial-coding group was highest at R-squared 0.12. (independent measurement, sponsor ARF (member-funded industry body; companies analysed ads in kind; initiative funding not stated in the report read)) [4, 268]; In Kantar's LIFT+ database of 873 multi-media campaigns over five years (over $3.2bn media spend), channel attention levels per impression do not seem directly related to channel-level cost-effectiveness. (vendor assertion, sponsor Kantar) [258]

### Lumen Research (active)

*Lumen Research Ltd (England and Wales, company 08682432)*

Eye-tracking panel specialist in both cohorts; its campaign numbers are model estimates.

**SWOT (2026-09-29).** Lumen is a London eye-tracking specialist whose model-based attention estimates are widely integrated, including upstream of IAS and DoubleVerify, while its outcome validation and accreditation remain weak. Its unaudited accounts show two loan covenant breaches at 31 December 2025, and the directors report no material going-concern uncertainty.

- **Strengths.** Lumen holds European patent EP3861424B1 on webcam eye-tracking data collection, granted 29 November 2023, and runs its own opt-in webcam panels rather than licensing gaze data. (record) [270, 271, 272] Lumen eye-tracking data or models feed IAS Quality Attention (January 2024), IAS social and Snap attention products, and DoubleVerify's social attention product on Snap (June 2025). (vendor claim) [273, 115, 274, 116] The PwC review hosted by Lumen publishes the model's formulas, features and validation statistics, at November 2022 vintage. (vendor documentation) [8] Documented routes include The Trade Desk billing (March 2024), Teads Ad Manager (2023; CTV HomeScreen May 2026), Omnicom's Omni planning tools (May 2024) and Netflix (March 2026). (vendor claim) [275, 276, 277, 278, 110] One eye-tracking instrument serves both media measurement (LAMP) and participant-based pre-launch creative testing (SPOTLIGHT). (vendor documentation) [20, 279, 280]
- **Weaknesses.** Every LAMP campaign, planning and bidding figure is a model estimate from placement signals; no gaze of the exposed person is observed. (vendor documentation) [281, 282] In PwC's Lumen-commissioned review, attention and viewability both related to three memory metrics, but attention was not significantly better; PwC could not recreate Lumen's validation. (vendor documentation) [8, 283] The published model uses viewable time as a core input, and Lumen's own plots showed weaker fit for skins and in-stream video. (vendor documentation) [8] Lumen has no MRC accreditation and no MRC audit in progress, and is absent from the CESP attention grid, at 29 September 2026. (record) [2, 48, 50, 3, 284] Group loss rose from £1.13m (FY2024) to £3.76m (FY2025, as filed); by inference, a £2.72m non-cash share-based payment charge explains almost all the rise. Turnover is not filed. (record) [285]
- **Opportunities.** Because IAS and DoubleVerify products draw on Lumen data, Lumen-derived attention reaches buyers through verification suites Lumen does not sell itself. (inference) [116, 274, 115] Teads' exclusive Lumen CTV HomeScreen measurement (May 2026) covers the US, EMEA, APAC and LATAM on a model that does not rest on TVision. (inference) [277, 286] The March 2026 Netflix release names five European client markets; adding further client markets is an open expansion route. (inference) [110] The April 2026 Cint partnership targets the link between attention and brand outcomes, the gap PwC's review left open. (inference) [287, 8]
- **Threats.** The Salica and Innovate UK loans were in covenant breach at 31 December 2025, leaving net current liabilities of £63,945; directors report no material going-concern uncertainty. (record) [288] TVision, Lumen's US CTV data partner, has been owned by Viant, a DSP, since May 2026; no later Lumen source mentions TVision. (inference) [219, 289, 290] DoubleVerify holds MRC accreditation for attention metrics on desktop, mobile web and in-app; buyers requiring accredited metrics may pass over unaccredited Lumen. (inference) [2] Lumen calls the PwC work an independent audit, which PwC's own wording does not support; buyers who check may discount Lumen's claims. (inference) [20, 8]

- **What is sold.** Attention measurement, planning and optimisation for media exposures under the Lumen Attention Measurement & Planning Platform (LAMP): a client-side JavaScript measurement tag that applies Lumen's attention model to impression data and estimates 'attentive seconds' per impression (ATT-LUMEN-S11); a self-service planning platform and dashboard (ATT-LUMEN-S07); a 'LAMP DSP plug-in' that embeds attention predictions as a custom algorithm in the buyer's DSP (ATT-LUMEN-S12), marketed as Lumen Attention Bidding (ATT-LUMEN-S08); attentive private marketplaces (aPMPs) with named partners (ATT-LUMEN-S31, S33); bespoke and brand-specific attention models; and bespoke cross-media evaluation studies (TV, BVOD, digital, print, OOH, point of sale) (ATT-LUMEN-S08). Separately, Lumen SPOTLIGHT sells pre-launch creative testing with webcam eye tracking of opt-in participants (ATT-LUMEN-S18, S19). Headline metric: APM, attentive seconds per 1,000 impressions, defined as % viewed x average eyes-on dwell time x 1,000 (ATT-LUMEN-S08).
- **Who buys it.** Advertisers and agencies, and media owners and platforms that integrate or resell Lumen measurement. Named agency relationships in the sources opened: Omnicom Media Group (Lumen benchmarks integrated into Omni channel-planning tools, May 2024, ATT-LUMEN-S49), dentsu (eye-tracking panels of 6,000 built with Lumen in its Attention Economy programme, ATT-LUMEN-S48) and Havas Media Network (brand-lift analysis, ATT-LUMEN-S26). Named media-owner and platform partners: Teads (Teads Ad Manager, 2023; CTV HomeScreen, 2026), Snapchat, Netflix (European clients), LG Ad Solutions, Nexxen, The MediaGrid (Criteo), The Trade Desk (billing), IAS (social attention reporting), News UK, Newsworks, Silverpush (ATT-LUMEN-S30 to S41, S23, S50).
- **How it earns.** Not disclosed in public filings: the company files unaudited small-company accounts without an income statement, so turnover and revenue mix are not public (ATT-LUMEN-S04). The documented commercial routes are measurement fees billable through The Trade Desk's 'Manage Fees & Rates' toggle (ATT-LUMEN-S30), platform integrations with media owners (Teads, Snapchat, Netflix) (ATT-LUMEN-S35 to S38), agency data licences (Omnicom, ATT-LUMEN-S49) and bespoke studies and SPOTLIGHT tests (ATT-LUMEN-S08, S18). Prices are not published in the sources screened.
- **Supply and data dependencies.** Reference data: Lumen's own webcam eye-tracking panels (opt-in browser extension or mobile app; UK and US panel sizes stated on a 2021 page) and lab or study eye tracking (ATT-LUMEN-S09, S14). CTV: TVision panel data and pixel (the 2023 partnership; LG Ad Solutions in the US via the TVision pixel, 2024; Nexxen measurement 'powered by TVision data') (ATT-LUMEN-S32, S33, S34); the Teads CTV HomeScreen model instead rests on MediaMento research with Teads (ATT-LUMEN-S36). Social: lab eye tracking with Snapchat since 2021 (ATT-LUMEN-S38). Production signals come from the LAMP tag on the buyer's creatives or from platform integrations (ATT-LUMEN-S11). Downstream, IAS and Snap-related products draw on Lumen eye-tracking data (ATT-LUMEN-S41), so Lumen is itself a supply dependency for other vendors. Financing: a Salica Growth Debt Fund I SPV venture-debt facility secured by a fixed and floating charge, and an Innovate UK loan (ATT-LUMEN-S04, S06).
- **Differentiation.** Owns a webcam eye-tracking collection platform (EU patent announced February 2024) and panels running since the early 2020s or before, rather than licensing gaze data from others (ATT-LUMEN-S09, S17). Its eye-tracking data are used by other vendors: IAS social attention reporting and IAS pre-bid segments draw on Lumen (ATT-LUMEN-S41), making Lumen an upstream reference source as well as a seller. Publishes more of its model than most peers: the PwC review hosted by Lumen gives the model formulas, features and validation statistics (ATT-LUMEN-S20). Has both a media measurement row (LAMP) and a participant-based creative testing row (SPOTLIGHT) on the same eye-tracking instrument (ATT-LUMEN-S18, S19). Documents platform integrations across The Trade Desk (billing), Teads Ad Manager, Snapchat, Netflix (Europe), LG Ad Solutions and Nexxen (ATT-LUMEN-S30, S33 to S38).
- **Risks.** Financing risk (record): the Salica venture-debt and Innovate UK loans were in covenant breach at 31 December 2025, leaving the group with net current liabilities of £63,945 after reclassification; the directors state no material uncertainty on going concern (ATT-LUMEN-S06). Rising losses (record): group loss £1.13m (FY2024) to £3.76m (FY2025, originally filed; £3.81m restated), with £2.72m of it share-based payment; cash fell from £1.92m to £1.15m in the three months to December 2025 (ATT-LUMEN-S04, S06). US CTV dependency (inference): documented US CTV measurement for LG Ad Solutions and Nexxen runs through TVision data or pixels (ATT-LUMEN-S33, S34). Whether that supply continues after TVision's change of ownership is not documented in the Lumen sources screened. Governance change (record): the co-founder and the director named in 2022 as Avocet's CEO left the board in July and August 2026; no reason is stated (ATT-LUMEN-S02, S47). Accreditation gap (inference): with a DoubleVerify attention product already accredited by MRC for some environments, buyers requiring accredited metrics may pass over an unaccredited specialist; this is a monitoring point, not a documented effect. Claim risk (record): Lumen calls the PwC work an independent 'audit', which PwC's own limitations wording does not support (ATT-LUMEN-S07, S20).
- **Best fit.** Buyers wanting a single attention unit (APM) across display, online video, social and CTV for planning and post-campaign comparison, who accept model estimates (ATT-LUMEN-S08, S30, S36, S38). UK and European campaigns, where the panel and partner integrations are strongest (Netflix in five European markets; LG EMEA tag) (ATT-LUMEN-S34, S37). Teads, Snapchat and The Trade Desk buyers who want measurement switched on inside the platform (ATT-LUMEN-S30, S35, S38). Pre-launch creative teams wanting participant eye tracking with heat maps and view order across many formats (SPOTLIGHT) (ATT-LUMEN-S18, S19).
- **Poor fit.** Buyers requiring MRC-accredited attention metrics or an accredited billing unit (ATT-LUMEN-S42, S43). Buyers needing proof of added value over time in view on held-out data (ATT-LUMEN-S20). US CTV buyers needing Netflix coverage (the Netflix solution is not offered to US clients) (ATT-LUMEN-S37). Audio, in-game and DOOH measurement as an ongoing product: only bespoke studies were documented (ATT-LUMEN-S08).
- **Limits.** All LAMP campaign, planning and bidding numbers are model estimates of attention from placement signals; none is observed gaze of the exposed person (ATT-LUMEN-S10, S11). The published model (PwC review, November 2022 vintage) predicts viewing and dwell from ad size, viewable time and domain terms, so LAMP estimates are partly built from time in view; the published comparisons reviewed here do not establish incremental predictive value beyond time in view (ATT-LUMEN-S20). No MRC accreditation and no MRC audit in progress at 29 September 2026 (ATT-LUMEN-S42, S43, S44). The only third-party method review (PwC) was commissioned by Lumen, used a Lumen-cleaned subset, could not recreate Lumen's validation, and found no significant memory advantage over viewability (ATT-LUMEN-S20). Panel sizes by country (UK 1,000; US 1,000 mobile) are published only on a 2021 page; current panel size, recruitment and weighting are not public (ATT-LUMEN-S09). No LAMP production volume with a period is published; scale claims are cumulative totals (ATT-LUMEN-S08, S30, S41). Financial disclosure is thin: unaudited small-company accounts, no income statement, no turnover (ATT-LUMEN-S04).

| Dimension | Availability | Evidence | Observed |
|---|---|---|---|
| Method validity | live | C | Campaign attention figures are model estimates, not observations of the exposed person. Reference method: webcam eye tracking of opt-in panellists (a browser extension or mobile app records on-screen events; the webcam estimates gaze 'within a circle c 2.5cm or 1 inch diameter', processed on device) (S09, S13, S14). Reference sample: a panel page published June 2021 states UK 1,000 (250 desktop, 750 mobile) and US 1,000 mobile users, 'nationally representative', consented and compensated, 'refreshed frequently' (S09); later vendor totals give 650,000+ eye-tracking sessions over 10+ years (S08, S17) and coverage of '30+ countries' (December 2024, S41) or '50+ countries' (April 2026, S40), without current panel sizes by country or recruitment detail. Model: placement features (format, % pixels in view, viewable time, screen real estate, scroll speed, competing ads, position, page geometry, domain) predict attention every 100ms (S10). PwC's commissioned review of the November 2022 model found a logistic regression for P(viewed, gaze of at least 100ms) and a quasi-Poisson dwell model on size, log viewable time and domain terms, without ad position or clutter (S20). Agreement statistics: Lumen's own validation, with plots shared with PwC that PwC 'was not able to recreate', reports McFadden pseudo-R2 0.26 (global) and 0.29 (domain), accuracy 81% (50% threshold) and 78% (30%), specificity 56% and 68%; PwC's re-performance on a Lumen-cleaned sample of 35,000 banner and 5,000 outstream impressions found view predicted correctly about 70% of the time, 68% for viewed impressions, with a dwell mean error of 1.2 s; skins and in-stream were not tested (S20). Whether Lumen's validation used held-out data is not stated. Channel models are described separately for CTV (TVision panel data, S32, S34; a HomeScreen model 'built on MediaMento research', S36) and Snapchat (lab eye tracking since 2021, S38), with no per-channel agreement statistic published in the sources screened. [392, 271, 281, 393, 394, 272, 8, 295, 290, 277, 395, 287, 115] |
| Outcome validation | live | D | Memory: in PwC's review (commissioned by Lumen; fieldwork June-October 2022), a brand-lift pilot for one advertiser (2,400+ panelists, six studies, 12 metrics) found attention 'not significantly better than view' on first-mention awareness, any-mention awareness and digital ad recall and no significant result for either measure on the other metrics listed; PwC also called the lift calculation 'slightly unconventional' (S20). Lumen's release conceded the brand-lift relationship was 'in line with viewable time metrics' (S21). The Teads and Lumen meta-analysis (16 selected campaigns, 14 advertisers, Dynata brand lift) reports online ad recall lift 40% higher for 5+ s than 1+ s, single exposure only, in-sample (S27). The Havas, Lumen and Brand Metrics analysis (9,089 display campaigns) is descriptive, with no control group, using Lumen attention predicted from viewability data (S26). Brand: Lumen and Cint claim luxury campaigns on high-attention media gained 9 pp spontaneous awareness and 22 pp favourability, with no design stated (S40); Newsworks, Lumen and Peter Field report 17% higher brand effects and 12% higher market-share growth for 'high-attention media campaigns', a media-class comparison (S50). Sales and conversion: PwC found attention 'correlated better than viewability' with CTR and conversion on a Lumen-supplied subset, correlation only (S20). Snap and Lumen claim 1 s of attention per impression 'is correlated to a 20X increase in ad recall' across 285 brand-lift sessions (S38). No independent, out-of-sample validation of LAMP estimates against memory, brand or sales outcomes was found in the sources screened. The independent Simonov et al. study used Lumen webcam eye tracking directly, not LAMP model estimates, so it does not validate this product (S46). [8, 283, 10, 68, 395, 287, 51, 70] |
| Granularity and timing | live | C | Every stage outputs a model estimate. In flight and post-campaign: the LAMP tag 'Collects impression-level data' and 'Applies Lumen attention model to impression data to estimate ‘attentive seconds’ each impression generates' (S11); a CPG case describes measuring all impressions 'in real-time via Lumen Attention Measurement Dashboard' (S28). Pre-bid: 'Lumen attention predictions embedded as a ‘custom algorithm’ within your DSP' with attention thresholds (S12), and aPMPs with 'pre-bid filters' (S31). Planning: a self-service platform to 'generate custom reports to plan your next campaign' (S07). So impression-level output is documented at planning or pre-bid, in-flight and post-campaign stages. Lumen's pages describe the model as a prediction ('Attention to ad predicted every 100ms of visual exposure', S08, S10), but release copy also says the tag 'can measure whether an ad was actually seen' (S30), so the vendor does not consistently label the measurement basis at each stage. [20, 392, 281, 282, 396, 69, 275, 397] |
| Activation | live | D | Lumen Attention Bidding is 'a pre-bidding custom algorithm' that integrates 'with your chosen DSP'; no DSP is named on the product page (S08, S12). Named buying platforms with documented Lumen activation: The MediaGrid (Criteo subsidiary), attentive PMPs launched 31 May 2023 (S31); Nexxen DSP, CTV aPMPs 'activated through Nexxen’s demand-side platform', 8 August 2023 (S33); Teads Ad Manager, described as 'a one-touch system for measuring and optimizing attention', 14 June 2023 (S35); the Avocet DSP, reported in 2022 as combined with Lumen's eye tracking, with dentsu and Havas using it to optimise bidding (S47; Avocet Systems Ltd is a subsidiary, S04, so it counts as one parent-owned platform). IAS markets pre-bid segments using Lumen data inside its own Quality Attention product (S41), a partner route. The Trade Desk integration is measurement billing, not a buying use (S30). Status at the cutoff of the 2022-2023 named routes was not reconfirmed in the sources screened. [285, 392, 396, 275, 397, 289, 276, 115, 398] |
| Creative diagnostics | unconfirmed | E |   |
| Channel coverage | live | D | Live as a product, per vendor and partner releases: display and online video (the LAMP tag; The Trade Desk billing for 'both display ad and video ad groups', S11, S30; YouTube campaign measurement against 'standard YouTube advertising attention benchmarks', February 2026, S39); CTV (US measurement via the TVision pixel processed through the Lumen platform for LG Ad Solutions, February 2024, S34; Teads CTV HomeScreen across the US, EMEA, APAC and LATAM, May 2026, S36; Netflix CTV, desktop and mobile for clients in five European markets, March 2026, S37); social (Snapchat integration, March 2024, S38; social attention reporting through IAS stated to go live from January 2025, S41). That is four of the seven scored channels. Audio, in-game and DOOH appear only as bespoke studies in the sources opened ('Lumen conducts bespoke media evaluation studies across a variety of media', S08). Retail media and print are attributes only. [392, 282, 275, 290, 277, 110, 395, 399, 115] |
| Data access and integrations | live | D | Named output routes: The Trade Desk native billing toggle that switches on Lumen measurement per ad group, March 2024 (S30); Teads Ad Manager, native Lumen tagging and metrics from June 2023, and CTV HomeScreen measurement 'available now globally in Teads Ad Manager' from May 2026 (S35, S36); IAS, Social Attention reporting with Lumen for IAS customers from January 2025 as announced (S41); Omnicom's Omni channel-planning tools, where Lumen benchmarks are integrated (May 2024, S49); LG Ad Solutions CTV ads (S34); Lumen's own self-service platform and dashboard (S07). No public API reference or log-level export documentation was found in the sources screened. [20, 275, 290, 276, 277, 115, 278] |
| Production scale | live | D | Only totals without a period or start date are stated for LAMP: 'more than 20 billions attentive impression' (The Trade Desk release, March 2024, S30); a modelling dataset of '30B+ programmatic ad impressions and 300B+ social ad impressions' (S08); 'over 300 billion impressions across multiple social platforms' measured by Lumen's models (IAS release, December 2024, S41). Period-bound figures are study-sized: 60M+ YouTube impressions across 30+ campaigns (S39) and 9,089 display campaigns in the Havas dataset, collection period not stated (S26). Teads reported '103 advertisers onboarded and over 200 campaigns run since the launch of TAM', without saying these were Lumen-measured (S35). No annual volume of impressions or campaigns measured with LAMP was found. [392, 10, 275, 276, 399, 115] |
| US availability and coverage | live | D | Sold in the US: a US subsidiary, Lumen Attention Solutions Inc. (131 Continental Dr., Suite 305, Newark, DE), is 100% owned (accounts, S04); US measurement through the TVision pixel for LG Ad Solutions (S34); Teads CTV HomeScreen includes the US (S36). US calibration: the panel page states 'US: 1000 mobile users' as a nationally representative sample, a page published in June 2021 with no later revision date (S09). US production use: 'A US telco company drove 8X higher conversion with high-attention PMPs', unnamed (S29). No named US client of LAMP was found in the sources opened. Lumen's Netflix solution is not offered to US clients (S37). [285, 271, 400, 290, 277, 110] |
| Independent assurance | live | A | No MRC accreditation and no audit in progress: on 29 September 2026 MRC's 'Accredited Services: Digital' page and its digital and television 'Services Under Review' pages contain no Lumen entry (S42, S43, S44). CESP's attention grid (10 June 2026) does not list Lumen (S45). Lumen describes its technology as 'the only solution validated and audited by an independent third-party in the industry' on the strength of the 2023 PwC report (S07), but PwC states that Lumen requested the review and that PwC 'has not independently verified, validated, or audited the data' (S20). This amounts to a vendor self-description, not an accreditor record. [20, 8, 2, 48, 50, 3] |
| Method disclosure | live | C | Against the seven elements: (1) construct published: attentive seconds, 'whether an ad was actually seen and for how long', APM = % view x average eyes-on dwell x 1,000 (S08); (2) signals published: placement features listed (S10, S11); (3) reference labels published: eye gaze on the ad for at least 100ms defines 'viewed', and dwell time (S20); (4) panel size and recruitment partly published: UK and US panel sizes and 'nationally representative' on a 2021 page, no current figures (S09); (5) model type published through the PwC review hosted by Lumen: logistic regression and quasi-Poisson formulas (S20); (6) validation statistics published in that review (accuracy, sensitivity, specificity, pseudo-R2), as of November 2022 (S20); (7) update cadence not stated beyond 'continuously updated' (S41) and 'Panels refreshed frequently' (S09). So six of the seven elements are published, at a 2021-2022 vintage for items 4 to 6, and no data are released. Lumen's metric definitions are also inconsistent: its February 2023 blog defines 'Attention seconds per 1000 impressions' as 'The number of attentive seconds divided by total impressions' (S22). [392, 271, 281, 282, 8, 44, 115] |
| Privacy posture | live | C | Panel consent model (vendor-stated): 'All panellists fully informed, consented and compensated for participation' (S09). Data minimisation (vendor-stated): webcam processing 'is done on the users’ device, and no video data or personally identifiable information is uploaded' (S13); 'No personally identifiable information (PII) collected, stored or reported'; 'All data collection, storage and processing fully GDPR-compliant' (S08, S09). Production tag data: Teads describes the metrics as 'inherently cookieless' (S35); no Lumen document on what the LAMP tag collects per impression beyond placement signals (S11) was found. The website privacy policy (last updated June 2026) covers the website only and addresses GDPR and US state laws, including the CCPA, for website users (S16). No panel or participant privacy notice, and no independent certification or regulator record, was found in the sources screened. [392, 271, 282, 393, 401, 276] |

Performance evidence: Attention correlated better than viewability with CTR and conversion for three advertisers; in a brand-lift pilot attention was not significantly better than viewability on first-mention awareness, any-mention awareness or digital ad recall. (commissioned measurement, sponsor Lumen (PwC review at Lumen's request)) [8, 283]; Online ad recall lift was 40% higher for exposures of 5+ seconds than 1+ second. (vendor assertion, sponsor Teads and Lumen (joint white paper)) [68]; More aggregate attention is associated with higher awareness, consideration, preference and action intent across 9,089 display campaigns. (vendor assertion, sponsor Havas Media Network with Lumen and Brand Metrics (funding of the analysis not stated in the white paper)) [10]; An attention-bidding campaign delivered +104% viewed ads, +110% view time and 40% lower CPM than a control campaign on the same budget. (vendor assertion, sponsor Lumen (case study, unnamed CPG brand)) [69]; One second of visual attention per impression is correlated with a '20X increase in ad recall' on Snapchat, and AR lens and commercial formats have 40% lower cost of attention. (vendor assertion, sponsor Lumen, with Snap marketing science quoted) [395]; High-attention media campaigns showed a 17% uplift in brand effects and a 12% increase in market-share growth, and 58% more attentive seconds per pound. (vendor assertion, sponsor Newsworks (UK newsbrands' marketing body) with Lumen and Peter Field) [70]; Luxury campaigns on high-attention media averaged +9 pp spontaneous awareness, over 10 pp prompted awareness and +22 pp favourability. (vendor assertion, sponsor Lumen and Cint) [287]; Relevant editorial context raised attention per 1,000 impressions by 31% (and average viewing time by 28%) across 980,000 impressions on The Sun's website (REG-002). (vendor assertion, sponsor News UK with Lumen) [62, 402]; Eye-tracked ad dwell is associated with aided recall and gift-card choice, and visibility 'still produces reliable results'. (independent measurement, sponsor Imperial College COVID-19 Response Fund and NET Institute; Lumen supplied the eye-tracking technology in kind) [51]

### TVision (active; wholly owned subsidiary of Viant Technology Inc. since 1 May 2026)

*TVision Insights, Inc. (Delaware corporation; 'TVision Insights Inc.' on its Form D)*

The one documented US in-home panel observing TV viewers; owned by a DSP since May 2026.

**SWOT (2026-09-29).** TVision runs a directly observed in-home US TV panel that feeds other vendors' models, but it is unaccredited, commercially small and, since May 2026, owned by a DSP.

- **Strengths.** TVision's delivered TV and CTV attention figures are weighted projections of direct observation from a continuous, passive, opt-in US panel of 5,000 homes. (vendor documentation) [45, 291, 292] Other measurement and attention products license or build on TVision panel data, including VideoAmp's CTV calibration licence (February 2026), Adelaide, Lumen and xpln.ai. (vendor claim) [293, 294, 295, 296] The panel measures presence and co-viewing passively across linear TV and CTV, including walled-garden apps viewed on the TV screen. (vendor documentation) [45, 297, 298] Independent academics have used TVision panel data; an MSI working paper found attention explained slightly more brand-search lift than tuning, in-sample, and fixed effects nearly closed the gap. (inference) [52, 299, 300, 301]
- **Weaknesses.** No TVision or Viant metric is MRC-accredited or on an MRC services-under-review list at 29 September 2026. (record) [2, 302, 50, 48, 303] TVision publishes no reference labels or validation statistics, and no agreement or error statistic against a gaze reference was found. (vendor documentation) [45, 292] Outcome evidence beyond the MSI study is vendor grade: the Upwave regression reports no sample or holdout, and Viant's 42-pilot 1.4x lift gives no design or control. (vendor claim) [74, 304] TVision's CRO cited a 2.5-second attentive-session threshold in March 2026, against the 2-second rule in its published Attention % definition. (vendor claim) [305, 292] The panel policy collects face geometry and snapshot images and allows re-identifiable hashed sharing; TVision's ISO 27001 announcement covers information security, not panel privacy. (vendor documentation) [306, 307]
- **Opportunities.** Viant began testing TVision pre-bid attention intelligence in August 2026; a completed rollout could make its announced attention-adjusted CPM a working buying unit. (inference) [308, 309, 304] The June 2026 Havas and Viant Attention Marketplace, scoring Viant inventory by attention decile, gives TVision data an agency route to buyers. (inference) [310] Continued post-closing sales of calibration data and licensing, and Adelaide's present-tense use in July 2026, can preserve TVision's upstream supplier role. (inference) [297, 106, 311] The June 2026 non-infringement verdict and August 2026 Federal Circuit ruling against Nielsen's '243 patent reduce part of TVision's litigation overhang. (inference) [82, 312] TVision said in June 2025 it supported the draft IAB/MRC attention guidelines; an MRC audit would close its main assurance gap. (inference) [313, 2]
- **Threats.** Viant markets TVision integration as exclusive while its filings are silent on outside access, so rival DSPs and vendors may lose TVision data. (inference) [82, 308, 314, 309] Several vendors' CTV attention models share TVision's panel as input, so errors in its data would propagate across them. (inference) [294, 295, 296, 315] Nielsen was contesting the June 2026 verdict in August 2026, a 2025 action remains active, and indemnification from former equityholders is limited. (inference) [82, 316] Viant's CFO called TVision revenue 'pretty de minimis' with an EBITDA-margin drag near 150 bp in Q2 2026, inviting parent reprioritisation. (inference) [304, 82] Face-geometry collection under a policy that disclaims CCPA and VCDPA applicability leaves biometric-privacy exposure, though no regulator action was found. (inference) [306]

- **What is sold.** Television and CTV viewing, presence, co-viewing and attention data from TVision's in-home US panel, sold as: (1) SaaS dashboards: Ad Scoreboard (Ad Scoreboard 2.0 announced 25 August, year inferred 2026), TV Planning & Measurement, CTV Analytics, Linear TV Insights and Total View; (2) CTV Campaign Measurement reports built from a buyer's log files or a TVision pixel matched to panel data; (3) Digital Activation: presence and attention scores for CTV inventory at daypart and channel level for DSP bidding, plus viewing-based audience segments through LiveRamp; (4) calibration data and data licensing feeds (sessionized data, calculated probabilities, behavioural projections) for measurement and data companies; and (5) syndicated reports such as State of Streaming. Since May 2026, the data are also being integrated into Viant's DSP. Viant describes its TVision pre-bid attention targeting as in testing (10 August 2026). Viant announced an 'attention-adjusted CPM' (April and May 2026) and has not documented a launch.
- **Who buys it.** Named in vendor and partner sources: brands and agencies (Anheuser-Busch quote on TVision's site; Havas Media Network North America through the Viant Attention Marketplace; dentsu for Anheuser-Busch in 2020-21); networks and streaming services (AMC Networks, The Weather Channel, Vevo and Scripps are quoted on TVision's pages; A+E Networks); measurement and currency companies (VideoAmp and iSpot as calibration-data licensees); attention and verification vendors (Adelaide, Lumen, DoubleVerify, xpln.ai); and an SSP (OpenX). Customer counts and revenue by buyer type are not disclosed.
- **How it earns.** Not disclosed. TVision's pages describe SaaS platforms and data licensing. That these are paid as subscriptions and licence fees is an inference; no price or contract terms are public. Viant's 10-Q says TVision's results were not material to Viant's consolidated results. Viant's CFO said on the Q2 2026 call that TVision is 'pretty de minimis from a revenue perspective' and that it was a drag of about 150 basis points on Q2 adjusted EBITDA margin (about 200 basis points in the Q3 guide). Within Viant, attention signals feed a DSP that earns platform fees and fees for proprietary data and measurement (424B5).
- **Supply and data dependencies.** In-home hardware installed by panellists: an optical sensor using facial recognition, a microphone array for audio-fingerprint content recognition, and a digital meter using network-traffic analysis to detect streaming apps. The panel is compensated and recruited from top DMAs; TVision's 2026 TV Attention & Creative Impact Report states 'an opt-in panel of 5,000 homes across the United States', in line with trade reporting. Panel data are projected to the US population with universe estimates from the US Census and the American Community Survey, applied through daily weighting. Content recognition relies on third-party ACR: Viant's 10-Q describes Nielsen's 2022 complaint as concerning 'content recognition tools provided by ACRCloud'. Program and ad reference databases are also needed. Downstream, several attention vendors (Adelaide, Lumen, DoubleVerify, xpln.ai) and an SSP (OpenX) have used TVision panel data as an input to their own models or products. That gives the category a common-source dependency on one panel.
- **Differentiation.** Direct observation: among the cohort A products screened, TVision is the only one whose CTV attention numbers come from observing people in their homes (vendor methodology and an independent 2021 paper; verified). Single-source linear TV and CTV panel with presence and co-viewing measured passively, used as a calibration truth set by measurement and currency companies (VideoAmp, iSpot). Upstream supplier: other attention vendors (Adelaide, Lumen, DoubleVerify, xpln.ai) and an SSP (OpenX) have built models or products on TVision panel data. Since May 2026, the only attention panel owned by a DSP, with pre-bid use in that DSP in testing.
- **Risks.** Ownership conflict: a buy-side DSP now owns an independent measurement panel. Viant markets TVision both as 'independent, market-wide measurement' and as an 'exclusive' buying advantage. Rival DSPs and vendors may lose access; this is unresolved in public sources. Litigation: Nielsen's 2025 patent action is active and Nielsen may appeal the June 2026 verdict; indemnification from former equityholders is limited (10-Q). Biometric and privacy exposure: the panel collects face geometry and periodic snapshot images and allows re-identifiable hashed sharing (panel policy). The policy disclaims CCPA and VCDPA applicability. No regulator action was found. Scale: a panel of about 5,000 households supports market and daypart aggregates. Whether it supports stable impression-level or campaign-level signals for small buys is not addressed in public sources (PPC Land raises the same question). Concentration: several vendors' CTV attention models share one panel as input, so errors in TVision data would propagate across them (inference). Commercial scale is small: Viant calls TVision revenue 'de minimis', and the unit carries an EBITDA-margin drag.
- **Best fit.** US TV and CTV planning and post-campaign reads where buyers want person-level presence, co-viewing and eyes-on-screen evidence across linear, apps and walled-garden apps viewed on TVs. Measurement and currency companies needing person-level calibration of household-level big data. Viant DSP clients testing attention-informed CTV bidding (beta at the cutoff).
- **Poor fit.** Buyers needing MRC-accredited attention metrics or a contractual billing unit with an audited definition. Channels other than TV screens (display, mobile, social feeds, audio, in-game, DOOH). Buyers requiring published classifier validation, reference labels or replicable data. Buyers on DSPs other than Viant who need TVision-based pre-bid signals after May 2026 (availability undocumented).
- **Limits.** Panel only: attention is observed only for panellists' TVs (5,000 US homes per TVision's 2026 report). Any figure for a campaign's non-panel impressions is a projection or a model estimate. No published validation statistic for the eyes-on-screen classifier against a gaze reference; the 2021 independent description is a head-orientation inference. One scored channel (CTV) plus linear TV; no display, mobile, social (outside YouTube on TV), audio, in-game or DOOH measurement. No MRC accreditation, and no MRC-listed audit, for any TVision metric at 29 September 2026. Outcome evidence is in-sample, vendor or partner, or uses a behavioural proxy (search); nothing is replicated or held out. Post-acquisition availability of TVision data to third parties is not stated in Viant's filings, while Viant's releases call the integration 'exclusive'. The Viant row 'attention-adjusted CPM' is announced only; it is not scored and has no launch documentation.

| Dimension | Availability | Evidence | Observed |
|---|---|---|---|
| Method validity | live | C | Construct: visual attention to the TV screen, defined as 'the amount of time that a viewer is looking at the TV screen during a piece of content'. Presence and co-viewing are measured separately. Observation: optical sensors in panel homes identify who is in the room and when they look at the TV. TVision's pages describe the classification only generically ('in-house developed and trained models', 'facial recognition'). The independent MSI paper describes the 2016-17 method: household members' faces enrolled at set-up, frames processed five to six times per second, and attention as the co-occurrence of an eyes-open detection and an eyes-on-screen inference from head orientation. The authors call it 'only one possible measure of attention'. Sampling: recruitment from top DMAs, compensated and opt-in, with daily in-tab quality checks and daily weighting to Census and ACS universe estimates on six household and five individual characteristics (vendor methodology page, apparently from March 2021). Panel size is stated by TVision itself: its 2026 TV Attention & Creative Impact Report (Methodology & Metrics, p. 21) gives 'an opt-in panel of 5,000 homes across the United States and weighted to US Census'. Trade reporting matches: about 5,000 US households in 25 DMAs and about 14,000 people (AdExchanger, November 2022, citing the CEO); 5,000 homes and 13,000 people (Next TV, April 2023); 5,000 households (PPC Land, 2026). Installed versus in-tab counts are not stated. The 2016-17 academic sample was 1,155 households and 3,659 viewers. Validation: no agreement or error statistic for eyes-on-screen or presence classification against a human-coded or eye-tracker gaze reference, on held-out data or otherwise, was found in the sources screened. No calibration by environment (linear versus CTV, or device) is documented. The CIMM and TVision active-versus-passive simulation used TVision's passive monitoring as the comparator for an active meter; it does not validate TVision's own classification. Measurement basis: direct observation of panellists, processed by a classifier. Population-level figures are weighted projections. [45, 291, 403, 52, 404, 298, 405, 406, 292] |
| Outcome validation | live | B | Memory: no study of TVision attention against recall or recognition with a published method was found. Brand: TVision and Upwave (vendor and partner release, 18 July 2022; funding not stated) ran a linear regression over 40+ US campaigns from Q4 2019 to Q1 2022. It reports that each percentage-point increase in Ad Attention (the share of ad impressions with two or more seconds of looking at the screen) for high-frequency ads went with 'close to a 1%' lift in Upwave aided awareness, with a 'similar' relation for ad recall. No sample of respondents, uncertainty, holdout or comparison with time in view is reported, and the report itself was not read. Search, a behavioural proxy: in the independent MSI working paper (the authors report no funding or conflicts; TVision provided the data), across 180 national ads in two NFL playoff games on 22 January 2017, eyes-on-screen attention explained the most variance in brand search lift in single-measure regressions (R-squared 0.116, against 0.082 for exposure tuning, 0.070 for presence and 0.035 for OTS tuning), with no holdout. Sales and conversions: Viant's CEO said on the 10 August 2026 call that over 80% of 42 pilot campaigns targeting TVision high-attention segments beat a benchmark conversion rate, with an average lift of 1.4x. No design, benchmark definition or cost is given (vendor assertion). This review found no replicated or out-of-sample validation, and no comparison of TVision attention with time in view or viewability on held-out data. [52, 74, 304] |
| Granularity and timing | live | C | The panel records second-by-second, person-level presence and attention for every program and ad a panellist's TV shows (direct observation, panel only). Outputs to buyers: planning dashboards by network, daypart, program, app and creative, with benchmarks and competitive comparisons (historical panel data, weighted); 'Get Mid-Campaign Indicators of Performance' and campaign reports of viewability, attention, in-view time and attention time (panel exposures, weighted); and CTV Campaign Measurement, which matches a buyer's log files or a TVision pixel to panel data, with presence, attention, co-viewing and demographics reported by app. Pre-bid: TVision's own Presence and Attention Scores are 'based on historic viewer attention' at daypart and channel level for DSP bidding. These are aggregates of past panel data, not impression-level scores. Impression-level pre-bid scores exist only in partner products: OpenX's 'predictive attention models' (TVision panel data combined with OpenX bidstream signals) and Viant's pre-bid attention intelligence (testing at 10 August 2026). Both are model estimates. TVision's pages do not label outputs by measurement basis. [45, 407, 408, 409, 311, 315, 309] |
| Activation | live | D | Named buying routes documented: (1) OpenX (SSP; curation in OpenXSelect): OpenX Attention Targeting launched 11 March 2026, 'Powered by TVision' and exclusive to OpenXSelect. Buyers apply predictive attention scores to curated CTV deals through self-serve or managed-service workflows (partner release). Its status after Viant's acquisition is not documented. (2) Viant DSP (the parent's platform, so one platform): Viant 'Began testing' TVision pre-bid attention intelligence (Q2 release, 10 August 2026). Separately, the Havas and Viant Attention Marketplace (24 June 2026) says inventory in Viant is 'scored by attention decile' for pre-bid targeting and in-flight optimisation for Havas Media Network North America clients (agency and DSP release). (3) Partner-mediated routes from before the acquisition: Adelaide pre-bid CTV segments using TVision data (January 2023); Lumen attentive PMPs on CTV activated in Nexxen with TVision data (August 2023). TVision's own Digital Activation page says 'Integrations with DSPs' allow bids to be adjusted with its scores but names no DSP. TVision audience segments through LiveRamp are viewing-behaviour segments, not attention. TVision's CTV Activation page offers platforms and publishers the option to 'Create tailor-made PMP', naming no platform. Historical: A+E optimised a guaranteed buy in flight in its own serving stack on TVision data (Q4 2020, Q1 2021). [315, 309, 310, 294, 289, 407, 410, 24] |
| Creative diagnostics | unconfirmed | E |   |
| Channel coverage | live | C | TVision measures only what plays on the panel household's TV screen. Of the seven scored channels, CTV is live: 'hundreds of apps including CTV walled gardens', including YouTube and Prime Video on the TV, per Viant's call. Linear TV is also live but is not one of the seven scored channels (recorded as an attribute). No display, online video on other devices, social (other than YouTube watched on a TV), audio, in-game or DOOH measurement is documented. Trade reporting in 2022 cites 900 streaming apps and publishers measured. [45, 297, 304, 404] |
| Data access and integrations | live | D | Named routes by which buyers get TVision output: (1) TVision's SaaS dashboards (Ad Scoreboard, TV Planning & Measurement, CTV Analytics, Linear TV Insights) with a client login; (2) data licensing and calibration feeds: 'data feeds so your engineering teams can develop and support their own solutions', with visible, attentive and co-viewing seconds, in three calibration output types (calculated probabilities, sessionized data, behavioural projections). No public API reference or developer documentation was found. (3) Summary presentations and 'detailed data for your own analysis' for CTV Campaign Measurement, which ingests a buyer's log files or a TVision pixel; (4) third-party products that embed TVision data: measurement and currency companies (VideoAmp, licence expanded 11 February 2026; iSpot), a verification suite (DoubleVerify, June 2023 announcement of a closed beta; no GA date found) and attention vendors (Adelaide, in AU per its 29 July 2026 blog; Lumen, 2023-24; xpln.ai, April 2026). Which of these third-party routes continue after the 1 May 2026 acquisition is not documented except by Adelaide's present-tense blog. [411, 297, 412, 408, 293, 220, 106, 290, 296] |
| Production scale | live | D | No impressions-per-year or campaigns-per-year figure for TVision's attention metric was found. Figures stated: '24000000000 Measured Seconds of TV Viewing Across Both Linear & CTV' (home-page counter, cumulative, no start date or period, in seconds rather than impressions); 'Millions of program and ad impressions' (data licensing page, no period); '71 Client Case Studies' (home page); 3,000+ TV advertisers in Ad Scoreboard (Lumen release, April 2023); 4,257,112 panel ad exposures of 3,659 viewers from July 2016 to June 2017 (MSI paper, panel exposures, not campaign impressions); 42 pilot campaigns in Viant (August 2026 call); and 40+ campaigns in the Upwave study (Q4 2019 to Q1 2022) and 50+ in the Wurl study (2025). Viant's CFO called TVision 'pretty de minimis from a revenue perspective'. [411, 297, 295, 52, 304, 74, 18] |
| US availability and coverage | live | C | Sold in the US, calibrated on US data and used in US production. The panel is US households ('recruits households from across the country'), and data are projected with US Census and ACS universe estimates. The 2026 creative report says data come from 'an opt-in panel across the United States and weighted to US Census'. US production use is documented: A+E, dentsu and Anheuser-Busch (Q4 2020); the Upwave study across 40+ campaigns (Q4 2019 to Q1 2022); Havas Media Network North America through Viant (June 2026); and US clients quoted on TVision's pages. International attributes (no points): UK panellists (panel privacy index) and TVision Japan and India links on the site. [45, 413, 24, 74, 310, 411] |
| Independent assurance | live | A | On 29 September 2026, MRC's accredited-services lists for television, digital and multimedia contain no TVision or Viant attention entry. The only 'Attention Metrics' accreditation listed is DoubleVerify's, for desktop, mobile web and mobile in-app. MRC's services-under-review lists for television and digital contain no TVision entry. No TVision or Viant statement of an MRC audit, or of conformance with the November 2025 IAB and MRC attention guidelines, was found in the sources screened. No accreditation and no confirmed audit is recorded. Absence is bounded to these lists and sources. [2, 302, 50, 48, 303] |
| Method disclosure | live | C | Against the seven elements, on TVision's own public pages: (1) construct: published (attention and presence definitions); (2) observations: published (optical sensors with facial recognition, microphone-array audio fingerprints, digital meter network-traffic analysis); (3) reference labels: not published; (4) panel sample and recruitment: recruitment and weighting are published (top DMAs, compensated opt-in, six household and five individual characteristics, Census and ACS universe estimates, daily weighting and in-tab rules), and panel size is stated in TVision's 2026 TV Attention & Creative Impact Report (5,000 US homes), though installed versus in-tab counts are not; (5) model type or scoring procedure: described only generically ('in-house developed and trained models'), with the head-orientation method described in the independent 2021 paper; (6) validation statistics: not published; (7) update cadence: daily in-tab qualification and daily weighting are published, while model recalibration cadence is not. The methodology page ends: 'For more detailed information about TVision's methodology, please contact us'. Tally: fully published, elements 1, 2 and 4; partly published, elements 5 (generic only) and 7 (weighting cadence, not model recalibration); not published, elements 3 and 6. No data are released for replication, although panel data have been provided to academics (MSI 2021; the JARxTVision Data Release behind 2025 JAR papers). [45, 291, 403, 52, 301, 292] |
| Privacy posture | live | C | Consent model, product-specific and documented: the panel is '100% opt-in' and compensated. The US panellist privacy policy (effective 18 December 2018, updated 9 September 2024) says that by using the System participants consent to its information practices, including for their children's information, and it covers information on guests in the System's vicinity. It lists collection of 'photos, images, and face geometry (a biometric identifier used for facial recognition)', 'periodic snapshot images of the webcam's entire field of view', colour and depth sensor data, household IP address and device information. It permits sharing personal information with clients and partners, 'in some cases in hashed form that may be re-identified', including for audience segments and lookalike models. It states that face geometries are destroyed within three years of the last interaction, that media-consumption data are kept for one year after participation ends, and that consent to third-party sharing can be withdrawn. Separate COPPA and UK panellist policies exist. The policy states that TVision is not subject to the CCPA or VCDPA. Data minimisation claims on marketing and method pages: 'No video streams are stored on the device or transmitted back to TVision'; 'TVision policy prohibits any audio and images from leaving the home once a household is InTab'; after 'an initial training period all data leaving the panelists' home is completely anonymous'. These sit in tension with the policy's snapshot collection and re-identifiable sharing clauses. No independent certification or audit of the panel's privacy practices was found. No regulator action or lawsuit naming TVision on privacy grounds was found in the sources screened (this paper's bounded searches, verified). [306, 414, 415, 45, 291] |

Performance evidence: Eyes-on-screen attention explained more variance in brand search lift (R-squared 0.116) than exposure tuning (0.082), presence (0.070) or OTS tuning (0.035) across 180 national ads in two NFL playoff games on 22 January 2017; attention was significant with quarter-hour fixed effects (coefficient 0.106). (independent measurement, sponsor None: the authors report no funding or conflicts of interest; TVision provided the data and feedback) [52]; For every percentage-point increase in Ad Attention to high-frequency ads, there is 'close to a 1%' lift in aided awareness (TVision and Upwave). (vendor assertion, sponsor TVision and Upwave (preferred partners); funding not stated) [74]; CTV ads emotionally aligned with surrounding content captured '2.1 times' more attention: attentive time of 71% versus 33% of the ad duration (as reported); the release's own key finding is 66% versus 27% (2.4x) for 30-second ads. (vendor assertion, sponsor Wurl (then owned by AppLovin, which sells the contextual targeting tested) with TVision as the attention data supplier) [17, 18, 16]; Over 80% of 42 Viant pilot campaigns targeting TVision high-attention inventory segments achieved higher conversion rates than the benchmark, with an average lift of 1.4x (vertical examples of 14x, 3.7x and 3.1x). (vendor assertion, sponsor Viant (owner of TVision)) [304]; OpenX's release subtitle claims that the TVision-powered integration improves 'brand recall and message retention'. (vendor assertion, sponsor OpenX and TVision) [315]

### Vevo (active, private)

*Vevo LLC (incorporated in the United States, per Universal Music Group's 2025 annual report; the site footer reads '© 2026 Vevo LLC')*

The media owner in the trading table: an attention guarantee with no public threshold or remedy.

**SWOT (2026-09-29).** Vevo is a music video media owner that announced an Adelaide-backed attention guarantee across screens in January 2026, but has published no terms, adoption or audit for it.

- **Strengths.** Attention Guaranteed is the only guarantee in the register stated as launched that names CTV, mobile and desktop in one offer. (vendor claim) [78, 94, 93, 97, 91, 95] The offer sits inside Vevo Evolve, bundled with targeting, reach planning, real-time dashboards and one cross-screen performance report. (vendor claim) [78, 317] Vevo states 154 million US monthly viewers and 22 billion global monthly content views; method and period are not stated. (vendor claim) [318] A Vevo-released Amplified Intelligence webcam study of 150 households observed Vevo at 13.1 against 9.4 seconds of active attention among 18-34s versus mid- and low-tier channels. (vendor claim) [319, 320, 321] Vevo runs separate partners for outcomes and reach: DISQO as preferred US brand lift partner since May 2026, AudienceProject for UK reach since February 2025. (vendor claim) [322, 323]
- **Weaknesses.** No threshold, remedy, audit, price or market list is published for Attention Guaranteed. (vendor claim) [78, 324] The release never names the unit of the guaranteed minimum; that it is Adelaide AU is a strong inference, not a stated term. (inference) [78] Adelaide AU is not MRC-accredited: MRC lists it as pre-audit reviewed May 2026, audit not begun. (record) [48] The only attention figure tied to the offer, AU 28% above undefined Adelaide benchmarks, covers YouTube CTV only, not mobile, desktop or FAST. (vendor claim) [78] No advertiser, campaign, impression count or spend under the guarantee is disclosed, and no product documentation page for the offer was found in the pages screened. (vendor documentation) [78, 324, 318]
- **Opportunities.** Kargo published a minimum average AU of 25 and Kargo and the WSJ published added-value remedies; Vevo could publish comparable terms to make its guarantee readable. (inference) [93, 91, 92] If Adelaide's MRC audit proceeds past the May 2026 pre-audit review, the metric behind Vevo's offer could gain an accreditor record. (inference) [48, 66] The May 2026 DISQO partnership gives Vevo a route to publish analysis linking attention on its inventory to brand lift; none has been published. (inference) [322] The Guardian launched AU-based attention packages in June 2026, widening the pool of buyers already trading on AU-based offers. (inference) [325]
- **Threats.** A guarantee announced without published terms is exposed if buyers read it as a contract, since peer guarantees publish remedies. (inference) [78, 93, 91] The offer was last marketed in the April 2026 upfront release and was absent from the September 2026 advertiser page; the cause is unknown. (inference) [324, 318] Adelaide sells AU and is the implied arbiter of guarantee delivery, while its MRC audit has not begun. (inference) [78, 48] Since shutting its own apps and website in 2018, Vevo depends on YouTube; MRC lists Google's YouTube Relay Log as pre-audit reviewed March 2026, audit not begun. (inference) [326, 48]

- **What is sold.** Vevo sells advertising on its music video network across YouTube (including YouTube on CTV), its FAST channels on CTV platforms, and mobile and desktop. Its buying routes include Vevo Evolve targeted media (advanced targeting), programmatic (PMP, programmatic guaranteed, YouTube/GAM via DV360, and YouTube Instant Reserve), sponsorships (Artist Franchise), inclusive networks (Somos Vevo, Vevo BLK), and originals and custom content. 'Attention Guaranteed' is an offer inside Vevo Evolve. Its 27 January 2026 release, announcing a partnership with Adelaide, says it 'secures a minimum attention score' for a campaign across all screens, with one performance report. The release does not name the unit of the minimum; reading it as Adelaide's AU is this paper's inference. It is a media offer. Vevo does not sell attention measurement, and no product documentation page for the offer was found; it appears only in press releases.
- **Who buys it.** Brands and agencies buying premium music video and CTV inventory, direct or through upfront deals and programmatic routes. The April 2025 Evolve release offers 'advanced attention tracking, optimization, and guarantees' to 'top advertisers'. No advertiser, agency or campaign bought under Attention Guaranteed is named in any source opened.
- **How it earns.** Advertising: Vevo describes itself as '100% Ad-Supported' (May 2025 event release). Revenue, pricing and rate cards are not disclosed. The pricing basis for Attention Guaranteed is not stated, and this review found no attention-priced unit from Vevo.
- **Supply and data dependencies.** Distribution: YouTube (including Instant Reserve and DV360 buying that counts toward YouTube commitments) and 35+ CTV partners, including FAST platforms such as Roku, Samsung TV Plus, Pluto TV and Sling Freestream. Measurement: Adelaide for attention (AU; the offer's metric); DISQO as preferred US brand lift and outcomes partner (named 20 May 2026); iSpot, cited in the April 2026 upfront release; AudienceProject for UK cross-media reach (February 2025). Vevo's own attention research used Amplified Intelligence (UK webcam panel) and Flood + Partners (survey, with Channel 4).
- **Differentiation.** Synthesis: among the guarantee and attention-priced offers in chapter 6's trading table, Vevo's is the only one stated as launched (not pilot or beta) that names CTV, mobile and desktop in one offer. Teads' pilot and beta use AU 'across screens'; TripleLift started with CTV; Duolingo is in-app; Impact Media lists CTV only as planned; the WSJ, Kargo, SeenThis, A+E and Adlook cases do not state screens (ATT-VEVO-S01, S20, S21, S22, S23, S25; ATT-NFA-S01, S03; ATT-F-S25; ATT-U-S36). Bundled with Vevo Evolve targeting, reach planning and real-time dashboards, with a single cross-screen performance report (ATT-VEVO-S01, S02). Vevo also holds a direct-observation attention study of its own YouTube-on-TV inventory (Amplified Intelligence, UK). It is a different vendor, construct and unit from the AU guarantee and must not be read as corroborating it (ATT-VEVO-S10, S11).
- **Risks.** Inference: a guarantee named as such but without published terms is exposed if buyers read it as a contract; the WSJ and Kargo precedents published a remedy (added-value impressions) and Vevo has not. The metric is not accredited, and Adelaide's MRC audit had not begun by the time this review read the record (29 September 2026) (ATT-VEVO-S16). Inference, cause unknown: the offer's last documented mention is the April 2026 upfront release, and it is absent from the September 2026 advertiser page. This may reflect de-emphasis or simply page design; no source establishes a retirement. Inference: dependence on YouTube for distribution and buying routes (Instant Reserve, DV360) may constrain third-party measurement on that inventory. No source opened says how AU is captured there.
- **Best fit.** Buyers already using Adelaide AU as a planning or media-quality metric who want one cross-screen AU report for a Vevo buy. Upfront or direct buyers able to negotiate the threshold and remedy in their own contract.
- **Poor fit.** Buyers who need a published threshold, remedy and independent audit before treating attention as a contractual term. Buyers who require MRC-accredited attention metrics. Buyers comparing Vevo's attention with other vendors' units, which are not comparable with AU. Buyers outside the US: no market list is given.
- **Limits.** No threshold, remedy, audit, price or geography is published for Attention Guaranteed (ATT-VEVO-S01, S03). The metric's vendor, Adelaide, both sells AU and, by implication, measures compliance. AU is not MRC-accredited; the pre-audit was reviewed in May 2026 and the audit has not begun (ATT-VEVO-S16). The only attention figure tied to the offer (+28%) covers YouTube CTV, not the mobile, desktop or FAST inventory the guarantee also spans (ATT-VEVO-S01). No advertiser, campaign, impression count or spend under the guarantee is disclosed. No product documentation page exists in the pages screened: the offer appears only in press releases and was absent from the For Advertisers page on 29 September 2026 (ATT-VEVO-S05). Vevo's own attention research is vendor-released: UK-only (Amplified Intelligence, Flood + Partners), self-report (Flood, 'Then is Now'), and with no outcome link beyond survey measures.

| Dimension | Availability | Evidence | Observed |
|---|---|---|---|
| Method validity | unconfirmed | E |   |
| Outcome validation | unconfirmed | E |   |
| Granularity and timing | unconfirmed | E |   |
| Activation | unconfirmed | E |   |
| Creative diagnostics | unconfirmed | E |   |
| Channel coverage | unconfirmed | E |   |
| Data access and integrations | unconfirmed | E |   |
| Production scale | unconfirmed | E |   |
| US availability and coverage | unconfirmed | E |   |
| Independent assurance | unconfirmed | E |   |
| Method disclosure | unconfirmed | E |   |
| Privacy posture | unconfirmed | E |   |

Performance evidence: Vevo's YouTube CTV inventory 'delivers attention scores 28% higher than average benchmarks', according to Adelaide. (vendor assertion, sponsor Vevo (seller release); measurement by Adelaide, the metric's vendor; who paid for the measurement is not stated) [78]; Amplified Intelligence found premium YouTube channels drive 40% more active attention than non-premium, and Vevo drives 38% more active attention among 18-34s. Premium channels averaged 17 seconds of active attention against a UK benchmark of 14 seconds. (vendor assertion, sponsor Vevo (released the study; egta says Vevo 'partnered with' Amplified Intelligence). Funding is not stated, so independence is unknown, and the study is vendor-affiliated.) [319, 320]; In a 2023 survey, respondents who liked or loved 'fit for TV' content reported higher attention paid to the ad, 'more than 2x that of non-premium'. (commissioned measurement, sponsor Vevo and Channel 4; egta states Vevo 'commissioned Flood & Partners') [416, 319]; In Vevo's 'Then is Now' survey, 67% of respondents said nostalgic messaging makes them more likely to pay attention to the brand. (vendor assertion, sponsor Vevo; the fieldwork provider is not named in the release) [417]



## 12. What a buyer should ask


*No attention metric sold in the US meets the conditions a buyer would need before paying on it or accepting it as a guarantee. The one accredited metric carries no guarantee or billing use, and none of the guaranteed or priced metrics is documented as accredited. Where a guarantee says who measures delivery, it is the metric's own vendor; the others do not say, and no independent audit of delivery was found. On Q1's registered rule, no metric-environment pair codes as a currency or partial currency at the cutoff. Each codes as diagnostic or unresolved, and a diagnostic code marks missing assurance, not an absence of billing. Seven conditions, most of them contract terms, would change that.*

> **In short.**
> On 28 September 2026 the Media Rating Council (MRC) listed one accredited attention product, DoubleVerify's, for desktop, mobile web and mobile in-app, and it carries no guarantee or billing use [2] [63, 222]. Among the guarantees, one publishes a numeric threshold, a minimum average of 25 AU [93, 91, 94, 97, 95]; Impact Media's billing suite publishes seconds seen instead [101]. Where a guarantee says who measures delivery, it is the metric's own vendor; the others do not say, and this review found no independent audit of delivery [94, 93, 91, 66]. No public study screened compares an attention metric with time in view on held-out data [46, 51, 52] [8, 10, 71, 64], and none discloses measurement fees or the media premium paid [69, 13, 67, 15]. No independent study screened reports how a commercial model performed after an update [52, 51, 46]. Three findings would change the verdict: an independent, replicated study showing on held-out data that attention predicts sales beyond time in view; an accredited metric used as a billing unit or guarantee with stated remedies, at scale; and documented production use of a named vendor's attention signal in agentic buying, where only test-mode handling with a placeholder vendor is documented [142, 141, 151, 33].

### Seven conditions before attention carries money

A billing unit or a guarantee turns a measurement into money. Q1's registered rule codes a metric-environment pair as a currency only when four conditions are documented: accreditation for the environment traded, a guarantee or billing use with stated threshold and remedy, independent audit, and disclosed adoption. On the evidence in [chapter 6](#trading), no pair codes as a currency or partial currency; pairs code as diagnostic or, where the terms cannot be read, unresolved [63, 222]. Adelaide's AU and Lumen's metric, which underpin most of the guarantees and priced offers found, held no MRC accreditation on 29 September 2026 [94, 93, 91, 66] [48, 78] [2, 48, 3]. A diagnostic code records missing assurance, not the absence of a billing function: SeenThis prices media on Lumen's measure, and that pair codes as diagnostic [99, 100].

The seven conditions below keep three of Q1's four and add four that any single buy needs. Adoption is a market test and returns in the last section. One question comes before all seven: is the number an observation of people or a model estimate? Direct observation appears only in panel work, and the attention figures behind Adelaide, Lumen and IAS campaign results are model estimates [10, 63, 73, 74, 9].

### Accreditation for the environment bought

Accreditation is scoped by product, metric, environment and stage, and the vendor's page is not the record. On 28 September 2026 MRC listed one entry under "Attention Metrics": DoubleVerify Quality Analytics Platform, for desktop, mobile web and mobile in-app [2]. Connected TV (CTV) attention is not accredited, and no pre-bid attention offering is [42, 205]. MRC denied accreditation to pre-bid Authentic Attention in March 2024 and gave no reason [210]. DoubleVerify's own pages describe its scope differently from MRC's record [209, 214, 47, 215].

XPLN's audit, in process since November 2025, covers mobile web and mobile in-app only [48]. Adelaide's AU had a pre-audit review in May 2026, and on 29 September 2026 its audit had not begun [48]. MRC removed IAS's attention measurement from in-process status in January 2026 [49]. On 29 September 2026 IAS was accredited for viewability and invalid-traffic filtering, among other metrics, and for no attention metric [2, 205]. Accredited inputs do not accredit the score built on them.

Ask for MRC's letter to the vendor and match it to the buy on four points: the metric, the environment, the stage (post-bid reporting or pre-bid segment) and the standards listed. The standards MRC lists for DoubleVerify's April 2026 accreditation do not include the Interactive Advertising Bureau (IAB) and MRC Attention Measurement Guidelines [42, 41], which is why H3 codes as contradicted, on a narrow basis: whether the guidelines informed the audit is not public [42, 2, 425, 563]. The condition is met when the letter names the metric, the environment and the stage of the buy.

### A published threshold and a remedy

Among the named guarantees, only the 2022 Kargo campaign for MediaCom (now EssenceMediacom) publishes a numeric threshold: a minimum average of 25 AU, Adelaide's attention unit. Among the AU guarantees, only Kargo and The Wall Street Journal publish a remedy, and Vevo publishes neither [93, 91, 94, 97, 95]. SeenThis agrees a threshold before launch without publishing the values, and promises extra delivery at no cost if it is missed [99, 100]. No level or remedy is public for A+E's 2020 and 2021 guarantee to dentsu for Budweiser, built on TVision data [24]. Every published remedy is more media [93, 91, 94, 97, 95] [99, 100].

A score is not yet a threshold. DoubleVerify's indices are normalised to 100 across its own impressions over a rolling 28-day window [71, 211, 370], so, by this paper's inference, a floor stated as an index moves with the benchmark. Adelaide's tiers in The Trade Desk are relative ranks, according to AdExchanger [161], and IAS's cut-off between high and low attention varies by analysis [11, 228, 235]. A worked example in AdCP v3.1.24, the agentic buying protocol maintained by AgenticAdvertising.org, sets an Adelaide threshold of 65 against a schema that defines the threshold as a rate between 0 and 1 [561, 145, 144].

Ask for the number, its unit and its model version; whether it applies per impression, to a campaign average or to a share of impressions; the denominator; whether the base is fixed or relative; the measurement window; and the remedy, in media or money, with a delivery date. The condition is met when a third party could read the contract and tell whether the threshold was hit.

### An audit of delivery by someone other than the metric vendor

Where a guarantee says who measures delivery, it is the vendor whose metric defines it: Adelaide for the Journal, Kargo and TripleLift, and Lumen for SeenThis. Vevo, Duolingo, Teads and A+E do not state who measures, and this review found no independent audit of guarantee delivery [94, 93, 91, 66]. The guidelines strongly encourage independent third-party auditing of attention methods used in buying and selling, without requiring it [41, 425]. The third-party reviews found were requested by the vendor. MediaSense rebuilt AU-to-outcome curves from sampled data Adelaide supplied, without verifying the data or reviewing source code [66, 170]. PwC says of its review for Lumen that it "has not independently verified, validated, or audited the data" [20, 8]. Some sellers fund the measurement: Adelaide says Condé Nast and Hearst cover the cost of AU on their inventory [167].

Inputs matter as much as scores. The guidelines require full filtering of sophisticated invalid traffic (SIVT), and base inputs audited for accreditation [41], and no Adelaide viewability or invalid-traffic input appeared on MRC's accredited digital list on 29 September 2026 [2]. IAS's own analysis found that Papyrus fraud traffic had about 13% higher attention scores than other traffic before IAS filtered it as invalid [240].

Ask for an audit right; impression-level delivery logs carrying the score and the model version; the invalid-traffic filtering applied before scoring, and whether MRC accredits it; and who pays the measurement vendor. The condition is met when a party independent of both seller and metric vendor can recompute delivery.

### Held-out evidence against time in view

A guarantee on a number that does not track the outcome moves risk onto the wrong number. The buyer needs H1a's test: a gain over time in view on data the model has not seen. In the sources screened, no study makes that comparison for memory, brand or sales [46, 51, 52]. Vendor and commissioned comparisons are in-sample and mostly against viewability or completion rate, and several vendor models take viewable time as an input [8, 10, 71, 64]. The one commissioned review with published statistics, PwC for Lumen, found attention not significantly better than viewability on two awareness measures or digital ad recall [8]. By the vendors' own accounts, AU is trained on outcome data and IAS's model on conversion events [168] [237, 236, 115]. Inference: a vendor chart showing that high scores go with better outcomes may partly restate what the model was trained to predict.

H1b's buying test is missing too. This review found no randomised or holdout test of buying on attention against a standard-metric arm that reports a memory, brand or sales gain with its cost [51, 4, 5] [69, 65, 11, 10].

Ask the vendor for its metric against time in view on the same impressions, scored on held-out campaigns and a later period, for the outcome bought, measured by someone other than the vendor. If none exists, run the randomised split with a holdout that [chapter 3](#validation) sets out. The condition is met by a held-out gain over time in view for the outcome that matters and, before a premium is paid, a cost per incremental outcome at or below the standard arm's.

### Cost and premium, disclosed

No study screened discloses measurement fees, and none reports the media premium a campaign paid for attention-bought inventory. The only premium figure is an estimate: the report by Newsworks, the UK national newsbrands' marketing body, and Lumen puts UK news-brand display at a 24% premium in cost per thousand impressions (CPM) over general display [23] [69, 13, 67, 15]. The fees exist. Google's Display & Video 360 documentation lists an additional media fee for its "Maximize Adelaide AU" algorithm [330, 159]. DoubleVerify's average fee per thousand media transactions measured was $0.07 in 2025, across all products, and its filings give no attention-specific price [47, 206, 207] [47, 207]. A higher score at a lower price is no saving until an outcome is measured. Lumen's consumer-goods case reports 104% more viewed ads at 40% lower CPM than a control on the same budget, and its outcomes were Lumen's own attention estimates [69].

Ask for the measurement fee per thousand impressions and who pays it; any data or algorithm fee in the buying platform; the CPM of attention-selected inventory against the same seller's standard inventory over the same dates; and cost per incremental outcome for both. The condition is met when the first three are written down before the buy and cost per incremental outcome is reported for both arms after it.

### Model versions and drift

A guarantee settles on a model's output, and models change. No independent study screened reports a commercial attention model's performance after an update, calibration drift or transfer to new channels [52, 51, 46]. The guidelines say model validation should run at least annually on fresh deterministic data [41]. Adelaide's CEO said in October 2024 that the seventh version of AU was under MRC scrutiny; which version is under review in 2026 is not stated [339, 48]. MRC's July 2026 report labels DoubleVerify's newly accredited reporting "Authentic Attention V4" [42, 205]. In the Adelaide documentation screened, no held-out agreement statistic between AU and gaze reference data is published [66, 168], and Lumen's panel page, published in June 2021, carries no later revision date [271, 392, 115, 287].

Ask for the model version on every report and every settlement; a change log; notice before any model change during a flight, with the settling version fixed in the contract; the validation statistic on fresh reference data after each update; and, for indexed or relative scores, the base and how often it resets. The condition is met when the settling version is named and was validated after its last update.

### Privacy posture for the signals used

Adelaide's AU Ratings policy treats what its tracking technology collects as personal information, and DoubleVerify's notice lists the viewport, focus, time-in-view, scroll and interaction signals its tags collect [340, 372]. Reading the European Data Protection Board's guidelines, this review infers that such tag signals collected in the EU need consent unless an exemption applies [126]. California's privacy regulations, effective 1 January 2026, require a written risk assessment before a business sells or shares personal information. Processing already under way before that date must be assessed by 31 December 2027 (section 7155(b)), and whether a given attention vendor is a covered business was not established [131, 528].

Panels carry more. TVision's US panellist policy lists face geometry, described as a biometric identifier, and periodic snapshot images of the webcam's field of view [306, 45, 291]. Illinois' Biometric Information Privacy Act requires written notice and a written release before a scan of face geometry is collected, with a private right of action of $1,000 or $5,000 per violation [128]. Amplified's TV camera can capture guests [184], and this review found no Lumen panel privacy notice in the sources screened [393, 401]. In this paper's reading of the European Commission's guidelines, gaze or facial data used to infer emotion, interest or intent fall inside the AI Act's emotion-recognition definition [30, 124], and from 2 August 2026 EU deployers of such systems must inform the people exposed [30, 123]. In the searches run, this review found no privacy or biometric-data regulator action or lawsuit naming an attention-measurement vendor; the searches named Realeyes, Affectiva, TVision, Tobii and iMotions, not Adelaide, Lumen, Amplified, DoubleVerify or IAS [507, 136]. Nielsen's patent actions against TVision are a separate matter [82].

Ask which signals the metric uses and whether any can identify a person; whether any output infers emotion or intent; the panellists' consent text, where their images are processed and how long they are kept; and, for California data, whether the vendor is a covered business and whether its risk assessment exists. Ask which deadline applies before treating a missing assessment as a failure: new processing needs the assessment first, and processing already under way has until 31 December 2027. The condition is met when consent and data minimisation are documented for both the panel and the tag.

The decision tool sets out each check with the test that settles it and what it cannot tell you.

| Buyer question | Test that fits | What the inference needs | Where it fails | What it cannot tell you |
|---|---|---|---|---|
| Is this attention metric accredited for the environment I am buying? | Read the accreditor's own record (MRC's accredited list and its letter to the vendor) for the product, the metric and the environment. A vendor page or an announced audit does not count. | Accreditation is scoped to product, metric and environment, and the standards the accreditor lists are the ones applied. | When a vendor page states a broader or narrower scope than the accreditor's record, or when the buy is in an environment the record excludes. CTV attention holds no accreditation in the records screened. | Tell you that the metric predicts memory, brand or sales outcomes. Accreditation certifies process and disclosure. [42, 2, 48] |
| Is the number an observation of people or a model estimate? | Ask for the metric card: the construct, the observation used, the reference labels the model was trained on, the sample, the validation statistic on held-out reference data, and the application stage. | The vendor will state its measurement basis for each number it reports; the IAB and MRC guidelines ask that observed measurement be distinguished from modelled. | When a campaign report labels a model output as measured attention, or when reference labels, sample and validation are not disclosed. | Settle whether the model transfers to your channel, audience or creative, or whether it held after the last model update. [41, 43, 44] |
| Does it predict my outcome better than time in view? | A comparison on held-out data, randomly or temporally split, of the attention metric against viewability and time in view, for the outcome you buy for, with the outcome measured by a party other than the metric vendor. | The outcome is measured independently, the held-out data were not used to build the model, and the comparison uses each measure's scale fairly (per standard deviation, not per second). | In-sample correlations, vendor-published case studies and comparisons against viewability rates alone. In the sources screened, no public study made the held-out comparison. | Show that buying on the metric pays back once media premiums and fees are counted. [46, 52, 51, 8] |
| Does buying on attention pay back? | A randomised or geographic holdout test: an arm bought on the attention metric against an arm bought on your standard metrics, with outcome lift and cost per incremental outcome reported for both, media premiums and measurement fees included. | The arms differ only in the buying rule, the test is powered for the outcome, and costs are counted in full. | Two-arm case studies without randomisation, tests that report lift without cost, and comparisons of high- and low-attention impressions inside one campaign. | Generalise beyond the channels, formats and categories tested. [4, 8] |
| Is an attention guarantee enforceable? | Read the contract for four terms: the numeric threshold, the remedy if it is missed, who measures delivery, and whether an independent party audits it. | The public offer and the contract say the same thing, and the buyer can see the delivery data. | When the metric's own vendor measures delivery and no audit right exists. Where a guarantee names who measures delivery, it is the metric's own vendor; Vevo, Duolingo, Teads and A+E do not say. Every published remedy is extra delivery. | Tell you whether the guaranteed attention changed an outcome. [93, 91, 78, 100, 94, 97, 95, 24] |
| Will the signal work when agents buy media? | Record which of the five protocol states the attention signal reaches in the protocol and version your buying agent uses: schema representation, implemented connector, interoperability test, production use, transaction. | The protocol version and the vendor's connector are the ones in production, not a release candidate or a test fixture. | When only schema representation or a test with a placeholder vendor exists. In AdCP 3.1.24, no named vendor's attention signal is documented beyond schema representation; the only connector evidence found is a conformance test with a placeholder vendor and simulated delivery. | Show that an agent weights attention sensibly, or that attention still means what it did when a person made the decision. [138, 33, 141] |
| Is the data behind the metric collected lawfully where I buy? | Ask whether the signals can identify a person and whether they are used to infer emotion, and read the vendor's consent model and minimisation for panels and device signals against the rules where the ads run. | The vendor documents what it collects, and panels are opt-in. | When facial or gaze data can identify people (a trigger under several US biometric laws) or feed emotion inference (high-risk under the EU AI Act for advertising). | Replace legal advice on a specific deployment. [123, 124, 418] |

### What would change this verdict

Three findings would change it.

An independent, replicated study showing, on held-out data, that a vendor's attention metric predicts sales beyond time in view. The nearest sales evidence is IAS's Kimberly-Clark case, two NCSolutions tests with impressions split after the fact. The first found no significant difference in sales lift; the second, which added formats in another quarter, found 313% higher sales lift in above-average-attention households [11]. No independent replication of any vendor claim was found [8, 71, 66]. TVision panel data have been released to independent academics, and Crossref records two 2025 Journal of Advertising Research articles on attention in CTV and linear TV, which this paper did not read [300, 301]. Such a study would move H1a for sales only.

An accredited attention metric used as a billing unit or guarantee with stated remedies, at scale. MRC could accredit AU, whose audit had not begun on 29 September 2026 [48]; that would lift the Journal and Kargo pairs, which carry published remedies, to partial currency ([chapter 6](#trading)). Or DoubleVerify could attach a guarantee or a price to its accredited metrics. Scale also needs disclosed US spend, impressions, campaigns or buyers against a denominator, which no offer found discloses [91, 92, 334, 93, 25].

Attention signals used in production agentic buying. AdCP v3.1.24 can represent an attention value, including a floor paired with a makegood, as schema only [144, 138, 140, 142]. In the registries, code search and media-vendor newsrooms screened, this review found no public documentation of a Model Context Protocol (MCP) server or other agent-facing API that returns attention metrics from an attention vendor [157, 141]. Viomba, Adverteyes and Amplified advertise MCP access to attention predictions, and none documents production use [562, 344]. DoubleVerify's account of an AdCP buy for Butler/Till names suitability, fraud, viewability and contextual controls, and does not mention attention [148]. Beyond the schema, the only handling documented is in test mode, where sales agents handled a placeholder vendor's metric with simulated delivery, and no production use of a named vendor's attention signal is documented. H5 moves when a named vendor's attention signal reaches documented production use or a transaction [142, 141, 151, 33]. The author's interest in AdCP is declared in the abstract and in full in [chapter 13](#method).

The codes below are dated to screening on 28 and 29 September 2026. The H3 and H5 codes are status-sensitive and were not rechecked before the 1 October 2026 cutoff [42, 2, 425, 563] [142, 141, 151, 33].

| Test | Code at the cutoff | What would move it |
|---|---|---|
| H1a, predictive value | Unresolved for memory, brand and sales [46, 51, 52] [8, 10, 71, 64] | A held-out gain over time in view, per outcome |
| H1b, optimisation value | Unresolved [51, 4, 5] [69, 65, 11, 10] | A randomised or holdout buying test reporting cost per incremental outcome |
| H2, trading use | Supported; scale undocumented [91, 92, 334, 93, 25] | Disclosed US spend, impressions, campaigns or buyers, with denominator and period |
| H3, accreditation | Contradicted, on a narrow basis [42, 2, 425, 563] | An MRC record accrediting an attention metric that lists the Attention Measurement Guidelines |
| H4, bundling | Unresolved [47, 231, 285, 232] | A specialist's revenue or pricing series, or buyers documented switching because of a bundle |
| H5, agentic use | Unresolved: no documented production use of a named vendor's signal; only test-mode handling with a placeholder vendor [142, 141, 151, 33] | Documented production use or a transaction |
| Q1, currency or diagnostic | No pair codes as currency or partial currency [63, 222] | Accreditation plus a guarantee or billing use (partial currency); all four conditions (currency) |



## 13. Method, limits and corrections


*How the paper was built, what the checks caught and what could not be reached. It is published so a reader can check the paper rather than trust it.*

> **In short.**
> AI research agents worked from one brief and one set of rules. They produced 9 research workstreams, 9 vendor-universe and product sweeps and 8 company profiles. A separate AI verifier, told to refute, checked each file. The research opened and logged 985 sources and put 584 claims in the ledger, every one of them checked by a verifier or a review. Verifiers made 1520 checks of claims, capability records, profile fields and table rows, counting both passes where a file was verified twice: 800 found the item supported as stated, 658 supported it with a correction and 62 contradicted it. Another AI model reviewed the briefs, then the evidence, then the full draft, and two internal passes read the full draft. Those reviews and a consistency audit of the scores produced the 260 entries in the corrections ledger, 23 of them scoring corrections. 20 products were scored on 12 dimensions, one of which does not apply to each cohort. This is AI-assisted research with a documented process. It is not an external audit, not peer review and not independently validated.

### What was fixed before the research

Three registers were written on 28 September 2026, before any research agent ran. The hypothesis register records 10 entries: six hypotheses (H1a, H1b, H2, H3, H4, H5) and four coded questions (Q1 to Q4). Each hypothesis states its population, comparator and observable outcome, and the rules for supported, contradicted and unresolved. Absence of accessible evidence yields unresolved, never contradicted. The product register lists every product scored, listed or treated as historical, with the entity that sells it and the reason. It now holds 44 records, of which 20 are listed and 4 historical. The scoring rules fixed the two cohorts, their applicability masks and the anchors for every dimension, with numeric thresholds where a scale is measured. The registers are published in the downloads.

The product register changed twice after the research and before scoring, both times from verified evidence: splits, moves and additions from the vendor-universe sweep and workstream E (for example, Amplified's planning and measurement modules became two rows, GumGum Attentive Lift moved to the outcome-comparator table as a lift product, and xpln.ai, Kantar LINK AI and Dragonfly AI were added as scored products), and additions from the second outside review (SeenThis Attention and Impact Media's suite, both listed, not scored). The anchors did not change.

### How the evidence was gathered

Each workstream agent read the shared rules, the brief and the registers, and wrote one structured file: sources with their publication, event and access dates, geography, period, funding and access limits; findings with the supporting passage and locator, evidence class, confidence, limitations and counterevidence; and data tables in which every row names its source, period and claim status. The rules require every cited source to be opened; a search snippet is a lead, not support. Where a page refused an automated fetch, agents retried with a browser user agent and recorded the route. 82 registered sources were only partly readable (paywalled, gated, abstract only or an access error), and claims resting on them say so.

Evidence classes follow the shared rules: direct record, independent measurement, commissioned measurement, vendor assertion, synthesis, inference, forecast and illustrative scenario. A vendor's own documentation is a vendor assertion, however technical. A third party paid by a vendor is not independent, and a study whose funding is unknown is not treated as independent. The second outside review applied that rule to the ARF's Phase 3 release, which states no funding; its finding was recoded from independent measurement to synthesis ([chapter 3](#validation)). The author's own work is labelled as author's work and never counts as corroboration: 2 registered source records are the author's, one essay registered under two URLs.

Every absence claim is bounded to what was searched, and the searches are logged in each file.

### How it was checked

Every workstream, product sweep and profile went to a separate verifier agent that had not seen the research process and was told to refute it. Verifiers opened the cited sources themselves and returned one of four verdicts per item: supported as stated, supported with a qualification (with the corrected wording), contradicted (with what the source says) or unresolved. Eight of the workstreams were verified twice, by two independent verifier passes, because a verification run was restarted after a usage limit; both records are kept, and the ledger takes the more severe verdict for any claim checked twice. A claim's final wording is its verifier's or a review's correction where one exists. All 584 claims were checked by a verifier or a review: 315 as stated, 254 with a qualification and 15 contradicted. Contradicted claims appear only in their corrected wording.

![Every stage was built to remove things, and the counts are computed, not typed.](figures/fig-pipeline.svg)

*Figure fig-pipeline. **Every stage was built to remove things, and the counts are computed, not typed.** The five stages of this project with the actual outputs of each. Every claim in the ledger was checked by a verifier or a review; eight workstreams were verified twice and the more severe verdict kept. Corrections include review findings and score changes from the consistency audit. Units: counts. Scope: this research project, as of the cutoff. Caveat: Counts describe process volume, not evidence quality; a claim that survived verification is supported by its source, not proven true. Data: data/stats.json*

There were six review passes. Before research began, the author ran an outside review of the briefs with another AI model; every factual finding was checked against its primary source before any change, and the dispositions are kept with the research record for the three papers in the series. After the workstreams, the vendor-universe sweep and the profiles were verified, and before the product sweeps were run, a second outside review spot-checked that evidence. It returned six findings; all six were checked against their primary sources, all held, and each is in the corrections ledger (R2-01 to R2-06). The largest was a missing offer: SeenThis Attention, launched in June 2026, which changed two claims about trading ([chapter 6](#trading)). A third outside review read the full draft as a hostile expert would. It returned 19 findings; each was checked against its source or the paper's own data before any change, all held, two had already been fixed by the consistency audit, and each is in the corrections ledger (R3-01 to R3-19). They narrowed the accreditation verdict to what MRC's letter lists, bounded two absence claims, corrected how the paper described who measures guarantee delivery, and fixed a display error that showed superseded claims in the evidence panels. Two internal review passes of the full draft ran alongside it: an integrity pass, which checked every number in the text against its claim and source, every marker, every absence bound and every label; and an adversarial pass, which read the paper as a measurement scientist, a vendor and a buyer would. Their changes were checked against sources before being applied and are logged in the corrections ledger. A fourth outside review checked the built paper before publication. It returned eight findings; each was checked at source, all held (one with qualification), and each was fixed before publication and is in the corrections ledger (FC-01 to FC-08). The most serious was a profile and SWOT line that said Adelaide measures delivery in every AU guarantee, where the sources show it only for the offers that name a measurer.

### How the products were scored

The scoring unit is a named product that serves a stated buying task, mapped to one record in the product register. Cohort A, media attention measurement, covers products that measure, plan or optimise attention on media exposures. Cohort B, creative attention testing, covers products that test or predict creative before launch. Each cohort has an applicability mask: creative diagnostics does not apply to cohort A, and activation does not apply to cohort B. n/a comes only from the mask, never product by product.

Each cell is scored from 0 to 5 against the dimension's anchors for that cohort, and then capped by the evidence: vendor documentation caps a cell at 4 and vendor marketing at 3. A 0 means a source documents absence or failure; missing evidence is n/e, never 0, and a partner's name alone is n/e. Only cells whose facts a verifier checked count as numeric evidence; an unverified cell is shown but counts as n/e. A product enters the matrix with at least four counted cells. The caps fold the type of evidence into the score: a 4 can mark a well-documented capability resting on vendor documentation, not a weaker capability than a 5. The numeric thresholds in the anchors, such as impressions a year for production scale, are the author’s assumptions, fixed before scoring.

Each row was scored once, by an AI scoring agent working from the fixed rules and the verified evidence, after the verifier records and the second outside review were read. A separate AI consistency audit then compared every cell across rows against the anchors and caps and proposed changes; those accepted are logged as scoring corrections. The third outside review led to two further cell changes, also logged. The rows nearest a band line were then scored a second time, blind: a separate AI scorer, given the rules and each row’s cited sources and verified claims but not the first scores, grades or rationales, re-scored every applicable cell of the six rows within 2 points of a band line or whose band moved under a reweighting. The two passes matched exactly on 78.8% of the 66 cells and within one point on 98.4% of the cells both scored (quadratic-weighted kappa 0.88), and no row changed band. The two material disagreements, both on Lumen SPOTLIGHT, were adjudicated by re-reading the deciding sources, and both went to the second pass (AJ-01, AJ-02). Both passes are published in data/second-pass.csv. Every other row was scored once. The bands were named Front rank, Strong, Mid and Narrow when the rules were fixed; before publication they were renamed Most, Well, Partly and Least documented, to say what they measure, with the same thresholds. [Chapter 11](#fq) sets out the Focus Quotient, its field mean, bands and sensitivity tests. Its field mean is the mean of a cohort's Raw scores, which differs from the pooled construction of the author's Contextual and Native Quotients, so the numbers are not comparable across papers.

### Limits

The paper scores what is documented in public by the cutoff, not what products do in private. Pricing, contract terms, realised prices, adoption and market shares are almost never public, and the paper lists them as unanswered rather than estimates them. A vendor with a thin public record scores lower than its product may deserve; the tests in chapter 11 show how much single cells move a quotient.

Status-sensitive facts (accreditation, ownership, transactions, protocol versions) were read on 28 and 29 September 2026 and are dated in the text. The registered evidence freeze was 2 October 2026. The author set the cutoff at 1 October 2026 and published on 6 October 2026 without a separate freeze recheck, so status-sensitive facts stand as read on 28 to 30 September. The hypotheses, product register and scoring rules were fixed against the registered dates and did not change; anything that changes after the cutoff will be logged as a dated correction, not folded in silently.

The United States is the primary market. Geography scores count documented US availability only; international breadth is recorded but earns nothing.

Several material documents could not be read in full: the ARF's Phase 3 report (members only), Adelaide's case-study guides and MediaSense's review of AU (both gated), and MRC's December 2022 accreditation letter to DoubleVerify (not opened; that date rests on MRC's quarterly report). Claims resting on summaries of those documents say so.

### Interests

The author has worked at Samba (TV measurement and advertising data) since May 2025. Adelaide, which is scored and profiled, names Samba among the tune-in attribution providers whose data train its CTV model [106]; in the sources screened, this review found no attention-based trading offer from Samba [106]. The author co-leads the Signals & Measurement working group of AgenticAdvertising.org, which maintains AdCP, one of the protocols assessed in chapter 10. No Fluff Advisory, the author's firm, operates two agents listed in the AgenticAdvertising.org registry on 28 and 29 September 2026; one, a sales agent, is among those cited in chapter 10 as passing a test-mode conformance storyboard. The companion native-advertising paper disclosed two further links, which this paper's own sources do not cover: AgenticAdvertising.org listed Samba among its members, and Teads named Samba among its measurement-partner integrations. Teads is not scored here; its pilot and beta guarantees sit in the trading table in [chapter 6](#trading) [95]. Earlier employers include Verve Group (Moments.AI), Hearts & Science and Lotame. No vendor funded, reviewed or saw the paper before publication. Every vendor is scored under the same published rules. The author's own essays argue positions on attention that this paper tests as hypotheses, not as evidence.

### The claims

Every claim marked in the text is listed here with its evidence class, review outcome, confidence and supporting passages.

**Claim ledger**

| ID | Claim (final wording) | Ch. | Evidence class | Review | Confidence | Sources |
|---|---|---|---|---|---|---|
| ATT-B-F12 | Orquin and Holmqvist (2018) state that eye trackers report eye movements and gaze while attention is always inferred, and advise against total dwell time as a dependent variable because a difference in it cannot be attributed to its components (fixation likelihood, fixation count, fixation or dwell duration), which they class as a threat to construct validity; extending this caution to commercial dwell-based or eyes-on-screen attention metrics is the stream's inference. | 1 | synthesis | supported with qualification | high | [419, 55] |
| ATT-U-F05 | Brainsights describes itself as a neuroanalytics company using EEG, sells a named Creative Optimization service (second-by-second analysis with general and category benchmarks) whose page does not name an attention construct, and names attention only in SI Maps, a neuro-and-attention score for billboards across Canada; no US availability is documented in the pages opened. | 1 | vendor assertion | supported as stated | medium | [420, 421, 422] |
| ATT-U-F11 | Lumen SPOTLIGHT is participant-based creative testing using webcam eye tracking on laptops and mobile devices across opt-in participants, offered for social, digital, in-game, TV/VOD, OOH, point-of-sale, print and packaging, with post-test surveys for recall. | 1 | vendor assertion | supported as stated | medium | [279] |
| ATT-U-F28 | ARF's survey of attention providers (Phase 1, October 2023) found that 61% focus on both creative and media, 92% collect data from human subjects at some point, and methods include eye tracking (50%), survey-based ad recall (46%), dwell time (42%) and facial coding (42%); its Phase 3 release (June 2026) says methods diverge most at platform and placement level and 'attention is not a single universal metric'. | 1 | independent measurement | supported as stated | medium | [423, 5] |
| ATT-AMPLIFIED-F06 | Campaign Strategy predicts attention before launch from panel-derived benchmarks, with outputs filterable by age and gender, a channel list spanning social, web and programmatic, OOH and DOOH, TV/BVOD/CTV, audio (Spotify and radio) and cinema, and SaaS pricing from pay-as-you-go to enterprise. | 1 | vendor assertion | supported as stated | high | [119] |
| ATT-AMPLIFIED-F13 | Amplified claims its in-flight methodology is 'independently validated, with 99% data accuracy' and, on its investor page, '99% accuracy vs. competitors at 70%', and lists 'Peer-reviewed attention science' as an Amplified feature, but none of the pages names a validator, metric, reference set or publication. | 1 | vendor assertion | supported as stated | high | [186, 191, 187] |
| ATT-AMPLIFIED-F31 | Amplified contrasts its 'measured behaviour' with competitors' 'modelled probability', yet by its own account the in-flight and creative products deliver model estimates: it told the ARF in 2023 that tag data are combined with panel ground truth to predict in-flight attention, and AttentionAI simulates 20,000 personas per ad. In September 2026 it called AttentionAI output 'measuring ... attention at impression level'. | 1 | inference | supported with qualification | medium | [187, 186, 424] |
| ATT-IAS-F10 | IAS documents attention activation as a beta pre-bid optimisation product (December 2024), sell-side packaging into segments and PMPs (October 2024), and, announced 26 March 2026, Mastercard-enriched pre-bid segments that use attention among other signals (US availability forecast for Q2 2026; no general-availability record found). In the sources screened by 29 September 2026 no DSP, SSP or curation platform is named for any attention-based segment. | 1 | vendor assertion | supported with qualification | medium | [115, 232, 226, 231] |
| ATT-KANTAR-F03 | Kantar's active attention is a facial-expression emotion measure rather than a gaze measure, and Kantar's own publications report that it barely relates to passive attention: 'essentially non-existent' in Affectiva data (February 2024) and 'surprisingly little correlation' at ad level (July 2024). | 1 | vendor assertion | supported as stated | medium | [259, 258, 249] |
| ATT-KANTAR-F06 | Kantar's July 2025 North America playbook describes LINK+ as a monadic online survey of N=150 respondents per study (10-20 minutes, mobile-friendly), benchmarked on 250K+ creative tests with category and country norms, with optional neuro modules (facial coding, webcam eye tracking, AI predictive eye tracking, intuitive associations) and AI-generated heatmaps and fog maps; the LINK+ page lists facial coding as an optional module that affects price. | 1 | vendor assertion | supported as stated | high | [263, 246] |
| ATT-KANTAR-F11 | LINK AI's attention outputs are model estimates: the product page combines brand metrics 'with measures of attention, engagement and likelihood to act' benchmarked to channel norms and offers predictive eye tracking of which elements stand out, and Kantar says individual LINK AI metrics 'can also predict' attention and skip rates, without publishing an accuracy statistic for the attention outputs. | 1 | vendor assertion | supported as stated | high | [247, 391] |
| ATT-LUMEN-F22 | Lumen defines its headline unit as attentive seconds per thousand impressions = % view x average eyes-on dwell time x 1,000, but its February 2023 definitions page also defines 'Attention seconds per 1000 impressions' as 'The number of attentive seconds divided by total impressions', omitting the x1,000 scaling. | 1 | vendor assertion | supported as stated | high | [392, 44] |
| ATT-LUMEN-F23 | The News UK and Lumen study (REG-002, released 22 March 2022) reports that relevant editorial context raised attention per 1,000 impressions by 31% across 980,000 impressions on The Sun's website, but gives the baseline as '25 minutes per 1,000 impressions' in one sentence and '25 seconds' in another, so only the percentage should be cited; it measured no memory, brand or sales outcome. | 1 | vendor assertion | supported as stated | high | [402, 62] |
| ATT-TVISION-F07 | TVision itself states its panel size. Its 2026 TV Attention &amp; Creative Impact Report gives 'an opt-in panel of 5,000 homes across the United States', its recruitment site shows 5,000 homes and 15,000 people, and its CRO gave 4,500 to 5,000 households (Digiday, March 2026). Trade reporting matches: about 5,000 US households in 25 DMAs and about 14,000 people (AdExchanger, November 2022); 5,000 homes and 13,000 people (Next TV, April 2023); 5,000 households (PPC Land, 2026); '5000 households' (dentsu, 2021). The 2016-17 academic extract covered 1,155 households and 3,659 viewers. Installed versus in-tab counts are not stated anywhere. | 1 | synthesis | contradicted | medium | [404, 298, 405, 24, 52, 292, 415] |
| ATT-TVISION-F20 | TVision and Upwave (18 July 2022) report a linear regression across 40+ US campaigns from Q4 2019 to Q1 2022. Each percentage-point increase in Ad Attention (share of ad impressions with two or more seconds of looking at the screen) for high-frequency ads went with 'close to a 1%' lift in aided awareness. The release calls this a '1:1 correlation'. No respondent sample, uncertainty or holdout is reported. | 1 | vendor assertion | supported as stated | medium | [74] |
| ATT-A-F03 | The guidelines define viewability as the opportunity to see or hear an ad, and attention as 'whether the ad was seen or heard by a person and to what depth'. MRC treats attention as a measure of audience (persons, not devices). Section 4.3 names three key measurement categories: Exposure (with Attentive Response, meaning gaze tracking and peripheral focus, as a sub-item), Cognitive Load (with Emotional Response as a sub-item that is not required for attention measurement), and User Interaction and Behavior. | 1, 2 | direct record | supported with qualification | high | [41] |
| ATT-A-F04 | The guidelines recognise four measurement approaches (data signals, visual tracking, physiological and neurological observation, panel and survey methods), plus predictive models that combine them to predict visual attention from logged data signals. They do not say which approach is best. | 1, 2 | direct record | supported as stated | high | [41] |
| ATT-A-F08 | The guidelines define no common attention unit. A text search of the Version 1.0 PDF found no definition of attention seconds, attentive seconds per 1,000 impressions (APM), a 0-100 attention unit or percentage attended. The text lists per-measure units (seconds, milliseconds, scores, percentages) and 'strongly encourage[s]' reporting attentive time relative to viewable time. | 1, 2, 3 | direct record | supported as stated | medium | [41] |
| ATT-A-F06 | The guidelines say observed measurement 'should be distinguished' from modelled measurement; that models 'must have empirical support' that they are fit for their intended purpose, including a link to validated attention and, where a model is used to predict outcomes, to campaign objectives and outcomes; that model validation 'should' be performed at least annually on fresh deterministic data and training data 'should' be deterministically collected; and that error margins and standard errors around sample-based projections must be disclosed. Randomised or A/B validation is strongly encouraged, not required. Outcome validation is encouraged in general but required for any model used to predict outcomes. | 1, 2, 3, 4, 12 | direct record | supported with qualification | high | [41] |
| ATT-VEVO-F05 | The guaranteed metric is, by clear implication (the release does not state the unit of the minimum), Adelaide's AU, which Vevo's release describes as a 0 to 100 media-quality metric that 'predicts' a placement's likelihood to capture attention and drive outcomes, built on eye-tracking research, placement-level signals and outcome data. It is a model estimate of the placement, not observed attention of the exposed audience. | 1, 2, 6 | synthesis | supported with qualification | high | [78, 94, 48] |
| ATT-A-F07 | The guidelines say attention 'should not be considered or used as a measure of outcomes'. They only 'strongly encourage' independent third-party auditing of attention methods used in buying and selling. IAB's page adds that attention 'is not a standalone measure of ad effectiveness'. | 1, 2, 6, 12 | direct record | supported as stated | high | [41, 425] |
| ATT-DOUBLEVERIFY-F03 | MRC's letter of 13 April 2026 continues accreditation of display and video Authentic Attention metrics in desktop, mobile web and mobile in-app and grants initial accreditation of Authentic Attention reporting (Indexed Impressions and Direct-Indexed-Rate, and metrics based on DV's Authentic Non-viewable and Authentic Modeled methodologies) in the same environments, which MRC's July 2026 report labels 'Authentic Attention V4'. | 1, 2, 6, 12 | direct record | supported as stated | high | [42, 205] |
| ATT-DOUBLEVERIFY-F07 | The accredited Authentic Attention statistics are indices and device-event rates, among them Attention, Exposure, Engagement, Intensity, Prominence, User Presence and Ad Interaction indices, ad-touched, scroll, tab-into-background and volume-change rates, and quartile audible and viewable rates, all for display or video in desktop, mobile web and mobile in-app. | 1, 2, 8 | direct record | supported as stated | high | [42] |
| ATT-DOUBLEVERIFY-F13 | DV's only documented CTV attention data source is TVision (June 2023 announcement; closed beta anticipated H2 2023; joint Q1 2024 CTV report from a 5,000-home US panel). DV's June 2025 Snap release states its attention measurement spans CTV, and its current CTV pages describe CTV attention, but none names a data source or gives a dated GA release. Viant completed its acquisition of TVision on 1 May 2026, and its closing release of 5 May 2026 does not address third-party access. | 1, 2, 8 | synthesis | supported with qualification | medium | [220, 426, 364, 219, 82] |
| ATT-A-F15 | XPLN (xpln.ai) is the only attention product found that MRC lists as in audit. MRC records it as 'Pre-audit reviewed 9/2025, Audit in process as of 11/2025', for display and video impressions, viewable impressions and attention metrics, including SIVT filtration, in mobile web and mobile in-app. | 1, 2, 11, 12 | direct record | supported as stated | high | [48] |
| ATT-B-F02 | In a German lab study at the University of Wuppertal (30 graduate students assessed, four excluded; mock Instagram feed on participants' own smartphones; 806 post exposures across ads and content posts), viewport time correlated .89 with mobile eye-tracked gaze duration across all posts, and in separate single-predictor logistic regressions on ad exposures (sample size and R-squared type not reported) viewport time predicted unaided brand recall (OR 1.12 per second, p = .045, R-squared .027) with the same explained variance as gaze duration (OR 1.14, p = .047, R-squared .027). | 1, 3 | independent measurement | supported with qualification | medium | [46] |
| ATT-B-F03 | In the same study, the MRC 1-second viewability rule identified posts that received at least 1 second of gaze in 94% of viewable cases, but 41% of non-viewable cases also received more than 1 second of gaze, and where viewport time and gaze diverged by more than one standard deviation viewport time was the larger in 77 of 85 cases. | 1, 3 | independent measurement | supported as stated | medium | [46] |
| ATT-B-F13 | An early open-source webcam system (WebGazer, 2016) showed about 172 px (3.94 degrees) of offset in the lab and 207 px online, with 52 of 84 online sessions incomplete (Semmelmann and Weigelt 2018); a 2022 comparison (n = 83) co-authored by three employees of the webcam vendor RealEye reported higher measurement error than a remote tracker while replicating expected effects; and a 2023 lab comparison co-authored by the Labvanced developers reported 1.4 degrees accuracy against an EyeLink 1000. None tests a commercial ad-attention panel, the best recent accuracy evidence is vendor-authored, and the whole-ad versus element-level conclusion is an inference from error sizes. | 1, 3 | synthesis | supported with qualification | medium | [58, 59] |
| ATT-B-F21 | TVision's attention, as described in independent academic use, is algorithmic classification of camera images of consenting panellists ('co-occurrence of eyes-open and eyes-on-screen inferences'), so its numbers are direct observation of people processed by a model, measured post-campaign on the panel, not a prediction for non-panel impressions. | 1, 3 | independent measurement | supported with qualification | medium | [52] |
| ATT-LUMEN-F11 | PwC's review (commissioned by Lumen) found that Lumen's view and dwell models were a logistic regression and a quasi-Poisson model on ad size, log viewable time and domain terms, with inputs of device, size, viewable time, domain and % in view but not ad position or clutter, and attention computed as expected dwell times expected view. | 1, 4 | commissioned measurement | supported as stated | high | [8] |
| ATT-ADELAIDE-F06 | Adelaide's display method is a four-step build: eye-tracking research identifies placement characteristics (coverage, in-view duration, position, clutter, device); a JavaScript tag and page scanning collect those signals at scale; an initial model produces preliminary AU; and the model is trained and calibrated on historical outcome data (brand lift, behavioural signals, conversions, sales). | 1, 4, 12 | vendor assertion | supported as stated | high | [168] |
| ATT-U-F16 | xpln.ai sells attention metrics built from eye-tracking panel data combined with machine learning to give real-time estimates, claims coverage of social media, walled gardens, CTV and streaming, is a Task Force participant in the IAB and MRC guidelines, and is not in the brief's seed list; this stream proposes adding it to cohort A. | 1, 5 | vendor assertion | supported as stated | medium | [122, 41] |
| ATT-ADELAIDE-F05 | Adelaide names its gaze reference data as licensed from Amplified Intelligence, Lumen, Tobii, TVision and Viomba plus its own eye-tracking tools, with TVision panel data used for CTV and linear TV and, from a November 2024 partnership, Viomba's Tobii infrared panels (operated in Europe) merged into AU models that previously used only webcam-based eye-tracking data. | 1, 5 | vendor assertion | supported as stated | medium | [169, 327] |
| ATT-DOUBLEVERIFY-F12 | DoubleVerify made Authentic Attention for Social globally available on Snapchat on 30 June 2025, combining DV exposure data with Lumen Research eyes-on-screen signals and reporting Ad Focus, Dwell Time and an Attention Index, and extended Authentic Attention to TikTok on 10 December 2025 as the first badged TikTok Marketing Partner for attention. | 1, 5 | vendor assertion | supported as stated | medium | [116, 117] |
| ATT-ADELAIDE-F13 | On 9 June 2026 Adelaide made AU Media Quality tiers (High, Average, Low) and an AU Quality Floor (excluding MFA-designated sites and the bottom 10% of AU-scored placements) available as pre-bid segments in Amazon DSP across display, online video and streaming TV; the segments are built from historical placement-level AU data, and Adelaide says Amazon DSP already offered post-campaign impression-level AU visibility. | 1, 5, 8 | vendor assertion | supported as stated | medium | [162] |
| ATT-IAS-F07 | IAS describes Quality Attention as a machine-learned composite of visibility, situation and interaction signals from its core products, weighed into a single attention score with three sub-scores, and says the combination 'can predict if an impression is likely to lead to a business result'; its signals include passive eye tracking 'powered by Lumen data'. | 1, 5, 8 | vendor assertion | supported as stated | high | [73, 238] |
| ATT-TVISION-F04 | After the 1 May 2026 closing, vendor sources show TVision still selling outside Viant's DSP. TVision's site markets calibration data and data licensing to measurement and data partners. On 25 August (year inferred as 2026 from 'a Viant Technology company'), it introduced Ad Scoreboard 2.0 for brand marketers. Adelaide's method blog of 29 July 2026 says, in the present tense, that AU for CTV 'incorporates TVision data' and that linear TV AU is matched with TVision panel data. | 1, 5, 8 | vendor assertion | supported as stated | medium | [297, 311, 106] |
| ATT-E-F15 | DoubleVerify launched pre-bid attention segments on 4 May 2023 in Amobee, MediaMath and Invest DSP, and DV Algorithmic Optimizer with Scibids on 19 June 2023 across DV360, MediaMath, The Trade Desk and Xandr; both are optimisation uses. Several of those DSPs no longer operate as named (MediaMath entered Chapter 11 on 30 June 2023; Microsoft's Invest DSP stopped supporting buying on 28 February 2026). DV's 2026 product page still offers Universal Attention Segments and Scibids attention optimisation without naming DSPs, and MRC's accreditation of DV attention metrics is not shown as covering pre-bid data. | 1, 6 | vendor assertion | supported with qualification | medium | [222, 213] |
| ATT-U-F10 | Lumen describes Attention Bidding as a pre-bid custom algorithm that integrates with the buyer's DSP and is 'Powered by the Lumen Attention Measurement &amp; Planning Platform' (components: LAMP measurement tag, LAMP DSP plug-in, bespoke attention models); the platform predicts attention every 100ms from placement data using a model trained on 650,000+ eye-tracking sessions; the operating entity is Lumen Research Ltd (UK company 08682432, active, incorporated 9 September 2013, London). | 1, 6 | synthesis | supported as stated | medium | [392, 427, 20, 428] |
| ATT-VEVO-F04 | Vevo announced on 27 January 2026 that, through a partnership with Adelaide, it had launched 'Attention Guaranteed', part of Vevo Evolve, which 'secures a minimum attention score' for a campaign across all screens (CTV, mobile and desktop) with one performance report. The release states no threshold, remedy, audit, price or geography, and does not say who measures compliance. | 1, 6 | vendor assertion | supported as stated | high | [78] |
| ATT-ADELAIDE-F04 | Adelaide defines AU as a 0-100 score representing 'a placement's probability of attention and subsequent impact', generated by a machine-learning model from media-quality signals and full-funnel outcome data, normalised across channels; it is a model estimate of placement quality that mixes attention with outcome prediction, not an observation of whether a person looked at an ad. | 1, 7, 8 | vendor assertion | supported as stated | high | [43, 106, 1] |
| ATT-G-F07 | The three vendors that attach attention metrics to audio use three different constructs, none of which is documented as observing a listener's auditory attention to an ad: Adelaide's podcast AU is a model estimate built from placement exposure data (genre, ad position, ad load, ad type, licensed transcription and Magellan data) with, per Adelaide, eye-tracking and full-funnel outcome data in the model; Amplified classifies active, passive and non-attention from listener tasks and surveys, alongside STAS (a brand-choice measure); and Lumen's 2023 audio score for dentsu was modelled to be equivalent to visual attention from listening data, survey results and exposure type. | 1, 8 | synthesis | supported with qualification | medium | [107, 105, 120, 112] |
| ATT-G-F09 | Dentsu's audio attention research with Lumen, released 1 August 2023, comprised three studies (podcasts and radio with named US audio publishers; music streaming conducted separately by Amazon Ads on Amazon Music) and reported an average of 10,126 attentive seconds per 1,000 impressions across audio formats, against dentsu norms of 6,501. These are modelled visual-equivalent scores, not observed auditory attention; sample size and geography are not stated. | 1, 8 | vendor assertion | supported with qualification | low | [112] |
| ATT-AMPLIFIED-F08 | The description of Campaign Performance is right. The limitation should read: the product page does not label the basis, but the vendor's 2023 ARF submission describes in-flight figures as model predictions trained on panel data. | 1, 8 | vendor assertion | supported with qualification | high | [186, 350] |
| ATT-LUMEN-F10 | In production, LAMP does not observe gaze: its JavaScript tag collects impression-level placement data (domain, format, page geometry, scroll speed, % viewable, time in view) and applies Lumen's attention model to estimate attentive seconds per impression, predicting attention every 100ms of exposure. | 1, 8 | vendor assertion | supported as stated | high | [282, 281] |
| ATT-LUMEN-F18 | Lumen documents US CTV measurement through TVision: for LG Ad Solutions (2 February 2024) the Lumen tag runs in LG CTV ads in EMEA while in the US measurement runs through the TVision pixel processed in the Lumen platform, and Nexxen's CTV offer (August 2023) uses measurement 'powered by TVision data'. | 1, 8 | vendor assertion | supported as stated | medium | [290, 289, 295] |
| ATT-TVISION-F08 | What TVision observes, per its methodology: optical sensors in panel homes identify who is in the room while the TV is on and when each person looks at the TV. A microphone array matches audio fingerprints to program and ad databases. Attention is defined as 'the amount of time that a viewer is looking at the TV screen'. The independent 2021 paper describes the 2016-17 implementation: facial recognition trained on each household member at set-up, frames processed five to six times per second, and attention as 'the co-occurrence of eyes-open and eyes-on-screen inferences', with head orientation inferred from facial features. The authors footnote it as 'only one possible measure of attention'. | 1, 8 | synthesis | supported as stated | high | [45, 52] |
| ATT-TVISION-F09 | Only panel exposures are directly observed. Outputs for the wider market are projections or models. TVision's DSP scores are 'based on historic viewer attention' at daypart and channel level. OpenX's targeting uses 'predictive attention models' built from TVision panel data combined with OpenX bidstream signals. xpln.ai extends 'panel-based observations' into 'real-time, scalable attention predictions'. | 1, 8 | synthesis | supported as stated | high | [407, 315, 296, 45] |
| ATT-DOUBLEVERIFY-F09 | DV's attention indices are normalised to 100 as the average across DV's own impressions over a 28-day rolling window, so an index of 125 means 25% above DV's moving benchmark. | 1, 8, 12 | vendor assertion | supported as stated | high | [71, 211, 370] |
| ATT-LUMEN-F09 | Lumen's panel page (published June 2021, no later revision date) states eye-tracking panels of 1,000 in the UK (250 desktop, 750 mobile) and 1,000 mobile users in the US, each a nationally representative sample, 'fully informed, consented and compensated', combining a browser extension with webcam eye tracking; later vendor statements give only cumulative or coverage totals: 650,000+ sessions over 10+ years, '30+ countries' (December 2024) and '50+ countries' (April 2026). | 1, 9, 12 | vendor assertion | supported as stated | medium | [271, 392, 115, 287] |
| ATT-AMPLIFIED-F25 | Amplified launched AttentionAI on 20 May 2026: a model that simulates 20,000 viewing personas per ad, trained on data from over 2 million people across 21 markets, returning second-by-second predicted attention, with Connect as first API partner running 500 creative tests a month and validation described only as continuous, without statistics. | 1, 10 | vendor assertion | supported as stated | high | [192, 429] |
| ATT-TVISION-F18 | In the independent MSI working paper (McGranaghan, Liaukonyte and Wilbur, 2021; the authors report no funding or conflicts), TVision panel data from 1,155 households (July 2016 to June 2017) show viewers absent from the room during 30% of ads played in active viewing sessions. Across 180 national ads from two NFL playoff games, attention explained the most variance in brand search lift in single-measure regressions without quarter-hour fixed effects (R-squared 0.116, against 0.082, 0.070 and 0.035), with no holdout. | 1, 11 | independent measurement | supported as stated | high | [52] |
| ATT-ADELAIDE-F08 | Adelaide states that it does not publicly disclose the algorithm weights used to calculate AU, and in the Adelaide documentation screened (glossary, method posts, gaze and data-use posts, comparison post) no reference-panel size, recruitment method or held-out agreement statistic between AU and gaze reference data is published. | 1, 12 | synthesis | supported as stated | medium | [66, 168] |
| ATT-AMPLIFIED-F10 | Amplified's reference panel, per its privacy policy, is recruited only through Cint, paid, and collects facial images via device cameras on mobile and PC and via an in-home TV micro-camera that can capture guests, with facial landmarks used only on TV to determine face orientation, and EU/UK panel data processed on AWS servers in Australia. | 1, 12 | vendor assertion | supported as stated | high | [184] |
| ATT-IAS-F08 | IAS names Lumen Research as its eye-tracking partner and, from January 2024, says its model was trained on 'billions of impressions and millions of conversion events'; the report-level definition of the Attention Score (reproduced by Funnel) describes it as click or conversion propensity. No IAS source opened states Lumen panel size, recruitment or geography, or a held-out agreement or error statistic against gaze. | 1, 12 | vendor assertion | supported with qualification | medium | [237, 236, 115] |
| ATT-IAS-F21 | IAS publishes no scale definition, but an IAS primary case (Pernod Ricard, Japan) segments scores into 1–57 (low) and 58+ (high), which confirms a scale starting at 1 and shows that the high/low cut-off IAS uses varies by analysis (58 there, not PPC Land's 65). Case averages as listed. Scores remain vendor-defined model outputs, not comparable with other vendors' units. | 1, 12 | vendor assertion | supported with qualification | low | [11, 228, 235] |
| ATT-KANTAR-F02 | Kantar's attention framework, launched on 14 November 2023 in LINK+ and Context Lab, uses facial coding with gaze monitoring to report passive attention (time eyes are on screen while the ad is visible) and active attention (time facial expressions indicate emotional engagement), with a summary score, norm comparison and second-by-second trace, and norms built on 50,000 attention ad tests. | 1, appendix | vendor assertion | supported as stated | high | [249] |
| ATT-A-F02 | The exact phrase 'serve as the basis for MRC accreditation audits' appears on IAB's web page, not in the Version 1.0 PDF. The PDF nonetheless frames several provisions in MRC accreditation terms: base inputs must be audited for MRC accreditation consideration, non-viewable and aggregate attention measures are eligible for accreditation consideration with proper disclosure, MRC audits will verify upstream processes, and users are told to encourage providers to be independently audited for compliance against the guidelines (pp. 8, 9, 11, 60). | 2 | direct record | supported with qualification | high | [425, 41] |
| ATT-A-F09 | Comparing the May 2025 draft with the November 2025 final shows five additions in the final: an MRC requirement that content attention be measured when attention is used to evaluate properties or channels; the sentence 'These guidelines strongly require validation for models and measurement approaches'; rules on synthetic training data; a reference to the ARF's Attention Initiative; and the encouragement to report attentive time relative to viewable time. | 2 | synthesis | supported as stated | medium | [41, 430] |
| ATT-A-F17 | Staff of DoubleVerify, Integral Ad Science, Adelaide, GumGum, Amplified Intelligence, Lumen Research and TVision were named key contributors to Version 1.0, alongside MRC's Senior Vice President for Digital Research and Standards. IAS has three key contributors listed (Anurag Pramod, Vanessa Turner and Vidhan Srivastava, the last under the misspelling 'Integrated Ad Science'). DoubleVerify is also a board member organisation of IAB's Measurement, Addressability &amp; Data Center. | 2 | direct record | supported with qualification | high | [41] |
| ATT-A-F20 | The other bodies screened issue guidance or comparisons, not attention standards or accreditation. IAB Europe's updated Guide to Attention (October 2025; first published May 2023) cites the IAB/MRC draft as 'the first step toward an industry standard'. IAB's page says CIMM collaborated on the framework and co-published a Playbook for Marketers (November 2025). France's CESP publishes an attention grid of vendor answers and MRC statuses, updated 10 June 2026. | 2 | synthesis | supported as stated | medium | [431, 432, 425, 3] |
| ATT-U-F17 | The CESP attention grid (updated 10 June 2026) lists five attention solutions in the French market: Adelaide, Adloox-Peer39, DoubleVerify, IAS and Xpln.ai; only DoubleVerify answers that its attention metrics are MRC-accredited, and Adloox-Peer39 answers that its attention solution has not been independently audited and its attention metrics are not MRC-accredited. | 2 | vendor assertion | supported as stated | medium | [3] |
| ATT-XPLN-F19 | XPLN holds no MRC accreditation. It does not appear on MRC's Accredited Services: Digital page (accessed 29 September 2026). In CESP's grid (10 June 2026), xpln.ai answered 'No' to whether its attention solution had been audited by an independent third party, and 'No***' to MRC accreditation of its SIVT filtration, viewable impressions and attention metrics, with a footnote relaying MRC's 'under review' status. | 2 | synthesis | supported as stated | high | [2, 3] |
| ATT-AMPLIFIED-F23 | On 29 September 2026 MRC lists no Amplified product among accredited digital services or digital services under review; the only 'Attention Metrics' accreditation is DoubleVerify Quality Analytics Platform (desktop, mobile web, mobile in-app) and XPLN's attention audit is in process. | 2 | direct record | supported as stated | high | [2, 48] |
| ATT-AMPLIFIED-F24 | Amplified Intelligence is listed as a company participant of the IAB Attention Task Force and its founder as a key contributor to the IAB/MRC Attention Measurement Guidelines v1.0 (November 2025), while Amplified's own pages self-attest alignment with the guidelines and describe them as 'Validation from the MRC and IAB' and as an 'IAB and ARF' framework. | 2 | synthesis | supported as stated | high | [41, 355, 193] |
| ATT-DOUBLEVERIFY-F01 | MRC first granted accreditation for DoubleVerify's 'Authentic Attention' metrics suite in desktop, mobile web and mobile in-app in December 2022. | 2 | direct record | supported as stated | high | [215, 433] |
| ATT-IAS-F18 | Three IAS staff are named key contributors to the IAB and MRC Attention Measurement Guidelines Version 1.0 (November 2025), which state that attention 'should not be considered or used as a measure of outcomes' for evaluating campaign performance and place unvalidated predictive models outside their scope, while IAS markets Quality Attention as predicting business results. | 2 | synthesis | supported as stated | medium | [41, 73] |
| ATT-A-F19 | The ARF runs an Attention Measurement Validation Initiative as research, not as a standard or an accreditation. Phase 1 (October 2023) found vendors' definitions shaped by their own tools. Phase 3 (9 June 2026, four national campaigns) found that attention methods agree at channel level, diverge at platform and placement level, and show 'no simple one-to-one relationship' with upper-funnel brand lift. The final guidelines point users to the ARF's work. | 2, 3 | synthesis | supported as stated | medium | [423, 5, 41] |
| ATT-B-F09 | The ARF's Phase 3 release (9 June 2026), covering four live national campaigns across TV, online video, social and display measured by multiple providers, reports 'no simple one-to-one relationship between attention and upper-funnel brand lift' and agreement between methods mainly at channel level, diverging at platform and placement level; the full report is member-only. | 2, 3 | synthesis | supported as stated | low | [5, 434] |
| ATT-A-F05 | The guidelines require a base of compliant viewable or audible impressions (one second for display, two continuous seconds for video), full SIVT filtration and presence of a user, and require base inputs to be audited for MRC accreditation consideration. Attention measured without or before viewability may also be reported if four conditions are met: the standard viewable version is also reported, the non-viewable version is presented separately as non-standard and diagnostic, it covers only visible ads, and it rests on compliant rendered or playback-initiated impressions. Both presentations, and aggregate measures combining viewable and non-viewable inputs, are eligible for MRC accreditation with proper disclosure. | 2, 3, 12 | direct record | supported with qualification | high | [41] |
| ATT-D-F14 | Viant Technology agreed on 14 April 2026 to acquire TVision Insights for $22.5 million in cash plus 1,656,701 Class A shares (agreed equity value $17.5 million) and completed the deal on 1 May 2026; its Q2 2026 10-Q estimates consideration at about $37.5 million with $13.4 million of goodwill and $12.3 million of intangibles, cites the 'integration of TVision's attention signals', and records that Viant assumed Nielsen's pending patent litigation against TVision, while Viant's 11 May 2026 release describes the integration of TVision's attention insights as 'exclusive'. | 2, 4, 5, 10 | direct record | supported with qualification | high | [435, 82] |
| ATT-C-F08 | Adelaide states that it asked MediaSense, a media advisory firm, to audit AU; between February and June 2025 MediaSense reviewed the methodology and rebuilt AU-to-outcome curves for aided awareness, familiarity, recall and intent from sampled, aggregated data that Adelaide supplied, while Adelaide does not publicly disclose its algorithm weights; the audit report is gated and is not an accreditation. | 2, 4, 6 | vendor assertion | supported as stated | medium | [66, 22] |
| ATT-LUMEN-F14 | Lumen markets its technology as 'the only solution validated and audited by an independent third-party in the industry' on the basis of the 2023 PwC report, but the report says Lumen requested it and that PwC 'has not independently verified, validated, or audited the data'. | 2, 4, 12 | synthesis | supported as stated | high | [20, 8] |
| ATT-D-F09 | DV's FY2025 10-K calls DV Authentic Attention 'MRC-accredited', a sentence DV has carried in every 10-K since the FY2022 report filed on 1 March 2023, which states that DV was accredited by the MRC in 2022 for 'Display and Video DV Authentic Attention' metrics; the wording therefore predates, and does not refer to, MRC's 13 April 2026 letter. The full 10-K discloses no attention-specific revenue, price, attach rate or customer count, and DV reports revenue by customer type (activation, measurement, supply-side). | 2, 5 | direct record | supported with qualification | medium | [47, 207, 436] |
| ATT-IAS-F04 | Novacap's take-private of IAS was announced on 24 September 2025 at $10.30 per share in cash, in a deal the parties said values IAS at approximately $1.9 billion (basis not stated), and completed on 23 December 2025, when IAS became a wholly owned subsidiary of Igloo Group Parent, Inc., an affiliate of funds managed by Novacap Management Inc.; the aggregate price paid for outstanding shares was approximately $1.6 billion, Nasdaq was asked to file Form 25, and IAS filed Form 15-12G on 2 January 2026. | 2, 5 | direct record | supported as stated | high | [83, 437, 244] |
| ATT-IAS-F01 | In January 2026 MRC removed 'IAS Attention Measurement', and separately 'IAS In-Game Measurement', from in-process status, stating the reason as IAS' lack of intent to engage in third-party audit and the resulting withdrawal from the accreditation process (MRC report dated 17 April 2026). | 2, 5, 11, 12 | direct record | supported as stated | high | [49] |
| ATT-IAS-F03 | On 29 September 2026 MRC's accredited digital list shows attention metrics only for DoubleVerify's Quality Analytics Platform; IAS's entries cover tracked ads, rendered and viewable impressions, viewability, SIVT, property-level verification, SSAI and pre-bid viewability and IVT, and MRC's June 2026 decisions grant IAS pre-bid viewability and IVT accreditation, with no attention metric in any IAS entry. | 2, 5, 12 | direct record | supported as stated | high | [2, 205] |
| ATT-DOUBLEVERIFY-F06 | No pre-bid attention offering is accredited: the April 2026 letter's pre-bid table lists brand safety and suitability, fraud/IVT, viewability, fully on-screen, contextual, app-store, language, site/app and URL-keyword offerings but no attention offering, and MRC's digital list shows DV's 'Attention Metrics' line without the 'Inclusive of pre-bid data' annotation its viewability lines carry. | 2, 6 | direct record | supported as stated | high | [42, 2] |
| ATT-TVISION-F10 | On 29 September 2026, no TVision or Viant attention product appears on MRC's accredited-services lists for television, digital or multimedia, or on MRC's services-under-review lists for television or digital. The only 'Attention Metrics' accreditation listed is DoubleVerify's, for desktop, mobile web and mobile in-app. | 2, 6 | direct record | supported as stated | high | [2, 302, 50, 48] |
| ATT-F-F25 | The CIMM/IAB Attention Measurement Playbook for Marketers (November 2025) tells buyers not to treat attention as a currency, summarises IAB and MRC's position as 'never as a stand-alone currency', quotes the ARF's Scott McDonald that attention-based currencies are 'premature', and repeats an ARF figure of about 50% of advertisers and 60% of agencies using attention metrics without giving a sample, date or question. | 2, 6, 7 | direct record | supported as stated | high | [438] |
| ATT-ADELAIDE-F01 | On 29 September 2026, MRC lists Adelaide's AU among digital services under review with 'Pre-audit reviewed 5/2026, Audit not begun', scoped to AU across digital display and video, audio (podcasts, streaming and terrestrial), linear TV, cinema and digital OOH. | 2, 6, 12 | direct record | supported as stated | high | [48] |
| ATT-VEVO-F09 | No independent audit of Attention Guaranteed delivery was found, and the guaranteed metric is not accredited: on 29 September 2026 MRC lists Adelaide AU as 'Pre-audit reviewed 5/2026, Audit not begun'. | 2, 6, 12 | synthesis | supported as stated | high | [48, 78] |
| ATT-LUMEN-F15 | On 29 September 2026 Lumen has no entry on MRC's Accredited Services: Digital page or on MRC's digital and television Services Under Review pages, and is not among the vendors in CESP's attention grid of 10 June 2026. | 2, 6, 12, appendix | direct record | supported as stated | high | [2, 48, 3] |
| ATT-XPLN-F18 | MRC's own record lists XPLN as 'Pre-audit reviewed 9/2025, Audit in process as of 11/2025' for 'Display and Video Impressions, Viewable Impressions and viewability metrics, and Attention metrics, all inclusive of SIVT filtration - Mobile Web and Mobile In-App' (grade A). This is an audit in progress confirmed by the accreditor. It covers mobile environments only, not desktop, CTV or social, and the record does not name the standards applied. | 2, 8 | direct record | supported as stated | high | [48] |
| ATT-DOUBLEVERIFY-F04 | DoubleVerify's CTV attention is not MRC-accredited: the April 2026 letter accredits CTV only for impressions, video viewability, completion and quartile metrics and property-level verification, lists attention metrics only for desktop, mobile web and mobile in-app, and on 29 September 2026 MRC's digital and television under-review lists carry no DoubleVerify entry. | 2, 8 | direct record | supported as stated | high | [42, 2, 48] |
| ATT-IAS-F02 | As stated, plus: IAS Attention was not on the page on 20 June 2025, so it entered MRC's pre-audit list between 20 June and 16 September 2025 and was gone by 20 January 2026; it was listed for at most about seven months and never reached audit. | 2, 8 | direct record | supported with qualification | high | [233, 48] |
| ATT-ADELAIDE-F03 | In October 2024 MRC told Digiday that Adelaide was 'in the pre-assessment phase (not yet audit)', and Adelaide's CEO said the seventh version of AU was the version under MRC scrutiny, then available to a limited number of clients with a full rollout hoped for by Q1 2025; MRC's record shows the pre-audit reviewed in May 2026, about 19 months later (derived), with the audit not begun at 29 September 2026. | 2, 12 | synthesis | supported as stated | high | [339, 48] |
| ATT-DOUBLEVERIFY-F02 | In March 2024 MRC continued accreditation of DoubleVerify's display and video Authentic Attention metrics in desktop, mobile web and mobile in-app, and denied accreditation for pre-bid Authentic Attention in the same environments. | 2, 12 | direct record | supported as stated | high | [210] |
| ATT-DOUBLEVERIFY-F05 | The standards MRC lists for DoubleVerify's April 2026 accreditation are viewability, invalid-traffic, impression, audience-measurement, outcomes and data-quality, and ad-verification standards, and do not include the IAB/MRC Attention Measurement Guidelines (Version 1.0, November 2025), to which a DV vice president of attention product management is a named key contributor. | 2, 12 | direct record | supported as stated | high | [42, 41] |
| ATT-DOUBLEVERIFY-F22 | DoubleVerify describes its own accreditation scope inconsistently with MRC's record: the product page says 'only direct indexed impressions', a November 2024 release dates first accreditation to 2023, and the 10-K calls the product 'MRC-accredited' without environment scope, while MRC's records show December 2022 accreditation in D, MW and MA and, from April 2026, modelled and non-viewable metrics as well. | 2, 12 | synthesis | supported as stated | high | [209, 214, 47, 215] |
| ATT-B-F04 | In a US and UK online experiment (1,013 participants, July-August 2020; random article-brand pairing; Lumen webcam eye tracking supplied in kind), one extra second of eye-tracked ad dwell was associated with a 3.43 percentage-point higher aided-recall probability and one extra second of MRC-standard ad visibility with 0.32 points; because the measures sit on different scales (SD 3.16 s versus 17.37 s), the per-standard-deviation effects are 10.83 versus 5.55 points. The full-model R-squared values (.143 on 3,925 eye-tracked observations versus .091 on 5,707) include extensive fixed effects and controls and come from different samples, so they are not an increment test; the authors conclude that visibility 'still produces reliable results'. | 3 | independent measurement | supported with qualification | medium | [51] |
| ATT-B-F05 | In the same experiment the purchase-proxy association was 0.73 percentage points per second of eye-tracked dwell (R-squared .136) against 0.13 points per second of visibility (R-squared .130), and the only cost figures offered are Lumen's own attentive-CPM estimates of about 21.88 GBP (desktop) and 13.54 GBP (mobile), not a tested buying outcome. | 3 | independent measurement | supported with qualification | medium | [51] |
| ATT-B-F06 | Using TVision panel data (1,155 paid consenting US households, 3,659 viewers, July 2016 to June 2017), McGranaghan, Liaukonyte and Wilbur found viewers absent from the room during 30% of ads played in active viewing sessions, and in 180 national ad slots from two NFL playoff games on 22 January 2017 eyes-on-screen attention explained slightly more brand-search lift (R-squared .116) than exposure tuning (.082), presence (.070) or opportunity-to-see tuning (.035), in separate single-measure regressions without a hold-out. | 3 | independent measurement | supported as stated | medium | [52] |
| ATT-B-F10 | In Venkatraman et al. (2015), the only screened study linking lab measures to TV advertising elasticities (37 30-second TV ads; market-response data for 26 ads from five firms, four of which supplied sales and GRP data and one its own elasticity estimates; eye tracking on 29 participants), eye-tracking measures added no significant explanatory power once traditional self-report measures were controlled (only fMRI did), although on their own eye tracking and EEG were moderate predictors; Zhang, Wedel and Pieters (2009), also screened, link gaze duration on print feature ads to sales. | 3 | independent measurement | supported with qualification | medium | [6] |
| ATT-B-F11 | The Kwon et al. 2021 meta-analysis carried from the earlier papers analyses media-context effects on attitudes and purchase intention (96 studies, 597 effect sizes, 139,233 participants, 31 years) and, on its abstract, includes no attention metric, so it is not evidence on attention-metric validity; its full text was not obtainable. | 3 | independent measurement | supported as stated | medium | [60] |
| ATT-B-F14 | Earlier independent academic eye-tracking work found fixations on the pictorial and brand, but not the text, promoted brand memory for print ads (88 consumers, 65 ads), and that gaze duration on grocery feature ads affected sales beyond the mere presence of the ad; neither compared attention with a time-in-view baseline or used a hold-out. | 3 | synthesis | supported as stated | low | [55, 56] |
| ATT-B-F16 | A frequently cited neuro and eye-tracking study reporting 82.9% test accuracy in classifying ads' YouTube view counts used 8 Super Bowl ads and a random 70/30 split of cases, which places the same ads in training and test sets for an ad-level outcome, so the accuracy cannot be read as prediction for new ads. | 3 | inference | supported as stated | medium | [54] |
| ATT-B-F17 | Held-out validation is feasible but rare in academic eye-tracking modelling: Martinovici et al. (2023) validated attention trajectories out of sample and ahead of period on 325 US consumers (85% brand-choice hit rate against 20% chance) and report that only one of the 21 prior attention-and-choice studies they reviewed used a classic two-group validation; none of the advertising-attention outcome studies screened here validated on held-out ads, people or time periods, except one whose case-level split leaks ad identity. | 3 | synthesis | contradicted | medium | [57, 46, 54] |
| ATT-B-F20 | Inference: in the 2016-17 TVision panel, attention decreased in the early slots of a break (confidence intervals overlap from the third slot on) and fell slightly with ad duration (an absolute 0.7 points for 30-second and 1.5 points for 60-second ads; ad duration explains 0.1% of attention variance), so optimising to a per-second attention rate could favour early slots and shorter ads without an outcome gain, whereas a total-attentive-seconds metric would favour longer ads; no screened study tests whether attention optimisation improves exposure quality or only the score. | 3 | inference | supported with qualification | low | [52, 5] |
| ATT-LUMEN-F29 | An independent academic experiment (Simonov, Valletti and Veiga; 1,013 US and UK participants, July-August 2020; funded by Imperial College and the NET Institute) used eye-tracking technology supplied by Lumen Research (terms not stated), retained only high-quality eye-tracking data (about 70% of observations), and concluded that the simpler visibility measure 'still produces reliable results'. | 3 | independent measurement | supported with qualification | high | [51] |
| ATT-C-F07 | Adelaide's page announcing its Lucid integration (16 March 2022) states that, across seven brands and 15 campaigns, high-AU media delivered on average +40% ad recall, +28% familiarity and +27% unaided awareness lift compared with low-AU media, and that 'structured A/B tests' of in-flight optimisation toward AU gave brand impact increases of up to 22% compared with viewability; no design, sample, cost or uncertainty is given. | 3, 4 | vendor assertion | supported as stated | medium | [65] |
| ATT-ADELAIDE-F27 | Figures as stated; the release does not establish a commissioning party or funding, so the record is a joint release by an agency trials unit and an interested DSP with funding unknown. | 3, 4 | vendor assertion | supported with qualification | low | [342] |
| ATT-AMPLIFIED-F15 | A joint Amplified-Chalice release (9 October 2023) says proof-of-concept trials in the US and Australia used Amplified's attention dataset to fuel a Chalice custom bidding algorithm and that Lucid measured a 70% ad-recall increase and a 14-point brand-lift increase versus the standard non-attention line item, with no sample, uncertainty or cost reported. | 3, 4 | vendor assertion | supported as stated | low | [195] |
| ATT-C-F13 | Lumen's CPG case (4 March 2024) reports a head-to-head test against a traditionally bought control campaign within the same budget in which its attention-bidding algorithm delivered +104% viewed ads and +110% view time at 40% lower CPM and 44% lower aCPM; the outcomes are Lumen's own attention estimates, so the test shows the score moving at lower cost, not a memory, brand or sales gain. | 3, 4, 12 | vendor assertion | supported as stated | high | [69] |
| ATT-C-F17 | IAS's Kimberly-Clark case (29 September 2025) reports two NCSolutions sales-lift tests with impressions split post hoc into above- and below-average Quality Attention: in Test 1 (Q2 2024, browser display only) there was no significant difference in sales lift between the segments, although IAS says higher attention correlated with stronger ROAS in both tests (no Test 1 ROAS figure is given); in Test 2 (Q4 2024, adding video and mobile in-app) above-average-attention households showed a 313% higher sales lift, and ROAS rose 78% 'with the addition of high-attention formats', which reads as Test 2 against Test 1 rather than high against low. | 3, 4, 12 | vendor assertion | supported with qualification | high | [11] |
| ATT-C-F29 | No study screened discloses measurement fees, and none reports a campaign's media premium for attention-bought inventory; cost appears as CPM or cost-per-completion changes (Lumen CPG, 40% lower CPM; MAGNA/Adlook, lower display CPM and 42% lower cost per completed video for AU pre-bid), cost-per-acquisition or cost-per-conversion claims (Adelaide: the 2026 guide's bank case, later given as 69% lower CPA; Haleon, 25% lower cost per conversion), undefined 'cost savings' (Adelaide 2024 guide), 'cost per success' (IAS page), ROAS (IAS/Kimberly-Clark; Adelaide pre-bid beta, 58% vs control) or a fixed budget (dentsu, Amplified/Audience Group); the Newsworks/Lumen report estimates, rather than observes, a 24% CPM premium for news-brand display over general display; iFIT acknowledges higher CPMs without quantifying them, and the Havas/Lumen study had no campaign CPMs. | 3, 4, 12 | synthesis | supported with qualification | medium | [69, 13, 67, 15] |
| ATT-B-F19 | None of the independent studies screened reports a commercial attention model's performance after model updates, calibration drift, calibration by channel or population, or transfer to new channels; their data vintages are 2016-17 (TVision panel), 2020 (Lumen webcam) and a recent lab feed study, so they validate only those vintages. | 3, 12 | synthesis | supported as stated | medium | [52, 51, 46] |
| ATT-TVISION-F24 | TVision panel data have been released to independent academics beyond the 2021 MSI paper. Crossref records two 2025 Journal of Advertising Research articles on attention in CTV and linear TV: Bhan, Rallabandi and Lee (online 23 September 2025; funder Yonsei University Research Fund) and Lee and Choi (online 18 December 2025; funder field 'JARxTVision Data Release and Call for Ideas'). The articles were not read. | 3, 12 | direct record | supported as stated | medium | [300, 301] |
| ATT-C-F01 | The News UK and Lumen release of 22 March 2022 (REG-002) reports that, across 980,000 impressions on The Sun's website (UK), ads in a relevant editorial context received 31% more attention per 1,000 impressions (as calculated by Lumen's model) and 28% longer average viewing time (1.8 to 2.3 seconds) than ads in a non-relevant context; the release reports no memory, brand or sales outcome. | 4 | vendor assertion | supported with qualification | high | [402, 62] |
| ATT-C-F02 | The News UK/Lumen release states its non-relevant-context baseline as '25 minutes per 1,000 impressions' in one sentence and '25 seconds' in the branded-context sentence, and the syndicated copy carries the same wording, so the inconsistency is in the release itself; only the percentage change (31%) should be cited, not a level. | 4 | inference | supported as stated | high | [402, 62] |
| ATT-C-F03 | Outbrain's page of 25 November 2019 (REG-001) reports in its body that native ads on premium news sites were more likely to be trusted (+31%), clicked (+16%) and lead to future purchases (+18%) than ads on social platforms, from responses of over 900 consumers in the UK, France and Germany, while its headline claims native ads on premium publisher sites are '44% more trusted' and a format comparison gives native recommendations 28% higher trustworthiness than social ads; none of the three trust figures is an attention measure, and none may be merged with the News UK/Lumen +31% attention figure (REG-002). | 4 | vendor assertion | supported with qualification | high | [21] |
| ATT-C-F04 | Nativo's release of 26 August 2025 (REG-003) reports that an Adelaide audit of about 25 million impressions over roughly six weeks in July-August 2025 gave its native formats an average 28.21 AU, 40% above Adelaide's native-format benchmark and 21% above its display benchmark; the release calls the audit 'independent', but Adelaide sells the AU metric, the funding and method are not stated, no report is linked, and no outcome was measured. | 4 | vendor assertion | supported as stated | high | [63] |
| ATT-C-F05 | Adelaide's annual Outcomes Guide releases report headline averages over selected case studies: 45 case studies with 40% upper-funnel lift, 53% lower-funnel lift and 37% cost savings (January 2024, building on an earlier 2023 edition); 52 case studies with 41% higher brand lift and 55% stronger lower-funnel impact for 2024 campaigns (January 2025); and 60 case studies with 33% upper-funnel and 53% lower-funnel lift for 2025 campaigns (January 2026). Adelaide cited 31% and 56% in August 2023. The averages cover different campaign years and undisclosed case sets and designs, so no trend can be read from them; the releases disclose no designs, holdouts, uncertainty or averaging method, and the guides are gated. | 4 | vendor assertion | supported with qualification | high | [13, 14, 15] |
| ATT-C-F06 | Adelaide's undated-on-page post (structured-data date 18 March 2020) claims that for a yogurt brand AU was 'over +180% more correlated' with Nielsen ROI than viewability, that video completion rate showed no correlation with purchase consideration for a technology brand, and that AU per impression correlated with conversion rate for an insurance brand; no sample sizes, coefficients, held-out data or uncertainty are given. | 4 | vendor assertion | supported as stated | medium | [64] |
| ATT-C-F09 | The Trade Desk-owned publication The Current reported on 3 October 2025 that iFIT, buying through The Trade Desk with Adelaide segments and Upwave brand-lift measurement across CTV, OTT, online video, audio and DOOH, achieved a 61% increase in aided awareness, 17% in consideration and 10% in favourability; Adelaide's own page of 30 July 2026 gives consideration as 18% and describes the lifts as 'compared with the control group' without saying whether the control was unexposed or bought on standard metrics; iFIT's executive says higher CPMs had to be justified by outcome lifts, without quantifying the premium. | 4 | vendor assertion | supported with qualification | medium | [67] |
| ATT-C-F10 | Havas Media Network commissioned, and published on 17 September 2024 with Lumen Research and Brand Metrics, a descriptive analysis matching Lumen's model-estimated attention (predicted from viewability data collected by a JavaScript tag) to Brand Metrics survey responses across a display dataset of 9,089 campaigns (9,087 in the cost note), 1,879,764 respondents and 5,618,088 viewable impressions, mostly collected in the UK; it found more aggregate attention associated with higher awareness, consideration, preference and action intent, had no control group, compared results with campaign averages, and had no campaign CPM data. | 4 | vendor assertion | supported with qualification | high | [10] |
| ATT-C-F12 | A Teads and Lumen whitepaper (2023) combined Lumen-predicted dwell time with Dynata brand lift for a selected sample of 16 campaigns from 14 advertisers across North America, APAC and EMEA, reporting online ad recall lift 40% higher for 5+ seconds of exposure than for 1+ second; the analysis covers a single exposure on Teads inventory, and the paper itself notes that the PwC review did not strongly support the attention-over-viewability hypothesis. | 4 | vendor assertion | supported as stated | high | [68] |
| ATT-C-F15 | DoubleVerify's 2021 Mondelez whitepaper reports a US static-display campaign in which Kantar's control/exposed brand-lift study, split by DV Authentic Attention quadrants, showed high-Exposure impressions lifting unaided awareness by 5 points versus 3 points (80% significance) for all DV Authentic Ads, with significance at 90% unless marked; no respondent count or cost is given, and the Exposure index itself includes viewable time. | 4 | vendor assertion | supported as stated | high | [71] |
| ATT-C-F19 | dentsu's Attention Economy Phase 2 report (2021) describes eye-tracking panels of 3,400 respondents built with Amplified Intelligence and 6,000 with Lumen, plus purchased TVision data from 5,500 US households, reports statistically significant recall and brand-choice uplifts against control groups that rise with dwell time, and claims one client's plan, re-optimised toward attention at the same budget, gained 3.7 years of extra attention in a month; that gain is modelled attention, not a measured outcome. | 4 | commissioned measurement | supported as stated | medium | [9] |
| ATT-C-F20 | TVision and Upwave (18 July 2022) report a linear regression across over 40 US campaigns from Q4 2019 to Q1 2022 in which each percentage-point increase in TVision ad attention (share of ad impressions with 2+ seconds of looking at the screen) for high-frequency ads went with close to a 1% lift in Upwave aided awareness; no uncertainty, sample of respondents or held-out test is reported. | 4 | vendor assertion | supported as stated | high | [74, 439] |
| ATT-C-F21 | The Wurl/TVision '2.1x' carried from earlier papers traces to a Wurl and TVision report published 31 July 2025 covering 50+ CTV advertiser campaigns: as reported by The Measure, attentive time rose to 71% of the ad from 33% when ads were emotionally aligned with surrounding content rather than mismatched; the release's own key finding is a 2.4x rise (27% to 66% of the ad) for 30-second ads; TVision's co-founder and CEO later gave roughly 30% and 70% in a Wurl interview; the report is gated and the outcome is attention only. Wurl, which sells the contextual targeting tested, was owned by AppLovin at the time of the release. | 4 | vendor assertion | supported with qualification | medium | [16, 17, 18, 440] |
| ATT-C-F23 | Amplified's undated case page (tagged Australia, media buying) says Audience Group's custom bidding algorithm, built with Chalice AI and tied to an online brand study, first lifted ad recall 14%, and that after an Amplified attention pre-bid score was added, letting it bid higher amounts for placements with more seconds of attention, brand recall rose 74% 'without spending any extra budget'; no control group or uncertainty is described and the page says the tactic cannot be isolated in business results. | 4 | vendor assertion | supported with qualification | low | [76] |
| ATT-ADELAIDE-F11 | Adelaide reported that beta partners using its pre-bid segments in The Trade Desk (August-September 2024) saw a 58% increase in cumulative ROAS, a 12% improvement in attentive media quality and a 62% increase in CTR compared with control groups, with no disclosed design, sample, control definition or cost. | 4 | vendor assertion | supported as stated | medium | [328, 161] |
| ATT-KANTAR-F16 | In Kantar's LIFT+ database of 873 multi-media campaigns over five years (over $3.2 billion in media spend), Kantar reports that channel attention levels per impression do not seem to be directly related to channel-level campaign cost-effectiveness; TV had the highest brand contribution but lower cost-effectiveness. | 4 | vendor assertion | supported as stated | medium | [258] |
| ATT-LUMEN-F13 | As stated, adding that both attention and viewability were significantly related to the three memory metrics, with attention showing higher z-values but no significant advantage. | 4 | commissioned measurement | supported with qualification | high | [8, 283] |
| ATT-LUMEN-F25 | The Outbrain and Lumen study published 25 November 2019 (REG-001) reports that native ads on premium news sites were more likely to be trusted (+31%), clicked (+16%) and lead to future purchases (+18%) than ads on social platforms, from over 900 consumers in the UK, France and Germany; it is a trust survey with no attention measure. | 4 | vendor assertion | supported as stated | high | [21] |
| ATT-TVISION-F17 | The Wurl/TVision '2.1x' figure comes from a Wurl and TVision report on 50+ CTV advertiser campaigns, released 31 July 2025. Its denominator is the share of ad duration attended: as reported by The Measure, attentive time was 71% of the ad for emotionally aligned placements against 33% for mismatched ones. The release's own key finding is a 2.4x rise (27% to 66%) for 30-second ads. The outcome is attention only, measured on the TVision panel; the full report is gated and was not read. | 4 | vendor assertion | supported as stated | medium | [17, 18] |
| ATT-ADELAIDE-F17 | Adelaide's AU Ecosystem had 54 participating partners on 23 October 2025, including publishers offering guaranteed minimum AU scores (the WSJ named, with six attention-guaranteed campaigns in the 2024 election cycle) and publishers (Condé Nast, Hearst) covering the cost of AU measurement; by 8 July 2026 Adelaide stated the ecosystem had more than 70 sell-side members and over 50 buy-side partners. | 4, 5 | vendor assertion | supported as stated | medium | [167, 176] |
| ATT-ADELAIDE-F15 | Google announced AU custom bidding in DV360 at the 2025 NewFronts and announced a packaged 'Adelaide AU' algorithm for mid-July 2026; by 29 September 2026 Google's DV360 Help Center documents the live activation steps for the 'Maximize Adelaide AU' algorithm and the AU multiplier, with an additional media fee. The packaged integration is live per platform documentation; only the exact go-live date is undocumented. | 4, 5, 6, 7, 12 | synthesis | contradicted | medium | [330, 159] |
| ATT-TVISION-F15 | Viant's CEO said on the 10 August 2026 call that 42 pilot campaigns targeted TVision high-attention inventory segments, that over 80% of them achieved higher conversion rates than the benchmark, and that the average lift was 1.4x. No design, benchmark definition, control or cost was given. | 4, 6 | vendor assertion | supported as stated | low | [304] |
| ATT-F-F19 | Haleon, with its agency Publicis, evaluated Amazon DSP inventory for Flonase in the US in mid-2023 with Adelaide AU (high at 60 or more for CTV, low under 20). AdExchanger (22 Feb 2024) paraphrases Haleon's position as attention being 'just another tool' for assessing media quality. Its effectiveness lead Phil Jackson is quoted directly that 'we need to prove [that] attention drives incremental business outcomes' and warned that the highest-attention CTV carries very high CPMs. | 4, 7 | vendor assertion | supported with qualification | medium | [441] |
| ATT-C-F28 | Of the 29 study records coded in this stream, none qualifies as independent measurement, three are commissioned measurement (PwC for Lumen; dentsu's Attention Economy programme; the Havas-commissioned study with Lumen and Brand Metrics) and 26 rest on vendor, partner, agency or media-owner publications; outcomes were often measured by named third parties (Kantar, NCSolutions, Dynata, Upwave, Lucid, Brand Metrics, Nielsen), but in every case the result reaches the public through an interested party, and no independent replication of any vendor claim was found in the sources screened. | 4, 12 | synthesis | supported with qualification | medium | [8, 71, 66] |
| ATT-ADELAIDE-F09 | Between February and June 2025, MediaSense, at Adelaide's request, reviewed AU's methodology and its AU Curve Materiality Analysis and rebuilt AU-to-outcome curves for aided awareness, familiarity, recall and intent from sampled data supplied by Adelaide; MediaSense did not independently verify Adelaide's data or documents, did not review AU source code, and noted limited data at the ends of the AU scale. | 4, 12 | commissioned measurement | supported as stated | medium | [66, 170] |
| ATT-D-F01 | Oracle ended ('End-Of-Life') all Oracle Advertising products and support, including Moat Analytics and Moat Reach, effective 30 September 2024. Its first dated end-of-life FAQ is Version 1 of 17 June 2024, and Moat Analytics access ceased at each customer's contract end or no later than 30 September 2024. | 5 | vendor assertion | supported as stated | high | [87, 88] |
| ATT-D-F02 | GumGum announced its acquisition of Playground xyz, an attention measurement and optimisation platform, on 16 December 2021, for a mix of cash and stock at an undisclosed price. The release described the platform's AI as trained and verified using real eye-tracking data. | 5 | vendor assertion | supported as stated | high | [80] |
| ATT-D-F03 | GumGum announced on 23 April 2024 (London dateline) that it had merged Playground xyz's APAC media business into GumGum while keeping the Attention Intelligence Platform as a standalone data solution, and on 3 October 2024 (Santa Monica dateline, PRWeb) it launched the GumGum Platform, in which the platform was renamed GumGum Attention. | 5 | vendor assertion | supported with qualification | medium | [442, 85] |
| ATT-D-F04 | At 28 September 2026 GumGum no longer presents GumGum Attention as a standalone product on its website. gumgum.com/attention redirects to a results page that presents attention through 'Attentive Lift' and receptivity, and playgroundxyz.com redirects to the GumGum homepage. | 5 | inference | supported as stated | medium | [443] |
| ATT-D-F06 | IAS's FY2024 10-K, its last 10-K before the take-private, describes Quality Attention as a metric built from time-in-view, share of screen and ad interaction, launched 'with eye tracking' in 2024. IAS charges a CPM on the volume of ads analysed. In the passages read the 10-K gives no attention-specific revenue, pricing or customer count. | 5 | direct record | supported as stated | medium | [231] |
| ATT-D-F11 | DoubleVerify's revenue grew from $452.4 million (2022) to $748.3 million (2025); 2025 measurement revenue was $249.7 million, up 10% partly because of the Rockerbox acquisition; in Q2 2026 revenue of $193.8 million (up 3%) split into activation $107.7 million, measurement $66.8 million and supply-side $19.3 million, while DV's average Measured Transaction Fee fell 3% in 2025 and 7% in Q2 2026; customers contributing at least $1 million a year rose from 93 (2023) to 110 (2024) and 131 (2025). None of these series is attention-specific. | 5 | direct record | supported with qualification | high | [444, 206, 207, 47] |
| ATT-D-F13 | IAS completed its take-private on 23 December 2025 and deregistered its stock (Form 15, 2 January 2026), and DV agreed on 6 August 2026 (226 days later) to a cash merger after which, per the parties, it will be privately held and delisted; if that deal closes, neither company is expected to publish SEC periodic reports, which would remove the main public financial series for testing H4 (an inference about future disclosure). | 5 | synthesis | supported with qualification | medium | [83, 84] |
| ATT-D-F15 | TVision Insights Inc., a Delaware company based in Cambridge, Massachusetts, reported selling $12.0 million of a $16.0 million equity offering (first sale 9 November 2022) to five investors in a Form D filed on 23 November 2022, its last Form D before the Viant acquisition. | 5 | direct record | supported as stated | high | [445, 436] |
| ATT-D-F16 | Life360 announced its acquisition of Nativo on 10 November 2025, valued at about $120 million in cash and stock, after an agreement dated 9 November 2025, and closed on 2 January 2026. Life360 recorded purchase consideration of $104.0 million ($75.9 million cash and $28.0 million, or 435,599 shares, of stock) and $39.0 million of goodwill. Its acquisition disclosures do not mention attention. | 5 | direct record | supported as stated | high | [446, 447] |
| ATT-D-F17 | Outbrain Inc. changed its corporate name to Teads Holding Co. effective 6 June 2025 and planned to change its ticker from OB to TEAD on or about 10 June 2025. | 5 | direct record | supported as stated | high | [448] |
| ATT-D-F18 | Realeyes announced the launch of Adverteyes as a creative intelligence company on 18 November 2025, with Realeyes' CEO as Adverteyes' executive chairman; the release discloses no ownership, stake or funding. Adverteyes' own legal pages identify the operating entity as Adverteyes Inc., a Delaware corporation, which explains the absence of an active UK company of that name; who owns Adverteyes Inc. is not disclosed in any source read. Realeyes' UK operating company is controlled (75% or more) by Realeyes (Holdings) Limited. | 5 | vendor assertion | supported with qualification | medium | [86, 449, 450] |
| ATT-D-F19 | At 28 September 2026 SPARK Neuro's homepage sells dementia diagnosis and care ('It's Time to Transform Dementia Care') and makes no advertising offer, which fits a pivot away from neuro-based ad testing. The pivot date was not established. | 5 | vendor assertion | supported as stated | medium | [451] |
| ATT-D-F20 | Smart Eye AB, listed on Nasdaq First North since 7 December 2016, owns 100% of Affectiva, Inc. (Boston) and iMotions A/S (Copenhagen). In January 2025 iMotions fully integrated Affectiva's Media Analytics business. Net sales of the Behavioral Research business area (Applied AI and iMotions) were SEK 208.8 million in 2025, against SEK 221.2 million in 2024, out of group net sales of SEK 403.7 million. | 5 | direct record | supported as stated | high | [452] |
| ATT-D-F22 | Lumen Research Ltd has no registrable person with significant control (statement notified 15 February 2024) and reports no ultimate controlling party. For the year to 30 September 2025 the group loss shown in the consolidated statement of changes in equity was £3.76 million (prior year £1.13 million) and the company loss £3.57 million (prior year £2.27 million); group net assets were £1.38 million (restated prior year £2.42 million) and average employees 68 (59). Turnover is not disclosed because the income statement was not filed. An outstanding charge in favour of Salica Growth Debt Fund I SPV (created 14 March 2025) indicates debt financing. | 5 | direct record | supported with qualification | high | [427, 285] |
| ATT-D-F23 | Lumen Research Ltd shortened its accounting period from 30 September 2026 to 31 December 2025 (filed 24 December 2025) and filed group accounts for the three months to 31 December 2025, showing a group loss of £240,527 (company £212,498) with a £289,421 share-based payment credit, average employees of 77, and still no ultimate controlling party; amended accounts for that period were filed on 27 June 2026. | 5 | direct record | supported with qualification | high | [427, 453] |
| ATT-D-F24 | DAIVID Limited (UK 12583663, incorporated 4 May 2020) has one person with significant control, an individual holding 75% or more of shares and votes. It files micro-company accounts (latest to 31 May 2025), which disclose no revenue. | 5 | direct record | supported as stated | high | [454] |
| ATT-D-F26 | Tobii acquired substantially all the assets of Sticky, a San Francisco-based webcam eye-tracking research provider in which it already held a minority stake, in a deal announced on 17 May 2017. At 28 September 2026 the product is sold as 'Tobii Sticky', a webcam-based platform for testing ads, packaging and UX. | 5 | vendor assertion | supported as stated | high | [79, 455] |
| ATT-D-F27 | Amplified Intelligence, headquartered in Adelaide, South Australia, rebranded as Amplified on 29 April 2025 (its own announcement); neither Amplified nor Neurons (Tåstrup, Denmark) discloses revenue in the sources read, and Neurons' last round found is a EUR 6 million seed reported on 25 August 2022, led by Fairpoint Capital with Vaekstfonden and Finance Zealand. | 5 | vendor assertion | supported with qualification | low | [456, 457] |
| ATT-U-F04 | GumGum Attentive Lift is a brand-lift reporting product that joins GumGum's attention data to On Device's impression-level brand-lift data via On Device's API. The December 2025 release documents early adopters in Australia, and GumGum's 28 July 2026 release announces a Canadian launch and cites momentum in Australia and the UK; no US availability is documented in the sources opened. gumgum.com/attention redirects to a Campaign Results page that sells Mindset targeting and says impressions, viewability and attention scores 'measure what happened to an ad' but not whether the audience was receptive. | 5 | vendor assertion | supported with qualification | medium | [458, 459] |
| ATT-U-F15 | MRC's digital services-under-review page (accessed 28 September 2026) lists XPLN with an attention audit in process since 11/2025 (display and video impressions, viewability and attention metrics, mobile web and in-app) and Adelaide AU with a pre-audit reviewed 5/2026 and the audit not begun (digital display and video, audio, linear TV, cinema and digital OOH); no other attention vendor appears on that page. | 5 | direct record | supported as stated | high | [48] |
| ATT-U-F20 | Adverteyes was spun out of Realeyes (announced 18 November 2025 in the PRWeb release; the adverteyes.ai blog copy is dated 26 November 2025). It sells PreView, which uses AI/ML models built on a human database of attention, emotion and sales data to predict which creative will drive sales; PreView offers an API, a Chrome extension and an MCP product, is marketed as 'purpose-built for agentic advertising', and is a partner in Nielsen's Outcomes Marketplace. Realeyes' own site now sells VerifEye human verification and Vision AI training-data licensing, not creative testing. | 5 | vendor assertion | supported with qualification | medium | [460, 461] |
| ATT-ADELAIDE-F19 | Adelaide announced on 19 August 2024 an all-stock acquisition of Rita, a consumer-data app company with about 200,000 members that uses EU data-access rights to collect behavioural data, to add deterministic behavioural, exposure and outcome data to AU models; no price was disclosed. | 5 | vendor assertion | supported as stated | medium | [341] |
| ATT-ADELAIDE-F20 | Adelaide announced a 7M USD seed (June 2022; total raised stated as 9.5M) and a 1.4M USD extension (August 2024), about 10.9M USD announced (derived); its Form D notices report equity offerings of 9.48M USD (2022) and 3.0M USD (2024), with 4.59M USD sold as of the filing dates, and decline to disclose revenue. | 5 | vendor assertion | supported with qualification | medium | [174, 81] |
| ATT-ADELAIDE-F21 | Adelaide Metrics Inc. is a Delaware corporation incorporated in 2020 (Form D), with its principal place of business in New York (622 Greenwich St per the Form D; 228 Park Ave S PMB per site data); Adelaide dates its launch to 2019. | 5 | vendor assertion | supported with qualification | medium | [105, 334] |
| ATT-ADELAIDE-F25 | AU became available through Nielsen's Outcomes Marketplace in Nielsen ONE on 7 October 2025, with Nielsen-built metrics combining AU with Nielsen reach data for mid-flight efficiency optimisation; this is a documented data-access route for AU outside Adelaide's own dashboard. | 5 | vendor assertion | supported as stated | medium | [331] |
| ATT-AMPLIFIED-F01 | Amplified's ABN 21 636 344 360 is registered to Amplified Intelligence Technologies Pty Ltd, an Australian private company (ACN 636 344 360) whose ABN has been active since 20 September 2019 with main business location SA 5000, and Amplified's site footer carries the same ABN. | 5 | direct record | supported as stated | high | [462, 456] |
| ATT-AMPLIFIED-F02 | Amplified traces its origin to Centre for Amplified Intelligence Pty Ltd, whose ABN has been active since 6 July 2017 (ABR). That company was renamed Amplified Intelligence Pty Ltd on 1 October 2019 and still exists alongside the operating entity. The brand changed from Amplified Intelligence to Amplified with effect from 29 April 2025. | 5 | vendor assertion | supported with qualification | medium | [456, 193, 463] |
| ATT-AMPLIFIED-F03 | Amplified's disclosed funding is vendor-stated only and internally inconsistent: A$3M post-seed led by TEN13 (30 September 2021, its first external capital), a '$6 million funding round' in 2024 per its December 2024 release, '$5M Funding Secured' placed under 2025 on its timeline, and '$10M+ secured in funding over the last 5 years' on its investor page. | 5 | vendor assertion | supported as stated | low | [204, 201, 193, 191] |
| ATT-AMPLIFIED-F30 | Amplified acquired AttentvAds on 13 December 2022, describing it as a company tracking and verifying impression-level data across CTV, mobile, web and in-app whose team had built mobile verification at IAS and the framework for the Open Measurement SDK; the acquisition preceded the 2023 launch of Live Attention Measurement (inference on sequence only). | 5 | vendor assertion | supported as stated | medium | [464, 193] |
| ATT-DOUBLEVERIFY-F16 | On 6 August 2026 DoubleVerify agreed to be acquired for $13.60 per share in cash (enterprise value about $2.15 billion) by Neptune BidCo US Inc., the parent of Nielsen Company (US), LLC and an affiliate of funds managed by Elliott and Brookfield; closing requires stockholder approval, HSR expiry and specified foreign clearances, termination fees are $60 million (DV) and $144 million (Parent), and the initial termination date is 6 May 2027, automatically extendable to 6 August 2027. | 5 | direct record | supported as stated | high | [84, 465, 466] |
| ATT-IAS-F05 | The 10-K names attention in the measurement-revenue definition and, separately, lists attention among its optimization (programmatic buying) solutions; attention revenue is not separable and, for pre-bid Quality Attention Optimization, would fall under optimization revenue by the 10-K's own definitions (inference). No attention-specific revenue, price, volume or customer count is disclosed. | 5 | direct record | supported with qualification | high | [231] |
| ATT-IAS-F06 | IAS's last periodic report, the Q3 2025 10-Q filed 4 November 2025, mentions attention once as a product capability ('viewability and attention' among what its pre-bid and post-bid solutions measure) and gives no attention-specific disclosure; FY2024 revenue was $530.1 million and nine-month 2025 revenue $437.6 million, company-wide. | 5 | direct record | supported as stated | high | [383, 467] |
| ATT-KANTAR-F01 | At 29 September 2026 Kantar Marketplace lists LINK+, Context Lab and LINK AI as separate products with separate US list prices: LINK+ from $4,500 self-serve and $8,080 serviced; LINK AI from $500 self-serve and $1,780 serviced; Context Lab from $13,950, serviced only; LINK+ and LINK AI can also be bought as a bundled subscription. | 5 | vendor assertion | supported as stated | high | [246, 247, 248] |
| ATT-KANTAR-F22 | As stated for the 2019 agreement and WPP's FY2025 figures; the split at the freeze is stated in Kantar's audited Annual Report and Accounts 2025: Bain Capital indirectly 60%, WPP affiliates 40% (with senior management equity at Midco level), ultimate controlling party Bain Capital Europe Fund V, SCSp; no change reported in the H1 2026 results. | 5 | direct record | supported with qualification | high | [468, 269] |
| ATT-KANTAR-F23 | Kantar Global Holdings S.à r.l. completed the sale of Kantar Media (audience measurement, data analytics and media planning) to H.I.G. Capital on 1 August 2025 for total consideration of $924.1 million (gain on disposal $244.4 million), announcing completion on 4 August 2025; Kantar Media is a different business from creative testing. | 5 | vendor assertion | contradicted | high | [265] |
| ATT-LUMEN-F03 | For the year to 30 September 2025 the group loss was £3,761,807 as originally filed (prior year £1,134,773), the parent company loss £3,567,595 (prior year £2,265,902), share-based payments credited to reserves £2,720,974 (prior year £78,592), group net assets £1,383,134 and average employees 68 (prior year 59); turnover is not disclosed because the income statement was not filed, and the accounts are unaudited. | 5 | direct record | supported as stated | high | [285] |
| ATT-LUMEN-F04 | Derived: share-based payments equal about 72% of the FY2025 group loss as filed; if the whole £2,720,974 credit to reserves was charged to profit or loss in the year, as the accounting policy describes, the group loss before that non-cash charge was about £1.04 million (about £1.09 million on the restated loss). | 5 | inference | supported as stated | medium | [285, 288] |
| ATT-LUMEN-F05 | As stated, adding that the Innovate UK loan was already in breach at 30 September 2025, so the FY2025 comparatives were also restated to show it as due within one year. | 5 | direct record | supported with qualification | high | [288] |
| ATT-LUMEN-F06 | Group cash fell from £1,920,236 at 30 September 2025 to £1,147,844 at 31 December 2025, while average employees rose from 59 (FY2024) to 68 (FY2025) and 77 (October-December 2025); the accounting period was shortened to 31 December 2025 'to align the year end with other group companies'. | 5 | direct record | supported as stated | high | [288, 453, 469] |
| ATT-LUMEN-F08 | Digiday reported on 15 November 2022 that Avocet (a DSP) and Lumen were combining under the Lumen name with a £3 million ($3.5 million) investment led by FirstPartyCapital, aiming to expand attention measurement in the US; Lumen's FY2025 accounts list Avocet Systems Ltd as a 100% subsidiary. | 5 | synthesis | supported as stated | medium | [398, 285] |
| ATT-TVISION-F01 | On 14 April 2026, Viant Technology agreed to acquire TVision Insights Inc. by merging its subsidiary TII Merger Sub Inc. into TVision. Consideration at closing is $22.5 million in cash plus 1,656,701 Viant Class A shares, a number set by an agreed equity value of $17.5 million. The shares are locked up: 50% become transferable after 6 months and 50% after 12 months. Viant's release put the total at $40.0 million, subject to adjustments and hold-backs. | 5 | direct record | supported as stated | high | [435, 308] |
| ATT-TVISION-F03 | As stated, describing the 424B5 as a preliminary prospectus supplement (16 September 2026) for a selling-stockholder offering withdrawn on 17 September 2026 (8-K Item 7.01). | 5 | synthesis | supported with qualification | high | [82, 470, 308, 314, 309] |
| ATT-TVISION-F05 | Before the acquisition, TVision panel data were licensed or integrated into other attention and measurement products. Adelaide (January 2023): TVision linear and CTV attention data in AU, plus pre-bid CTV segments. Lumen (April 2023): TVision metrics in the Lumen platform; in the US, LG Ad Solutions measurement runs through the TVision pixel (February 2024); Nexxen CTV aPMPs with TVision data (August 2023). DoubleVerify (June 2023): planned fusion with TVision eyes-on-screen signals, closed beta anticipated in H2 2023. xpln.ai (April 2026): TVision data as an input to its CTV models. VideoAmp: a linear-data licence expanded to CTV calibration data on 11 February 2026. | 5 | vendor assertion | supported as stated | medium | [294, 295, 290, 289, 220, 296, 293] |
| ATT-TVISION-F16 | Viant's CFO said TVision is 'pretty de minimis from a revenue perspective', and estimated its drag on 'EBITDA margins' at about 150 bp in Q2 2026 (actual) and about 200 bp in the Q3 2026 guide (forecast). The 10-Q says TVision's results were not material. | 5 | synthesis | supported with qualification | medium | [304, 82] |
| ATT-D-F10 | DoubleVerify's filings describe two acquisitions in 2023 to 2025: Scibids Technology SAS (August 2023, AI campaign optimisation) and Rockerbox, Inc. (13 March 2025, marketing attribution, $82.3 million net of cash acquired). The filings describe neither as an attention-measurement acquisition, and no DV acquisition was made in 2026 before the merger agreement. | 5, 6 | direct record | supported as stated | medium | [47, 207] |
| ATT-IAS-F11 | Quality Attention itself is documented for open-web display and video, including mobile in-app. Social attention is a Lumen model reported through IAS (platforms unnamed; stated live from January 2025) and Snap Attention Measurement is a separate bespoke score (announced June 2025, Q3 2025 target, live status unconfirmed); both rest on Lumen's model rather than the Quality Attention model. No IAS source documents attention measurement in CTV, audio, in-game, DOOH or retail media. | 5, 8 | synthesis | supported with qualification | medium | [115, 237, 381, 379, 380] |
| ATT-DOUBLEVERIFY-F14 | DoubleVerify prices its solutions as a Measured Transaction Fee per thousand impressions measured, mostly fixed, and reports revenue only by customer type; its FY2025 10-K, Q2 2026 10-Q and merger proxy disclose no attention-specific revenue, price or adoption, and the Q2 2026 10-Q does not mention Authentic Attention. | 5, 12 | direct record | supported as stated | high | [47, 207] |
| ATT-DOUBLEVERIFY-F15 | DoubleVerify's revenue was $748.3 million in 2025 (measurement $249.7 million, activation $427.3 million, supply-side $71.3 million), on about 9.5 trillion Media Transactions Measured at an average fee of $0.07 per thousand; Q2 2026 revenue was $193.8 million (measurement $66.8 million). | 5, 12 | direct record | supported as stated | high | [47, 206, 207] |
| ATT-D-F29 | Coded per shared rule 11, the attention-market events found for 2017 to the cutoff include: acquisitions of specialists by non-verification buyers (Tobii-Sticky 2017, Smart Eye-Affectiva June 2021, Smart Eye-iMotions November 2021, GumGum-Playground xyz 2021, Viant-TVision 2026); acquisitions by specialists (Lumen-Avocet 2022, Adelaide-Rita 2024); the acquisition of IAS by Novacap funds (a take-private, 2025); a pending acquisition of a verification company by a measurement company (Nielsen-DV, 2026); product retirements (Oracle Moat 2024; GumGum Attention as a standalone product, inferred from the website); a restructuring announced as a spin-out (Adverteyes, 2025); a pivot inferred from the current website (SPARK Neuro, date unknown); rebrands (GumGum Attention, October 2024; Amplified, April 2025; Teads Holding Co., June 2025); and trading-table acquisitions (Outbrain-Teads, completed February 2025; Life360-Nativo, January 2026). No source found attributes any of these events to a failure of attention metrics. | 5, appendix | synthesis | supported with qualification | medium | [87, 80, 82, 83, 465, 86] |
| ATT-E-F02 | The Wall Street Journal offered advertisers (the sources say 'advertisers', not 'US advertisers') an Adelaide AU guarantee for qualified campaigns with impressions measured between 15 October and 15 November 2024, timed to the US election; campaigns missing it would be offered a makegood of added-value impressions. The publisher later reported that all five participating clients, running six campaigns, exceeded the benchmark. Adweek described it as a promotion; no numeric threshold or benchmark is readable in the public sources. | 6 | vendor assertion | supported with qualification | medium | [90, 91, 92, 167] |
| ATT-E-F04 | On 13 September 2023 TripleLift announced Guaranteed AU deals for CTV, 'certified by Adelaide'. It called itself the first SSP to offer guaranteed deal packages based on AU, sold through its sales representatives. The release states no numeric threshold, remedy or price. | 6 | vendor assertion | supported as stated | medium | [94] |
| ATT-E-F05 | Adelaide announced on 15 September 2026 that Duolingo is launching AU Guarantees, described as a contractual commitment to a minimum level of media quality across its in-app inventory. The announcement follows a March 2026 audit that scored Duolingo's mobile in-app OLV inventory 34% above Adelaide's benchmark. It states no threshold or remedy, and AU reporting is only planned. | 6 | vendor assertion | supported as stated | medium | [97] |
| ATT-E-F08 | Teads Holding Co.'s FY2025 10-K (filed 16 March 2026) and Q2 2026 10-Q (filed 7 August 2026) contain no mention of an attention guarantee or an attention-priced unit. The filings list CPC, CPM, CPV, CPA, video completion-based pricing and other outcome-based formats as buying and monetisation models, none defined on attention, and describe attention mainly as something the platform predicts or optimises. | 6 | direct record | supported with qualification | high | [96, 471] |
| ATT-E-F10 | Yahoo launched Adelaide AU high-attention pre-bid segments in the Yahoo DSP on 13 June 2023, as a 'global solution', with Verizon as the brand launch partner. This is an optimisation use in a named US buying platform. | 6 | vendor assertion | supported as stated | medium | [334] |
| ATT-E-F11 | Adelaide AU pre-bid segments are documented as live in two more named DSPs. The Trade Desk offered low, medium and high attention tiers from 8 October 2024, after an August to September 2024 beta. Amazon DSP offered 'AU Media Quality' tiers and an 'AU Quality Floor' from 9 June 2026; the Floor excludes MFA sites and the bottom 10% of AU-scored placements across Display, Online Video and Streaming TV. | 6 | vendor assertion | supported as stated | medium | [161, 162] |
| ATT-E-F13 | Sell-side attention curation is documented: Adelaide High Attention Markets launched on 7 March 2023 with Magnite, PubMatic, Xandr, DV360, Teads, Sovrn, TRUSTX and Sharethrough (display and OLV, 20+ countries including the US); Comscore-Certified Deal IDs with AU on PubMatic followed on 10 June 2025; and AU became available in Index Marketplaces on 7 January 2026, which Index describes as part of its approach of giving access to data signals through any DSP 'without added fees or custom integrations'. That phrase describes Index's access model and does not state that Adelaide's data carries no data fee. All three are optimisation or curation uses, not guarantees. | 6 | vendor assertion | supported with qualification | medium | [183, 165, 166] |
| ATT-E-F25 | The outcome figures attached to trading announcements are vendor-reported and include no cost per incremental outcome. Adelaide reported that Trade Desk beta testers (August to September 2024) saw 58% higher ROAS and 62% higher CTR than control audience groups. DoubleVerify reported 63% higher attention levels, a score change rather than an outcome. | 6 | vendor assertion | supported as stated | low | [161, 213] |
| ATT-U-F24 | Two attention-guaranteed or attention-priced offers not in the register were found. Duolingo announced AU guarantees as launching on 15 September 2026: a contractual commitment to a minimum AU level across its in-app inventory, following a March 2026 audit that found its mobile in-app OLV 34% above Adelaide's benchmark. Adlook's AU-CPM buying model (12 April 2023) was described as the first DSP to offer guaranteed media quality using attention metrics; its status at the cutoff was not checked. Vevo's 'Attention Guaranteed' is a named offer that launched per its 27 January 2026 release. Teads' May 2025 post describes a pilot, 'Attention-to-Outcome Guarantees' (also called 'Effective Attention Guarantees'), run as a beta program on Adelaide AU. None of the four states a threshold, remedy or independent audit. | 6 | vendor assertion | supported with qualification | medium | [97, 98, 78, 95] |
| ATT-ADELAIDE-F14 | Yahoo launched AU pre-bid segments in Yahoo DSP on 13 June 2023; the latest status evidence is Adelaide's statement, reported by AdExchanger on 10 June 2026, that AU pre-bid targeting is available in Yahoo DSP; no Yahoo-side confirmation at the cutoff was found. | 6 | vendor assertion | supported with qualification | medium | [334] |
| ATT-TVISION-F13 | Viant announced an 'attention-adjusted CPM' on 15 April 2026, and on 5 May 2026 said its platform 'will be able to deliver' it. An EDGAR full-text search of filings from 1 January to 29 September 2026 finds 'attention-adjusted' only in the 15 April release. On 10 August 2026, Viant said it had 'Began testing' TVision pre-bid attention intelligence. On the call the same day, management described a planned rollout 'across the course of this year' and a technical integration still in progress. The metric is therefore announced, not launched, and its use as a billing unit is undocumented. | 6 | synthesis | supported as stated | high | [308, 219, 472, 309, 304] |
| ATT-VEVO-F07 | Vevo's 2026 upfront release (dateline 22 April 2026) still lists 'Attention Guaranteed' within Vevo Evolve, but describes it as a capability to help brands understand 'how audiences engaged' with campaigns. The release also claims its inventory outperforms CTV benchmarks on brand lift and attention 'according to' DISQO, iSpot and Adelaide, with no figures. | 6 | vendor assertion | supported as stated | high | [324] |
| ATT-VEVO-F08 | In the Vevo pages screened on 29 September 2026, the For Advertisers page lists Vevo Evolve (Advanced Targeting) but does not mention Attention Guaranteed or attention. No newsroom post after the 22 April 2026 upfront release mentions the offer, including the 9 July 2026 Evolve nostalgia release. No dedicated product page was found: the release links point only to press posts. | 6 | inference | supported as stated | medium | [318, 473, 417] |
| ATT-VEVO-F14 | Synthesis: Vevo's is the only guarantee in the register that is stated as launched and names CTV, mobile and desktop in one offer. Only the WSJ (five clients, six campaigns) and Kargo (one campaign, more than 6 million impressions) disclose completed campaigns sold under a guarantee; Teads describes one Q1 2024 campaign analysed against an AU threshold 'as part of' its beta, without saying it was sold under a guarantee. Vevo discloses no adoption. | 6 | synthesis | supported with qualification | medium | [78, 95, 94, 92, 93] |
| ATT-VEVO-F16 | Inference: under H2, Vevo's offer is an announced contractual guarantee whose terms cannot be read from public sources, so it is coded 'unresolved', and it does not meet the currency-grade test (threshold, remedy and independent audit all stated). Under Q1 the Vevo and AU pair is coded 'unresolved' by the literal rule (terms unreadable), but it could not reach 'currency' or 'partial currency' whatever its terms, because AU is not accredited. This is consistent with workstream E's ATT-E-F24, which its verifier rated supported as stated. | 6 | inference | supported with qualification | medium | [78, 48] |
| ATT-E-F03 | An Adelaide case study documents a 2022 MediaCom (now EssenceMediacom) and Kargo campaign, run before October 2022 on DV360 PMPs, in which the parties 'agreed upon' a minimum average AU of 25 (the sources call it an 'AU guarantee'; they do not say 'contractual'). Added-value PMP impressions were served when the average dipped below 25, delivery was monitored in Adelaide dashboards, and the campaign ran across more than 6 million impressions and 262 domains. Geography is not stated, and the campaign length is not stated (the only charted real flight spans 13 weeks; the 9-week chart is hypothetical). | 6, 7 | vendor assertion | supported with qualification | medium | [93] |
| ATT-E-F17 | Viant announced an 'attention-adjusted CPM' on 15 April 2026, and its 5 May 2026 closing release said the platform 'will be able to deliver' a metric that prices CTV and linear inventory on TVision attention. No later filing mentions the attention-adjusted CPM (EDGAR full-text search to 29 September 2026), and its role as a billing unit is undocumented. Viant's TVision-based pre-bid attention targeting is described inconsistently: 'testing' in the 10 August 2026 Q2 release, 'integrating' in the 10-Q and in a 16 September 2026 prospectus supplement, but offered to Havas clients through the Havas and Viant Attention Marketplace announced on 24 June 2026. | 6, 7 | direct record | supported with qualification | high | [308, 219, 309, 474] |
| ATT-F-F06 | Digiday reported on 12 March 2024 that media groups at five holding companies (Havas Media Network, GroupM, IPG Mediabrands, Omnicom Media Group and Publicis Media) had signed on to Adelaide's Flight Control, a planning tool that outputs a target AU and flags media below an AU floor for reduction; no terms, fees or payer were reported. | 6, 7 | vendor assertion | supported as stated | medium | [333] |
| ATT-F-F14 | Buyer-side attention guarantees in the sources are: (a) a live A+E guarantee to dentsu for AB InBev's Budweiser (Q4 2020, repeated in Q1 2021). It was based on TVision panel data counting eyes on screen for at least two seconds, optimised in flight in A+E's serving stack, and priced on the principle that attentive inventory costs more; (b) AB InBev's May 2021 aim to guarantee attention like in-demo reach; (c) dentsu's 2023 testing; and (d) an unnamed agency cited by Adelaide's CEO. None discloses a guarantee level, remedy or independent audit. | 6, 7 | synthesis | supported with qualification | medium | [475, 89, 476, 24] |
| ATT-F-F28 | No buyer-side source screened states who pays for attention measurement, its fee, or a specific contract term (guarantee level, remedy, audit or make-good) for an attention-based buy. Buyers do describe commercial principles: AB InBev and A+E priced attentive inventory higher and shared the efficiency (2020-21), JLR uses attention scores to negotiate media premiums (2025), and Haleon flags CPM premiums on the highest-attention CTV (2024). | 6, 7 | inference | supported with qualification | medium | [477, 333] |
| ATT-ADELAIDE-F28 | The Trade Desk-owned publication The Current reported (3 October 2025) that iFIT, using Adelaide segments in The Trade Desk with Upwave brand-lift measurement, achieved a 61% increase in aided awareness, 17% in consideration and 10% in favourability, without stating the comparator; iFIT's executive said outcome lifts were needed to 'rationalize the higher costs'. | 6, 7 | vendor assertion | supported as stated | medium | [67] |
| ATT-TVISION-F14 | On 24 June 2026, Havas Media Network and Viant announced the Attention Marketplace. Inventory 'within Viant' is 'scored by attention decile' for pre-bid targeting and ongoing optimisation, and Viant's CEO names TVision's eyes-on-screen measurement as a component. This is an optimisation use asserted by an agency and a DSP, and it sits uneasily with Viant's 'Began testing' status of 10 August 2026. | 6, 7 | vendor assertion | supported as stated | medium | [310] |
| ATT-E-F06 | Adelaide's AU Ecosystem had 54 participating partners on 23 October 2025. Adelaide says several participants offer guaranteed minimum AU scores, naming only the WSJ, and that Condé Nast and Hearst cover the cost of AU measurement on their inventory. This is one of few public statements on who pays for attention measurement; others found are Adelaide's free publisher audit tier (AFP, April 2024) and Index Exchange's statement on data-signal access without added fees. No fee level is disclosed in any of them. | 6, 7, 12 | vendor assertion | supported with qualification | medium | [167] |
| ATT-TVISION-F23 | The only executed guarantee on TVision data found in the sources screened: A+E guaranteed dentsu's Q4 2020 Budweiser buy for Anheuser-Busch, repeated in a second wave. The advertiser describes the guarantee as based on people being present in the room (TVision 'viewability'). The article adds that the data register eyes on screen for two seconds or more, so whether the guaranteed metric was presence or attention is ambiguous. The buy was optimised in flight in A+E's serving stack. No guarantee level, remedy or audit is given. | 6, 7, 12 | vendor assertion | supported with qualification | medium | [24] |
| ATT-ADELAIDE-F12 | AU pre-bid segments in The Trade Desk are relative tiers, per AdExchanger: high attention is generally the top 30% of TTD inventory by Adelaide's score, medium the middle 40% and low the bottom 30%, scored separately for display, online video and CTV, with buyer-customisable thresholds, available in the self-serve platform from 8 October 2024. | 6, 12 | vendor assertion | supported as stated | medium | [161] |
| ATT-VEVO-F13 | Among the six named attention guarantees in the register, all built on Adelaide AU (Teads' combines AU with its own benchmarks and purchase data; Vevo's names AU by implication), only Kargo/EssenceMediacom publishes a numeric threshold (minimum average AU of 25). Only Kargo and the WSJ publish a remedy, both added-value impressions. Teads' offers were a pilot and a beta as of May 2025 (no later status found), TripleLift's targets 'desired attention thresholds' with no number, and Duolingo gives a 'contractual commitment' with no number. Vevo publishes neither threshold nor remedy. | 6, 12 | synthesis | supported with qualification | medium | [93, 91, 94, 97, 95] |
| ATT-R2-F02 | Impact Media announced a Guaranteed Attention Suite from New York on 18 April 2024, billing clients only when a display ad is seen for one second or a video ad for three seconds, with CTV and audio planned; the release does not say how &quot;seen&quot; is measured, and this review found no later source on its availability. | 6, 12 | vendor assertion | supported as stated | medium | [101] |
| ATT-R2-F01 | SeenThis launched SeenThis Attention on 30 June 2026: a video offer measured by Lumen Research, bought on attentive CPM (aCPM) or cost per attentive second (CPAS), with an attentive-outcome threshold agreed before launch and, per the product page, extra delivery at no cost if the threshold is missed. The release says it is available globally; US sales, adoption, the threshold values and any independent audit of settlement are not documented. | 6, 12, appendix | vendor assertion | supported as stated | medium | [99, 100] |
| ATT-E-F07 | Teads presented 'Guaranteed Attention' at the IAB NewFronts in May 2025 and describes 'Attention-to-Outcome Guarantees' as a pilot and 'Effective Attention Guarantees' as a beta program. Both combine Adelaide AU scores, Teads benchmarks and engagement and purchase data. The blog states no threshold, remedy, price or live date. | 6, 13 | vendor assertion | supported as stated | high | [95] |
| ATT-E-F14 | In the sources screened, agency and holding-company attention commitments amount to access, planning and optimisation, not published contract terms. Xaxis (then a GroupM unit) said in March 2023 it could 'offer guaranteed attention-based outcomes' on its curated Adelaide marketplace, without terms. IPG Mediabrands' 25 April 2024 global deal with Adelaide covers planning, measurement, custom bidding and auction packages. | 7 | vendor assertion | supported with qualification | low | [183, 477] |
| ATT-F-F01 | WPP renamed its global media company from GroupM to WPP Media on 28 May 2025; Mindshare, Wavemaker and EssenceMediacom continue as brands within WPP Media. | 7 | vendor assertion | supported as stated | high | [478] |
| ATT-F-F02 | Omnicom completed its merger with IPG on 26 November 2025; its FY2025 10-K (signed 20 February 2026) lists Omnicom Media as OMD, PHD and Hearts &amp; Science plus UM, Acxiom, Initiative and Mediahub acquired from IPG, and does not name IPG Mediabrands or KINESSO. | 7 | direct record | supported as stated | high | [479, 480] |
| ATT-F-F04 | Havas N.V. became an independent listed company on Euronext Amsterdam (ticker HAVAS) on 16 December 2024, after its spin-off from Vivendi SE was approved by Vivendi shareholders on 9 December 2024. | 7 | vendor assertion | supported as stated | high | [481] |
| ATT-F-F05 | In February 2026 dentsu said it was no longer considering a sale of its international business and would pursue its midterm plan instead, while staying open to local divestments. | 7 | vendor assertion | supported as stated | medium | [482] |
| ATT-F-F07 | IPG Mediabrands announced a global Adelaide partnership on 25 April 2024 (brokered by KINESSO) covering planning, measurement, in-platform custom bidding and activation on SSP-delivered auction packages. In February 2023 Digiday reported that Mediahub had used attention metrics for clients including the NBA, and recalled an earlier Digiday report that Mediahub was working to bring attention into programmatic buying; the NBA work is not described as programmatic, and no vendor is named. | 7 | vendor assertion | supported with qualification | medium | [477, 476] |
| ATT-F-F08 | Havas Media Group said in November 2022 that its Converged planning platform added a Meaningful Media Planner that uses Lumen eye-tracking-panel data to score over 9,000 domains by format for the average attention they are likely to deliver, plus a social tool that compares attention with cost. | 7 | vendor assertion | supported as stated | medium | [483] |
| ATT-F-F09 | Havas Media North America and Adelaide announced the Meaningful Ad Unit for US clients on 19 June 2023, for planning and for programmatic activation through custom algorithms, pre-bid segments and custom PMPs. No buying platform is named. In February 2023 Digiday characterised Havas's approach as treating attention more as a KPI than a currency; this was the reporter's framing, not a Havas quote. | 7 | vendor assertion | supported with qualification | medium | [484, 476] |
| ATT-F-F10 | A Havas-commissioned study (released 16 September 2024; white paper by Lumen, Brand Metrics and Havas) applied Lumen's predictive attention model to 9,089 display campaigns (1,879,764 respondents; 5,618,088 viewable impressions), mostly in the UK. Its descriptive analysis, with no control group, reports that attention and brand-survey outcomes are closely correlated. Because the attention estimate is itself derived from viewable time, the study makes no comparison against time in view and uses no held-out data. | 7 | commissioned measurement | supported with qualification | low | [485] |
| ATT-F-F11 | Dentsu UK&amp;I launched Effective Attention Forecasting (working title) with Lumen on 7 April 2022. It uses a predictive model built on Lumen's UK eye-tracking panel to optimise programmatic campaigns, with a companion measurement product that extends to Facebook, Instagram, YouTube and Snap. Social use is described as measurement, not optimisation. The use is UK only. | 7 | vendor assertion | supported with qualification | medium | [486] |
| ATT-F-F12 | Dentsu's US Attention Economy page (dated 12 January 2024) says dentsu built proprietary attention prediction models with Lumen to 'plan, measure, and activate attention across platforms', using eye-tracking panels with TVision and Amplified Intelligence; in February 2023 dentsu told Digiday it was 'doing some testing around guarantees around attention'. | 7 | vendor assertion | supported as stated | medium | [487, 476] |
| ATT-F-F13 | Dentsu launched Global Media Planner on 18 March 2026 inside dentsu.Connect across 15 markets, powered by an attention dataset built with Lumen that lets planners compare formats on a common attention-based metric; US inclusion is not stated. | 7 | vendor assertion | supported as stated | medium | [488] |
| ATT-F-F15 | OMD has used attention in planning since 2021 (AttentionTRACE with Amplified Intelligence, 10 countries, 500 to 3,000 users per country). Its Attention Requirement Calculator, built with Annalect inside Omni, was reported by Adweek in July 2022. In February 2023 OMD's CEO said OMD had applied attention in 11 markets, 13 categories and 25 brands, that 'optimizing to attention lifts KPIs', and that it is 'a planning metric'. That count and quote describe OMD's attention work generally, not the calculator. Lumen benchmarks were integrated into Omni channel planning from May 2024. | 7 | vendor assertion | supported with qualification | medium | [343, 476, 278, 489] |
| ATT-F-F16 | Omnicom's FY2025 Form 10-K contains no reference to attention metrics, attention vendors (Adelaide, Lumen) or viewability; the word 'attention' appears once, about management attention during integration, and the description of the next-generation Omni platform (unveiled January 2026) does not mention attention. | 7 | direct record | supported as stated | high | [479] |
| ATT-F-F17 | WPP's GroupM (now WPP Media) was reported in February 2023 to use Adelaide data in Xaxis Switzerland for a customised attention algorithm that optimised media purchases; WPP Media's Open Intelligence launch (3 June 2025) lists Adelaide and Lumen Research as data and measurement partners, and Adelaide's CEO says in that release that AU is 'embedded in the Large Marketing Model'. | 7 | vendor assertion | supported as stated | medium | [476, 490] |
| ATT-F-F18 | Publicis records found are Publicis Media's reported sign-on to Adelaide's Flight Control (2024), the Haleon/Flonase Amazon DSP evaluation with Adelaide (US, mid-2023), and Starcom's AU measurement of Visa's German campaign (vendor release, August 2023; reporting use, PMP activation under consideration). This review found no Publicis-branded attention product. The 2025 URD (filed 30 April 2026) contains no attention-metric reference. | 7 | synthesis | supported with qualification | medium | [491, 476, 333] |
| ATT-F-F20 | Mars said in March 2020 that it had not bought media on attention. In May 2023 its senior director of global media told AdExchanger that Mars had seen 'promising signs that attention scores are highly correlated with sales' across attention tests with several vendors (Adelaide, Realeyes, Lumen, IAS, DoubleVerify). Its layering of attention metrics onto Nielsen Catalina Solutions sales-lift panels tested a link between publisher sustainability scores and attention scores, not sales. Mars was also testing Adelaide's attention-based PMPs on Xandr Curate. No design is published. | 7 | vendor assertion | contradicted | medium | [492, 26] |
| ATT-F-F21 | Nespresso began testing attention metrics in 2019. In Q2 2023 it began optimising to Adelaide's AU as a KPI and was testing Adelaide's high-attention (and low-emissions) PMPs built on Xandr Curate, which its media manager preferred to attention targeting in open auctions (AdExchanger, 2 May 2023). The geography is not stated. | 7 | vendor assertion | supported with qualification | medium | [26] |
| ATT-F-F23 | DoubleVerify's 2024 Global Insights release (18 June 2024) says the report incorporates a Sapio survey of 1,000 advertisers across North America, LATAM, EMEA and APAC, and states that 47% of media buyers planned to integrate attention metrics into their strategies in 2024, which DV headlines as a '47% adoption rate'; the release does not explicitly attribute the 47% to the Sapio survey or give its base, question wording or US subsample. | 7 | commissioned measurement | supported with qualification | medium | [371] |
| ATT-F-F24 | IAS's 2026 Industry Pulse survey (YouGov; nearly 300 US media experts across brands, agencies, publishers and ad tech; published 8 December 2025) reports that on social media 77% name attention and 85% viewability as essential metrics. IAS's 'The Attention Payoff' (YouGov; 200+ US digital media experts from both the buy side and the sell side; published 20 November 2023) reported 88% using attention measurement in some capacity, 54% via proxy signals, 39% in-house solutions and 36% a third-party attention vendor. Both samples mix buyers and sellers. | 7 | commissioned measurement | supported with qualification | low | [27, 493, 494] |
| ATT-F-F26 | A TVision blog (May 2022) relays the ARF Attention Report (26 April 2022, with the UK's Attention Council). That report is a qualitative study of 63 participants surveyed in January 2022 among ARF members and the Wonks community. TVision reports that more than 75% of buyers would consider attention metrics for planning and buying in future, and that 94% of buy-side respondents see them as an improvement rather than a replacement. These buy-side percentages rest on a subsample of fewer than 63 people, so they are not generalisable shares. | 7 | synthesis | supported with qualification | low | [475] |
| ATT-F-F30 | Inference: in the sources screened, holding-company attention announcements were mostly standalone vendor partnerships in 2021-2024; in 2025-2026 attention appears as one input inside broader AI or planning platforms (WPP Open Intelligence, June 2025; dentsu Global Media Planner, March 2026), while Omnicom's description of its January 2026 Omni platform does not mention it. | 7 | inference | supported as stated | low | [490, 488, 479] |
| ATT-ADELAIDE-F16 | As stated, with Xandr's DSP retired on 28 February 2026 (Microsoft Curate continues; any AU route there is undocumented at the cutoff). | 7 | vendor assertion | supported with qualification | medium | [183, 164, 165, 166, 1] |
| ATT-LUMEN-F17 | From 25 March 2024, Lumen attention measurement could be switched on and billed per ad group inside The Trade Desk for display and video; this is a measurement and billing route, not attention-priced media. | 7 | vendor assertion | supported as stated | medium | [275] |
| ATT-LUMEN-F34 | Agencies have used Lumen data for planning (Omnicom's Omni benchmarks, May 2024; dentsu's 6,000-person panels, 2021) and for buying (Dentsu Aegis Network's Google custom-algorithm buy for Co-op, 2019; dentsu and Havas using Lumen signals in the Avocet DSP, reported 2022). | 7 | vendor assertion | supported with qualification | medium | [278, 9] |
| ATT-B-F15 | In an in-game study (79 participants, Turkey, mobile tennis game, Tobii Pro Glasses 2), total fixation duration on in-game ads had no significant effect on visual-aided brand recall once brand placement was controlled (B = -0.003, SE 0.006, p = .672), though brands were not counterbalanced across placements. | 8 | independent measurement | supported as stated | low | [53] |
| ATT-G-F03 | MRC's 13 April 2026 letter accredits DoubleVerify's Display and Video Authentic Attention metrics only in desktop, mobile web and mobile in-app; CTV attention is not accredited, the Meta, YouTube (via Ads Data Hub) and TikTok accreditations cover impressions and viewability only, and the standards the letter lists do not include the IAB/MRC Attention Measurement Guidelines. | 8 | direct record | supported as stated | high | [42] |
| ATT-G-F04 | DoubleVerify announced CTV attention with TVision on 13 June 2023 (closed beta anticipated in H2 2023); the combined DV and TVision data were used in a joint CTV report released 30 May 2024; by 30 June 2025 DV described its attention measurement as spanning CTV. No dated general-availability release was found in DV's newsroom, and DV's own Authentic Attention product page (accessed 28 September 2026) describes no CTV attention capability. | 8 | vendor assertion | supported with qualification | medium | [220, 116] |
| ATT-G-F08 | Audio attention measurement is documented at two cohort A products: Adelaide announced show-level podcast AU in Magellan AI on 10 January 2024 (availability stated as from early 2024; current availability not confirmed) and extended Adelaide for Publishers to podcasts, streaming audio and terrestrial radio on 9 May 2024 with NPR, SiriusXM Media and Westwood One as launch partners; Amplified's audio coverage is planning-stage prediction only (Spotify and radio in Campaign Strategy; audio is not listed for its in-flight measurement). Audio is thin under the Q3 rule. | 8 | vendor assertion | supported with qualification | medium | [107, 108, 119] |
| ATT-G-F11 | DOOH attention evidence rests on methods that differ in kind: Lumen's study with Ocean Outdoor (UK, 11 June 2025) tracked 2,500 online participants' eyes by webcam while they watched filmed walks past OOH sites; Amplified's study with QMS (Australia) used in-situ eye, gaze and pose tracking across 1.2 million exposures at twelve sites; Quividi's explainer defines DOOH Attention Time as time spent watching, measurable only with camera-based anonymous video analytics; Adelaide lists DOOH in its coverage without a documented method; and Amplified lists OOH and DOOH in its planning tool. The OOH audience currencies already apply eye-tracking-derived likelihood-to-see adjustments (Geopath in the US, Route in the UK). These are visibility adjustments to audience counts, not per-ad attention measurement. | 8 | synthesis | supported with qualification | medium | [114, 495, 105, 119] |
| ATT-G-F13 | IAS Quality Attention measurement has been generally available since January 2024 for open-web display and video (with optimisation in beta from 17 December 2024) and was extended to mobile in-app on 30 July 2024 (IAS release). IAS announced Social Attention reporting with Lumen on 17 December 2024, stated as live from January 2025 with no platforms named, and Snap Attention Measurement on 30 June 2025 ('will be available' in IAS Signal). This review found no IAS source documenting Quality Attention on CTV, audio, in-game, DOOH or retail media. | 8 | vendor assertion | supported with qualification | medium | [115, 111, 73] |
| ATT-G-F14 | Lumen's Netflix attention solution (5 March 2026) measures Netflix's global ad-supported CTV, desktop and mobile inventory but is offered only to clients in the UK, Germany, France, Italy and Spain (for local or global campaigns), so it adds no US client availability. Lumen does document US CTV measurement through other routes: with TVision for LG Ad Solutions (February 2024) and Nexxen (August 2023), and through Teads' CTV HomeScreen placement (May 2026). | 8 | vendor assertion | supported with qualification | high | [110] |
| ATT-G-F15 | Amplified's Campaign Strategy tool lists social (Meta, TikTok, Snapchat, X, Pinterest), web and programmatic, OOH and DOOH, linear TV, BVOD and CTV, audio (Spotify and radio) and cinema, and predicts attention before launch from '59 billion human attention data points collected across 16 countries'; it lists no in-game or retail media channel. | 8 | vendor assertion | supported as stated | medium | [119] |
| ATT-G-F16 | Creative testing documents formats where media measurement is thin: Lumen SPOTLIGHT tests social, display and video, in-game, TV and VOD, OOH, print, point of sale and packaging with webcam eye-tracking of opt-in participants. Kantar extended LINK+ survey-based creative testing on 6 May 2025 to social, websites, retail media, audio, VOD, console gaming, TV and cinema trailers, and LINK AI to static display, print and outdoor, but that release does not state that attention is measured in any of the new formats. | 8 | vendor assertion | supported with qualification | medium | [279, 260] |
| ATT-G-F17 | GumGum launched Attentive Lift in Australia on 3 December 2025 with On Device's impression-level brand-lift data; the release names no channel and no US availability, so Attentive Lift's channel coverage is unresolved from the sources opened. | 8 | vendor assertion | supported as stated | medium | [458] |
| ATT-XPLN-F06 | For CTV, xpln.ai's models take TVision's second-by-second, person-level attention data as an input (joint release carried by Advanced Television on 23 April 2026; PPC Land, 24 April, calls it a calibration input and describes TVision's panel as 5,000 US households). Viant completed its acquisition of TVision on 1 May 2026 (Viant Form 10-Q, a filing); whether xpln.ai's access continued after closing is undocumented. | 8 | vendor assertion | supported with qualification | medium | [405, 121] |
| ATT-XPLN-F16 | xpln.ai announced its North American expansion on 28-29 January 2026. The USA is listed among the countries of its training observations (CESP), and TVision's US panel was a CTV model input as of April 2026. A syndicated copy of the launch release names US-headquartered advertisers (General Motors, Levi's, KFC) among clients 'across markets', but the same release places its clients in six countries in Europe and APAC, so no US campaign or US production use is documented. | 8 | vendor assertion | supported with qualification | medium | [496, 121, 405] |
| ATT-ADELAIDE-F22 | Adelaide claims AU covers about 95% of an advertiser's spend across digital, online video, social, CTV, linear TV, cinema, audio and DOOH and '19 channels'; in the sources screened, measurement methods are documented for display, online video, CTV, linear TV and podcasts, social coverage rests on licensed Amplified data (since September 2020) and report uploads, and no method is documented for DOOH or cinema; the same September 2026 post also says '18 channels'. | 8 | vendor assertion | supported as stated | medium | [105, 1, 329] |
| ATT-ADELAIDE-F23 | The largest volume with a stated period is 'over 32B impressions from 395 US campaigns' (2022-2024) in a Google-commissioned US meta-analysis spanning 30 placement types, including competitor streaming platforms and similar premium placements (so not YouTube-only); derived about 10.7B impressions and 132 campaigns a year. A company-wide statement exists (eMarketer, Q2 2026, vendor-supplied: over 173B impressions across more than 1,000 global campaigns) but gives no period, so no annual company-wide rate can be derived. | 8 | commissioned measurement | supported with qualification | medium | [338] |
| ATT-AMPLIFIED-F22 | Amplified opened a New York office on 7 November 2022, described as its US headquarters, and told the ARF in 2023 that it operates in the US. Its in-flight product is described only as available in '20+ markets', without naming the US. | 8 | vendor assertion | supported with qualification | medium | [354, 119, 186] |
| ATT-DOUBLEVERIFY-F24 | DoubleVerify presents Authentic Attention in its commerce-media offering and reported, from 2023 data, that retail media network owned-and-operated inventory showed engagement rates 183% above DV's attention baseline, without naming the networks or placements. | 8 | vendor assertion | supported as stated | low | [366, 367] |
| ATT-IAS-F09 | General availability was announced on 4 January 2024 (IAS release of that date). Snap Attention Measurement was announced on 30 June 2025 with a stated Q3 2025 availability target; no later IAS source confirms it went live. The rest of the sequence stands. | 8 | vendor assertion | supported with qualification | medium | [238, 115, 226, 118] |
| ATT-LUMEN-F19 | Lumen's Netflix solution (5 March 2026) measures Netflix's global ad-supported CTV, desktop and mobile inventory and is available to clients in the UK, Germany, France, Italy and Spain (the release names no other markets); Teads' exclusive Lumen HomeScreen measurement (6 May 2026) covers the US, EMEA, APAC and LATAM in Teads Ad Manager, with InStream measurement announced to follow in the same quarter. | 8 | vendor assertion | supported with qualification | high | [110, 277] |
| ATT-TVISION-F26 | TVision's coverage is the TV screen: linear TV and CTV, including walled-garden apps viewed on the TV. Viant describes TVision signals as quantifying attention across linear TV, CTV and 'the walled garden platforms of YouTube and Prime Video'. TVision reports on 'All national and major linear TV channels' and 'Hundreds of apps including CTV walled gardens'. No other device or channel is documented. | 8 | vendor assertion | supported as stated | high | [304, 297] |
| ATT-G-F12 | One cohort A product documents attention measurement on retail media: DoubleVerify presents Authentic Attention within its commerce-media offering and reported, in May 2024, aggregate attention-engagement results for retail media network owned-and-operated inventory (2023 data), without naming the networks or placements measured or publishing a method for them. No other cohort A product in the sources screened documents it. Adelaide's AU pre-bid segments in Amazon DSP (9 June 2026) cover display, online video and streaming TV inventory, and the release does not say whether Amazon-owned retail placements are included. Kantar extended LINK+ creative testing to ads in retail media networks on 6 May 2025, but that release does not mention attention. Lumen's in-store work (point of sale, packaging, an in-store retail media study) is study-based and covers physical stores, not retail media network inventory. | 8, appendix | synthesis | contradicted | medium | [162, 260] |
| ATT-ADVERTEYES-F14 | The company-level policy (Adverteyes Inc., effective 2 February 2026) documents, for market-research participants, explicit consent for facial images and biometric data (research and model training) and some data minimisation (derived face, attention and emotion data points; truncated IP; aggregated, anonymised client outputs). It excludes under-18s unless local law sets a lower age or parental consent is obtained, and names GDPR, CCPA as amended by CPRA, and US state laws. Retention for model training and training to tell whether the same person appears in two recordings conflict with 'No images stored'. Nothing is product-specific or independently certified. Anchor 2 is the defensible reading because the documentation is company-level, not product-specific. Anchor 3 would need the participant flow to be accepted as product-specific. | 9 | vendor assertion | supported with qualification | medium | [497, 498, 499] |
| ATT-EYESQUARE-F12 | Anchor 4 is fully met for browser InContext studies (S17). For InContext Eye Tracking the consent model is documented (S18, Art. 6(1)(a)), but data minimisation is contradicted across vendor documents (S18: face video, with camera and audio access; S05: 'Only X/Y gaze coordinates leave the device. No raw video.'). A strict scorer could hold the eye-tracking variant at anchor 3 (consent documented, minimisation not reliably documented). The cell should state that the anchor-4 reading rests on the browser policy plus the eye-tracking consent policy despite the conflict. | 9 | vendor assertion | supported with qualification | medium | [500, 501, 502, 503, 504] |
| ATT-H-F01 | The AI Act's biometric-data definition omits GDPR's unique-identification requirement (confirmed by the Commission guidelines, para 252 and footnote 160), but it is still confined to personal data, that is, data relating to an identified or identifiable person; recital 14 asks for it to be read in light of GDPR Art 4(14). Data that are truly anonymous fall outside it. | 9 | direct record | supported with qualification | high | [30, 125] |
| ATT-H-F06 | Under GDPR, biometric data is a special category requiring an Article 9(2) condition, such as explicit consent, only when processed 'for the purpose of uniquely identifying a natural person'. Recital 51 says photographs count as biometric data only when processed by specific technical means allowing unique identification. | 9 | direct record | supported as stated | high | [125] |
| ATT-H-F08 | The UK ICO's biometric guidance says personal information is biometric data only if it can uniquely identify the person, and special-category biometric data only if used to identify them. It adds that transient processing lasting 'a fraction of a second' is still processing, and that the guidance is under review because of the Data (Use and Access) Act. | 9 | direct record | supported as stated | high | [127] |
| ATT-H-F11 | In Zellmer v. Meta Platforms (9th Cir., 17 June 2024), the court held on the record that Facebook 'face signatures' 'cannot identify and therefore are not biometric identifiers or information as defined by BIPA'. It described a 'broad consensus' that BIPA biometric identifiers 'must be able to identify'. | 9 | direct record | supported as stated | high | [418] |
| ATT-H-F12 | Texas CUBI, as amended by HB 149 (effective 1 January 2026), defines a biometric identifier to include a 'record of hand or face geometry'. The amendment exempts 'the training, processing, or storage of biometric identifiers involved in developing, training, evaluating, disseminating, or otherwise offering artificial intelligence models or systems, unless a system is used or deployed for the purpose of uniquely identifying a specific individual'; an identifier captured to train an AI system and later used for another commercial purpose falls back under the possession, destruction and penalty provisions (subsection (f)). The Texas Attorney General's first CUBI action ended in a $1.4 billion settlement with Meta on 30 July 2024. | 9 | direct record | supported as stated | medium | [129, 505] |
| ATT-H-F13 | The CCPA defines biometric information by use, as characteristics used or intended to establish individual identity. Its sensitive personal information covers biometric processing only 'for the purpose of uniquely identifying a consumer', and it also covers neural data. Its 'unique identifier' definition covers device identifiers, pixel tags and 'probabilistic identifiers'. | 9 | direct record | supported as stated | high | [130] |
| ATT-H-F15 | Washington's My Health My Data Act treats biometric data that identifies a consumer as consumer health data. Collection needs the consumer's opt-in consent for a specified purpose, or must be necessary to provide a product or service the consumer requested (RCW 19.373.030(1)(a)), and consent may not be obtained by hovering, muting, pausing or closing content. Compliance began on 31 March 2024 (30 June 2024 for small businesses). Washington's separate biometric identifier statute excludes photographs, video or audio recordings and data generated from them. | 9 | direct record | supported with qualification | high | [132, 133] |
| ATT-H-F16 | Colorado HB24-1130 (effective 1 July 2025) splits its rules. Consent before collection applies to biometric data, meaning biometric identifiers used or intended for identification. Collecting any biometric identifier, meaning data that can be processed to uniquely identify a person, including iris scans and facial geometry, requires prior notice of the collection, its purpose, the retention period and any disclosure to a processor (6-1-1314(4)(a)), and a written retention and deletion policy. These provisions reach any controller that processes any amount of biometric identifiers or biometric data, whatever the Colorado Privacy Act volume thresholds, except that the consumer access right in subsection (5)(a) applies only to entities meeting its own conditions (100,000 consumers a year, or 25,000 with revenue from selling personal data, or common-branding affiliation). | 9 | direct record | supported with qualification | high | [134] |
| ATT-H-F17 | The FTC's Policy Statement on Biometric Information (18 May 2023) defines biometric information to include images or recordings of facial features, iris or retina and 'characteristic movements or gestures'. It covers derived data where it 'would be reasonably possible to identify the person', and names surreptitious and unexpected collection as a potentially unfair practice. The statement was still listed on the FTC's policy-statements page on 28 September 2026. | 9 | direct record | supported as stated | medium | [135, 506] |
| ATT-H-F18 | Synthesis: in the statutes screened, biometric-data status for webcam gaze, landmark or facial-coding signals turns mostly on identification or identifiability. The EU AI Act (emotion recognition from biometric personal data) and California's profiling definition (data reasonably linkable to a consumer) reach emotion inference without identification. GDPR and UK GDPR can also reach such inferences as special-category health data where they support a health inference (ICO guidance; NAIH left the route open), and Washington's MHMDA reaches physiological measurements linked to a consumer. | 9 | synthesis | supported with qualification | medium | [130, 134, 132, 128, 30, 131] |
| ATT-H-F19 | Roberts v. Cooler Screens (Cook County, Illinois, filed 3 March 2022, No. 2022CH01824) is a putative BIPA class action. It alleges that in-store 'Smart Cooler' advertising screens used a camera and an 'iris tracker to detect exactly where the customer is looking', and a 'facial profiling system' said to detect the age, gender and emotional response of more than 3 million daily viewers, without written consent. | 9 | direct record | supported as stated | medium | [507] |
| ATT-H-F20 | Canada's federal, Alberta and British Columbia privacy commissioners found (28 October 2020) that Cadillac Fairview collected sensitive biometric information without valid consent. It used 'Anonymous Video Analytics' in mall directories to estimate visitors' age range and gender. Its service provider had stored about 5 million numerical face representations. The commissioners recommended express opt-in consent. | 9 | direct record | supported as stated | high | [136] |
| ATT-H-F21 | On 8 February 2022, Hungary's data protection authority fined Budapest Bank HUF 250,000,000. It ordered the bank to stop analysing emotions in its AI voice analysis of customer-service calls, which detected 'emotional/mood elements' in recordings. | 9 | direct record | supported as stated | high | [508] |
| ATT-H-F22 | A São Paulo court (judgment released 7 May 2021) found that ViaQuatro's interactive advertising doors on Metro Line 4 captured passengers' images to detect characteristics, emotions and reactions to the advertising, without notice. The operator had described the system as counting views, 'tempo de atenção' (attention time) and emotions. The court ordered it to stop capturing data without prior consent and awarded R$100,000 in collective moral damages. Trade press reports that an appellate chamber raised this to R$500,000 in May 2023. | 9 | synthesis | supported as stated | medium | [137, 509] |
| ATT-H-F23 | Amplified's privacy policy (June 2026) says its panel collects facial images through device cameras (mobile and PC) and through an in-home micro camera for TV. It says the TV images are processed in the home and 'promptly deleted', and that facial landmarks are used 'solely for attention measurement' and not to identify anyone. It says Amplified does not use facial recognition and requires explicit prior consent for sensitive data. | 9 | vendor assertion | supported as stated | medium | [184] |
| ATT-H-F24 | iMotions' privacy policy (22 June 2022) says that for test persons the client, as data controller, must collect consent under GDPR Articles 6(1)(a) and 9(2)(a) 'due to' the processing of biometric data for analysing test results. Tobii's eye-tracking data transparency policy requires developers to obtain active user acceptance before eye-tracking data is stored or transferred. Tobii defines 'psychographic eye tracking data' to include insights about a person's emotional state. | 9 | vendor assertion | supported as stated | medium | [510, 511] |
| ATT-H-F25 | Lumen's eye-tracking page says its software turns a phone or computer webcam into an eye tracker, that 'all processing is done on the users’ device', and that 'no video data or personally identifiable information is uploaded'. It says the software 'records what was on the screen'. The pages opened do not state Lumen's consent model for panellists. | 9 | vendor assertion | supported as stated | medium | [393] |
| ATT-H-F26 | TVision states that its TV panel is '100% opt-in', that panelists are compensated, and that after an initial training period all data leaving the home is 'completely anonymous', referring to households and viewers 'by numbers not names'. | 9 | vendor assertion | supported as stated | medium | [291] |
| ATT-H-F28 | IAS announced ISO/IEC 27001:2022 certification of its information security management system on 22 May 2024, with a scope that names 'Attention' among in-scope services. ISO 27001 certifies information-security management, not privacy compliance, consent or measurement validity. | 9 | vendor assertion | supported as stated | medium | [384] |
| ATT-H-F29 | Adverteyes says its models are trained on '18 million+ webcam observations' measuring 'Visual Attention (eyes on screen) and Emotions (facial expressions)', and describes its testing as '100% GDPR compliant'. The phrase 'with consumer consent' appears only in its description of a patent (U.S. 11,812,105), not as a statement about its own data collection. Its SOC 2 statement names Realeyes, not Adverteyes. | 9 | vendor assertion | supported with qualification | medium | [512, 499] |
| ATT-IMOTIONS-F07 | As stated, plus: face counts now run from 18 million (solution page) to 19 million (May 2026 release; product page '19M+') to 'more than 20 million' (1 September 2026 launch release). The vendor describes the 100,000-ad base as tested with 'Affectiva's technology' using facial expression analysis over about 15 years, so the base is facial-coding tests and may include tests run through partners that license the technology (for example Kantar's LINK); its attribution to this product row, and its attention content, are not established. The top-10 markets are named but no per-market or per-category counts are published. | 9 | vendor assertion | supported with qualification | medium | [513, 514, 515, 452] |
| ATT-IMOTIONS-F14 | As stated, with two changes: (1) the iMotions policy assigns test-person consent to 'the Data Controller' without naming the client; reading the client as controller (with iMotions as processor) is a reasonable inference from section 5, not a quotation. (2) Affectiva, Inc.'s privacy policy (version 14 July 2022) documents an informed-consent model and a de-identification statement for ad-testing facial videos, retention of video for algorithm training, and a CCPA compliance statement; whether it still governs Media Analytics data after the January 2025 integration is not stated. No retention period, BIPA or other US state biometric-law documentation was found in either policy. | 9 | vendor assertion | supported with qualification | medium | [510, 514, 516, 517] |
| ATT-TOBII-F11 | No accreditor, certifier or independent audit record of Sticky's testing method was found. The product page's ISO/IEC 27001 claim is backed by a certificate: Tobii AB (which operates Sticky) holds ISO/IEC 27001:2022, certificate SE556613-9654 from the SWEDAC-accredited Scandinavian Business Certification AB, valid to 8 May 2028. It covers the company's information security management and does not name Sticky, the attention method or fieldwork. 'Academic validation' still names no study. | 9 | vendor assertion | supported with qualification | medium | [455, 518, 519] |
| ATT-TOBII-F13 | Sticky's data protection white paper (23 March 2026) documents a product-specific consent model (participants informed and consenting; legal basis 'legitimate interest, or consent') and data minimisation. Webcam framing is limited to face and head; no names, addresses, phone numbers or social security numbers are collected; reports are scrubbed of identifying information; webcam and screen recordings are deleted after up to 90 days; data is hosted at AWS US East. GDPR is the only legal framework it names. Emotion analysis is opt-in. | 9 | vendor assertion | supported with qualification | high | [520, 521] |
| ATT-XPLN-F21 | xpln.ai documents data minimisation for its measurement tag in a Solutions Privacy Notice that covers its tag domains (xpln.tech, log.xpln.tech, cdn.xpln.tech, vast.xpln.tech). The notice lists the categories collected, states no third-party cookies and no cross-site tracking or profiles, de-identified impression-level reporting, retention of personal information of no more than 30 days, and storage in the EEA, and it names GDPR, CCPA and TCF. For the eye-tracking panel, the only public consent statement is one line in xpln.ai's CESP answer: 'Panel respondents are opted-in and explicitly consented.' The sources are vendor documents (grade C). | 9 | vendor assertion | supported as stated | medium | [522, 121, 523] |
| ATT-XPLN-F22 | IAB Europe's Global Vendor List (version 178, updated 24 September 2026) records vendor 1142 as 'xpln.ai SAS'. The entry declares legitimate interest for purposes 7 (measure advertising performance) and 10 (develop and improve services) and special purpose 1 (security, fraud prevention, fixing errors), declares no cookie use and a standard retention of 90 days, and lists the data categories IP addresses, device characteristics, browsing and interaction data, non-precise location data and privacy choices (grade A registry record of the vendor's declarations). | 9 | direct record | supported as stated | high | [190] |
| ATT-ADELAIDE-F24 | Adelaide's AU Ratings Privacy Policy (effective 13 September 2024) says its tracking technology collects device type, IP address, location to city, DMA and zip code, and inferences, does not sell or share personal information under the CCPA and processes no sensitive personal information, and states that AU Ratings process no personal data under the GDPR and use no cookies; Adelaide separately says its eye-tracking data come from fully opted-in panels. | 9 | vendor assertion | supported as stated | high | [340, 168] |
| ATT-AMPLIFIED-F09 | Amplified's June 2026 privacy policy says the attentionPROVE tag is covered by its IAB Europe TCF registration as vendor 1195 with legitimate interest as the default basis for Purposes 7, 9 and 10 and does not access device storage, and IAB Europe's vendor list (version 178, 24 September 2026) records vendor 1195 with those legitimate-interest purposes, no cookie or non-cookie device access, four data categories and a standard retention of 3,650 days. | 9 | direct record | supported as stated | high | [184, 190] |
| ATT-DOUBLEVERIFY-F21 | DV's Solutions Privacy Notice (effective 9 October 2025) lists the exposure, engagement and viewability attributes its tags collect and classes them as Technical Information. It does not name Authentic Attention. The GDPR 'processor' role it states applies to Page Level Engagement Metrics, a separate publisher content-performance solution, not to ad attention measurement. | 9 | vendor assertion | supported with qualification | medium | [372, 47] |
| ATT-IAS-F19 | IAS's Technology Solutions privacy policy (IASLG-17, most recent revision 01.12.2023) describes pixel and SDK collection of impression data and IP addresses but does not name Quality Attention or its interaction signals; the website privacy policy (updated 9 March 2026) excludes Technology Solutions; and IAS's ISO/IEC 27001:2022 certification (announced 22 May 2024) names 'Attention' among in-scope services for information security only. | 9 | vendor assertion | supported as stated | medium | [385, 386, 384] |
| ATT-KANTAR-F21 | Kantar's Global Surveys Privacy Notice (May 2026) says facial-expression and eye-tracking data are collected only with the respondent's consent, may be analysed by partners eye square GmbH and Affectiva Inc., that Affectiva stores facial-expression video for seven years, that standard survey retention is 12 to 18 months, and that anonymised data may be used to train AI algorithms. | 9 | vendor assertion | supported as stated | high | [267, 387] |
| ATT-EYESQUARE-F10 | Austrian Standards' certification database lists certificate S 001201 for eye square GmbH (Berlin) against ISO 20252:2019, scope 'Services of market, opinion and social research - Full-Service Provider', issued 13 November 2020 and expiring 13 November 2026, so it is valid at the cutoff. It certifies the company's research-process quality system, not the InContext attention metric or SEAL. The vendor's claims of annual privacy and quality audits by 'our partner PSW' have no public record. This review found no MRC accreditation, independent validation audit or academic audit of the InContext method. The facts fit creative anchor 4 (ISO 20252 covering part of the method, not the attention metric). Grade A. | 9, 11 | direct record | supported as stated | high | [524, 525, 526, 502, 527] |
| ATT-H-F03 | Inference, restated: a gaze-only product falls outside the emotion-recognition definition only while its outputs stay at where and how long people look. If gaze or facial data is used to infer emotions, interest, arousal or intentions (for example 'engagement' or purchase intent), the Commission's reading places it inside the definition, because eye-tracking data is biometric data under the AI Act. | 9, 12 | inference | supported with qualification | medium | [30, 124] |
| ATT-H-F04 | From 2 August 2026, anyone deploying an emotion recognition or biometric categorisation system in the EU must inform the people exposed to it (AI Act Article 50(3), which applies from the Act's general application date). The July 2026 Digital Omnibus added a grace period only for Article 50(2) marking and did not amend Article 50(3) in the text screened. | 9, 12 | direct record | supported as stated | high | [30, 123] |
| ATT-H-F07 | The EDPB's Guidelines 2/2023 (v2.0, 7 October 2024) state that ePrivacy Article 5(3) protects information on a user's device whether or not it is personal data, and that JavaScript instructing a browser to send information 'clearly' falls within Article 5(3). Inference: tag-based attention signals (viewport, scroll, focus, time in view) collected in the EU need consent unless an Article 5(3) exemption applies. | 9, 12 | inference | supported as stated | medium | [126] |
| ATT-H-F10 | Illinois BIPA defines biometric identifiers to include a retina or iris scan and a 'scan of hand or face geometry'. Before collection it requires written notice, a stated purpose and term, and a written release (informed written consent or an electronic signature). It gives a private right of action of $1,000 (negligent) or $5,000 (intentional or reckless) per violation. A 2024 amendment (P.A. 103-769, effective 2 August 2024) limits repeated collection from the same person by the same method to one recovery. | 9, 12 | direct record | supported as stated | high | [128] |
| ATT-H-F14 | California's CPPA regulations (approved 22 September 2025, effective 1 January 2026) require a written risk assessment before a business sells or shares personal information, processes sensitive personal information, or processes personal information it intends to use to train emotion-recognition technology. Their definition of physical or biological profiling includes using facial expression 'to identify or infer emotion' and excludes characteristics that cannot reasonably be linked to a consumer. Processing already under way must be assessed by 31 December 2027, with submissions due by 1 April 2028. | 9, 12 | direct record | supported as stated | high | [131, 528] |
| ATT-H-F27 | Adelaide's AU Ratings privacy policy (effective 13 September 2024) treats data collected by its 'attention measurement tracking technology' as personal information: device type, IP address, location to city, DMA and zip code, and inferences. It says Adelaide does not process sensitive personal information. DoubleVerify's Solutions Privacy Notice (9 October 2025) lists the tag signals it collects as viewport and screen size, tab and browser focus, time in view, scroll position and ad interactions, and describes IP address and user agent as pseudonymous identifiers. The notice does not use the word 'attention'. | 9, 12 | vendor assertion | supported as stated | medium | [340, 372] |
| ATT-H-F30 | In the sources screened by the cutoff, this review found no public regulator action or lawsuit naming a seed attention-measurement vendor (searches named Realeyes, Affectiva, TVision, Tobii and iMotions, plus general BIPA searches on eye tracking, facial coding and webcam research). The nearest matters found involve in-store and out-of-home advertising screens or non-advertising emotion AI (ATT-H-F19 to ATT-H-F22). | 9, 12 | inference | supported with qualification | low | [507, 136] |
| ATT-LUMEN-F32 | Lumen's privacy statements for its eye tracking are vendor-level only: webcam processing is on device with no video or PII uploaded, and panellists are consented and compensated; the site's privacy policy (last updated June 2026) covers the website, and no panel or participant privacy notice was found in the sources screened. | 9, 12 | vendor assertion | supported as stated | medium | [393, 401] |
| ATT-TVISION-F11 | TVision's US panellist privacy policy (updated 9 September 2024) lists collection of face geometry, described as a biometric identifier, and of periodic snapshot images of the webcam's whole field of view. It permits sharing personal information with clients and partners, in some cases in hashed form that may be re-identified. It sets destruction of face geometries within three years of a panellist's last interaction and states that TVision is not subject to the CCPA or VCDPA. TVision's marketing and method pages say that no video streams leave the home, that no audio or images leave once a household is in-tab, and that all data leaving the home is 'completely anonymous' after an initial training period. | 9, 12 | vendor assertion | supported as stated | high | [306, 45, 291] |
| ATT-TVISION-F21 | Per Viant's 10-Q (10 August 2026), Viant assumed Nielsen's patent litigation against TVision; Nielsen has filed several actions since 2021. A jury returned a unanimous non-infringement verdict for TVision in the 2022 action on 11 June 2026, and by 14 August 2026 Nielsen was contesting that verdict. A 2025 action remains active, with TVision's November 2025 Sherman Act counterclaims. On 14 August 2026 the Federal Circuit affirmed the PTAB's ruling, in TVision's IPR, that the challenged claims of Nielsen's US Patent 11,470,243 are unpatentable. Indemnification from former equityholders is limited. | 9, 12 | direct record | supported with qualification | high | [82] |
| ATT-ADVERTEYES-F02 | As stated, plus: the predicted attention signal is defined against forced-exposure viewing (the reference test condition), and the percentile norm is described as a forced-exposure norm set for all countries and the desktop device type in the OpenAPI schema, which differs from the Metrics page wording. Both remain model estimates. | 10 | vendor assertion | supported with qualification | high | [529, 530, 531, 532] |
| ATT-ADVERTEYES-F09 | Named routes: (1) public API with an OpenAPI 3 specification (version 1.28.2 at api.adverteyes.ai/openapi.json), API key or OAuth 2.0, batch endpoint, beta suffix dropped 6 February 2026. (2) Clinch Flight Control Predict IQ Attention Score, launch announced by Clinch on 18 December 2025. (3) CreativeX Datalink launch partner, confirmed by CreativeX on 26 February 2026. (4) Vidmob: 2024 announcement, still named as an 'existing partner' in the current FAQ; Vidmob-side status unconfirmed. (5) Nielsen ONE Outcomes Marketplace (July 2025, Realeyes as first provider; continuity per MediaPost), a measurement platform. (6) Read-only MCP server documented since 12 May 2026 and opened to a limited group, per the 21 July 2026 blog (beta). Three routes (API, Clinch, CreativeX) are documented from the partner side as launched, which meets the cohort-B anchor-5 description. The grade-C cap holds the cell at 4. | 10 | synthesis | supported with qualification | medium | [533, 534, 535, 536, 537, 538, 539, 540, 541, 542, 543] |
| ATT-DAIVID-F01 | DAIVID first announced the Creative Data Feed API on 24 June 2025, called the 24 July 2025 retail-media release an update and the 15 August 2025 competitor-benchmarking release an upgrade. The Creative Data Feed name still appears in DAIVID boilerplate in partner releases until at least 13 April 2026 (Kinetiq). By September 2026 DAIVID's own site sells 'Creative Intelligence' via 'API, MCP or Self Serve', and the public API reference is titled 'DAIVID Creative Analyzer API' v1. That the Creative Analyzer API is the current API route of the register product remains inference; no source states a rename. | 10 | vendor assertion | supported with qualification | medium | [544, 545, 546, 547, 548, 39, 549, 550] |
| ATT-I-F03 | AdCP's own issue tracker documents that attention could not be carried end to end in AdCP 3.0: a buyer could set an attention_seconds or attention_score optimisation goal, but delivery reporting had no field for it and the goal named no vendor (issues #4579 and #4644, 15 to 17 May 2026); AdCP 3.1 kept attention reporting in vendor_metric_values, added vendor-bound goals (PR #4668, merged 17 May 2026) and deprecated the seller-native attention values, which are slated for removal at the next major version and remain in v3.2.0-rc.7. | 10 | direct record | supported with qualification | high | [146, 32, 143, 150] |
| ATT-I-F04 | On 28 and 29 September 2026 the AgenticAdvertising.org registry of member-enrolled agents listed 34 agents (29 sales, 3 signals, 1 creative, 1 buying by self-declared type) and no measurement agent; queries for metric_id=attention_units and type=measurement returned zero agents, although AdCP's registry documentation gives Adelaide-style attention as the first example of measurement-vendor discovery. | 10 | direct record | supported with qualification | medium | [141, 551] |
| ATT-I-F05 | In the capability records the AdCP registry returned for its 34 agents on 28 September 2026, attention appears only inside tool input schemas inherited from the AdCP specification (attention_score and attention_seconds enums and the placeholder vendor domain attentionvendor.example, 27 occurrences); no record names adelaidemetrics, lumen-research or integralads, and the TVision and DoubleVerify mentions sit in schema description text. | 10 | direct record | supported as stated | medium | [141] |
| ATT-I-F06 | Prebid SalesAgent (prebid/salesagent), a community open-source AdCP seller agent that AdCP's docs recommend for self-hosting, records in its storyboard coverage map (commit 2a961349, 17 September 2026, mapped to AdCP 3.1.1 storyboards) coverage 'none' for the vendor-metric optimisation flow and catalog precondition and 'partial' for vendor-metric accountability; its end-to-end vendor-metric scenario (placeholder vendor, simulated delivery) is listed as dormant. | 10 | direct record | supported with qualification | high | [151, 152] |
| ATT-I-F08 | IAB Tech Lab's AAMP is an umbrella of independently versioned repositories: AAMP 2.3 was released on 30 July 2026 (buyer-agent and seller-agent v2.4.2 on 18 August 2026; ARTF v1.0 on 20 July 2026). IAB Tech Lab announced AAMP 3.0 on 22 September 2026, adding OpenProposal, which is in public comment until 22 October 2026 (draft 1 tagged on GitHub on 24 September 2026) and states that it is not a ratified IAB Tech Lab standard. | 10 | direct record | supported with qualification | high | [552, 553, 36, 554, 147] |
| ATT-I-F09 | No AAMP artifact screened names an attention metric: OpenProposal 3.0's committed_metrics examples are viewable_rate, completion_rate and an audience-modelled DOOH metric; ARTF and OpenProposal do not contain the word; buyer-agent and seller-agent use it as ordinary English and in bundled copies of the Agentic Audiences draft taxonomy, which lists attention embeddings only as a future extension; the reference buyer agent v2.4.2 reports viewability and video completion rate. | 10 | synthesis | supported with qualification | medium | [36, 153, 154, 147] |
| ATT-I-F11 | IAS announced IAS Agent on 16 December 2025 as an AI assistant integrated into the IAS user interface, expected to roll out globally in early Q1 2026, for brand safety and suitability recommendations, insights and campaign setup; the release does not mention attention, and the product page says some features reflect the product's future state. | 10 | vendor assertion | supported as stated | medium | [155, 555] |
| ATT-I-F14 | AdCP v3.1.24 places an attention metric's standard reference, accreditations, methodology and unit in the vendor's own measurement-agent catalog, a surface marked experimental in 3.1; the accreditation entry names the accrediting body, certificate ID, expiry and evidence URL but has no structured field for environment scope, and AAO treats such entries as vendor assertions. With no attention vendor catalog registered at the screening date, no accreditation claim for an attention metric travels through AdCP (inference). | 10 | inference | supported with qualification | medium | [139, 149, 141] |
| ATT-I-F16 | AdCP 3.1 graduates viewability to standard delivery reporting (adding viewability.viewed_seconds) while keeping attention a vendor-defined Tier 2 metric reported through vendor_metric_values; its taxonomy says a metric graduates when a standards body publishes a measurement specification. The specification names the IAB Attention Measurement Guidelines only as an example standard_reference that a vendor's attention metric may implement, and its maintainers state that no MRC-or-equivalent body has graduated a shared attention definition, so AdCP does not treat the November 2025 guidelines as grounds to graduate attention. | 10 | direct record | supported with qualification | high | [143, 556] |
| ATT-I-F18 | The author's 10 July 2026 essay (author corpus, declared interest) asks which metric breaks first 'once agents become the primary buyers' and argues that attention 'fails hardest and fastest' because the registration event it infers has no referent once an agent, not a person, makes the decision; that mechanism concerns agents as the decider being marketed to, not the carriage of attention measured on human viewers through agent-driven media buying, so this stream's protocol-state evidence neither supports nor refutes it (inference). | 10 | inference | supported with qualification | medium | [557] |
| ATT-AMPLIFIED-F26 | The finding stands. Add that the API hostname in the vendor's published sample does not resolve in public DNS, so no public endpoint was observable. | 10 | vendor assertion | supported with qualification | medium | [196, 429, 558] |
| ATT-DOUBLEVERIFY-F18 | The finding stands, with two additions. DV's Butler/Till post (24 July 2026) is a vendor-reported production use of DV data in an AdCP buying workflow, but for suitability, fraud, viewability and contextual controls, not attention. The DV Neura Activation Agent, announced in June for Q3, was still described as forthcoming ('in the months ahead') on 17 September 2026. | 10 | synthesis | supported with qualification | medium | [217, 559, 218, 148, 560] |
| ATT-IAS-F15 | IAS announced IAS Agent on 16 December 2025 as an AI assistant integrated into the IAS user interface, built with Databricks Agent Bricks and expected for all IAS UI customers in early Q1 2026; its listed features centre on brand safety and suitability recommendations and campaign insights, the release does not mention attention, and the product page says some features reflect the product's future state. | 10 | vendor assertion | supported as stated | high | [155, 555] |
| ATT-I-F10 | DoubleVerify's blog (24 July 2026) describes Butler/Till using Claude as a buyer agent through AdCP, connected to DV's Media AdVantage Platform via MCP, with execution on PubMatic's AgenticOS; it names brand suitability as the core control and fraud and invalid-traffic avoidance, viewability thresholds and contextual alignment as optional add-ons, does not say which protocol carries which control, and does not mention attention, although DoubleVerify sells Authentic Attention. | 10, 12 | vendor assertion | supported with qualification | medium | [148] |
| ATT-I-F13 | In the sources screened by 28 September 2026 (the official MCP Registry, the AdCP agent registry, GitHub repository and code search, and sitemaps or newsrooms of Adelaide, Lumen, Amplified, TVision, Viant, DoubleVerify and IAS), this review found no public documentation of an MCP server, AdCP measurement agent or other agent-facing API that returns attention metrics from any attention vendor. | 10, 12 | synthesis | supported with qualification | medium | [157, 141] |
| ATT-I-F15 | AdCP v3.1.24's worked example layers an Adelaide attention performance standard with threshold 65, but the performance-standard schema defines threshold as a rate between 0 and 1 and defines attention_score as an 'attention score rate' over measurable impressions without stating what makes an impression count; a contractual attention floor is therefore under-specified in the current specification (inference). | 10, 12 | inference | supported as stated | medium | [561, 145, 144] |
| ATT-U-F23 | Viomba markets 'Viomba MCP' as a human attention intelligence layer for AI agents that predicts, scores and optimises ad attention across display, social and video, free to start. It is the only vendor screened whose home page leads with an MCP server, but Adverteyes (a dedicated MCP product page) and Amplified (AttentionAI's 'API and MCP') also advertise MCP access. None documents interoperability, production use or transactions. | 10, 12 | vendor assertion | supported with qualification | low | [562, 344] |
| ATT-ADELAIDE-F02 | MRC's accredited digital services list (accessed 29 September 2026) shows 'Attention Metrics' only for DoubleVerify Quality Analytics Platform (desktop, mobile web, mobile in-app); Adelaide does not appear on the list, so no Adelaide attention metric, and no Adelaide viewability or invalid-traffic input, holds MRC accreditation in the sources screened. | 12 | direct record | supported as stated | high | [2] |
| ATT-IAS-F20 | The main claim stands, but the limitation is contradicted: IAS's own analysis reports that Papyrus fraud traffic had about 13% higher attention scores than non-Papyrus traffic (product not named; IAS's attention score is Quality Attention, so applying it to Quality Attention is a close inference), before IAS filtered the traffic as invalid. This is direct vendor evidence that synthetic interaction raised attention scores, not only a general manipulability point. | 12 | vendor assertion | supported with qualification | medium | [240] |
| ATT-E-F21 | For Samba (formerly Samba TV, the author's employer), this review found no attention-based trading offer in the sources screened by the cutoff. Samba appears in the attention market only as an outcome-data supplier: Adelaide names it among 'tune-in attribution' providers used to calibrate AU. | 13 | vendor assertion | supported as stated | medium | [106] |
| ATT-D-F12 | On 6 August 2026 DoubleVerify (NYSE: DV) agreed to be acquired for $13.60 per share in cash (enterprise value about $2.15 billion) by Neptune BidCo US Inc., the parent of Nielsen Company (US), LLC and an affiliate of funds managed by Elliott and Brookfield. Closing requires DV stockholder approval, HSR expiry and certain specified foreign antitrust or foreign-investment clearances; termination fees are $60 million (DV) and $144 million (Parent); the initial termination date is 6 May 2027, automatically extendable to 6 August 2027 if only regulatory conditions remain. At 29 September 2026 the deal was pending. DV filed its definitive proxy (DEFM14A) that day for a special meeting on 29 October 2026. The HSR waiting period had expired on 21 September 2026, and DV targets closing by the first quarter of 2027. | abstract | direct record | supported with qualification | high | [465, 84, 436, 221] |
| ATT-A-F01 | The final IAB and MRC Attention Measurement Guidelines are Version 1.0, published November 2025. IAB's web page for them is dated 11 November 2025, and MRC lists them as issued in November 2025. They follow a public-comment draft published 12 May 2025, with comments due by 12 July 2025. | abstract, 1, 2 | direct record | supported as stated | high | [41, 430, 425, 563] |
| ATT-A-F18 | Vendors publish different attention units with different constructs and bases. Adelaide's AU is a 0-100 score of 'a placement's probability of attention and subsequent impact' from a machine-learning model that blends media-quality signals with outcome data. Lumen's APM is total attention seconds per 1,000 impressions, produced from predictive models built on panel, survey and proxy data (Lumen blog, February 2023). Amplified reports active and passive attention seconds. TVision measures time a panelist is looking at the TV screen. IAS Quality Attention is a single machine-learned score that IAS says predicts whether an impression is likely to lead to a business result. DoubleVerify reports indices built from device signals, and since April 2026 also accredited metrics from a modelled methodology. AU and IAS Quality Attention are described by their vendors as outcome-predictive, so they mix the attention layer with response layers. None is comparable across vendors. | abstract, 1, 2, 6, 7 | vendor assertion | supported with qualification | medium | [43, 44, 564, 45, 73, 433] |
| ATT-DOUBLEVERIFY-F08 | DoubleVerify's 10-K describes Authentic Attention as built from exposure signals (viewable time, share of screen, video presentation, audibility) and engagement signals (touches, screen orientation, video playback, audio controls) that ladder into the DV Attention Index, and says customers use it 'to predict which ads will impact consumers'; no reference method is named there, and the 2026 product page adds 'eyes-on-screen' to the engagement events without describing how it is observed. | abstract, 1, 2, 8 | direct record | supported as stated | high | [47, 209] |
| ATT-B-F08 | In the ARF's Phase 2 test (June 2024), in which 12 attention-measurement companies scored a common set of 32 ads with their usual methods, two eye-tracking-only companies' scores for the same ads agreed at R-squared 0.08 and 0.005 (appendix: 0.0005), eye-tracking-plus-facial-coding companies at 0.78, survey companies at 0.66 and neuro companies at 0.10, and no method group's scores correlated with advertisers' own success ratings above R-squared 0.12. | abstract, 1, 3 | independent measurement | supported with qualification | medium | [4] |
| ATT-TVISION-F02 | Viant completed the acquisition on 1 May 2026. Its 10-Q for the quarter ended 30 June 2026 puts the estimated purchase consideration at about $37.5 million: $16.2 million cash, $18.9 million of Class A stock and an estimated holdback, with $4.2 million held back for working-capital adjustments and indemnification. The preliminary allocation records $13.4 million of goodwill and $12.3 million of intangibles, including $1.1 million of 'panelist relationships' amortised over 4 years. The 10-Q states that TVision's results were not material. | abstract, 1, 5, 6 | direct record | supported as stated | high | [82] |
| ATT-G-F05 | Third-party attention measurement on social is documented through platform integrations: Lumen announced a Snapchat attention measurement integration on 18 March 2024 (model trained on eye-tracking data gathered with Snap since 2021); DoubleVerify's Authentic Attention for Social became globally available on Snapchat on 30 June 2025 (DV exposure data plus Lumen eyes-on-screen signals); IAS announced Snap Attention Measurement the same day (Lumen eye-tracking plus IAS media-quality data, 'will be available' in IAS Signal); and DoubleVerify extended Authentic Attention to TikTok on 10 December 2025 as a badged TikTok Marketing Partner using direct impression-level signals. All three Snap products rest on Lumen's eye-tracking model. | abstract, 1, 8 | vendor assertion | supported with qualification | high | [116, 118, 117] |
| ATT-ADELAIDE-F07 | For CTV, Adelaide says AU combines an image pixel fired at ad serve, ACR matching to program and pod, and TVision panel data that sets variable weights, and is trained on brand lift (Kantar, Lucid, Dynata), tune-in attribution (Samba, VideoAmp) and advertiser sales data; it reports CTV placements averaging 67 AU, about 23% above its linear TV benchmark. Samba is the author's employer. | abstract, 1, 8, 13 | vendor assertion | supported as stated | medium | [106] |
| ATT-G-F02 | In CTV, TVision is the only cohort A product whose attention data come from direct observation of people (optical sensors in opt-in panel homes, projected to the US population by daily weighting). Adelaide, DoubleVerify and Lumen score CTV impressions with models, and all three document TVision panel data as an input (Adelaide in its July 2026 method blog; DoubleVerify only in its June 2023 partnership announcement; Lumen in its 2023 to 2024 TVision, Nexxen and LG Ad Solutions releases). Lumen's Netflix scores (European clients) and its Teads HomeScreen model are machine-learning predictions from eye-tracking data. Whether TVision data remain licensed to these three vendors after Viant completed the acquisition on 1 May 2026 is not documented. | abstract, 1, 8, appendix | synthesis | supported with qualification | medium | [45, 106, 220, 110] |
| ATT-A-F10 | MRC first granted accreditation for DoubleVerify's 'Authentic Attention' metrics suite in desktop, mobile web and mobile in-app in December 2022, nearly three years before the attention guidelines were finalised. | abstract, 2 | direct record | supported as stated | high | [215, 433] |
| ATT-A-F12 | On 28 September 2026, MRC's list of accredited digital services has one entry under the 'Attention Metrics' category: DoubleVerify Quality Analytics Platform, for desktop, mobile web and mobile in-app. The Integral Ad Science entry lists no attention metric. No accredited attention product from Adelaide, Lumen, TVision or Amplified appears on MRC's digital, television or multimedia accredited lists. | abstract, 2, 6, 8, 11, 12, appendix | direct record | supported as stated | high | [2] |
| ATT-A-F13 | In January 2026, MRC removed 'IAS Attention Measurement', and separately 'IAS In-Game Measurement', from in-process status, citing 'IAS' lack of intent to engage in third-party audit'. Digiday had reported in October 2024 that IAS was in pre-assessment for attention. | abstract, 2, 8 | direct record | supported as stated | high | [49, 339] |
| ATT-A-F14 | On 28 September 2026 MRC lists Adelaide's AU under review with 'Pre-audit reviewed 5/2026, Audit not begun', covering digital display and video, audio, linear TV, cinema and digital OOH. The same five-channel scope, including digital OOH, has been on MRC's under-review list since at least 2 November 2024, days after MRC told Digiday that Adelaide was in pre-assessment. The CESP grid's omission of digital OOH is a CESP transcription difference, not evidence of a scope change. | abstract, 2, 8, 11 | direct record | supported with qualification | high | [48, 339] |
| ATT-A-F11 | MRC's letter of 13 April 2026 continues DoubleVerify's accreditation for display and video Authentic Attention metrics in desktop, mobile web and mobile in-app, and grants initial accreditation for 'Authentic Attention reporting' in the same environments (Indexed Impressions and Direct-Indexed-Rate, and metrics based on DV's Authentic Non-viewable and Authentic Modeled methodologies), which MRC's July 2026 quarterly report labels 'Authentic Attention V4'. CTV is accredited for impressions, verification and video viewability and completion, not for attention, and no pre-bid attention offering is accredited. The standards the letter lists do not include the Attention Measurement Guidelines. | abstract, 2, 10, 12 | direct record | supported with qualification | high | [42, 205] |
| ATT-A-F16 | Under the registered rule, H3 codes as contradicted at 28 September 2026, on a narrow basis: the only MRC attention accreditation found is DoubleVerify’s, granted in December 2022 and continued and extended on 13 April 2026, and the standards its letter lists do not include the Attention Measurement Guidelines. The letter’s stated basis also includes the procedures described in DoubleVerify’s answers to MRC’s questionnaire, so whether the guidelines informed the audit is not public; only MRC could settle it. IAB’s page calls the guidelines ‘the basis for MRC accreditation audits’. The status is as read on 28 September 2026 and was not rechecked before publication. | abstract, 2, 10, 12, appendix | synthesis | supported as stated | high | [42, 2, 425, 563] |
| ATT-C-F27 | No commissioned or vendor study screened makes a held-out comparison of an attention metric against time in view; the comparisons published are in-sample and mostly against viewability or completion rate, several vendor attention models take viewable time as an input, and the one commissioned review with published statistics (PwC for Lumen) found no significant memory advantage over viewability; H1a is unresolved on this evidence. | abstract, 3, 4, 11, 12, appendix | synthesis | supported as stated | medium | [8, 10, 71, 64] |
| ATT-C-F26 | In the vendor, agency and media-owner sources screened by 28 September 2026 (Adelaide, Lumen, Amplified, TVision, Wurl, DoubleVerify, IAS, dentsu, Havas, Teads, Nativo, Vevo, Ogury, TripleLift, Newsworks), this review found no public study that compares media bought on an attention metric with a randomised or holdout control bought on standard metrics and reports a memory, brand or sales gain with cost per incremental outcome and uncertainty; H1b is unresolved on commissioned and vendor evidence. | abstract, 3, 4, 12, appendix | synthesis | supported with qualification | medium | [69, 65, 11, 10] |
| ATT-B-F01 | This review found no public study, in the sources screened by 1 October 2026 (the Journal of Advertising and Journal of Advertising Research via Crossref title filters, ARF reports, logged web searches; vendor and commissioned studies were not screened and the International Journal of Advertising was not in the screen), that compares an attention metric with a viewability or time-in-view baseline on held-out data for memory, brand or sales outcomes. The closest comparisons (Bruns et al. 2025 and McGranaghan et al. 2021 on the same data; Simonov et al. 2025 on different subsamples) are all in-sample, so H1a is coded unresolved for each outcome on independent and academic evidence. | abstract, 3, 11, 12, appendix | synthesis | supported with qualification | medium | [46, 51, 52] |
| ATT-B-F18 | This review found no randomised or hold-out test, by independent or academic researchers, of buying media on an attention metric against a standard-metric control with costs reported, in the sources screened by the cutoff, so H1b is unresolved from this stream. | abstract, 3, 12, appendix | synthesis | supported as stated | medium | [51, 4, 5] |
| ATT-E-F22 | Where an attention guarantee discloses who measures delivery (the WSJ, Kargo and TripleLift offers, all on Adelaide’s AU, and SeenThis Attention on Lumen’s measure), it is the vendor whose metric the offer uses. For Vevo, Duolingo and Teads the measurer is implied but not stated, and for A+E’s 2020 and 2021 guarantee to dentsu for Budweiser, built on TVision panel data, it is not stated. Most guarantees found are defined on Adelaide’s AU; the exceptions are SeenThis (Lumen) and A+E (TVision). This review found no independent third-party audit of guarantee delivery in the sources screened; the only third-party review found is MediaSense’s February to June 2025 methodology review of AU, which Adelaide requested. | abstract, 4, 6, 7, 12 | synthesis | contradicted | medium | [94, 93, 91, 66] |
| ATT-C-F11 | PwC's review, commissioned by Lumen (report pages dated November 2022, cover dated March 2023, released by Lumen on 27 April 2023), re-performed Lumen's campaign analysis for three advertisers on a Lumen-supplied subset of data (Lumen's own analysis covered about 1 billion impressions; PwC could not verify Lumen's data cleaning or recreate the exact results) and found attention correlated better than viewability with click-through and conversion; in a brand-lift pilot for one advertiser (2,400+ panelists, six studies, 12 metrics) attention was not significantly better than viewability on first-mention awareness, any-mention awareness or digital ad recall, and PwC found no significant result for attention or viewability on the eight other metrics it lists. | abstract, 4, 6, 11, 12 | commissioned measurement | supported with qualification | high | [8] |
| ATT-C-F31 | Coding the numbers in this stream by measurement basis (Q2): the attention figures behind Adelaide AU, Lumen LAMP and IAS Quality Attention results are model estimates, applied pre-bid, in-flight or post-campaign; DV Authentic Attention in the 2021 Mondelez study is an index of device and on-screen signals (viewable time, share of screen, audibility, interaction) with no reference labels described, which the Q2 rule codes as a candidate proxy, not a validated model estimate; direct observation of people appears only in panel work (TVision's in-home panel, Amplified's participant panels, dentsu's eye-tracking panels), and TVision's campaign-level figures mix observation and projection. | abstract, 4, 6, 12, appendix | synthesis | supported with qualification | medium | [10, 63, 73, 74, 9] |
| ATT-D-F05 | Novacap's take-private of IAS was announced on 24 September 2025 at $10.30 per share in cash, in a deal the parties said values IAS at approximately $1.9 billion (basis not stated), and completed on 23 December 2025, when the aggregate price paid for the outstanding shares was approximately $1.6 billion; Nasdaq filed Form 25 the same day and IAS filed Form 15 on 2 January 2026 to deregister. | abstract, 5 | direct record | supported with qualification | high | [83, 437] |
| ATT-DOUBLEVERIFY-F17 | At 29 September 2026 the merger was pending. DV filed its definitive proxy (DEFM14A) on 29 September 2026 for a special meeting on 29 October 2026, and the HSR waiting period expired on 21 September 2026. No closing filing exists. DV targets closing by the first quarter of 2027, a forecast from the corrected release of 7 August 2026, repeated in the 10-Q, PREM14A and DEFM14A. DV has suspended earnings calls and withdrawn guidance for the pendency. | abstract, 5 | direct record | contradicted | high | [181, 465, 565] |
| ATT-D-F28 | H4 is unresolved from public sources: full-text reads of the DV and IAS 10-Ks, DV's Q2 2026 10-Q and its September 2026 preliminary merger proxy, UK registry accounts for Lumen and DAIVID, and vendor releases for Adelaide, Amplified and Neurons found no filing, revenue or pricing series, or buyer-level evidence linking verification companies' attention bundling to a change in specialists' demand, pricing or revenue, in the sources screened by 28 September 2026. | abstract, 5, 12, appendix | synthesis | supported with qualification | medium | [47, 231, 285, 232] |
| ATT-E-F18 | Two attention-priced billing models were found, neither with documented US sales. SeenThis Attention (launched 30 June 2026, offered globally) is bought on attentive CPM or cost per attentive second, measured by Lumen. Impact Media announced in April 2024 a US suite that bills only when a display ad is seen for one second or a video ad for three seconds, without saying how 'seen' is measured; its current availability is not documented. Otherwise this review found no attention-priced billing unit in the sources screened by the cutoff. Teads' filings list CPC, CPM, CPV, CPA, video completion-based and other outcome-based models, none defined on attention; Lumen presents 'cost of attention' as an analysis, and Lumen Attention Bidding is a DSP custom algorithm with no named DSP or price. | abstract, 6 | inference | contradicted | medium | [392, 96, 566] |
| ATT-E-F23 | On the registered H2 rule, attention is used in US trading beyond planning and reporting. It is marketed as a live optimisation input in named buying platforms. Yahoo DSP, a US platform, launched Adelaide AU segments in June 2023 with Verizon as brand launch partner, and Adelaide said in June 2026 that the targeting is available there. The Trade Desk has offered AU tiers since October 2024, and its own publication reported iFIT’s use of Adelaide segments there in October 2025. Amazon DSP has offered AU tiers since June 2026; neither release states geography. DV360’s Help Center documents a live Adelaide algorithm with no go-live date, and JLR North America described AU custom bidding in DV360 in May 2025. Guarantees ran but do not carry the verdict on their own: the Wall Street Journal’s from October to November 2024, for advertisers the sources do not place in the US, which Adweek described as a promotion, and A+E’s for Budweiser in the US in 2020 and 2021, on terms not disclosed. DoubleVerify’s segments and Scibids optimisation name no platform, and Viant described its TVision pre-bid attention targeting, offered through the Havas marketplace, as in testing in August 2026. Apart from JLR’s account to a trade publication, all of these rest on vendor or seller statements. No guarantee states threshold, remedy and independent audit together, and no offer discloses US spend, impressions or buyers against a denominator. H2 is therefore ‘supported’ with its scale undocumented: the rule’s ‘supported at scale’ needs disclosed volume, which no source gives, so the size of the trade is unknown rather than shown to be small. No use meets the currency-grade test. | abstract, 6, 7, 12, appendix | inference | supported with qualification | medium | [91, 92, 334, 93, 25] |
| ATT-E-F24 | Applying the Q1 coding rule, no metric-environment pair found by this stream is coded 'currency'. AU pairs whose guarantee terms are unpublished (Vevo, TripleLift, Duolingo, Teads) should be coded 'unresolved' under the rule's unresolved clause, not 'diagnostic'. Adelaide AU pairs with readable terms (WSJ 2024, Kargo) lack documented accreditation, and Adelaide asserts only that it entered MRC review in 2024. DoubleVerify's attention metric is used in optimisation segments, with no guarantee or billing use found. All codes are provisional until workstream A confirms accreditation scope. | abstract, 6, 7, 12, appendix | inference | supported with qualification | low | [63, 222] |
| ATT-I-F01 | At the screening date (28 September 2026) the current stable AdCP release was v3.1.24, published 23 September 2026, and v3.2.0 was a release candidate (rc.7, 27 September 2026); the 3.1 line began with v3.1.0 on 18 June 2026, and v3.0.0 was released on 22 April 2026. | abstract, 6, 10 | direct record | supported as stated | high | [567] |
| ATT-I-F02 | AdCP v3.1.24 can represent a vendor-bound attention value for forecasting (planning), as a vendor_metric optimisation goal (in-flight), in delivery reporting through vendor_metric_values (value required; unit and measurable_impressions optional), and as a vendor-named performance-standard floor (attention_score) that can be paired with a makegood remedy; this is schema representation only. | abstract, 6, 10, 12 | direct record | supported with qualification | high | [144, 138, 140, 142] |
| ATT-F-F22 | In IAB's US buy-side Outlook surveys, the share expecting to focus 'significantly/somewhat more' on attention metrics was 36% for 2023 (n=223, fielded Oct-Nov 2022), 47% for 2024 (n=203, fielded Oct-Nov 2023) and 47% for 2025 (n=200, fielded Nov-Dec 2024). The share rose once and then held flat, while the N/A share fell from 24% to 19% to 14%. | abstract, 7 | independent measurement | supported with qualification | medium | [103, 102] |
| ATT-F-F29 | Of 25 buyer-use records in the stream's table (plus four non-use rows), 14 involve planning, 8 optimisation, 5 reporting or evaluation, 4 guarantee-related and 0 billing units. With the missed records added, optimisation rises to at least 12 and the guarantee category includes one executed, repeated US guarantee (A+E for AB InBev via dentsu, Q4 2020 and Q1 2021). No record discloses US spend, impressions, campaigns or buyers with a denominator. | abstract, 7 | synthesis | supported with qualification | medium | [484, 476] |
| ATT-G-F06 | This review found no attention measurement product in Meta’s or YouTube’s measurement partner programmes, with a platform partner badge, or with a certified partner reporting integration for attention, in the sources screened by the cutoff: DoubleVerify’s MRC-accredited Meta and YouTube reporting covers impressions and viewability, and its June 2026 Meta expansion is Authentic AdVantage with no attention mention. xpln.ai states that it measures Facebook and Instagram through the platforms’ reporting interfaces or API, without a granular-data partnership (vendor disclosure). Attention on these platforms is also measured without platform integration: Adelaide has offered AU ratings for YouTube, Facebook and Instagram placements since September 2020 (from licensed Amplified panel data) and published a Google-commissioned AU analysis of YouTube placements (2022 to 2024); Lumen measured YouTube campaigns for Silverpush and Epitaph (February 2026); TVision’s panel observes attention in YouTube’s CTV app; and Amplified’s planning tool lists Meta. | abstract, 8 | synthesis | supported with qualification | medium | [42, 568, 569, 570, 109] |
| ATT-XPLN-F12 | Vendor sources document xpln.ai attention measurement as live in four of the seven scored channels: display (web and in-app), online video (VAST/VPAID; YouTube listed separately from social), CTV and social (Facebook, Instagram, Snapchat, TikTok, X, through platform reporting interfaces or API, with no granular-data partnership). xpln.ai answered 'No' to CESP for digital audio, DOOH, OOH, linear TV, terrestrial radio, print and cinema, and no in-game coverage was found. | abstract, 8 | vendor assertion | supported as stated | medium | [571, 121, 496] |
| ATT-G-F01 | Applying the Q3 rule to the cohort A register products, in the sources screened by 29 September 2026: display and online video are covered (Adelaide, Lumen, DoubleVerify, IAS and Amplified, with Amplified documenting in-flight tag measurement as well as planning); CTV is covered, with four products documenting US CTV measurement (TVision by panel observation; Adelaide, DoubleVerify and Lumen by models, all three documented as using TVision panel data) plus Amplified (planning and in-flight tags, US availability not stated); social is covered (Adelaide, Lumen, DoubleVerify, Amplified planning; IAS announced); audio and DOOH are thin; in-game is study only; retail media is thin, with one product (DoubleVerify) presenting Authentic Attention in its commerce-media offering and reporting aggregate attention-engagement results for retail media network owned-and-operated inventory, without naming the networks or placements measured. | abstract, 8, appendix | synthesis | supported with qualification | medium | [105, 119, 42, 45, 115, 110, 113] |
| ATT-G-F10 | No cohort A product documents live in-game attention measurement in the sources screened. In-game evidence consists of one-off studies: dentsu and Lumen's January 2024 report (gaming APM 6,736; rewarded video 10,043 APM vs 6,100 for online video; intrinsic in-game 3,442 vs 1,416, stated as 'seconds' and consistent with attentive seconds per 1,000 impressions), measured with eye-tracking while respondents played real games on mobile and PC; the Lumen and Anzu study released in June 2023; and Amplified's April 2023 study with Digital Turbine. One in-game ad seller, Frameplay, markets a proprietary viewable-time metric as attention, on the basis of one-campaign eye-tracking validation. | abstract, 8, appendix | vendor assertion | supported with qualification | medium | [113, 29] |
| ATT-H-F02 | The AI Act bans emotion inference only in the workplace and education (with medical and safety exceptions) from 2 February 2025. The Commission guidelines place customer-facing and advertising emotion inference, including 'intelligent billboards', outside the ban. Those that infer emotions from biometric data are emotion recognition systems and are high-risk under Annex III 1(c), subject to Art 6(3); emotion inference not based on biometric data (the guidelines' keystroke and text examples aside, e.g. text sentiment) is neither an emotion recognition system nor high-risk. | abstract, 9 | direct record | supported with qualification | high | [30, 124] |
| ATT-H-F05 | Emotion recognition systems are high-risk under Annex III 1(c), and Regulation 2026/1744 moved the Chapter III Sections 1-3 obligations for Annex III systems to 2 December 2027. Under Article 111(2) as amended, an emotion-recognition system already placed on the market or put into service before 2 December 2027 is subject to those obligations only if its design changes significantly after that date (public-authority uses excepted, with a 2030 deadline). Every facial-coding attention product on the market at the cutoff is therefore outside the high-risk regime until a significant design change, while the Article 50(3) transparency duty applies to deployers now. | abstract, 9 | direct record | supported with qualification | high | [30, 123] |
| ATT-I-F07 | AdCP 3.1 defines seller conformance storyboards for vendor-metric optimisation and accountability whose fixtures use the placeholder attentionvendor.example and delivery injected by a test controller; they test a seller's handling of vendor metrics, not interoperability with a real attention vendor. AAO's public compliance records show three registered sales agents (Addiscovery, Purrsonality Seller and No Fluff Advisory's sales agent, operated by the author's firm) passing the accountability storyboard between 15 and 23 September 2026 with the placeholder vendor, and no agent tested on the optimisation-flow or catalog-precondition storyboards; no pass involving a named attention vendor was found. | abstract, 10 | direct record | supported with qualification | medium | [33, 572] |
| ATT-I-F17 | Under the registered H5 rule, H5 is unresolved at 28 to 29 September 2026. No named attention vendor's signal is documented beyond schema representation in AdCP v3.1.24. Three registered AdCP sales agents, one operated by the author's firm, pass AdCP's vendor-metric accountability storyboard with a placeholder attention vendor and simulated delivery, which shows seller-side handling in test mode, not a connector to or interoperability test with a real attention vendor. DoubleVerify's live MCP server and claimed AdCP support carry no documented attention metric. No source documents an attention signal dropped or failing in an agent-driven buy, so the contradiction rule is not met. | abstract, 10, 12, appendix | synthesis | supported with qualification | medium | [142, 141, 151, 33] |

### The corrections

Every substantive change after the verification passes, with what questioned it, what was checked, the verdict, the change and how it was verified.

**Corrections ledger**

| ID | Original claim or score | Challenge | Evidence checked | Verdict | Change made | Reason | Affected |
|---|---|---|---|---|---|---|---|
| R2-01 | DV360 row: '... rolled out mid-July 2026'; ATT-V-ADELAIDE-S03 event_date 'mid-July 2026 (rollout per announcement)' | The Adelaide verifier table and source metadata date the packaged DV360 integration to mid-July 2026, while the go-live date is undocumented. | Google's announcement page says only 'Expected mid July'; the operating page documents 'Maximize Adelaide AU' with no date. [159, 158] | supported as stated | Both now say: live by 29 September 2026 per Google's Help Center; announced for mid-July 2026; go-live date not documented. | Shared rules: verify every external finding against its primary source; independence is unknown when funding is unknown; bound absence claims. | evidence/verify/adelaide.json (data_tables.adelaide_activation_routes correction; ATT-V-ADELAIDE-S03) |
| R2-02 | ATT-B-F09 evidence_class independent_measurement | ARF Phase 3 coded independent_measurement although the release states no funding. | The release's 'About the Study' section describes 'multiple independent measurement providers and methodologies' and contains no funding or participation statement. [5] | supported as stated | Recoded as synthesis with low confidence; limitation added that independence is unknown. | Shared rules: verify every external finding against its primary source; independence is unknown when funding is unknown; bound absence claims. | data/claims-overlay.json (ATT-B-F09); evidence/research/B.json (ATT-B-F09 limitations) |
| R2-03 | Limitation: 'The AI-training exemption covers training, not the capture of panel data for measurement.' | The H-F12 limitation says the Texas AI exemption covers training only. | HB 149 text: 'the training, processing, or storage of biometric identifiers involved in developing, training, evaluating, disseminating, or otherwise offering artificial intelligence models or systems, unless ...'; subsection (f) on later commercial use. [129] | supported as stated | Claim and limitation now quote 503.001(e)(2) in full and note subsection (f); applicability to panels stated as not established. | Shared rules: verify every external finding against its primary source; independence is unknown when funding is unknown; bound absence claims. | data/claims-overlay.json (ATT-H-F12); evidence/research/H.json (ATT-H-F12 limitations) |
| R2-04 | 'It applies whatever the Colorado Privacy Act volume thresholds.' | H-F16 omits the notice duties for biometric identifiers and says the law applies whatever the thresholds, although the subsection (5) access right has its own. | HB24-1130: '(4)(a) A CONTROLLER SHALL NOT COLLECT OR PROCESS A BIOMETRIC IDENTIFIER OF A CONSUMER UNLESS THE CONTROLLER FIRST ...'; '(5)(b) THE REQUIREMENTS OF SUBSECTION (5)(a) OF THIS SECTION APPLY ONLY TO ...' (100,000 consumers, or 25,000 with revenue from sales, or common branding). [134] | supported as stated | Final wording adds the 6-1-1314(4)(a) notice duties and the subsection (5)(b) conditions for the access right. | Shared rules: verify every external finding against its primary source; independence is unknown when funding is unknown; bound absence claims. | data/claims-overlay.json (ATT-H-F16) |
| R2-05 | ATT-E-F18: no attention-priced billing unit found; ATT-E-F22: every guarantee is defined on Adelaide's AU | SeenThis Attention (30 June 2026) is missing: an attention-priced (aCPM, CPAS) and guaranteed offer measured by Lumen, with a stated remedy. | Release (30 June 2026): 'Available through attentive CPM (aCPM) and Cost Per Attentive Second (CPAS) buying models'; 'available globally'; measured by Lumen. Product page: 'We run more. At no extra cost.' [99, 100] | supported as stated | New claim ATT-R2-F01 and register record; E-F18 and E-F22 marked contradicted with corrected wording; H2 and Q1 coding to include the offer (still no accredited metric, so Q1 unchanged). | Shared rules: verify every external finding against its primary source; independence is unknown when funding is unknown; bound absence claims. | data/claims-extra.json; data/claims-overlay.json (ATT-E-F18, ATT-E-F22); data/sources-extra.json; data/product-register.json; profiles/vevo.json guarantee comparison table (to be read with the register) |
| R2-06 | ATT-E-F18 and the trading table | Impact Media's Guaranteed Attention Suite (April 2024) is missing from the guarantee and billing inventory. | Release (New York, 18 April 2024): 'billing clients only when a display ad is seen for one second or a video ad is seen for 3 seconds'; CTV and audio planned; no measurement method named. Search on 29 September 2026 found only April 2024 coverage. [101] | supported as stated | New claim ATT-R2-F02 and register record (listed, vendor-advertised, measurement method and current availability not documented); E-F18 corrected. | Shared rules: verify every external finding against its primary source; independence is unknown when funding is unknown; bound absence claims. | data/claims-extra.json; data/claims-overlay.json (ATT-E-F18); data/sources-extra.json; data/product-register.json |
| SC-01 | Adelaide AU / channel_coverage: v 4, grade C, verified | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | Social and audio coverage rest on vendor posts and a homepage (grade D), which the profile calls live by vendor claim. [168, 106, 105, 336, 329, 332] | supported with qualification | Adelaide AU / channel_coverage: v 3, grade D, verified | Grade cap: vendor marketing caps a cell at 3, applied after the anchor; same standard as this row’s activation cell and both Lumen rows. | data/scores.json or data/matrix-extra.json (Adelaide AU) |
| SC-02 | Amplified Campaign Strategy / channel_coverage: v 4, grade C, verified | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | The channel list comes only from a vendor product page and a release, graded D elsewhere in the matrix (Lumen SPOTLIGHT). [119, 351] | supported with qualification | Amplified Campaign Strategy / channel_coverage: v 3, grade D, verified | Same standard for the same kind of source; grade cap 3. | data/scores.json or data/matrix-extra.json (Amplified Campaign Strategy) |
| SC-03 | DoubleVerify Authentic Attention / outcome_validation: v 2, grade D, qualified | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | The Mondelez study gives no respondent counts and no other DV case states a sample. [71, 72, 212, 361, 360, 5] | supported with qualification | DoubleVerify Authentic Attention / outcome_validation: v 1, grade D, qualified | Anchor 2 requires sample and outcome stated; same reading as IAS Quality Attention and Neurons AI. | data/scores.json or data/matrix-extra.json (DoubleVerify Authentic Attention) |
| SC-04 | Amplified Campaign Performance / outcome_validation: v 2, grade D, qualified | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | The OMD company study measures panel-observed attention and is not specific to the tag. [195, 76, 346, 189] | supported with qualification | Amplified Campaign Performance / outcome_validation: v 1, grade D, qualified | Evidence is not inherited between rows or from the company; outcome validation concerns the product’s own measure. | data/scores.json or data/matrix-extra.json (Amplified Campaign Performance) |
| SC-05 | Amplified Campaign Strategy / outcome_validation: v 2, grade D, qualified | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | Same OMD study; it validates the panel instrument, not the planning estimates (Lumen’s row excludes Simonov et al. on the same reasoning). [346, 119, 199, 156] | supported with qualification | Amplified Campaign Strategy / outcome_validation: v 1, grade D, qualified | No inheritance; outcome validation concerns the product’s own measure. | data/scores.json or data/matrix-extra.json (Amplified Campaign Strategy) |
| SC-06 | Kantar LINK AI / production_scale: v ne | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | This row’s creative-diagnostics cell credits a 21,000+ creator-content benchmark base (a verifier correction). [247, 391, 263, 264] | supported with qualification | Kantar LINK AI / production_scale: v 1, grade D, qualified | Anchor 1 covers a norm-base total with no period; n/e applies only when the record gives no basis to score. | data/scores.json or data/matrix-extra.json (Kantar LINK AI) |
| SC-07 | TVision / assurance: v 1, grade D, qualified | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | The cell rests on MRC’s own lists, like the Adelaide, Amplified and Lumen assurance cells, graded A. [313, 2, 302, 50, 48, 303] | supported with qualification | TVision / assurance: v 1, grade A, qualified | Evidence grade follows the source: an accreditor record is grade A. | data/scores.json or data/matrix-extra.json (TVision) |
| SC-08 | IAS Quality Attention / assurance: v 2, grade A, verified | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | The verifier record for this capability is supported_with_qualification. [2, 49, 205, 73] | supported with qualification | IAS Quality Attention / assurance: v 2, grade A, qualified | Verification status follows the verifier record. | data/scores.json or data/matrix-extra.json (IAS Quality Attention) |
| SC-09 | Kantar LINK+ / outcome_validation: v 1, grade D, verified | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | The verifier record for this capability is supported_with_qualification. [258] | supported with qualification | Kantar LINK+ / outcome_validation: v 1, grade D, qualified | Verification status follows the verifier record. | data/scores.json or data/matrix-extra.json (Kantar LINK+) |
| SC-10 | Kantar LINK+ / granularity: v 3, grade D, verified | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | The verifier record for this capability is supported_with_qualification. [249, 246, 263] | supported with qualification | Kantar LINK+ / granularity: v 3, grade D, qualified | Verification status follows the verifier record. | data/scores.json or data/matrix-extra.json (Kantar LINK+) |
| SC-11 | Adverteyes PreView / outcome_validation: v 3, grade D, qualified | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | The joint study is vendor-produced and confounds the two products’ scores; it is treated the same way in the Adelaide row, which stays at 2. [573, 86] | supported with qualification | Adverteyes PreView / outcome_validation: v 2, grade D, qualified | Same evidence, same treatment across rows; anchor 3 requires commissioned validation with a published method. | data/scores.json or data/matrix-extra.json (Adverteyes PreView) |
| SC-12 | Neurons AI / method_validity: v 4, grade C, qualified | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | The validation set’s size and separation from training are not stated. [574, 575, 576, 577] | supported with qualification | Neurons AI / method_validity: v 3, grade C, qualified | Anchor 4 requires an agreement or error statistic on held-out reference data. | data/scores.json or data/matrix-extra.json (Neurons AI) |
| SC-13 | Amplified Campaign Performance / channel_coverage: v 3, grade C, verified | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | Rests on the vendor’s product pages, graded D in other rows (Kantar, iMotions, eye square, Lumen SPOTLIGHT). [186, 350, 349] | supported with qualification | Amplified Campaign Performance / channel_coverage: v 3, grade D, verified | Same grade for the same kind of source; value unchanged (at or under the D cap). | data/scores.json or data/matrix-extra.json (Amplified Campaign Performance) |
| SC-14 | Amplified Campaign Performance / method_validity: v 3, grade C, qualified | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | Rests on the vendor’s product pages, graded D in other rows (Kantar, iMotions, eye square, Lumen SPOTLIGHT). [186, 184, 189, 4] | supported with qualification | Amplified Campaign Performance / method_validity: v 3, grade D, qualified | Same grade for the same kind of source; value unchanged (at or under the D cap). | data/scores.json or data/matrix-extra.json (Amplified Campaign Performance) |
| SC-15 | Amplified Campaign Performance / granularity: v 3, grade C, qualified | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | Rests on the vendor’s product pages, graded D in other rows (Kantar, iMotions, eye square, Lumen SPOTLIGHT). [347, 186, 189] | supported with qualification | Amplified Campaign Performance / granularity: v 3, grade D, qualified | Same grade for the same kind of source; value unchanged (at or under the D cap). | data/scores.json or data/matrix-extra.json (Amplified Campaign Performance) |
| SC-16 | Amplified Campaign Strategy / method_validity: v 3, grade C, verified | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | Rests on the vendor’s product pages, graded D in other rows (Kantar, iMotions, eye square, Lumen SPOTLIGHT). [119, 184, 189] | supported with qualification | Amplified Campaign Strategy / method_validity: v 3, grade D, verified | Same grade for the same kind of source; value unchanged (at or under the D cap). | data/scores.json or data/matrix-extra.json (Amplified Campaign Strategy) |
| SC-17 | Amplified Campaign Strategy / granularity: v 2, grade C, verified | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | Rests on the vendor’s product pages, graded D in other rows (Kantar, iMotions, eye square, Lumen SPOTLIGHT). [119, 351] | supported with qualification | Amplified Campaign Strategy / granularity: v 2, grade D, verified | Same grade for the same kind of source; value unchanged (at or under the D cap). | data/scores.json or data/matrix-extra.json (Amplified Campaign Strategy) |
| SC-18 | Kantar Context Lab / production_scale: rationale cited the LINK+ norm base | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | The 50,000-test norm base is LINK+’s. [249, 248] | supported with qualification | Kantar Context Lab / production_scale: rationale cites only Context Lab’s own total; value unchanged | Evidence is not inherited between product rows. | data/scores.json or data/matrix-extra.json (Kantar Context Lab) |
| SC-19 | Neurons AI, Tobii Sticky, DAIVID, xpln.ai (11 cells) / source lists: rationales cited verifier sources missing from src | Consistency audit of all scored cells (a separate AI pass comparing every cell across rows against the anchors, caps and verification records) | The sources were registered in the rebuild of 30 September 2026. | supported with qualification | Neurons AI, Tobii Sticky, DAIVID, xpln.ai (11 cells) / source lists: verifier sources added to src; no value changed | Every cell lists the sources it rests on. | data/scores.json or data/matrix-extra.json (Neurons AI, Tobii Sticky, DAIVID, xpln.ai (11 cells)) |
| R3-01 | H3 wording exceeds its rule: the letter not naming the guidelines does not show they were not applied. | H3 wording exceeds its rule: the letter not naming the guidelines does not show they were not applied. | MRC letter of 13 April 2026: basis includes ‘the procedures described in DoubleVerify’s answers to the MRC questionnaire’. | supported as stated | H3 narrowed everywhere to what the letter lists; the letter’s questionnaire basis noted; whether the guidelines informed the audit stated as not public. | Verified against the primary source or the paper’s own data before any change. | claims-overlay ATT-A-F16; abstract; chapter 2; appendix; plain |
| R3-02 | The Meta/YouTube absence claim conflicts with xpln.ai’s stated platform API access. | The Meta/YouTube absence claim conflicts with xpln.ai’s stated platform API access. | ATT-XPLN-F12 (verified); CESP methodology grid: ‘Platform reporting API: Yes’ for Facebook and Instagram. | supported as stated | Absence narrowed to partner programmes, badges and certified integrations; xpln.ai’s stated API access recorded as a vendor disclosure. | Verified against the primary source or the paper’s own data before any change. | claims-overlay ATT-G-F06; abstract; chapter 8; plain |
| R3-03 | “In every guarantee the metric’s own vendor measures delivery” is unsupported where the measurer is not disclosed. | “In every guarantee the metric’s own vendor measures delivery” is unsupported where the measurer is not disclosed. | ATT-E-F22 itself: Adelaide’s role ‘implied but not stated for Vevo, Duolingo and Teads’; Vevo’s release does not say who measures compliance. | supported as stated | Reworded to: where the measurer is disclosed, it is the metric vendor; otherwise not stated. The bounded no-independent-audit finding is kept. | Verified against the primary source or the paper’s own data before any change. | claims-overlay ATT-E-F22; hero; abstract; chapters 6 and 12; fig-guarantees; plain; site texts |
| R3-04 | Claim panels, search and backlinks show the original claim rather than the corrected wording. | Claim panels, search and backlinks show the original claim rather than the corrected wording. | showClaim() headed the panel with c.claim; the search index used c.claim. | supported as stated | Template shows final wording first, with the original under ‘As first written’ when it differs; search and backlinks use the final wording. | Verified against the primary source or the paper’s own data before any change. | build/paper.template.html (paper copy; kit bug) |
| R3-05 | The abstract turns the accreditation check date into the accreditation date. | The abstract turns the accreditation check date into the accreditation date. | MRC January 2023 status communication: accreditation granted for Authentic Attention in December 2022. | supported as stated | ‘As of 28 September 2026, MRC listed …’; initial accreditation December 2022. | Verified against the primary source or the paper’s own data before any change. | abstract |
| R3-06 | Deletion results were described as missing-evidence fill results. | Deletion results were described as missing-evidence fill results. | data/fq-missing-evidence.json holds deletion cases; the q_miss_* statistics are drawn from it. | supported as stated | The single-cell examples moved to the deletion paragraph and described as deleting a score of 1; the scenario paragraph now reports only the band-sensitive count. | Verified against the primary source or the paper’s own data before any change. | chapter 11 |
| R3-07 | Neurons AI method validity 4 without held-out separation. | Neurons AI method validity 4 without held-out separation. | The cell states the validation set’s separation from training is not stated. | supported as stated | Already rescored to 3 by the consistency audit before this review arrived (SC-12). | Verified against the primary source or the paper’s own data before any change. | matrix-extra (Neurons AI) |
| R3-08 | The 2/3 boundary on method validity was applied inconsistently to Adelaide and IAS. | The 2/3 boundary on method validity was applied inconsistently to Adelaide and IAS. | Adelaide names partner gaze references (Amplified, Lumen, Tobii, TVision, Viomba) and publishes no reference sample or held-out statistic. | supported as stated | Adelaide rescored 3 to 2 under anchor 2 (partner reference data, no agreement statistic of its own), the same as IAS. The anchors were fixed before scoring and are not changed. | Verified against the primary source or the paper’s own data before any change. | scores.json (adelaide/method_validity) |
| R3-09 | xpln.ai data access relabelled restricted access as beta. | xpln.ai data access relabelled restricted access as beta. | xpln.ai privacy notice: ‘in limited circumstances impression level reporting is provided … via FTP end points or APIs’; no beta stated. | supported as stated | Availability set to live (restricted); rescored 2 to 3 on one documented, live route (impression-level FTP export). | Verified against the primary source or the paper’s own data before any change. | matrix-extra (xpln.ai/data_access) |
| R3-10 | Decisive verifier sources missing from Neurons and Tobii cell source lists. | Decisive verifier sources missing from Neurons and Tobii cell source lists. | Register holds ATT-V-NEURONS-S01/S02 and ATT-V-TOBII-S01/S02. | supported as stated | Already added by the consistency audit (SC-19). | Verified against the primary source or the paper’s own data before any change. | matrix-extra (Neurons, Tobii) |
| R3-11 | The guarantee inventory omits the A+E guarantee documented in chapter 7. | The guarantee inventory omits the A+E guarantee documented in chapter 7. | ATT-F-F14: a live A+E guarantee to dentsu for Budweiser, Q4 2020, repeated Q1 2021, on TVision panel data. | supported as stated | A+E (Q4 2020 and Q1 2021, TVision data, metric ambiguity noted) added to the guarantee table and figure; the earliest-case claim, the count and the exceptions updated. | Verified against the primary source or the paper’s own data before any change. | chapter 6; fig-guarantees; claims-overlay ATT-E-F22 |
| R3-12 | Undocumented scale was presented as small scale. | Undocumented scale was presented as small scale. | The H2 rule’s ‘supported at scale’ requires disclosed volume; absence of disclosure does not show small volume. | supported as stated | ‘Not at scale’ and ‘small trade’ replaced by ‘scale undocumented’; planning predominance qualified as a feature of the records reviewed. | Verified against the primary source or the paper’s own data before any change. | claims-overlay ATT-E-F23; hero; abstract; chapters 6, 7, 14; plain |
| R3-13 | Q1’s ‘diagnostic’ code conflates commercial function with assurance. | Q1’s ‘diagnostic’ code conflates commercial function with assurance. | Q1 coding rule requires accreditation for both currency codes. | supported as stated | Text states that the registered rule codes an unaccredited metric ‘diagnostic’ even when it prices media (SeenThis), so the code records missing assurance, not the absence of a billing function. The registered rule is unchanged. | Verified against the primary source or the paper’s own data before any change. | chapters 6 and 12; appendix |
| R3-14 | The enforcement absence statement is broader than the search. | The enforcement absence statement is broader than the search. | ATT-TVISION-F21: Nielsen has filed patent actions against TVision since 2021 (Viant 10-Q). | supported as stated | Restricted to privacy and biometric proceedings in the searches run; Nielsen’s patent actions against TVision acknowledged. | Verified against the primary source or the paper’s own data before any change. | chapters 9 and 12 |
| R3-15 | Chapter 12 drops the California transition period for risk assessments. | Chapter 12 drops the California transition period for risk assessments. | CPPA regulations §7155(b): pre-existing processing assessed ‘no later than December 31, 2027’. | supported as stated | Chapter 12 restores the 31 December 2027 deadline for processing already under way and notes that covered-business status was not established. | Verified against the primary source or the paper’s own data before any change. | chapter 12 |
| R3-16 | “Agents can hold the number, not use it” overstates the finding. | “Agents can hold the number, not use it” overstates the finding. | ATT-I-F17 (corrected): three agents pass a test-mode storyboard with a placeholder vendor. | supported as stated | Reworded to: no documented production use of a named vendor’s attention signal; test-mode handling with a placeholder vendor recorded separately. | Verified against the primary source or the paper’s own data before any change. | abstract; chapters 10 and 12; fig-protocol-states title; plain |
| R3-17 | Chapter 3 reads equal rounded R-squared as equivalence, and different samples as conflict. | Chapter 3 reads equal rounded R-squared as equivalence, and different samples as conflict. | Bruns et al.: .027 for each single-predictor model; Simonov et al.: different sub-samples. | supported as stated | ‘Same reported in-sample R-squared’ and ‘results not directly comparable’; H1a remains unresolved. | Verified against the primary source or the paper’s own data before any change. | chapter 3 |
| R3-18 | The hero statistic counts products as vendors. | The hero statistic counts products as vendors. | The matrix rows are products; several vendors have more than one row. | supported as stated | Label changed to ‘products scored’ in the hero and the site scope text. | Verified against the primary source or the paper’s own data before any change. | paper.config.json |
| R3-19 | The survey figure and abstract omit the flat 2025 wave. | The survey figure and abstract omit the flat 2025 wave. | ATT-F-F22 (corrected): 36% (2023), 47% (2024), 47% (2025). | supported as stated | The figure shows all three waves; the text says intent rose once and then held flat. | Verified against the primary source or the paper’s own data before any change. | fig-surveys; abstract; chapter 7; plain |
| I1-001 | On 28 September 2026 MRC accredited one attention product, DoubleVerify's Authentic Attention, in desktop, mobile web and mobile in-app, not CTV (ATT-A-F11). The standards its letter lists do not include the attention guidelines, so the registered test of whether accreditation rests on them is contradicted at that date (ATT-A-F16). | Wrong date for the accreditation. The sentence reads as though MRC granted accreditation on 28 September 2026. 28 September is only the date MRC's list was read. The accreditation was first granted in December 2022 and was continued and extended on 13 April 2026. The marked claim ATT-A-F11 is about the 13 April letter; the 28 September listing is ATT-A-F12, which is not marked. 'Its letter' has no antecedent, and 'the registered test ... is contradicted' misnames the verdict: the hypothesis H3 is what is contradicted, not a test. | ATT-A-F12: 'On 28 September 2026, MRC's list of accredited digital services has one entry under the Attention Metrics category: DoubleVerify Quality Analytics Platform, for desktop, mobile web and mobile in-app.' ATT-A-F10: 'MRC first granted accreditation ... in December 2022'. ATT-A-F11: 'MRC's letter of 13 April 2026 continues ... and grants initial accreditation for Authentic Attention reporting'. ATT-A-F16: 'H3 codes as contradicted at 28 September 2026'. Chapter 2 table: 'Accredited December 2022; continued March 2024 and 13 April 2026'; chapter 2 dates the H3 verdict 28-29 September 2026. | contradicted | Finding 2 now says MRC listed one accredited product as of 28 September 2026, first accredited December 2022 and continued and extended 13 April 2026, citing ATT-A-F12, F10, F11; H3 named and dated. | Holds: 28 September 2026 is the listing date (ATT-A-F12), first accreditation was December 2022 (ATT-A-F10) and the verdict is on H3, not a test; merged with R3-01 and R3-05 wording. | chapters/00-abstract.md |
| I1-002 | Agentic protocols can represent an attention score and do nothing further with it. | Wrong on both counts, and it overstates the H5 verdict. (1) 'Agentic protocols' in the plural is wrong: AAMP names no attention metric, so only AdCP can represent one. (2) 'Do nothing further with it' states a capability limit. The record contradicts that: AdCP 3.1.24 can carry the score as an optimisation goal and as a performance floor paired with a makegood remedy. What is missing is documented use by a named vendor, which is an evidence gap, not a limit of the protocol. Under the registered rule H5 is unresolved, and absence of evidence does not count as refutation. This sentence reads like a contradicted verdict, or like support for the rival hypothesis. | ATT-I-F02: AdCP v3.1.24 can represent a vendor-bound attention value 'for forecasting (planning), as a vendor_metric optimisation goal (in-flight), in delivery reporting ... and as a vendor-named performance-standard floor (attention_score) that can be paired with a makegood remedy'. ATT-I-F09: 'No AAMP artifact screened names an attention metric'. ATT-I-F17: 'H5 is unresolved ... Absence of accessible evidence yields unresolved, not refuted.' Chapter 10 standfirst: 'Attention can be written into the leading agentic advertising protocol, and on the public record it goes no further.' | contradicted | Inshort: AdCP can represent a named vendor's score; beyond the schema only test-mode handling with a placeholder vendor is documented; no production use found. | Holds: only AdCP represents attention (ATT-I-F09) and the gap is evidence, not capability (ATT-I-F17); rewritten to the R3-16 wording. | chapters/00-abstract.md |
| I1-003 | **Agents can hold the number, not use it.** AdCP v3.1.24 represents a vendor-bound attention metric in its schemas (ATT-I-F02); no attention signal is documented beyond schema representation, so the agentic hypothesis is unresolved (ATT-I-F17). | The heading 'not use it' makes an unresolved verdict sound like a finding that agents cannot use the score. The schema can already express the score as an optimisation goal and a floor. 'No attention signal' is broader than the claim, which says no named attention vendor's signal; placeholder-vendor storyboard passes do exist. The protocol version is a status-sensitive fact and carries no date. | ATT-I-F17: 'No named attention vendor's signal is documented beyond schema representation in AdCP v3.1.24. Three registered AdCP sales agents ... pass ... with a placeholder attention vendor'; 'H5 is unresolved at 28 to 29 September 2026'. ATT-I-F01: 'current stable AdCP release was v3.1.24, published 23 September 2026'. ATT-I-F02 lists the forecast, optimisation-goal, reporting and floor representations. | supported with qualification | Finding 10 heading 'no production use was found'; AdCP v3.1.24 dated 23 September 2026 with the four representations; test-mode placeholder handling; H5 unresolved. | Holds: heading overstated an unresolved verdict and the version lacked a date (ATT-I-F01); body rewritten to the R3-16 wording. | chapters/00-abstract.md |
| I1-004 | Independent assurance and outcome validation are the weakest dimensions in both. | The capability matrix contradicts this for cohort B. There, production scale has the lowest mean score, below independent assurance, and outcome validation is only third lowest. The sentence holds for cohort A only. It has no stat or claim marker, and chapter 11 repeats it in its standfirst and inshort. | capability-matrix.json, means of counted (verified or qualified) cells. Cohort A: outcome_validation 1.62 (8 cells), assurance 1.75 (8), next activation and data_access 2.38. Cohort B: production_scale 1.25 (8 cells, 4 n/e), assurance 1.27 (11), outcome_validation 1.40 (10), next data_access 2.17. | supported with qualification | Finding 11: two lowest-scoring in cohort A and, with production scale, the three lowest in cohort B; cohort names added. | Holds on recomputed matrix means (cohort B: production scale 1.25, assurance 1.27, outcome validation 1.40); replaced with the round-3 item 8 wording that chapter 11 uses. | chapters/00-abstract.md |
| I1-005 | Two outside reviews by another AI model, one of the briefs and one of the evidence, and two internal review passes of the full draft led to the [corrections] entries in the corrections ledger. | This misstates where the ledger entries came from. Of the 25 entries, 6 (R2-01 to R2-06) come from the second outside review of the evidence. The other 19 (SC-01 to SC-19) come from the separate AI consistency audit of the scores, which the sentence does not name. None comes from the review of the briefs, whose dispositions are recorded with the briefs, and at present none comes from the internal passes. Chapter 13's inshort repeats the sentence. | corrections-ledger.json: 6 entries of kind 'review' with review_pass 'OpenAI round 2 (evidence spot-check)'; 19 of kind 'scoring' with challenge 'Consistency audit of all scored cells'. stats.json corrections 25, corrections_scores 19. Chapter 13: brief-review 'dispositions are recorded with the briefs'; scoring-audit changes 'are logged as scoring corrections ([corrections_scores] in the ledger)'. | supported with qualification | Three outside reviews (briefs, evidence, full draft) plus two internal passes; consistency audit named; ledger holds [corrections] entries, [corrections_scores] of them scoring corrections. | Holds: ledger entries come from outside reviews and the scoring audit, not from the brief review or internal passes; combined with round-3 item 9 (three outside reviews). | chapters/00-abstract.md |
| I1-006 | Display, online video and CTV are covered; this review found no attention product integrated with Meta or YouTube (ATT-G-F06). Audio, in-game, DOOH and retail media are thin or study-only. | Social is left out, although the registered Q3 rule codes it as covered. Placing the list of covered channels next to the Meta/YouTube absence suggests social is not covered. The coverage codes carry no marker: ATT-G-F06 supports only the Meta/YouTube absence. | ATT-G-F01: 'social is covered (Adelaide, Lumen, DoubleVerify, Amplified planning; IAS announced); audio and DOOH are thin; in-game is study only; retail media is thin'. figure-data.json channels coding: Social live 3, 'covered'. Chapter 8 inshort: 'display, online video, CTV and social are covered'; 'Social coverage runs through platform integrations on Snap and TikTok' (ATT-G-F05). ATT-G-F10: in-game is one-off studies only. | supported with qualification | Finding 8 adds social, Q3 framing, ATT-G-F01, F05, F10 markers; in-game study only. | Holds: social is covered under Q3 (ATT-G-F01) and coverage lacked markers; phrased as 'social has platform integrations on Snap and TikTok' because Adelaide's social coverage is not integration-based (ATT-G-F06). | chapters/00-abstract.md |
| I1-007 | It is live as a bidding input in named platforms (ATT-E-F23). | This presents vendor and seller statements as observed fact. The claim, an inference, says attention is marketed as a live optimisation input and that all of it rests on vendor or seller statements. | ATT-E-F23 (evidence_class inference): 'attention is marketed as a live optimisation input in named buying platforms ...; all of these rest on vendor or seller statements.' Chapter 6 inshort: 'Attention is marketed as a live optimisation input in named US buying platforms'. | supported with qualification | Finding 6: marketed as a live bidding input in named buying platforms, on vendor and seller statements; inshort says 'marketed as a bidding input'. | Holds: ATT-E-F23 is an inference resting on vendor and seller statements. | chapters/00-abstract.md |
| I1-008 | No randomised or holdout test of buying on attention with costs was found (ATT-B-F18). | The marked claim is narrower than the sentence. ATT-B-F18 covers independent and academic work only, and its limitation says vendor and commissioned tests may exist. The unqualified absence needs ATT-C-F26, which covers vendor and commissioned sources. Chapters 3 and 12 cite both claims for the same statement. | ATT-B-F18 limitation: 'Vendor and commissioned optimisation tests (workstream C) may exist; this finding covers independent and academic work only.' ATT-C-F26: 'In the vendor, agency and media-owner sources screened ... no public study that compares media bought on an attention metric with a randomised or holdout control ... with cost per incremental outcome'. Chapter 3 inshort and chapter 12 cite (ATT-B-F18) (ATT-C-F26). | supported with qualification | Absence now bounded to independent, academic or vendor sources with both markers. | Holds: ATT-B-F18 covers independent and academic work only; ATT-C-F26 covers vendor sources. | chapters/00-abstract.md |
| I1-009 | **Both verification suites left or agreed to leave public markets.** Novacap took IAS private in December 2025, and on 6 August 2026 DoubleVerify agreed to be acquired by the parent of Nielsen for $13.60 a share, pending at the cutoff (ATT-D-F12). | Three problems. (1) The IAS take-private has no supporting marker: ATT-D-F12 covers only the DoubleVerify deal. (2) 'Pending at the cutoff' is not supported: the status was read on 29 September 2026, and the claims say it must be rechecked at the 2 October freeze. (3) 'Verification suites' is a category error: companies leave public markets, suites are products. | ATT-D-F12: 'At 29 September 2026 the deal was pending'; limitation: 'Status must be rechecked at the 2 Oct 2026 freeze.' ATT-DOUBLEVERIFY-F17: 'At 29 September 2026 the merger was pending.' ATT-D-F05 and ATT-IAS-F04: Novacap take-private 'completed on 23 December 2025'. Chapter 13: status-sensitive facts 'were read on 28 and 29 September 2026 and are dated in the text'. | supported with qualification | Heading 'verification companies'; IAS private on 23 December 2025; DV deal pending on 29 September 2026 with ATT-DOUBLEVERIFY-F17. | Holds: IAS take-private needed ATT-D-F05, status was read on 29 September, and companies (not suites) went private. | chapters/00-abstract.md |
| I1-010 | Viant, a buying platform, bought TVision (ATT-TVISION-F02). | Ownership is a status-sensitive fact and is given here without a date. | ATT-TVISION-F02: 'Viant completed the acquisition on 1 May 2026.' Chapter 5 inshort: 'completed its purchase of TVision on 1 May 2026'. | supported with qualification | Viant completed its purchase of TVision on 1 May 2026. | Holds: ATT-TVISION-F02 dates completion to 1 May 2026. | chapters/00-abstract.md |
| I1-011 | No public source links verification bundling to specialists' demand or pricing (ATT-D-F28). | The absence claim has no bound; it should be limited to the sources screened. It also drops revenue and does not state that H4 is unresolved. | ATT-D-F28: 'H4 is unresolved from public sources: full-text reads of ... found no filing, revenue or pricing series, or buyer-level evidence linking verification companies' attention bundling to a change in specialists' demand, pricing or revenue, in the sources screened by 28 September 2026.' Limitation: 'Private pricing and contracts are out of reach.' | supported with qualification | Bounded absence, revenue added, H4 unresolved stated. | Holds: ATT-D-F28 is bounded to sources screened by 28 September 2026, includes revenue and codes H4 unresolved. | chapters/00-abstract.md |
| I1-012 | 10 wording, precision and labelling issues in 00-abstract | Integrity pass findings rated P2: 00-abstract-12, 00-abstract-13, 00-abstract-14, 00-abstract-15, 00-abstract-16, 00-abstract-17, 00-abstract-18, 00-abstract-19, 00-abstract-20, 00-abstract-21 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 10 passages reworded for precision; no verdict changed | Precision and consistency. | chapters/00-abstract.md |
| I1-013 | In the Advertising Research Foundation's (ARF) Phase 2 test of creative measurement, 12 companies scored a common set of 32 ads. Two eye-tracking-only companies' scores agreed at R-squared 0.08 and 0.005. Eye-tracking-plus-facial-coding companies agreed at 0.78, survey companies at 0.66 and neuro companies at 0.10 (ATT-B-F08). | Two companies cannot produce two agreement values. The verifier correction on ATT-B-F08 shows two pairs (Eye1 vs Eye2 at 0.08; Eye3 vs Eye2 at 0.005 in the figure label, 0.0005 in the appendix), and each company scored only the subset of the 32 ads it chose, not the whole set. The review note requires both values where the report disagrees. Chapter 3 already uses the corrected rendering, so chapter 1 contradicts it. The test is also undated here (June 2024). | ATT-B-F08 review_note correction: '12 attention-measurement companies each scored the subset of a common 32-ad set they chose ... R-squared 0.08 between two eye-tracking-only companies (Eye1 vs Eye2) and 0.005 in the figure label (0.0005 in the appendix table) between a third eye-tracking-only company and one of them (Eye3 vs Eye2) ... Any value quoted from this report should carry both the figure label and the appendix value where they differ.' 03-validation.md ('The industry's own validation work'): '12 attention-measurement companies scored ads from a common set of 32 ... Two pairs of eye-tracking-only companies agreed at R-squared 0.08 and 0.005 (0.0005 in the report's appendix).' The survey value also differs (figure label 0.61, appendix 0.66); chapter 3 gives only 0.66, so if both ch | supported with qualification | Rewrote the ARF sentence to match chapter 3 (June 2024, ads from a common set of 32, two pairs at 0.08 and 0.005 with appendix 0.0005, paired groups) and added the survey figure label 0.61 next to 0.66. | ATT-B-F08 review note confirms two eye-tracking pairs, a chosen subset of the 32 ads, a June 2024 date, and that values should carry both the figure label and the appendix value where they differ. | chapters/01-defining.md |
| I1-014 | In CTV, Adelaide, DoubleVerify and Lumen score impressions with models that document TVision panel data as an input. Whether that supply continues after Viant completed its acquisition of TVision on 1 May 2026 is not documented (ATT-G-F02) (ATT-TVISION-F02). | The present tense drops the dates that the final wording attaches. DoubleVerify's TVision input rests only on a June 2023 pre-launch announcement, and DV's current CTV pages name no data source, so the sentence overstates a current dependency. The absence claim ('is not documented') has no bound. Chapters 5 and 8 also note that Adelaide's July 2026 blog, published after the closing, still says in the present tense that CTV AU 'incorporates TVision data'. | ATT-G-F02 final_wording: 'all three document TVision panel data as an input (Adelaide in its July 2026 method blog; DoubleVerify only in its June 2023 partnership announcement; Lumen in its 2023 to 2024 TVision, Nexxen and LG Ad Solutions releases)'. ATT-DOUBLEVERIFY-F13: DV's current CTV pages name no data source. ATT-TVISION-F04: Adelaide's 29 July 2026 blog says AU for CTV 'incorporates TVision data'; limitation: 'Adelaide's wording could describe historical training data.' 08-channels.md ('The shared panel is a dependency') and 05-market.md ('Vendor sources suggest they do'). | supported with qualification | Dated each vendor's TVision input, bounded the continuity absence claim to the sources screened, and added the Adelaide blog caveat as its own sentence with ATT-TVISION-F04. | ATT-G-F02 final wording dates each vendor's TVision documentation, and ATT-TVISION-F04 records Adelaide's present-tense July 2026 wording with its historical-data caveat. | chapters/01-defining.md |
| I1-015 | Webcam facial coding with gaze monitoring of recruited participants, N=150 per study | N=150 is the base of the LINK+ survey. The ledger says it is not stated whether it is also the base for the attention measures, since facial-coding completes are usually a subset. Facial coding is also an optional module that affects the price (the playbook lists 'Facial Coding or Eye Tracking'). The cell attaches the number to the wrong denominator. | ATT-KANTAR-F06 final_wording: 'a monadic online survey of N=150 respondents per study ... the LINK+ page lists facial coding as an optional module that affects price'; limitations: 'Whether N=150 is the base for the attention measures (facial-coding completes are usually a subset) is not stated; recruitment is not described.' Review note: 'The playbook's module list reads 'Facial Coding or Eye Tracking', i.e. either/or.' | supported with qualification | Kantar LINK+ cell now names the optional module, gives N=150 as the survey base and says the base for the attention measures is not stated. | ATT-KANTAR-F06 gives N=150 as the LINK+ survey base, does not say it is the base for the attention measures, and lists facial coding as an optional module. | chapters/01-defining.md |
| I1-016 | Fields the public record does not state are coded "not published". | The chapter misquotes the registered code. The Q2 rule codes unstated fields 'not disclosed' and reports them under dimension 11 (method disclosure). The same label appears again at the end of the chapter ('most often coded "not published"') and in the table cells 'None published' / 'Not published'. If this sentence changes, those cells should follow the registered code. | hypotheses.json Q2 rule_unresolved: 'A field the public record does not state is coded 'not disclosed', which is itself reported (dimension 11).' scoring-rules.json: dimension 11 is Method disclosure. | supported with qualification | Changed the coding sentence to "not disclosed" with the method disclosure note, and aligned the TVision, DoubleVerify and Kantar reference-label cells from 'None/Not published' to 'Not disclosed'. | hypotheses.json Q2 rule_unresolved codes unstated fields 'not disclosed' and reports them under dimension 11, method disclosure. | chapters/01-defining.md |
| I1-017 | a 2023 lab comparison co-authored by the software's developers reported 1.4 degrees against an EyeLink 1000 | 'The software's developers' reads as the developers of the open-source system named in the previous clause (WebGazer). The 2023 study was co-authored by developers of Labvanced, a different, commercial webcam system. The wording hides that the best recent accuracy evidence is vendor-authored, which the final wording and chapter 3 both state. | ATT-B-F13 final_wording: 'a 2023 lab comparison co-authored by the Labvanced developers reported 1.4 degrees accuracy against an EyeLink 1000 ... the best recent accuracy evidence is vendor-authored'; review_note: Kaduk et al. 2023, 'co-authored by developers of the Labvanced webcam system (Scicovery GmbH)'. 03-validation.md: 'The accuracy evidence for webcam eye tracking is thin and partly vendor-authored'. | supported with qualification | Named Labvanced as a commercial webcam eye-tracking system in the accuracy sentence. | ATT-B-F13 final wording and review note show the 2023 study was co-authored by Labvanced's developers, not WebGazer's. | chapters/01-defining.md |
| I1-018 | xpln.ai, the only attention product found on MRC's list as in audit (ATT-A-F15), | Audit status is status-sensitive but is undated here. Chapter 13 says such facts 'are dated in the text', and chapter 2 dates the same status to 28–29 September 2026. | ATT-A-F15: MRC records XPLN as 'Pre-audit reviewed 9/2025, Audit in process as of 11/2025'; scope_time '2025-09 to 2026-09-28'. 13-method.md, Limits: 'Status-sensitive facts (accreditation, ownership, transactions, protocol versions) were read on 28 and 29 September 2026 and are dated in the text.' 02-standards.md in-short dates the status 'at 28-29 September 2026'. | supported with qualification | Split the xpln.ai sentence and dated the audit status to MRC's list as read on 28 September 2026, recorded as in process as of November 2025. | ATT-A-F15 records 'Audit in process as of 11/2025' with scope ending 28 September 2026, and chapter 13 requires status-sensitive facts to be dated. | chapters/01-defining.md |
| I1-019 | 12 wording, precision and labelling issues in 01-defining | Integrity pass findings rated P2: 01-defining-08, 01-defining-09, 01-defining-10, 01-defining-11, 01-defining-12, 01-defining-13, 01-defining-14, 01-defining-15, 01-defining-16, 01-defining-17, 01-defining-18, 01-defining-19 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 12 passages reworded for precision; no verdict changed | Precision and consistency. | chapters/01-defining.md |
| I1-020 | AU is the metric behind Vevo's Attention Guaranteed, so that guarantee rests on an unaccredited metric (ATT-VEVO-F09). | States as fact that AU is the guaranteed metric. The ledger labels the metric's identity as inference: Vevo's release does not state the unit of the minimum. Chapter 6 calls it an inference, so the two chapters disagree on the evidence label. | ATT-VEVO-F05 final_wording: 'The guaranteed metric is, by clear implication (the release does not state the unit of the minimum), Adelaide's AU'; review_note: 'Label: Metric identity is inference.' ATT-VEVO-F04: the release 'states no threshold, remedy, audit, price or geography'. 06-trading.md line 35: 'It does not state the unit of the minimum; reading it as AU is an inference from the partnership (ATT-VEVO-F05)'. 06-trading.md table: 'Adelaide (unit of the minimum not stated)'. | supported with qualification | Vevo sentence now says the unit of the minimum is not stated, that reading it as AU is inference, and that the unaccredited-metric point holds on that reading; added ATT-VEVO-F05 marker; split into two sentences. | ATT-VEVO-F05 labels the identity of the guaranteed metric as inference, and chapter 6 already treats it that way. | chapters/02-standards.md |
| I1-021 | Audits in progress are reported separately and do not change the verdict, which is dated ([chapter 13](#method)). MRC's records, as read, hold one attention accreditation, and its April 2026 letter does not list the guidelines. H3 is therefore contradicted as of 28-29 September 2026 (ATT-A-F16) (ATT-DOUBLEVERIFY-F05). An MRC decision published between those dates and the 2 October cutoff would not be reflected. | The paraphrase of the registered rule leaves out 'to the cutoff', and the verdict is then dated to the 28-29 September reading instead. The rule dates the verdict to the cutoff, and ATT-A-F16 says the verdict 'must be re-dated at the 2 October freeze'. Saying a pre-cutoff MRC decision 'would not be reflected' also conflicts with chapter 13, which says such changes are 'logged separately'. The verdict date differs across the paper too: the abstract and ATT-A-F16 say 28 September, while this chapter and the appendix say 28-29 September. The best fix is to run the freeze recheck and re-date the verdict to 2 October in the standfirst, the in-short, this paragraph and the appendix row. Until then, use the replacement. | hypotheses.json H3 rule_contradicted: 'Audits in progress confirmed by MRC are reported separately and do not change the verdict, which is dated to the cutoff.' ATT-A-F16 final_wording: 'H3 codes as contradicted at 28 September 2026 ... The verdict must be re-dated at the 2 October freeze.' 13-method.md Limits: 'anything that changes between the reading and the freeze, or after it, is logged separately rather than folded in silently.' 00-abstract.md item 2: 'contradicted at that date' (28 September). 14-appendix.md: 'Contradicted as of 28 and 29 September 2026'. | supported with qualification | Rule paraphrase restored ('confirmed by MRC', 'which the rule dates to the cutoff'); verdict stated on the records read on 28-29 September 2026; last sentence now follows chapter 13 (freeze 2 October 2026, later MRC decisions logged separately) without asserting a recheck that has not happened; merged with R3-01. | The paraphrase dropped 'confirmed by MRC' and 'dated to the cutoff', and 'would not be reflected' conflicted with chapter 13's rule that later changes are logged separately. | chapters/02-standards.md |
| I1-022 | 13 wording, precision and labelling issues in 02-standards | Integrity pass findings rated P2: 02-standards-03, 02-standards-04, 02-standards-05, 02-standards-06, 02-standards-07, 02-standards-08, 02-standards-09, 02-standards-10, 02-standards-11, 02-standards-12, 02-standards-13, 02-standards-14, 02-standards-15 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 13 passages reworded for precision; no verdict changed | Precision and consistency. | chapters/02-standards.md |
| I1-023 | no test of buying on attention reports what it cost. | Overstated and contradicted by the chapter's own H1b table. Several buying tests do report cost: Lumen's consumer-goods test reports 40% lower CPM and 44% lower aCPM against a same-budget control, IAS/Kimberly-Clark reports ROAS, and MAGNA/Adlook reports the CPM for each arm. What no screened test reports is cost per incremental memory, brand or sales outcome. | ATT-C-F13 final_wording: 'a head-to-head test against a traditionally bought control campaign within the same budget ... at 40% lower CPM and 44% lower aCPM'. ATT-C-F29 final_wording: 'cost appears as CPM or cost-per-completion changes ... ROAS'. ATT-C-F26 final_wording frames the gap as 'cost per incremental outcome and uncertainty'. Chapter 3 H1b table, attention-only row: '+104% viewed ads at 40% lower CPM'. | supported with qualification | Standfirst now says no test reports its cost per incremental memory, brand or sales outcome. | ATT-C-F13 and ATT-C-F29 show buying tests that report CPM, cost per completion or ROAS; what none reports is cost per incremental outcome (ATT-C-F26). | chapters/03-validation.md |
| I1-024 | because the tests that would decide them have not been published. | Unbounded absence claim in the standfirst. The review can say only that such tests were not found in the sources screened, not that none has been published. The screen left out the International Journal of Advertising and member-only reports such as ARF Phase 3. | ATT-B-F01 final_wording: 'in the sources screened by 2 October 2026 ... the International Journal of Advertising was not in the screen'; ATT-B-F01 limitations: 'Absence is bounded to the searches logged'; ATT-B-F18: 'in the sources screened by the cutoff'. H1a rule_unresolved: 'Absence of accessible studies is unresolved'. | supported with qualification | 'have not been published' became 'no screened source reports the tests that would decide them'. | ATT-B-F01 bounds absence to the sources screened; the wording was rephrased so the standfirst does not repeat 'in the sources screened' twice. | chapters/03-validation.md |
| I1-025 | For a purchase proxy, a voucher choice made right after exposure, the per-second figures were 0.73 points for dwell and 0.13 for visibility (ATT-B-F05). | Misleading. Two sentences earlier the chapter says 'The per-second gap misleads' and switches to per-standard-deviation effects for recall. It then gives only per-second figures for the purchase proxy, which suggests a roughly fivefold dwell advantage. Per standard deviation the two effects are almost identical, which the verifier flagged as changing the reading of F05. (The F05 final_wording also leaves out the correction and should take it.) | ATT-B-F05 review_note correction: 'the purchase-proxy association was 0.73 percentage points per second of eye-tracked dwell against 0.13 per second of visibility, which per standard deviation is almost identical (2.31 versus 2.26 points)'. evidence/verify/B.json: 'Simonov purchase effect per SD: +2.31 pp dwell vs +2.26 pp visible'; B.pass1.json: 'Juxtaposing per-second coefficients across differently scaled measures overstates the gap roughly fivefold'. | supported with qualification | Added the near-identical per-SD figures and 'again from different sub-samples'; the F05 final_wording fix in the ledger is outside this file. | ATT-B-F05 review note confirms per-SD purchase effects of 2.31 and 2.26 points on different samples; R3-17 also asks that the different sub-samples be stated. | chapters/03-validation.md |
| I1-026 | \| Outcome \| Closest independent evidence \| Baseline tested \| Hold-out \| Result \| H1a \| | The column labels every study in it as independent, including two whose independence the paper records as unknown. The chapter itself (line 31) codes ARF Phase 3 as 'independence unknown'. Zhang, Wedel and Pieters 2009 has unknown funding, so under the shared rules its independence is also unknown. This breaks the rule that unknown funding is not independent. | ATT-B-F09 limitations: 'Funding and participation terms are not stated in the release, so the study's independence is unknown (round 2, R2-02)'. corrections-ledger R2-02 recoded F09 from independent_measurement to synthesis. ATT-B-F14 review_note: 'funding is unknown for both, so independence is unknown under the shared rules'. Chapter 13: 'a study whose funding is unknown is not treated as independent'. | supported with qualification | Column header now reads 'Closest academic and industry-body evidence'. | ATT-B-F09 codes ARF Phase 3 independence as unknown and ATT-B-F14 records Zhang et al. funding as unknown, so the column cannot call them independent. | chapters/03-validation.md |
| I1-027 | [studies] | The study table rendered here tags the ARF Phase 3 row as 'independent' with sponsor 'ARF'. Under the table's 'Independent' filter chip it therefore appears among the independent studies. That contradicts this chapter (line 31), the round-2 correction and the evidence-ladder figure, which all treat Phase 3 as funding unknown. The fix belongs in data/study-evidence.json: drop the 'independent' tag and set the sponsor to 'ARF; funding not stated, independence unknown'. | data/study-evidence.json row 'ARF Phase 3 (press release)': sponsor 'ARF', tags ['brand','independent']. build/instruments.mjs renders tags as filter chips (chipLabels independent: 'Independent'). ATT-B-F09 limitations and corrections-ledger R2-02 (independence unknown). figures/fig-evidence-ladder.json caption: 'ARF Phase 3 is counted as funding unknown because its release states no funding (round 2)'. | supported with qualification | Tag replaced by funding_unknown and sponsor note added (data/study-evidence-overlay.json). | The mis-tag sits in data/study-evidence.json, which this pass may not edit; the chapter text already treats Phase 3 as independence unknown, and the data fix is flagged for the data owner. | data/study-evidence-overlay.json |
| I1-028 | The study table lists every study screened here and in chapter 4, one row per study and outcome, with design, sample, sponsor and uncertainty. | False description of the paper's own dataset. The table holds the 48 coded study records, and several studies screened in chapters 3 and 4 are not in it. Missing: the Amplified-Chalice trials cited in this chapter's own H1b table, the MAGNA Media Trials/Adlook trial from chapter 4's table, Viant's 42 TVision pilots (chapter 4) and Bellman et al. 2019 (screened, not used). Any count should come from the stat token. | data/study-evidence.json has 48 rows and none mentions Chalice, MAGNA, Adlook, Viant or Bellman. stats.json studies = 48. figures/fig-evidence-ladder.json caption: 'The 48 study records coded in workstreams B (independent and academic) and C (commissioned and vendor)'. Chapter 4 table row: 'MAGNA Media Trials and Adlook, July 2023 (ATT-ADELAIDE-F27)'. | supported with qualification | Sentence now cites [studies] study records from independent, academic, commissioned and vendor work and notes that some studies discussed, such as the Amplified and Chalice trials, are not in the table. | study-evidence.json holds 48 records and omits Chalice, MAGNA/Adlook, Viant and Bellman, so 'every study screened' is false; workstream letters are not used in the chapters. | chapters/03-validation.md |
| I1-029 | 10 wording, precision and labelling issues in 03-validation | Integrity pass findings rated P2: 03-validation-07, 03-validation-08, 03-validation-09, 03-validation-10, 03-validation-11, 03-validation-12, 03-validation-13, 03-validation-14, 03-validation-15, 03-validation-16 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 10 passages reworded for precision; no verdict changed | Precision and consistency. | chapters/03-validation.md |
| I1-030 | The one outside review of AU is MediaSense's, which Adelaide requested. | The absence claim is not bounded. 'The one outside review' says no other review of AU exists, but the paper only establishes that no other was found in the sources screened. Chapter 6 bounds the same fact correctly. | ATT-E-F22 final_wording: 'The only third-party review found is MediaSense's February to June 2025 methodology review of AU, which Adelaide requested.' 06-trading.md line 67: 'The only third-party review found is MediaSense's 2025 methodology review of AU, which Adelaide requested.' Check 3 requires absence claims bounded to what was searched. | supported with qualification | Rewrote to 'The only outside review of AU found in the sources screened is MediaSense's, which Adelaide requested' and added (ATT-E-F22). | ATT-E-F22 final wording bounds the fact to 'the only third-party review found'; the chapter stated it as an unbounded absence. | chapters/04-vendor-evidence.md |
| I1-031 | 16 wording, precision and labelling issues in 04-vendor-evidence | Integrity pass findings rated P2: 04-vendor-evidence-02, 04-vendor-evidence-03, 04-vendor-evidence-04, 04-vendor-evidence-05, 04-vendor-evidence-06, 04-vendor-evidence-07, 04-vendor-evidence-08, 04-vendor-evidence-09, 04-vendor-evidence-10, 04-vendor-evidence-11, 04-vendor-evidence-12, 04-vendor-evidence-13, 04-vendor-evidence-14, 04-vendor-evidence-15, 04-vendor-evidence-16, 04-vendor-evidence-17 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 16 passages reworded for precision; no verdict changed | Precision and consistency. | chapters/04-vendor-evidence.md |
| I1-032 | It names attention in both its measurement and its optimisation revenue definitions, so attention revenue cannot be separated (ATT-IAS-F05). | Wrong fact. The IAS 10-K names attention only in the measurement-revenue definition. Attention appears separately in the list of optimisation (programmatic buying) solutions, not in the optimisation-revenue definition, which covers 'pre-bid solutions'. The verifier corrected exactly this reading. | ATT-IAS-F05 final_wording: 'The 10-K names attention in the measurement-revenue definition and, separately, lists attention among its optimization (programmatic buying) solutions; attention revenue is not separable and, for pre-bid Quality Attention Optimization, would fall under optimization revenue by the 10-K's own definitions (inference).' Review note: 'The revenue note defines optimization revenue as "pre-bid solutions" and names attention only under measurement revenue.' | contradicted | IAS sentence now says attention is named in the measurement-revenue definition and, separately, listed among optimisation solutions. | ATT-IAS-F05 final wording names attention only in the measurement-revenue definition and lists it separately among optimisation solutions. | chapters/05-market.md |
| I1-033 | Two specialists bought companies of their own. | Contradicts the chapter's own events table. The table lists three acquisitions by specialists: Amplified buying AttentvAds (13 December 2022), Lumen combining with Avocet and Adelaide buying Rita. The paragraph drops Amplified, which the chapter classes as a specialist. | Events table row '13 December 2022 \| Amplified buys AttentvAds \| Acquisition \| (ATT-AMPLIFIED-F30)'. ATT-AMPLIFIED-F30 final_wording: 'Amplified acquired AttentvAds on 13 December 2022, describing it as a company tracking and verifying impression-level data across CTV, mobile, web and in-app'. Source ATT-AMPLIFIED-S19 is titled 'Amplified acquires attention verification company AttentvAds'. ATT-D-F29 lists only Lumen-Avocet and Adelaide-Rita, but the chapter table adds AttentvAds. | supported with qualification | Two became Three; added a vendor-labelled AttentvAds sentence with ATT-AMPLIFIED-F30. | The events table lists Amplified-AttentvAds, and the chapter treats Amplified as a specialist, so three specialists made purchases. | chapters/05-market.md |
| I1-034 | Two deals sit outside measurement. | Wrong and self-contradictory. The next sentence calls Kantar Media an audience-measurement business, and Life360 describes Nativo as providing measurement solutions. The point is that these deals sit outside attention measurement. The count is also off: the table marks three non-attention deals (Outbrain-Teads, Life360-Nativo, Kantar Media), and this paragraph covers two of them. | ATT-KANTAR-F23 final_wording: 'Kantar Media (audience measurement, data analytics and media planning)'. ATT-D-F16 review note: Life360 describes Nativo as providing 'advertising serving, content distribution, and measurement solutions'. Events table rows 'February 2025 (completed) \| Outbrain buys Teads \| Acquisition (trading table)', '2 January 2026 \| Life360 closes its purchase of Nativo \| Acquisition (trading table)', '1 August 2025 \| Kantar sells Kantar Media ... \| Acquisition (not an attention business)'. | supported with qualification | Now reads 'Outbrain's purchase of Teads and two other deals sit outside attention measurement' with ATT-D-F29 for the Teads mention. | Kantar Media and Nativo involve measurement, so the point is attention measurement; the table has three such deals, including Outbrain-Teads. | chapters/05-market.md |
| I1-035 | Lumen's losses widened and it breached two loan covenants while it hired (ATT-LUMEN-F05) (ATT-LUMEN-F06). | Misleading in the H4 test. Almost all of the widening of the group loss is the non-cash share-based payment charge. Before that charge the loss was about flat: about GBP 1.06m in FY2024 against about GBP 1.04m in FY2025 as filed, or GBP 1.09m restated. Read beside the covenant breaches, 'losses widened' suggests a commercial deterioration the accounts do not show. The marker also omits ATT-LUMEN-F03, the source of the year-on-year comparison. | ATT-LUMEN-F03: group loss GBP 3,761,807 (prior year GBP 1,134,773); share-based payments credited to reserves GBP 2,720,974 (prior year GBP 78,592). The loss rose by GBP 2.63m and the share-based charge by GBP 2.64m. ATT-LUMEN-F04: loss before the non-cash charge about GBP 1.04m (about GBP 1.09m on the restated loss). ATT-D-F28 review note: 'group loss widening largely with a GBP 2.72m share-based payment credit'. | supported with qualification | Split into two sentences: the loss widened almost entirely through the share-based charge (labelled inference, F03 and F04), then the covenant breaches while hiring (F05, F06). | Per ATT-LUMEN-F03/F04 almost all of the loss increase is the non-cash share-based charge, so the unqualified 'losses widened' misleads. | chapters/05-market.md |
| I1-036 | \| 3 October 2024 \| Playground xyz's platform renamed GumGum Attention \| Rebrand \| (ATT-D-F03) \| | The chapter's date is correct, but the embedded figure contradicts it. fig-consolidation plots the GumGum Attention rebrand on 2024-10-04, the stale ExchangeWire URL date that the verifier corrected to 3 October 2024. The same figure also plots Tobii-Sticky and Adelaide-Rita as 'close' events, while the table dates Tobii-Sticky as announced. The fix belongs in data/figure-data.json (events), not in this chapter. | ATT-D-F03 final_wording: 'on 3 October 2024 (Santa Monica dateline, PRWeb) it launched the GumGum Platform, in which the platform was renamed GumGum Attention'. Review note: 'Update the market_events rows for the restructuring (2024-04-23) and rebrand (2024-10-03) accordingly.' data/figure-data.json events: {date: '2024-10-04', lane: 'Rebrands, spin-outs', label: 'GumGum Attention'}; {date: '2017-05-17', kind: 'close', label: 'Tobii buys Sticky'}. | supported with qualification | Rebrand dated 3 October 2024; Tobii-Sticky and Adelaide-Rita plotted as announcements. | The chapter's 3 October 2024 date is correct; the stale 2024-10-04 date and the close/announce kinds sit in data/figure-data.json, outside this file. | scripts/derivations.mjs (events); fig-consolidation |
| I1-037 | Public sources cannot show whether verification bundles cost the specialists demand or revenue, so H4 is unresolved, and no source ties any event to a metric failing. | Two absence claims are unbounded. 'Public sources cannot show' and 'no source ties' go beyond the ledger, which bounds both to the sources screened or found. The sentence also misstates H4: it frames the test as one-directional ('cost the specialists') and leaves out pricing. The registered hypothesis is whether bundling changes specialists' demand, realised pricing or revenue. | hypotheses.json H4 statement: 'changes specialist attention vendors' demand, realised pricing or revenue'. ATT-D-F28 final_wording: '... in the sources screened by 28 September 2026'. ATT-D-F29 final_wording: 'No source found attributes any of these events to a failure of attention metrics.' | supported with qualification | Standfirst now says no public source screened shows whether bundles changed demand, pricing or revenue, and no source found ties an event to a metric failing, as two sentences. | Both absence claims were unbounded and H4 omitted pricing; ATT-D-F28 and ATT-D-F29 bound them to the sources screened and found. | chapters/05-market.md |
| I1-038 | 18 wording, precision and labelling issues in 05-market | Integrity pass findings rated P2: 05-market-07, 05-market-08, 05-market-09, 05-market-10, 05-market-11, 05-market-12, 05-market-13, 05-market-14, 05-market-15, 05-market-16, 05-market-17, 05-market-18, 05-market-19, 05-market-20, 05-market-21, 05-market-22, 05-market-23, 05-market-24 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 18 passages reworded for precision; no verdict changed | Precision and consistency. | chapters/05-market.md |
| I1-039 | The first documented case is an Adelaide case study of a 2022 campaign run by MediaCom (now EssenceMediacom) with Kargo on DV360 private marketplaces (PMPs). | Wrong fact and cross-chapter contradiction. The Kargo campaign (2022) is not the first documented attention guarantee: the ledger and chapter 7 record an executed, repeated US guarantee by A+E to dentsu for AB InBev's Budweiser in Q4 2020 and Q1 2021, on TVision panel data. Chapter 6 never mentions it, so its account of guarantees starts two years late and leaves out the only US guarantee on a non-Adelaide panel metric. The same gap runs into the guarantees table, fig-guarantees and the Q1 table (see 06-trading-02 and 06-trading-09). | ATT-F-F14 final_wording: '(a) a live A+E guarantee to dentsu for AB InBev's Budweiser (Q4 2020, repeated in Q1 2021). It was based on TVision panel data counting eyes on screen for at least two seconds ... None discloses a guarantee level, remedy or independent audit.' ATT-TVISION-F23 (scope 2020-Q4 to 2021, US): 'The only executed guarantee on TVision data found ... whether the guaranteed metric was presence or attention is ambiguous.' ATT-F-F29: 'one executed, repeated US guarantee (A+E for AB InBev via dentsu, Q4 2020 and Q1 2021)'. Chapter 7 line 47: 'A+E guaranteed dentsu's Q4 2020 Budweiser buy ... and repeated it in Q1 2021'; chapter 7 line 86: 'The A+E guarantee was live in 2020 and 2021'. | contradicted | Guarantees section now opens with A+E as the earliest documented case (TVision data, two-second eyes-on-screen count, presence-or-attention ambiguity, no level, remedy or audit, labelled as coming from an A+E-branded trade article (ATT-F-S25)); Kargo becomes the earliest case with published terms; inshort adds A+E as earliest. | ATT-F-F14 and ATT-TVISION-F23 record the A+E guarantee to dentsu for Budweiser in Q4 2020 and Q1 2021, so Kargo (2022) was not the first. | chapters/06-trading.md |
| I1-040 | The table lists every guarantee and attention-priced offer found; "none found" is bounded to the sources screened by the cutoff. | The completeness claim is false. The table (and fig-guarantees, titled 'Every attention guarantee or attention-priced offer found') leaves out (a) the A+E/dentsu guarantee of 2020 and 2021 on TVision data, recorded in chapter 7, and (b) Viant's attention-adjusted CPM. The product register lists Viant's CPM as a trading-table entry, and this chapter's own Q1 table codes it ('Announced price'). The table does include other announced-only offers (Duolingo, TripleLift, Adlook), so leaving Viant out is inconsistent. The fix needs a decision: add both rows here and in fig-guarantees, or narrow the sentence and say why each is excluded. | ATT-F-F14 and ATT-TVISION-F23 (A+E guarantee, US, Q4 2020 to Q1 2021). product-register.json 'Viant attention-adjusted CPM': 'Trading table (not scored): announced attention-based valuation; not live (workstream E)'. ATT-TVISION-F13 and ATT-E-F17 (announced 15 April 2026, billing role undocumented). figures/fig-guarantees.json caption 'Every attention guarantee or attention-priced offer found' (nine rows, neither included). | supported with qualification | Added an A+E row to the offers table; the table intro now says Viant's CPM is left out and coded under Q1, citing ATT-TVISION-F13. | The completeness claim was false for A+E; Viant's attention-adjusted CPM is announced, not launched, with no documented billing role, and the guarantees figure (ten offers) also leaves it out. | chapters/06-trading.md |
| I1-041 | Attention is marketed as a live optimisation input in The Trade Desk, Amazon DSP, DoubleVerify's segments and Scibids optimisation, and Viant's DSP through the Havas marketplace. All of it rests on vendor or seller statements (ATT-E-F23). | The H2 verdict paragraph cites evidence that the chapter qualifies elsewhere, and repeats none of those qualifications. The H2 rule needs a live (not announced) US offer 'in a named buying platform'. (1) DoubleVerify's 2026 page names no platform, as line 23 of this chapter says, so DV segments and Scibids optimisation do not meet the 'named buying platform' limb. (2) Viant called its TVision pre-bid targeting 'testing' on 10 August 2026 and 'integrating' in its 10-Q. Chapter 7 carries that caveat for the same marketplace; this paragraph does not. (3) The Trade Desk and Amazon DSP releases state no geography (line 19). The verdict still stands on The Trade Desk, Amazon DSP and the Journal, but as written the list overstates how firm the live evidence is. | hypotheses.json H2 rule_supported: 'At least one live (not announced) US offer documents attention as an optimisation input in a named buying platform'. ATT-E-F15: 'DV's 2026 product page still offers Universal Attention Segments and Scibids attention optimisation without naming DSPs'. ATT-DOUBLEVERIFY-F11: 'DV's 2026 product pages name no platform for attention segments or attention optimisation'. ATT-E-F17: ''testing' in the 10 August 2026 Q2 release, 'integrating' in the 10-Q and in a 16 September 2026 prospectus supplement'. ATT-TVISION-F14: 'sits uneasily with Viant's 'Began testing' status of 10 August 2026'. ATT-E-F11 limitations: 'Geography and pricing are not stated.' Chapter 7 line 86: 'The Havas and Viant marketplace was announced in June 2026 and described by Viant as testing i | supported with qualification | H2 paragraph names The Trade Desk and Amazon DSP (no geography stated) and DV360 (live per Help Center, ATT-ADELAIDE-F15) as named platforms, and qualifies DoubleVerify and Viant; inshort no longer lists DoubleVerify under named platforms. | ATT-E-F15 (no platform named), ATT-E-F17 (Viant 'testing') and ATT-E-F11 (no geography) qualify the H2 evidence list. | chapters/06-trading.md |
| I1-042 | Every one of these routes bids on a model estimate: Adelaide's segments are built from historical placement-level AU (ATT-E-F11). | Measurement-basis mislabel. 'These routes' include DoubleVerify's segments and Scibids optimisation. The paper's Q2 coding treats DV's attention figures as indices of device signals, a candidate proxy, not a model estimate. The Havas and Viant deciles involve TVision panel data. The cited claim supports only the Adelaide routes, and chapter 7 correctly limits the same point to 'Every AU-based route'. | ATT-C-F31: 'DV Authentic Attention ... is an index of device and on-screen signals ... which the Q2 rule codes as a candidate proxy, not a validated model estimate'. ATT-A-F18 limitations: 'DV's inputs are device signals, coded 'candidate proxy''. Abstract finding 1: 'DoubleVerify's are indices of device signals'. Chapter 7 line 43: 'Every AU-based route bids on a model estimate of placement quality'. ATT-E-F11 covers only Adelaide segments. | supported with qualification | 'Every one of these routes' became 'Every AU route'; added that DoubleVerify's attention indices are built from device signals, coded candidate proxy, citing ATT-A-F18 and ATT-C-F31. | ATT-E-F11 supports model estimates only for AU routes, and ATT-A-F18/ATT-C-F31 code DoubleVerify's device-signal indices as a candidate proxy. | chapters/06-trading.md |
| I1-043 | DoubleVerify launched pre-bid attention segments on 4 May 2023 and attention bidding through its Scibids optimiser on 19 June 2023. | Ownership anachronism. 'Its Scibids optimiser' says DV owned Scibids on 19 June 2023, but DV acquired Scibids in August 2023. The June 2023 release was a joint DV and Scibids launch of the DV Algorithmic Optimizer. Ownership is a status-sensitive fact and needs a date. | ATT-D-F10: 'DoubleVerify's filings describe two acquisitions in 2023 to 2025: Scibids Technology SAS (August 2023, AI campaign optimisation)'; D.json passage from the FY2025 10-K: 'In August 2023, we acquired Scibids Technology SAS'. ATT-E-F15: 'DV Algorithmic Optimizer with Scibids on 19 June 2023'. Source ATT-E-S28 title: 'DoubleVerify and Scibids Unveil ... DV Algorithmic Optimizer'. | supported with qualification | Now 'with Scibids, the DV Algorithmic Optimizer for attention bidding on 19 June 2023', plus 'DV acquired Scibids in August 2023' citing ATT-D-F10. | ATT-D-F10 dates DV's acquisition of Scibids to August 2023, after the 19 June 2023 joint launch. | chapters/06-trading.md |
| I1-044 | Guarantees are rarer. Only the Kargo case publishes a numeric threshold, a minimum average of 25 AU (ATT-VEVO-F13). | Contradicts the chapter's own figure and table. fig-guarantees (placed in this chapter) says 'Two publish a numeric threshold (Kargo's campaign ... and Impact Media's seconds seen)', and the chapter's table marks Impact Media 'Threshold public? Yes: 1 second display, 3 seconds video'. Read on its own, the unqualified 'Only the Kargo case' is wrong. VEVO-F13 supports the statement only for guarantees. | figures/fig-guarantees.json caption: 'Two publish a numeric threshold (Kargo’s campaign, at an average of 25 AU, and Impact Media’s seconds seen)'; figure-data.json guarantees.public_threshold = 2. Chapter table row: '\| Guaranteed Attention Suite \| Impact Media \| Not stated \| Yes: 1 second display, 3 seconds video; method not stated \|'. ATT-VEVO-F13 is scoped to 'the six named attention guarantees in the register'. | supported with qualification | Inshort now says 'Among the guarantees, only the Kargo case...' (adding ATT-F-F14 for A+E) and notes that Impact Media's billing suite publishes seconds seen instead. | The figure and table show Impact Media also publishes a numeric (seconds) threshold; ATT-VEVO-F13 covers guarantees only. | chapters/06-trading.md |
| I1-045 | The only third-party review found is MediaSense's 2025 methodology review of AU, which Adelaide requested. | Cross-chapter contradiction. Chapters 4 and 12 report a second vendor-commissioned third-party review, PwC's review for Lumen. It matters here because SeenThis Attention, one of the guarantees, runs on Lumen's metric. Chapter 12 line 35 says 'The third-party reviews found were requested by the vendor' and names both. | ATT-C-F11: 'PwC's review, commissioned by Lumen (report pages dated November 2022, cover dated March 2023, released by Lumen on 27 April 2023), re-performed Lumen's campaign analysis'. Chapter 12 line 35: 'The third-party reviews found were requested by the vendor. MediaSense ... PwC says of its review for Lumen ...'. Chapter 4 line 8 lists 'PwC's review for Lumen' as commissioned. ATT-R2-F01: SeenThis Attention is 'measured by Lumen Research'. | supported with qualification | Replaced 'The only third-party review found' with 'The third-party reviews found were requested by the vendors', keeping the MediaSense detail and adding PwC's Lumen-commissioned review on Lumen-supplied data. | ATT-C-F11 records PwC's Lumen-commissioned review, relevant because SeenThis runs on Lumen's metric. | chapters/06-trading.md |
| I1-046 | Unreadable accreditation or trading terms make it "unresolved"; offers that publish part of their terms are coded on what can be read (ATT-E-F24). | The sentence presents the paper's own gloss as part of Q1's registered rule. hypotheses.json has no partial-terms clause: 'A pair whose accreditation or trading terms cannot be read from public sources is coded unresolved, not diagnostic.' The Journal pair (threshold not readable) and the SeenThis pair (threshold values not published) get 'diagnostic' only through this gloss. On a literal reading they would be 'unresolved', giving 4 diagnostic and 13 unresolved instead of 6 and 11. The top-line verdict (no currency or partial currency) does not change, but the gloss should be labelled as this paper's interpretation. | hypotheses.json Q1 rule_unresolved: 'A pair whose accreditation or trading terms cannot be read from public sources is coded 'unresolved', not 'diagnostic'.' ATT-E-F02: 'no numeric threshold or benchmark is readable in the public sources'. ATT-R2-F01: 'the threshold values ... are not documented'. ATT-E-F24 (inference, confidence low) treats WSJ terms as 'readable' but states no partial-terms rule. | supported with qualification | Split the sentence: the registered unresolved clause cites ATT-E-F24, and the partial-terms coding is labelled as this paper's reading, noting that it places the Journal and SeenThis pairs in diagnostic and that neither reading yields currency or partial currency. | The Q1 rule_unresolved clause has no partial-terms provision, so the gloss is this paper's interpretation. | chapters/06-trading.md |
| I1-047 | Six pairs code "diagnostic" and eleven "unresolved". | The count is misleading against the chapter's own method. Line 75 says pairs meeting neither of the first two conditions 'are diagnostic and left out', so the paper's diagnostic count is more than six. Six is only the table's diagnostic count. Q1's rule requires the paper to report the count in each code. Separately, if the A+E/TVision guarantee (06-trading-01) is added, its pair codes 'unresolved' (terms not stated, TVision not accredited), which makes the unresolved count twelve. | Chapter line 75: 'Pairs meeting neither, such as AU segments in The Trade Desk, are diagnostic and left out.' hypotheses.json Q1: 'The paper reports the count in each code.' ATT-F-F14 (A+E guarantee, no level, remedy or audit); ATT-TVISION-F10 (TVision not accredited). | supported with qualification | Added an A+E/TVision row to the Q1 table coded Unresolved; the count now reads 'Of the pairs in the table, six code diagnostic and twelve unresolved; the pairs left out are also diagnostic and are not counted here'. | The count covered table pairs only, and adding the A+E pair (TVision not accredited, terms unreadable) makes twelve unresolved. | chapters/06-trading.md |
| I1-048 | Sellers often fund the measurement. | Unsupported frequency claim. The cited evidence is one Adelaide statement naming two publishers, Condé Nast and Hearst, plus a free audit tier. ATT-E-F06 itself calls these 'one of few public statements on who pays', and ATT-F-F28 finds no buyer-side statement of who pays. 'Often' implies a prevalence the sources cannot show. Chapter 12 line 35 repeats the phrase and should change with it. | ATT-E-F06 final_wording: 'This is one of few public statements on who pays for attention measurement; others found are Adelaide's free publisher audit tier (AFP, April 2024) and Index Exchange's statement ... No fee level is disclosed in any of them.' ATT-F-F28: 'No buyer-side source screened states who pays for attention measurement'. | supported with qualification | 'Sellers often fund the measurement' became 'The few public statements on who pays point to sellers'. | ATT-E-F06 calls these 'one of few public statements' on who pays, which does not support 'often'. | chapters/06-trading.md |
| I1-049 | 15 wording, precision and labelling issues in 06-trading | Integrity pass findings rated P2: 06-trading-11, 06-trading-12, 06-trading-13, 06-trading-14, 06-trading-15, 06-trading-16, 06-trading-17, 06-trading-18, 06-trading-19, 06-trading-20, 06-trading-21, 06-trading-22, 06-trading-23, 06-trading-24, 06-trading-25 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 15 passages reworded for precision; no verdict changed | Precision and consistency. | chapters/06-trading.md |
| I1-050 | the buyer-side record found holds one executed guarantee, from 2020 and 2021, no billing-unit use, no contract term and no statement of who pays. | The standfirst says the buyer-side record holds one executed guarantee and no contract term. The chapter's own WPP table row ('a guarantee received by MediaCom from Kargo (2022)') and its guarantees section describe a second executed guarantee with an agency buyer, a public threshold (minimum average 25 AU) and a make-good (added-value impressions). The difference is the kind of source (buyer-side sources versus the metric vendor's case study), but the text never says so. A reader will see the claim contradicted three sections later. | ATT-E-F03 final_wording: 'the parties "agreed upon" a minimum average AU of 25 ... Added-value PMP impressions were served when the average dipped below 25'. ATT-F-F28 is bounded to 'buyer-side source[s] screened'. ATT-F-F14 lists only buyer-side sources. Chapter 6 table row: 'Kargo, for MediaCom ... Yes: minimum average 25 AU \| Yes: added-value PMP impressions'. | supported with qualification | Standfirst now says buyer-side sources record one executed guarantee and that Kargo's guarantee to MediaCom, with a public threshold and remedy, comes from the metric vendor's case study. | ATT-E-F03 documents Kargo's 2022 guarantee to MediaCom with a 25 AU threshold and make-good, so the standfirst had to scope the one-guarantee claim to buyer-side sources and name the vendor-sourced exception. | chapters/07-buyers.md |
| I1-051 | The only executed buyer-side guarantee found, A+E's for Budweiser in 2020 and 2021, states no level, remedy or audit (ATT-F-F14). | The summary calls A+E's the only executed buyer-side guarantee found. The chapter also documents Kargo's executed 2022 guarantee to MediaCom, a WPP agency, with a stated level and remedy. A+E is itself a seller guaranteeing to an agency, just like Kargo, so 'buyer-side' does not separate the two cases unless it means buyer-side sources. As written, the sentence contradicts the WPP table row and the guarantees section. | ATT-E-F03 final_wording (MediaCom and Kargo 'agreed upon' a minimum average AU of 25; added-value PMP impressions below it). ATT-TVISION-F23 is scoped to 'the only executed guarantee on TVision data'. The chapter's own table row: 'a guarantee received by MediaCom from Kargo (2022)'. | supported with qualification | Inshort sentence rescoped to buyer-side sources and followed by the Kargo/MediaCom guarantee cited to ATT-E-F03. | Checked against ATT-E-F03 and ATT-F-F14: A+E is the only executed guarantee in buyer-side sources, and Kargo's comes from Adelaide's case study. | chapters/07-buyers.md |
| I1-052 | The performance claims attached to agency announcements, such as Havas's, have no published comparator or holdout (ATT-F-F09); [chapter 3](#validation) codes that evidence against H1b. | 'Codes that evidence against H1b' reads as saying the claims count as evidence against H1b, as if contradicting it. The ledger says they do not bear on H1b. Chapter 3 holds that case studies without a holdout, randomisation or cost leave H1b unresolved. The sentence also drops the missing cost. | ATT-F-F09 limitations: 'The performance claims (up to 132% upper-funnel lift, 2-3x) have no published comparator, holdout or cost, so they do not bear on H1b.' hypotheses.json H1b rule_unresolved: 'Published tests lack a holdout or randomisation, omit cost, or are case studies without design'. Chapter 3: 'Under the registered rule, H1b is unresolved for memory, brand and sales'. | supported with qualification | The sentence now says the claims have no published comparator or holdout and so leave H1b unresolved under the registered rule (chapter 3). | H1b's unresolved rule and ATT-F-F09 support 'leave H1b unresolved', but the F09 review note says the release does give a cost-per-conversion figure, so 'or cost' was left out. | chapters/07-buyers.md |
| I1-053 | The filings are silent. Omnicom's FY2025 Form 10-K contains no reference to attention metrics, attention vendors or viewability (ATT-F-F16), and Publicis's 2025 Universal Registration Document, filed 30 April 2026, contains no attention-metric reference (ATT-F-F18). | The absence claim is not bounded. 'The filings are silent' covers all five groups, but only two filings were searched. WPP's and Havas's annual reports were explicitly not searched, and no dentsu filing appears in the search log. The ledger frames each absence as the result of a term search of one filing. | evidence/research/F.json open_questions: 'WPP plc and Havas N.V. annual reports were not searched for attention references (Omnicom and Publicis were).' The search log's document screens list only the Omnicom FY2025 10-K and the Publicis 2025 URD. ATT-F-F16 limitations: 'Bounded absence: this review found no attention-metric reference in this one filing, searched by the terms listed'. ATT-F-F28 limitations name only 'the Omnicom 10-K, the Publicis URD'. | supported with qualification | The paragraph now says only Omnicom's and Publicis's filings were searched, describes the full-text search, and says WPP's, dentsu's and Havas's filings were not searched (ATT-F-F28). | ATT-F-F16 and ATT-F-F28 limit the search to the Omnicom 10-K and the Publicis URD. The F stream's open questions confirm WPP and Havas were not searched, and the search log has no dentsu filing. | chapters/07-buyers.md |
| I1-054 | Most buyer optimisation records are announcements from 2022 to 2024, with no named buying platform in the text read and no confirmation that they are live at the cutoff (ATT-F-F29). | The chapter's own enumeration contradicts this. Of the ten buyer optimisation records the section describes, six name a platform: dentsu/Co-op via a Google custom algorithm (2019), dentsu and Havas in the Avocet DSP (2022), Nespresso and Mars on Xandr Curate, JLR in DV360, and Havas and Viant in Viant's DSP. Only the four agency announcements (Havas MAU, Xaxis Switzerland, IPG Mediabrands, dentsu UK) name none. The sentence repeats an F29 limitation written for the uncorrected stream table, which the verifier found understated platform-named records. | ATT-F-F21 review_note: 'This names a buying platform (Xandr Curate), which contradicts the buyer_uses table note.' ATT-F-F29 review_note: the table misses JLR's DV360 use, EssenceMediacom France's DV360 bidding and dentsu's 2019 algorithm. ATT-LUMEN-F34 (Co-op 2019; Avocet 2022). ATT-TVISION-F14 (Viant DSP). The section's own heading: 'Brand records name platforms more often.' | supported with qualification | Replaced with: the agency announcements from 2022 to 2024, by dentsu UK, Havas, Xaxis Switzerland and IPG Mediabrands, name no buying platform and none is confirmed live (ATT-F-F29). | Brand and some agency records do name platforms (ATT-LUMEN-F34, ATT-F-F21, ATT-TVISION-F14). 'Most agency records' would still be shaky once EssenceMediacom France is counted, so the sentence names the platformless announcements instead. | chapters/07-buyers.md |
| I1-055 | The buyer records add none of the four beyond what chapter 6 coded, so no pair's code changes and none codes as currency (ATT-E-F24). | The claim that no pair's code changes skips the A+E guarantee on TVision data (2020 and 2021). It is a guarantee use whose threshold and remedy cannot be read, and chapter 6's Q1 table does not list it. Chapter 6 splits AU pairs by seller and codes guarantees with unpublished terms (Vevo, TripleLift, Duolingo, Teads, Adlook) as 'unresolved' under Q1's unresolved clause. On the same basis, the A+E TVision pair would be a new 'unresolved' pair, so chapter 6's count of eleven unresolved pairs would change. The fix needs an author decision: add the pair to chapter 6's table, or say why a 2020-21 guarantee is excluded. | hypotheses.json Q1 rule_unresolved: 'A pair whose accreditation or trading terms cannot be read from public sources is coded "unresolved", not "diagnostic".' ATT-E-F24: 'AU pairs whose guarantee terms are unpublished ... should be coded "unresolved"'. ATT-F-F14 and ATT-TVISION-F23: A+E guarantee live, no level, remedy or audit given. Chapter 6: 'Six pairs code "diagnostic" and eleven "unresolved"', with no A+E row. | supported with qualification | The Q1 paragraph now says the buyer records add none of the four conditions and none codes as currency, and adds that the A+E TVision guarantee, which is outside chapter 6's table, would code unresolved under the rule's unresolved clause (ATT-F-F14). Chapter 6's count of eleven unresolved pairs does not include it; that is for the chapter 6 editor. | The A+E guarantee has no stated threshold or remedy, so it does not meet Q1's second condition. Its trading terms are unreadable, though, so on chapter 6's practice it would code unresolved. The fix is kept to this chapter, stated as inference, and dropped 'no pair's code changes'. | chapters/07-buyers.md |
| I1-056 | A+E guaranteed dentsu's Q4 2020 Budweiser buy for AB InBev and repeated it in Q1 2021. | This chapter is correct, but it conflicts with chapter 6 and the guarantees figure. Chapter 6 says the Kargo/MediaCom 2022 campaign is 'The first documented case' of a guarantee, although this executed A+E guarantee dates from Q4 2020. The fig-guarantees caption ('Only SeenThis runs on a metric other than Adelaide's') and ATT-E-F22 ('Every attention guarantee found but one is defined on Adelaide's AU') leave out this TVision-based guarantee. This chapter also cites ATT-E-F22 for 'As in every guarantee found', a claim that does not cover A+E. The fix belongs in 06-trading line 29 (for example 'The first documented seller-offered AU guarantee') and in the figure caption's scope ('offers'). | ATT-F-F14 and ATT-TVISION-F23 final_wording: an executed A+E guarantee in Q4 2020, repeated in Q1 2021, on TVision data. Chapter 6 line 29: 'The first documented case is an Adelaide case study of a 2022 campaign run by MediaCom'. ATT-E-F03 review_note: the ANA recap calls Kargo 'the first programmatic campaign guaranteeing media quality with attention metrics' (programmatic only). figures/fig-guarantees.json caption. | supported with qualification | 'As in every guarantee found' now reads 'As in every guarantee that names who measures delivery, the metric's own vendor did so; A+E's guarantee does not name its measurer' (ATT-E-F22). Chapter 6's 'first documented case' wording is left to that chapter's editor. | Most of this finding concerns chapter 6 line 29 and the guarantees figure caption, which already lists A+E on TVision. The in-file part holds, because the new ATT-E-F22 wording says A+E's measurer is not stated. | chapters/07-buyers.md |
| I1-057 | 9 wording, precision and labelling issues in 07-buyers | Integrity pass findings rated P2: 07-buyers-08, 07-buyers-09, 07-buyers-10, 07-buyers-11, 07-buyers-12, 07-buyers-13, 07-buyers-14, 07-buyers-15, 07-buyers-16 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 9 passages reworded for precision; no verdict changed | Precision and consistency. | chapters/07-buyers.md |
| I1-058 | Lumen measures US CTV through TVision. For LG Ad Solutions, US measurement runs through the TVision pixel and is processed in the Lumen platform, and Nexxen's CTV offer uses measurement "powered by TVision data" (ATT-LUMEN-F18). Teads' exclusive Lumen HomeScreen measurement, announced in May 2026, covers the US, EMEA, APAC and LATAM (ATT-LUMEN-F19). | Present-tense 'Lumen measures US CTV through TVision' states a current, post-acquisition supply relationship that the evidence dates to 2023-2024 and that the chapter itself calls undocumented four paragraphs later ('whether TVision data remain licensed to Adelaide, DoubleVerify and Lumen is not documented'). The LG and Nexxen routes are undated. Also, because the HomeScreen sentence follows directly, it reads as another TVision route, but the ledger records HomeScreen as a machine-learning prediction from eye-tracking data, not TVision-based. | ATT-LUMEN-F18 final_wording: LG Ad Solutions '(2 February 2024)', Nexxen's CTV offer '(August 2023)'; limitation: 'No Lumen source after TVision's change of ownership confirms the TVision supply continues.' ATT-G-F02: 'Lumen's Netflix scores (European clients) and its Teads HomeScreen model are machine-learning predictions from eye-tracking data. Whether TVision data remain licensed to these three vendors after Viant completed the acquisition on 1 May 2026 is not documented.' Same chapter, CTV section, paragraph 5. | supported with qualification | 'Lumen measures' became 'Lumen has documented'; added February 2024 and August 2023 dates; marked Teads HomeScreen as a separate machine-learning route and added ATT-G-F02. | ATT-LUMEN-F18 dates the LG and Nexxen routes to 2024 and 2023 with no post-acquisition confirmation, and ATT-G-F02 records HomeScreen as a machine-learning prediction from eye-tracking data, not a TVision route. | chapters/08-channels.md |
| I1-059 | The xpln.ai rows rest on its own sources, which were not in the independent verification sample, and they change no code. | Factually wrong about the paper's own process: the xpln.ai claims behind these rows were verified. Every ledger claim has checked_independently = true, and there is a dedicated verifier file for the xpln.ai stream. The method has no 'verification sample'. The phrase 'independent verification' also edges toward describing the paper as independently validated, which chapter 13 explicitly denies. What is true is that the xpln.ai rows rest on vendor sources and are not part of the ATT-G-F01 Q3 count. | ATT-XPLN-F12 and ATT-XPLN-F06: checked_independently true, review_outcome supported_as_stated / supported_with_qualification; evidence/verify/XPLN.json ('Every cited source was re-opened on 2026-09-29'). stats.claims_checked = 584 of 584. Chapter 13: 'It is not an external audit, not peer review and not independently validated.' ATT-G-F01 lists no xpln.ai product. | supported with qualification | Replaced the 'independent verification sample' sentence with 'rest on its own statements and are not part of the Q3 count (ATT-G-F01). Adding them changes no code.' | All xpln.ai claims have checked_independently true and the method has no verification sample; what holds is that the rows rest on xpln.ai's own statements and ATT-G-F01 lists no xpln.ai product. | chapters/08-channels.md |
| I1-060 | [fig-channel-coverage] | The embedded figure contradicts the chapter in three places. (1) It codes IAS Quality Attention as 'beta' for display and online video, but the chapter and the ledger say it has been generally available since January 2024, and the chapter counts IAS in both rows. (2) It shows xpln.ai as 'none found' for CTV and social, but the table lists xpln.ai in both 'by its own account' (ATT-XPLN-F12, ATT-XPLN-F06), and the figure caveat defines a blank cell as none found in the sources screened. (3) Its title, 'Display, video and CTV are covered; audio, in-game, DOOH and retail media are thin or study-only', leaves out social, which the chapter and the figure's own coding table ('Social', live 3, 'covered') code as covered. | ATT-G-F13: 'IAS Quality Attention measurement has been generally available since January 2024 for open-web display and video'; ATT-IAS-F09: 'General availability was announced on 4 January 2024'. ATT-XPLN-F12: live in 'display ..., online video ..., CTV and social'. figure-data.json#channels: row 'IAS Quality Attention' cells ['beta','beta',...], row 'xpln.ai attention metrics' cells ['live','live','none','none',...]; fig-channel-coverage.json title. The capability-matrix IAS channel cell also counts display and video as live. | supported with qualification | IAS display and video live (general availability January 2024); xpln.ai CTV and social coded as vendor claims. | The finding holds (figure-data.json#channels still codes IAS display/video as beta and xpln.ai CTV/social as none, and the figure title omits social), but the fix belongs in figure-data.json and fig-channel-coverage.json, outside this file; the chapter text is correct. | scripts/derivations.mjs (channels); fig-channel-coverage |
| I1-061 | Under the registered Q3 rule, display, online video, CTV and social are covered, audio and DOOH are thin, and in-game is study only (ATT-G-F01) (ATT-G-F10). | Cross-chapter inconsistency. This sentence, which is correct, codes social as covered. The abstract's finding 8 ('Display, online video and CTV are covered; ... Audio, in-game, DOOH and retail media are thin or study-only') and the figure title both drop social from the covered list, so the Q3 verdict reads differently in the abstract and in this chapter. The fix belongs in chapter 00 finding 8 and in the figure title, not here. | ATT-G-F01 final_wording: 'social is covered (Adelaide, Lumen, DoubleVerify, Amplified planning; IAS announced)'. hypotheses.json Q3 rule: covered = at least three cohort A products documenting live measurement. figure-data.json coding: Social live 3, 'covered'. 00-abstract.md line 28. | supported with qualification | Title now names display, video, CTV and social as covered. | The chapter sentence is correct and the finding itself places the fix elsewhere; chapter 00 finding 8 already lists social as covered, and only the figure title (outside this file) still omits it. | scripts/paper-figures.mjs |
| I1-062 | On MRC's digital, television and multimedia lists, the only accredited attention metrics are DoubleVerify's, in desktop, mobile web and mobile in-app (ATT-A-F12). No CTV, social, audio, in-game or DOOH attention metric appears there. Two MRC reviews are open. | Accreditation status is given without a date (rule 6), in the chapter's closing statement on assurance. The claim is dated 28 September 2026, and chapter 2 and the abstract date it. The in-short uses the same undated framing ('The only accredited attention metrics cover display and video...') and should take the same date. Also, ATT-A-F12 supports only a single attention entry on the digital list and no Adelaide, Lumen, TVision or Amplified product on the television or multimedia lists. It does not support a general statement covering every vendor on those two lists. | ATT-A-F12 final_wording: 'On 28 September 2026, MRC's list of accredited digital services has one entry under the 'Attention Metrics' category: DoubleVerify ... No accredited attention product from Adelaide, Lumen, TVision or Amplified appears on MRC's digital, television or multimedia accredited lists.' Limitation: 'must be rechecked at the 2 October 2026 freeze'. 02-standards.md: 'On 28 September 2026 the "Attention Metrics" category of MRC's digital list has one entry'. | supported with qualification | Dated the closing accreditation passage, split it to match ATT-A-F12's scope, added ATT-DOUBLEVERIFY-F04 for DV's CTV attention, bounded the channel statement to 'no accredited attention metric found', and dated the in-short accreditation sentence (adding ATT-G-F03 for display and video). | ATT-A-F12 is dated 28 September 2026 and supports one digital attention entry plus no attention product from Adelaide, Lumen, TVision or Amplified on the three lists, not a general statement about every list entry. | chapters/08-channels.md |
| I1-063 | 8 wording, precision and labelling issues in 08-channels | Integrity pass findings rated P2: 08-channels-06, 08-channels-07, 08-channels-08, 08-channels-09, 08-channels-10, 08-channels-11, 08-channels-12, 08-channels-13 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 8 passages reworded for precision; no verdict changed | Precision and consistency. | chapters/08-channels.md |
| I1-064 | Tobii's Sticky page asserts ISO/IEC 27001 certification but shows no certificate (ATT-TOBII-F11). | This contradicts the claim's corrected wording and the matrix. The verifier found a certificate behind the claim: Tobii AB, which operates Sticky, holds ISO/IEC 27001:2022. The sentence implies the vendor's certification claim is unbacked, which misrepresents the vendor. | ATT-TOBII-F11 final_wording: 'The product page's ISO/IEC 27001 claim is backed by a certificate: Tobii AB (which operates Sticky) holds ISO/IEC 27001:2022, certificate SE556613-9654 from the SWEDAC-accredited Scandinavian Business Certification AB, valid to 8 May 2028. It covers the company's information security management and does not name Sticky, the attention method or fieldwork.' The Tobii Sticky privacy and assurance cells in capability-matrix.json say the same thing ('Tobii AB's ISO/IEC 27001:2022 certificate ... is company-wide information security and does not name Sticky'). The phrase 'without a certificate' comes from the original claim, and the correction replaced it. | contradicted | Replaced 'but shows no certificate' with two sentences naming Tobii AB's company-wide certificate, its validity date and that it does not name Sticky. | ATT-TOBII-F11 final_wording says Tobii AB's company-wide ISO/IEC 27001:2022 certificate, valid to 8 May 2028, backs the page's claim but does not name Sticky, so 'shows no certificate' was wrong. | chapters/09-privacy.md |
| I1-065 | In the sources screened by the cutoff, this review found no public regulator action or lawsuit naming a seed attention-measurement vendor. | Read literally, this is false by the paper's own record. The paper documents Nielsen's patent litigation against TVision, a seed vendor, which Viant took on when it bought TVision (chapter 5). The June 2026 verdict and the 2025 action are also in the ledger. The absence finding is about privacy and biometric-data matters, so it needs that scope. The same unscoped wording appears in the standfirst, the in-short and chapter 12, line 65. ATT-H-F30's final_wording carries the same unscoped phrase and should be scoped there too. | Chapter 5, line 60: 'Viant took on Nielsen's pending patent litigation against TVision when it bought the company (ATT-D-F14)'. ATT-U-F09: 'a June 2026 unanimous non-infringement verdict in the 2022 action (1:22-cv-00057); the 2025 action (1:25-cv-00575) remains active; TVision filed antitrust counterclaims in November 2025.' ATT-TVISION-F21: 'Viant assumed Nielsen's patent litigation against TVision; Nielsen has filed several actions since 2021.' ATT-H-F30's searches were privacy searches: vendor names plus 'general BIPA searches on eye tracking, facial coding and webcam research'. | supported with qualification | Body sentence now scoped to privacy or biometric-data actions, plus a sentence acknowledging Nielsen's patent actions against TVision since 2021. | Holds: ATT-H-F30's searches were privacy/BIPA searches and ATT-TVISION-F21 records Nielsen's patent actions against TVision; fixed per R3-14 using ATT-TVISION-F21 rather than ATT-D-F14 and a chapter 5 link. | chapters/09-privacy.md |
| I1-066 | In the sources screened by the cutoff, this review found no regulator action or lawsuit naming a seed attention vendor. | This is the standfirst version of finding 09-privacy-02. Without a privacy scope, 'no ... lawsuit naming a seed attention vendor' contradicts chapter 5's account of Nielsen's patent litigation against TVision. | Chapter 5, line 60 (ATT-D-F14); ATT-U-F09; ATT-TVISION-F21. ATT-H-F30 covers only privacy and BIPA searches. | supported with qualification | Standfirst now reads 'no privacy or biometric-data regulator action or lawsuit'. | The standfirst's unscoped 'no regulator action or lawsuit' contradicts the TVision patent litigation record. | chapters/09-privacy.md |
| I1-067 | No action found names a seed attention vendor (ATT-H-F30). | This is the in-short version of the same problem. It has no privacy scope and no bound to the searches run, and the searches named only five vendors (Realeyes, Affectiva, TVision, Tobii, iMotions). It contradicts chapter 5 on the TVision patent litigation. | ATT-H-F30 limitations: 'Adelaide, Lumen, Amplified, DoubleVerify, IAS, Neurons, DAIVID, Kantar and eye square were not named in any litigation search ... An access error or absent search hit is not evidence that no action exists.' Chapter 5, line 60 (ATT-D-F14). | supported with qualification | In-short now says no privacy or biometric-data action found in the searches run names a seed vendor, and that Nielsen's patent actions against TVision are a separate matter (ATT-TVISION-F21). | Holds; R3-14 also asks to acknowledge the patent actions, so the in-short adds that clause and marker. | chapters/09-privacy.md |
| I1-068 | iMotions' policy makes the client, as data controller, collect consent under GDPR Articles 6(1)(a) and 9(2)(a) (ATT-H-F24) | This misstates the vendor document. The iMotions policy gives test-person consent to 'the Data Controller' and does not name the client. The verifier corrected this exact reading in ATT-IMOTIONS-F14, and the matrix cell carries the correction. The chapter repeats the uncorrected wording from ATT-H-F24. The policy's date (22 June 2022, before the Affectiva integration) is also left out. | ATT-IMOTIONS-F14 final_wording: '(1) the iMotions policy assigns test-person consent to 'the Data Controller' without naming the client; reading the client as controller (with iMotions as processor) is a reasonable inference from section 5, not a quotation.' Source passage (ATT-H-S32): 'it is the Data Controller's responsibility to collect your consent in accordance with art. 6(1)(a) and 9(2)(a) of the GDPR'. The iMotions Media Analytics privacy cell in capability-matrix.json says: 'puts GDPR consent on 'the Data Controller', without naming the client.' | supported with qualification | Sentence now quotes 'the Data Controller', labels client-as-controller as this review's inference, dates the policy 2022 and adds (ATT-IMOTIONS-F14). | ATT-IMOTIONS-F14 corrects ATT-H-F24: the policy assigns consent to 'the Data Controller' without naming the client, and the policy dates from 2022. | chapters/09-privacy.md |
| I1-069 | 9 wording, precision and labelling issues in 09-privacy | Integrity pass findings rated P2: 09-privacy-06, 09-privacy-07, 09-privacy-08, 09-privacy-09, 09-privacy-10, 09-privacy-11, 09-privacy-12, 09-privacy-13, 09-privacy-14 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 9 passages reworded for precision; no verdict changed | Precision and consistency. | chapters/09-privacy.md |
| I1-070 | The contradicted rule needs an attention signal documented as dropped or failing in an agent-driven buy. The 3.0 carriage gap was a specification defect found in review and fixed in the next minor version, so it does not meet that bar (ATT-I-F03) (ATT-I-F17). | This misstates the registered contradiction rule. The rule also covers a signal that 'could not be carried', and it counts an interoperability test as well as a buying workflow. Those are the words ATT-I-F03 and the inshort use for the 3.0 gap ('could not be carried end to end'), so a reader who checks the rule will think the chapter dropped the one clause its own evidence touches. The reason given for setting the gap aside ('fixed in the next minor version') is not in the rule. The reasons the rule does support are these: the gap was found in a design review, not in a buying workflow or an interoperability test, and 3.0 is not a version current at the cutoff (H5's population). The verdict, unresolved, still holds. | hypotheses.json H5 rule_contradicted: 'A source documents that an attention signal was dropped, could not be carried, or failed in an agent-driven buying workflow or interoperability test, and no attention signal reaches documented production use.' H5 population: 'Agentic advertising protocols ... at the versions current at the cutoff'. ATT-I-F03 final_wording: 'attention could not be carried end to end in AdCP 3.0'. ATT-I-F03 limitations: 'A design review, not an observed failure in a live agent buy.' | supported with qualification | Verdict now restates the full contradicted rule and sets aside the 3.0 gap as a design-review finding on a non-current version; 'What would move the verdict' aligned to the same rule wording. | H5 rule_contradicted includes 'could not be carried' and interoperability tests, and ATT-I-F03 limitations call the 3.0 gap a design review; 3.0 is not a version current at the cutoff per ATT-I-F01 and H5's population. | chapters/10-agents.md |
| I1-071 | H5 is supported when an attention signal reaches production use or a transaction in an agent-driven buy, and contradicted when a source documents one dropped or failing there while none reaches production use. | This paraphrase of the registered H5 rules leaves out conditions. The supported rule also requires the protocol name and version to be recorded. The contradicted rule also covers a signal that 'could not be carried' and failures in an interoperability test. As written, the rule looks narrower than the one registered, and the 3.0 'could not be carried' evidence seems to fall outside it without argument. | hypotheses.json H5 rule_supported: '... documented production use or a demonstrated transaction in an agent-driven buying workflow, with the protocol name and version recorded.' rule_contradicted: '... dropped, could not be carried, or failed in an agent-driven buying workflow or interoperability test, and no attention signal reaches documented production use.' | supported with qualification | Replaced the rule sentence in 'The test' with two sentences matching hypotheses.json. | The paraphrase omitted the protocol-name-and-version condition and the 'could not be carried' and interoperability-test clauses of the registered H5 rules. | chapters/10-agents.md |
| I1-072 | Attention can be written into the leading agentic advertising protocol, and on the public record it goes no further. | No claim or source supports calling AdCP 'the leading' agentic advertising protocol. The superlative favours the body in which the author co-leads a working group and runs two registered agents. The paper's own data do not suggest AdCP leads: IAB Tech Lab's registry listed 47 agents against AAO's 34. 'It goes no further' is also flatter than ATT-I-F17's final wording, which records that sellers handle the attention fields in test mode and scopes the limit to a named vendor's signal. | No ledger claim calls AdCP leading (ledger searched for AdCP with leading, dominant or largest). ATT-V-I-S05: IAB Tech Lab Agent & MCP Server Registry, 47 agents. ATT-I-F04: 34 agents in AAO's registry. ATT-I-F17 final_wording: 'No named attention vendor's signal is documented beyond schema representation in AdCP v3.1.24 ... shows seller-side handling in test mode'. Interest declared in ch13 and in this chapter's callout. | supported with qualification | Standfirst now names AdCP as the protocol maintained by AgenticAdvertising.org, with the test-mode handling stated separately. | No claim or source supports 'leading', and 'goes no further' is flatter than ATT-I-F17; fixed together with R3-16. | chapters/10-agents.md |
| I1-073 | DoubleVerify is the one attention vendor with a live agent server. | This absence claim is unbounded, and the chapter contradicts it four sentences later. Adverteyes, which this chapter treats as an attention vendor, documents an MCP server with a live endpoint, opened to a limited group on 21 July 2026, which its claim records as an implemented connector. Viomba also markets a 'free to start' MCP server. The distinction the evidence supports is that DoubleVerify is the only attention measurement vendor found with a live agent server. Chapter 12 (line 81) makes a broader claim still, that no MCP server returning attention metrics was found at any attention vendor, which is inconsistent with this chapter's paragraph on vendor MCP servers. | ATT-ADVERTEYES-F09 final_wording: 'Read-only MCP server documented since 12 May 2026 and opened to a limited group, per the 21 July 2026 blog (beta)'. Its limitations: 'the MCP server shows an implemented connector (a documented endpoint)'. Evidence passage (ATT-ADVERTEYES-S18): 'The server is available at: https://mcp.adverteyes.ai/mcp'. ATT-U-F23: Viomba MCP 'free to start'. ATT-I-F13 verifier note: 'the paper cannot say no attention vendor has an agent-facing server'. | supported with qualification | Now: 'Of the media-attention vendors whose agent surfaces were screened (DoubleVerify, IAS, Adelaide, Lumen, TVision and Amplified), only DoubleVerify was found with a live agent server (ATT-I-F13).' | ATT-ADVERTEYES-F09 records a documented Adverteyes MCP endpoint opened to a limited group, so the unbounded 'the one attention vendor' is wrong; the bound uses ATT-I-F13's screened vendor list rather than 'attention measurement vendors', which would still include Adverteyes. Chapter 12 line 81 is outside this file. | chapters/10-agents.md |
| I1-074 | The register lists every agent surface screened, with its status and the observable behind it. | This is false. [agentic] renders data/agentic-map.json, which has no row for several surfaces this chapter discusses: DoubleVerify's DV Neura MCP endpoint (only the Butler/Till workflow row appears), Ask Neura, the MCP servers of Adverteyes, Viomba and Amplified, IAB Tech Lab's Agent & MCP Server Registry, AAO brand records and brand.json, and TripleLift's TL Direct. | build/instruments.mjs maps agentic to data/agentic-map.json (20 rows). Those rows cover AdCP, the AAO registry, Prebid, AAMP, the DV/Butler/Till workflow, PubMatic, IAS Agent, Amplified's planning API, the official MCP Registry, and Adelaide/Lumen/TVision. None covers the surfaces listed in the issue. | supported with qualification | Register sentence now scopes the register to protocol, registry and platform surfaces and says that surfaces covered only in the text, such as vendors' MCP servers and TL Direct, are not in it. | agentic-map.json has no rows for DV Neura's MCP endpoint, the vendor MCP servers, IAB Tech Lab's registry, brand.json checks or TL Direct, so 'every agent surface screened' is false. | chapters/10-agents.md |
| I1-075 | Three vendors advertise MCP access to attention predictions. | The count is wrong against the paper's own ledger and matrix. DAIVID, a scored creative attention-prediction vendor, also sells its product via 'API, MCP or Self Serve'. The sentence reads as a complete count and is not bounded to the vendors screened. If DAIVID is added, the later clause 'none of these states AdCP or AAMP support' needs checking for DAIVID. | ATT-DAIVID-F01 final_wording: 'By September 2026 DAIVID's own site sells 'Creative Intelligence' via 'API, MCP or Self Serve''. capability-matrix row daivid-creative-data-feed-api (note 'Predicted creative attention and emotion outcomes delivered by API'; data_access: 'MCP is claimed but undocumented'). | supported with qualification | Opening now bounded to 'the vendor pages screened' and names Viomba, Adverteyes, Amplified and DAIVID; added a DAIVID sentence with (ATT-DAIVID-F01) and (ATT-DAIVID-S17); 'None documents' is now 'None of the three'; 'none of these states AdCP' is now 'none of Viomba, Adverteyes or Amplified'. | ATT-DAIVID-F01 and the DAIVID data_access cell show DAIVID also sells via MCP, which is claimed but undocumented; a typed count was avoided, and the AdCP/AAMP clause was limited to the three vendors ATT-U-F23 and ATT-AMPLIFIED-F26 cover. | chapters/10-agents.md |
| I1-076 | 10 wording, precision and labelling issues in 10-agents | Integrity pass findings rated P2: 10-agents-07, 10-agents-08, 10-agents-09, 10-agents-10, 10-agents-11, 10-agents-12, 10-agents-13, 10-agents-14, 10-agents-15, 10-agents-17 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 10 passages reworded for precision; no verdict changed | Precision and consistency. | chapters/10-agents.md |
| I1-077 | The largest single-cell effect in cohort A is [q_miss_vendor_media]'s [q_miss_dim_media]: filling it moves the FQ from [q_miss_from_media] to [q_miss_to_media]. In cohort B it is [q_miss_vendor_creative]'s [q_miss_dim_creative], from [q_miss_from_creative] to [q_miss_to_creative]. | These numbers describe the wrong test. The q_miss_* stats come from the deletion test (data/fq-missing-evidence.json, which scripts/stats.mjs reads as Q.missing_json), not from the fill scenarios. The sentence says a cell was 'filled', but neither cell is n/e. Amplified Campaign Performance's independent assurance is scored 1 (grade A), and Amplified AttentionAI's outcome validation is scored 1 and the row has no n/e cells at all. Both effects come from deleting a low cell, which is the chapter's own point that 'a low cell left out can raise a quotient'. The only n/e cells in cohort A are the two Amplified rows' production scale, and filling them moves FQ 48 to 47/49 and 45 to 43/46. Also, 'Amplified' is ambiguous, because two Amplified products sit in cohort A. | fq-missing-evidence.json: {cohort media, vendor Amplified, dim assurance, value 1, from 47.55, to 51.12} and {cohort creative, vendor Amplified, dim outcome_validation, value 1, from 49.46, to 53.78, band Mid->Strong}. capability-matrix.json quotient.cohorts.media.sensitivity.deletion.max_up is the same record. Matrix rows: Amplified Campaign Performance assurance v1 (scored), production_scale ne; Amplified AttentionAI scenarios {ne_cells: 0}. Fill scenarios: Amplified CP low 47 / high 49; Amplified CS 43 / 46; DAIVID 45 / 52 (Strong); Lumen SPOTLIGHT 49 (Mid) / 56. stats: q_scenario_tested_media 2, q_scenario_tested_creative 6. | contradicted | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/11-fq.md |
| I1-078 | In cohort B, no product documents an independent audit or validation of its testing method; the best any scores is self-attestation. | This contradicts the matrix. eye square InContext scores 4 on independent assurance (grade A, verified), from an ISO 20252 certificate for its research process. Creative anchor 4 exists for exactly this case. The first clause holds, but 'the best any scores is self-attestation' is false. The absence is also not bounded to the sources screened, and the certificate is status-sensitive: it expires about six weeks after the cutoff. | capability-matrix.json eye-square-incontext assurance: v 4, grade A: 'Austrian Standards' own database lists certificate S 001201, ISO 20252:2019 ... valid to 13 November 2026 ... It certifies the research process, not the InContext attention metric'. ATT-EYESQUARE-F10 final_wording: '... valid at the cutoff. It certifies the company's research-process quality system, not the InContext attention metric or SEAL ... fit creative anchor 4'. scoring-rules assurance creative anchor 4: 'an independent audit or certification covering part of the method, for example ... ISO 20252, but not the attention metric'. | contradicted | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/11-fq.md |
| I1-079 | No product documents independent, replicated validation against brand or sales outcomes, and no product documents a held-out comparison against time in view ([chapter 3](#validation)). The highest scores rest on commissioned studies with a published method, such as PwC's review for Lumen ([chapter 4](#vendor-evidence)). | This passage has three problems. (1) The top outcome-validation scores are 3, held by Lumen and TVision in cohort A. TVision's 3 rests on an independent, peer-reviewed academic study (evidence grade B), not a commissioned one, so 'the highest scores rest on commissioned studies' misstates the evidence class. (2) Citing PwC's review as the example, without its result, can read as validation. PwC found attention not significantly better than viewability on awareness and recall. (3) Both absence statements are not bounded to the sources screened, although chapter 3 and the abstract bound them. | Matrix outcome_validation: Lumen 3, grade C, commissioned PwC review, no holdout; TVision 3, grade B, 'the independent peer-reviewed McGranaghan, Liaukonyte and Wilbur study (no funding; TVision supplied data) ... in-sample with no holdout'. ATT-B-F06 evidence_class independent_measurement. ATT-C-F11 final_wording: 'attention was not significantly better than viewability on first-mention awareness, any-mention awareness or digital ad recall'. ATT-B-F01 and ATT-C-F27 bound the absence to 'the sources screened'. | supported with qualification | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/11-fq.md |
| I1-080 | most of the difference between the Strong and Mid bands comes from a few cells: an MRC record, a named buying platform, a published validation statistic, a stated production volume. | The matrix contradicts two of the four drivers named. In cohort A the only MRC-accredited product (DoubleVerify, assurance 4) is in Mid, and mean assurance is higher in Mid than in Strong (2.0 vs 1.5). Production scale also averages higher in Mid than in Strong in both cohorts (A: 3.0 vs 2.25; B: 1.4 vs 1.0), and cohort B has no MRC record at all. What actually separates the bands is privacy posture, outcome validation and activation in cohort A, and privacy posture, outcome validation and granularity in cohort B. | Mean Strong minus Mid by dimension, computed from capability-matrix.json. Cohort A: privacy +1.5, outcome validation +1.25, activation +1.25, assurance -0.5, production scale -0.75. Cohort B: privacy +0.83, outcome validation +0.67, granularity +0.66, production scale -0.4. DoubleVerify Authentic Attention: assurance 4, q 49, band Mid. | supported with qualification | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/11-fq.md |
| I1-081 | The strongest cells are granularity, channel coverage and method disclosure, where vendor documentation describes the product in detail. | This is not what the matrix shows. In cohort A the highest-scoring dimension is privacy posture (mean 3.25), ahead of granularity (3.0); method disclosure (2.88) ties with US availability, and channel coverage (2.75) ties with method validity. In cohort B, creative diagnostics (3.33) is second after granularity (3.5), ahead of channel coverage (3.0) and method disclosure (2.92). | Dimension means over counted cells, computed from capability-matrix.json. Cohort A: privacy 3.25, granularity 3.0, method disclosure 2.88, US availability 2.88, channel coverage 2.75, method validity 2.75. Cohort B: granularity 3.5, creative diagnostics 3.33, channel coverage 3.0, method disclosure 2.92. | supported with qualification | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/11-fq.md |
| I1-082 | Independent assurance and outcome validation score lowest in both cohorts. | This is true for cohort A only. In cohort B, production scale averages lowest (1.25 over its 8 scored rows), just below independent assurance (1.27) and outcome validation (1.40). The abstract's point 11 repeats the same sentence ('the weakest dimensions in both'). | Cohort A means from capability-matrix.json: outcome validation 1.62, assurance 1.75, next lowest activation and data access 2.38. Cohort B: production scale 1.25 (n 8, 4 n/e), assurance 1.27 (n 11), outcome validation 1.40 (n 10). | supported with qualification | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/11-fq.md |
| I1-083 | Independent assurance is the weakest group. | 'Group' is a defined term in this chapter (Depth, Reach, Standing), and independent assurance is a dimension, not a group. Read as a dimension, the claim is still wrong for cohort A, where outcome validation averages lower (1.62 vs 1.75). The next paragraph's 'Outcome validation is next.' repeats the wrong ordering and should become 'Outcome validation is the other; in cohort A it averages lower still.' | capability-matrix.json quotient.groups defines Depth, Reach and Standing as the groups. Cohort A means: outcome validation 1.62, assurance 1.75. Cohort B means: assurance 1.27, outcome validation 1.40, production scale 1.25. | supported with qualification | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/11-fq.md |
| I1-084 | Removing the shrinkage (K = 0, which reports Raw) moves cohort A positions by up to [q_k0_max_shift_media] places and cohort B by up to [q_k0_max_shift_creative]. At K = 2 and K = 8 the largest moves are [q_k2_max_shift_media] and [q_k8_max_shift_media] places in cohort A, and [q_k2_max_shift_creative] and [q_k8_max_shift_creative] in cohort B. | The paragraph leaves out the one band change the shrinkage test produces. At K = 0, Amplified Campaign Strategy falls from Mid to Narrow (Raw offset -8.06), and the rule says the rerun reports 'how far positions and bands move'. This sits in a section titled 'How stable the bands are', right after the chapter says the tests rerun the index without K. The tokens also render as 'up to 0 places' and '0 and 0 places', which reads badly and suggests movement where there is none. (The replacement types 'unchanged' and 'no position or band moves' for the zero-valued stats, so re-check it if the stats change.) | capability-matrix.json quotient.cohorts.media.sensitivity.shrink: {k 0, max_shift 0, band_changes ['Amplified']}, {k 2, max_shift 0, band_changes []}, {k 8, max_shift 0, band_changes []}; creative {k 0, max_shift 1, band_changes []}. stats q_k0_bands_media 1, q_k0_bands_creative 0. Amplified Campaign Strategy raw 2.1375, F 2.5406, 20 x (2.1375 - 2.5406) = -8.06, below the Mid floor of -8. scoring-rules shrink_rationale: 'reports how far positions and bands move'. | supported with qualification | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/11-fq.md |
| I1-085 | In cohort A, one product holds an MRC accreditation for its attention metrics, and it covers three environments, not CTV; one is in an MRC-confirmed audit; the rest have at most a pre-audit or no record ([chapter 2](#standards)). | Accreditation status is status-sensitive and is given here with no date. Chapter 2 and fig-accreditation date it to 28-29 September 2026. 'At most a pre-audit or no record' also misdescribes IAS. IAS was in MRC's process until it withdrew in January 2026, and it scores 2, above the pre-audit product, because its viewability and invalid-traffic inputs are accredited. The sentence carries no claim markers. | ATT-A-F12 final_wording: 'On 28 September 2026, MRC's list of accredited digital services has one entry under the Attention Metrics category: DoubleVerify ... desktop, mobile web and mobile in-app'. ATT-A-F15 (XPLN in audit). ATT-A-F14 (Adelaide pre-audit reviewed 5/2026). ATT-IAS-F01 (removed from in-process status, January 2026). Chapter 2 line 39: 'Status is as of 28-29 September 2026'. Matrix: IAS assurance 2 (accredited inputs), Adelaide 1. | supported with qualification | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/11-fq.md |
| I1-086 | 5 wording, precision and labelling issues in 11-fq | Integrity pass findings rated P2: 11-fq-10, 11-fq-11, 11-fq-12, 11-fq-13, 11-fq-14 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 5 passages reworded for precision; no verdict changed | Precision and consistency. | chapters/11-fq.md |
| I1-087 | On Q1's registered rule, attention trades as a diagnostic at the cutoff. | Unsupported Q1 verdict. The registered Q1 rule codes each metric-environment pair and gives no category-level verdict. Under it, six pairs code as diagnostic and eleven as unresolved (chapter 6). The rule's unresolved clause, and the verifier correction in ATT-E-F24, say explicitly that pairs with unreadable terms are 'unresolved', not 'diagnostic'. Calling attention a diagnostic 'on Q1's registered rule' collapses most of the coded pairs into the wrong code. Line 13 of the same chapter states the result correctly. | hypotheses.json Q1: rule_supported ('partial currency' ... 'otherwise diagnostic'); rule_unresolved ('coded unresolved, not diagnostic'). ATT-E-F24 final_wording: 'AU pairs whose guarantee terms are unpublished ... should be coded unresolved under the rule's unresolved clause, not diagnostic'. Chapter 6: 'Six pairs code diagnostic and eleven unresolved'. | contradicted | Standfirst now says no pair codes as currency or partial currency and each codes as diagnostic or unresolved, split into two sentences with the R3-13 note that diagnostic marks missing assurance. | Holds: the Q1 rule codes pairs, not the category, and unreadable terms code unresolved (hypotheses.json Q1, ATT-E-F24). | chapters/12-next.md |
| I1-088 | The one accredited metric is not traded, the traded metrics are not accredited, and in every guarantee found the metric's own vendor measures delivery. | 'Not traded' contradicts the paper's own definition of trading. Under H2, optimisation is a trading use. Chapter 6 codes DoubleVerify segments and Scibids optimisation as live trading uses, and its Q1 table gives DoubleVerify's accredited pairs 'optimisation only'. ATT-E-F24 says DV's attention metric 'is used in optimisation segments'. What the claims support is that the accredited metric carries no guarantee or billing use, and that none of the guaranteed or priced metrics is accredited (the chapter 6 inshort). | hypotheses.json H2 statement: 'used in trading media ... (optimisation, billing unit or contractual guarantee)'. ATT-E-F24: 'DoubleVerify's attention metric is used in optimisation segments, with no guarantee or billing use found'. ATT-E-F23: attention marketed as a live optimisation input in 'DoubleVerify segments and Scibids optimisation'. Chapter 6, Q1 table, DV row: 'None found; optimisation only'. | supported with qualification | Standfirst now says the accredited metric carries no guarantee or billing use and none of the guaranteed or priced metrics is documented as accredited; the delivery clause uses the R3-03 wording. | Holds: DV's accredited metric is used in optimisation, a trading use under H2, so 'not traded' was wrong. | chapters/12-next.md |
| I1-089 | Only Kargo and The Wall Street Journal publish a remedy, and Vevo publishes neither (ATT-VEVO-F13). | The count is wrong for the set the paragraph names ('Among the named guarantees'). SeenThis also publishes a remedy, and the very next sentence says so, which leaves the paragraph contradicting itself. ATT-VEVO-F13 covers only the six AU guarantees. Chapter 6's table marks SeenThis 'Remedy public? Yes'. The fig-guarantees caption counts three offers that state a remedy. | ATT-VEVO-F13 final_wording: 'Among the six named attention guarantees in the register, all built on Adelaide AU ... Only Kargo and the WSJ publish a remedy'. ATT-R2-F01: 'extra delivery at no cost if the threshold is missed'. Corrections ledger R2-05 (SeenThis added 'with a stated remedy'). Chapter 6 table, SeenThis row: 'Yes: extra delivery at no cost'. fig-guarantees.json caption: 'three state a remedy'. | supported with qualification | Scoped the remedy sentence to the AU guarantees and added that A+E's TVision guarantee has no public level or remedy (ATT-TVISION-F23), so 'among the named guarantees' matches the ten-offer set. | Holds: ATT-VEVO-F13 covers only the AU guarantees, and SeenThis publishes a remedy. | chapters/12-next.md |
| I1-090 | \| Q1, currency or diagnostic \| No pair codes as currency or partial currency (ATT-E-F24) \| Accreditation plus a guarantee with threshold and remedy (partial currency); all four conditions (currency) \| | This row misstates the registered partial-currency rule. Under Q1, 'partial currency' needs only accreditation plus a documented guarantee or billing use, with threshold, remedy, audit or adoption allowed to be missing. The row also conflicts with line 79 and chapter 6, which say AU accreditation would lift the Journal pair to partial currency. That pair publishes a remedy but no threshold, so it would not qualify under the row's wording. | hypotheses.json Q1 rule_supported: 'partial currency when accreditation and a guarantee or billing use are documented but remedy, audit or adoption is not'. Chapter 6: 'MRC accreditation of AU would lift the Journal and Kargo pairs to partial currency'. Chapter 6 Q1 table, Journal row: 'Remedy only'. | supported with qualification | Q1 table row now reads 'Accreditation plus a guarantee or billing use (partial currency)'. | Holds: the registered partial-currency rule needs accreditation plus a guarantee or billing use only. | chapters/12-next.md |
| I1-091 | An independent, replicated study linking a vendor's attention metric to sales beyond time in view. | Under the registered H1a rule, only a held-out comparison against a time-in-view baseline can move the code. A study that links attention to sales beyond time in view in-sample leaves H1a unresolved ('only in-sample associations are published'). The paragraph's closing sentence says 'Such a study would move H1a for sales only', which is therefore true only if the study is held-out. The rule also does not require independence or replication, though the paper may add those for the buyer. The table row for H1a states the test correctly. | hypotheses.json H1a rule_supported: 'positive increment over a time-in-view baseline on held-out data'; rule_unresolved: 'or only in-sample associations are published'. Table row H1a in this chapter: 'A held-out gain over time in view, per outcome'. | supported with qualification | First 'what would change' finding now requires a held-out study predicting sales beyond time in view. | Holds: only a held-out comparison against time in view moves H1a under the registered rule. | chapters/12-next.md |
| I1-092 | In the sources screened, this review found no regulator action or lawsuit naming a seed attention vendor (ATT-H-F30). | The absence claim is not bounded to what was searched, and 'seed attention vendor' is undefined in this chapter. The sentence follows discussion of TVision, Amplified and Lumen, so a buyer will read it as covering the vendors this chapter discusses. According to the claim, the litigation searches named Realeyes, Affectiva, TVision, Tobii and iMotions. Adelaide, Lumen, Amplified, DoubleVerify and IAS were not named in any litigation search. Chapter 9 (line 54) states that bound; this chapter drops it. | ATT-H-F30 final_wording: 'searches named Realeyes, Affectiva, TVision, Tobii and iMotions'; limitations: 'Adelaide, Lumen, Amplified, DoubleVerify, IAS, Neurons, DAIVID, Kantar and eye square were not named in any litigation search'. Chapter 9 line 54 carries the full bound. | supported with qualification | Bounded to privacy and biometric-data proceedings in the searches run, listed searched and unsearched vendors, and noted Nielsen's patent actions against TVision as separate (ATT-TVISION-F21); also applies R3-14. | Holds: ATT-H-F30's searches named only five vendors, and Nielsen's patent suits against TVision mean the unbounded sentence was wrong. | chapters/12-next.md |
| I1-093 | Sellers often fund the measurement: Condé Nast and Hearst cover the cost of AU on their inventory (ATT-E-F06). | This presents a vendor assertion as fact and overgeneralises it. ATT-E-F06 is Adelaide's own statement, classed vendor_assertion, and it describes itself as 'one of few public statements on who pays'. Two publishers named by the metric vendor do not support 'often'. Chapter 6 attributes the same point ('Adelaide says'). | ATT-E-F06 (evidence_class vendor_assertion) final_wording: 'Adelaide says ... that Condé Nast and Hearst cover the cost of AU measurement on their inventory. This is one of few public statements on who pays for attention measurement'. Chapter 6: 'Adelaide says Condé Nast and Hearst cover the cost of AU measurement'. | supported with qualification | Now 'Some sellers fund the measurement: Adelaide says Condé Nast and Hearst cover the cost'. | Holds: ATT-E-F06 is a vendor assertion naming two publishers, which does not support 'often'. | chapters/12-next.md |
| I1-094 | 15 wording, precision and labelling issues in 12-next | Integrity pass findings rated P2: 12-next-08, 12-next-09, 12-next-10, 12-next-11, 12-next-12, 12-next-13, 12-next-14, 12-next-15, 12-next-16, 12-next-17, 12-next-18, 12-next-19, 12-next-20, 12-next-21, 12-next-22 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 15 passages reworded for precision; no verdict changed | Precision and consistency. | chapters/12-next.md |
| I1-095 | Two outside reviews by another AI model, one of the briefs and one of the evidence, and two internal review passes of the full draft led to the [corrections] entries in the corrections ledger. | The sentence gets the source of the corrections ledger wrong. The ledger's 25 entries are 6 from the second outside review (R2-01 to R2-06, kind 'review') and 19 from the AI consistency audit of the scores (SC-01 to SC-19, kind 'scoring'). The review of the briefs is not in the ledger: this chapter says its dispositions 'are recorded with the briefs' (research/_briefs/ROUND-1-CHANGES.md). The ledger holds no entry from either internal review pass. The consistency audit, which produced three quarters of the entries, is not named. The same sentence appears in 00-abstract.md ('How it was done'). The same inshort also calls the second review 'one of the evidence', but it did not cover the product sweeps (see 13-method-04). | data/corrections-ledger.json: 25 rows, kind counts {review: 6, scoring: 19}; stats.json corrections=25, corrections_scores=19; 13-method body: 'the dispositions are recorded with the briefs'; commit 16a65942 'consistency-audit changes applied and logged (SC-01 to SC-19)'. | supported with qualification | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/13-method.md |
| I1-096 | Verifiers made [verifier_checked] checks: [verifier_supported] found a claim supported as stated, [verifier_qualified] supported it with a correction and [verifier_contradicted] contradicted it. | The sentence says every check was a check of a claim, but nearly half were not. Of the 1,520 verifier checks, 785 target findings (claims). The other 735 target capability records (for example 'capabilities.method_validity'), product-capability rows, profile fields ('why_profiled', 'who_buys'), data-table rows and summaries. The total also counts both passes for the eight workstreams verified twice: 243 checks sit in the *.pass1.json files. | evidence/verify/*.json 'checked' arrays: 1520 total (stats.json verifier_checked); finding_id pattern *-Fnn = 785, other = 735; *.pass1.json = 243 checks. | supported with qualification | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/13-method.md |
| I1-097 | Several material documents could not be read in full: the ARF's Phase 3 report (members only), Adelaide's case-study guides (gated), the full Newsworks and Lumen 2025 report, and some MRC letters that returned errors. | Two items in this list are wrong, and one material item is missing. (1) The full Newsworks, Lumen and Peter Field 'Attention 2025' report was read in full: verifier C downloaded the 24-page PDF and read it with pdftotext, the register records it as full access, and chapters 4 and 12 cite its 24% CPM-premium estimate. (2) The one MRC letter whose URL returned an error (June 2025, 404) was then read in full from an archived capture by verifier A. The MRC letter that was not read is the December 2022 accreditation letter, which was never opened, not one that errored. (3) MediaSense's gated report, the one outside review of AU (chapter 4), was not read, and the list omits it. | source-register ATT-V-C-S08 'Attention 2025 (Newsworks, Lumen and Peter Field report)', access_limits 'full'; evidence/verify/C.json verifier_method_note: 'all six PDFs (... Newsworks 2025 ...) ... read with pdftotext'; overall_reliability: 'six PDFs read in full, plus the Newsworks report the stream did not open'; ATT-C-F29 final_wording cites the report's 24% estimate; evidence/verify/A.json: June 2025 letter 'The live URL returns 404 ... the archived PDF ... was read in full'; ATT-A-F10 limitations: 'The December 2022 accreditation letter was not opened'; ATT-ADELAIDE-F09 limitations: 'Report gated and not read'; ATT-C-S10 reliability: 'The MediaSense report itself is behind a lead form and was not opened'. | supported with qualification | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/13-method.md |
| I1-098 | After the evidence and profiles were verified, a second outside review spot-checked the evidence. | The order of events is wrong. The second outside review did not see all of the evidence: the product sweeps were researched and verified after it. The round-2 package held 10 streams (A to I and U), 8 profiles, 26 verifier records, 461 findings and 132 capability records. The ledger now holds 584 claims. Eight scored rows rest mainly on the later product sweeps and their verifiers: xpln.ai in cohort A, and Adverteyes, Neurons, Tobii, iMotions, DAIVID, eye square and Dragonfly in cohort B. A reader would assume the outside review covered them. | private/openai-round-2.md: 'every evidence file (10 workstreams, 8 profiles, 26 verifier records ...)', 'all 461 findings', '132 capability records'; commit c1219a3b (29 Sep 19:30) applied round 2; commit b851d0eb (30 Sep 06:46) added evidence/research/{ADVERTEYES,DAIVID,DRAGONFLY,EYESQUARE,IMOTIONS,NEURONS,TOBII,XPLN}.json and their verifiers; evidence/research/DRAGONFLY.json written 29 Sep 20:16, verify/DRAGONFLY.json 30 Sep 07:11; stats.json claims=584. | supported with qualification | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/13-method.md |
| I1-099 | Earlier employers include Verve Group (Moments.AI), Hearts & Science and Lotame. | The full interests statement leaves out the one documented commercial link between the author's employer and a scored vendor. Adelaide is scored and profiled, and it names Samba among the tune-in attribution providers whose data train its CTV AU model. Chapters 1 and 8 mention this in passing. The chapter the abstract calls 'the full statement' does not. The native paper's interests statement disclosed the equivalent link (Samba TV and Teads). The author may also want to decide whether Samba's AgenticAdvertising.org membership and the Teads link (Teads is in the trading table here) belong in this statement too; this paper's evidence does not verify either. | ATT-ADELAIDE-F07 final_wording: '... trained on brand lift (Kantar, Lucid, Dynata), tune-in attribution (Samba, VideoAmp) and advertiser sales data ... Samba is the author's employer.'; ATT-E-F21 final_wording: 'Samba appears in the attention market only as an outcome-data supplier: Adelaide names it among tune-in attribution providers used to calibrate AU'; 01-defining.md line 90 and 08-channels.md line 38; native-2026/chapters/00-abstract.md Interests. | supported with qualification | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/13-method.md |
| I1-100 | Their findings were also checked against sources before any change and are logged in the corrections ledger. | This is not true yet. The corrections ledger holds only round-2 review entries (R2) and consistency-audit entries (SC); it has no entry from the integrity pass or the adversarial pass. Before publication, either log each accepted internal-pass change in the ledger, ideally with a distinct kind so it can be counted, or reword the sentence. If entries are added, the inshort fix in 13-method-01 should name the internal passes as a source of entries. | data/corrections-ledger.json kinds: review 6 (R2-01 to R2-06), scoring 19 (SC-01 to SC-19); no other kind or id prefix. | supported with qualification | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/13-method.md |
| I1-101 | The author then chose to replace a third outside round with two internal review passes of the full draft: | The chapter may be out of date on this point. A third outside review of the full draft was prepared at 07:43 on 30 September, after this chapter was committed at 06:46. Its prompt says 'two internal review passes follow this one', so the plan is an outside round plus two internal passes, not a replacement. If round 3 is run, this sentence and several others become wrong: 'Two outside reviews' in the inshort and abstract, the footer's 'two outside AI reviews' and the pipeline figure's '2 outside AI reviews'. The author needs to decide, and then align every one of them. | private/openai-review-3.md (modified 2026-09-30 07:43): 'OpenAI critical review of the full draft ... two internal review passes follow this one'; paper.config.json footer and pipeline.review_lines; commit b851d0eb (13-method.md added 2026-09-30 06:46). | supported with qualification | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/13-method.md |
| I1-102 | [fig-pipeline] | The pipeline figure embedded here contradicts this chapter and stats.json. (1) Its caption says 'The verifier sample covers the material claims per file, not every claim; the rest are labelled unresolved', and its footnote says 'Claims outside the verifier sample are marked unresolved'. In fact all 584 claims were checked and none is unresolved. (2) The caption says corrections include 'withdrawn claims and score changes from the review passes'. No claim was withdrawn, and the 19 score changes came from the consistency audit, not the review passes. The alt text also says 'review passes produced 25 corrections'. (3) The Scored box says '12 dimensions each', but each cohort has 11 applicable dimensions. It says '20 vendors' where the scoring unit is products. (4) The Registered box reads '2 author works', but the two records are one essay at two URLs. (5) It prints '9 universe sweep'; the chapter counts 1 vendor-universe sweep plus 8 product sweeps. The text is generated in scripts/make-figures.mjs, so the fix goes there. | figures/fig-pipeline.json caption; figures/fig-pipeline.svg text and aria-label; stats.json claims_checked=584, claims_unresolved=0, corrections_withdrawn=0, corrections_scores=19; corrections-ledger kinds; scoring-rules.json cohorts_note 'Both cohorts keep eleven applicable dimensions'; source-register ATT-I-S32 and ATT-V-I-S10 (same essay). | supported with qualification | applied by the lead author (see decision log) | Finding checked against the claim ledger, sources, stats or matrix. | chapters/13-method.md |
| I1-103 | 7 wording, precision and labelling issues in 13-method | Integrity pass findings rated P2: 13-method-09, 13-method-10, 13-method-11, 13-method-12, 13-method-13, 13-method-14, 13-method-15 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 7 passages reworded for precision; no verdict changed | Precision and consistency. | chapters/13-method.md |
| I1-104 | Unresolved for memory, brand and sales: no held-out comparison against time in view was found (ATT-B-F01) | The verdict is not bounded to the sources screened, and its only marker covers independent and academic work. ATT-B-F01 says vendor and commissioned studies were not screened and the International Journal of Advertising was left out. The sentence reads as a statement about all evidence, which needs ATT-C-F27 as well. Chapters 3 and 12 pair the two claims for this verdict. | ATT-B-F01 final_wording: 'in the sources screened by 2 October 2026 (... vendor and commissioned studies were not screened and the International Journal of Advertising was not in the screen) ... H1a is coded unresolved for each outcome on independent and academic evidence.' ATT-C-F27: 'No commissioned or vendor study screened makes a held-out comparison of an attention metric against time in view.' Chapter 12's verdict table: 'Unresolved for memory, brand and sales (ATT-B-F01) (ATT-C-F27)'. Chapter 3 line 53 uses ATT-C-F27 in the same way. | supported with qualification | H1a row bounded to the sources screened, naming independent, commissioned and vendor studies, with ATT-C-F27 added. | ATT-B-F01 covers only independent and academic sources, and ATT-C-F27 is needed for commissioned and vendor studies, as in chapter 12. | chapters/14-appendix.md |
| I1-105 | Unresolved: no randomised or holdout test with cost reported was found (ATT-B-F18) | ATT-B-F18 covers only independent and academic tests, and its limitations say vendor tests may exist. The paper-level verdict also needs ATT-C-F26. The absence is not bounded to the sources screened. 'With cost reported' is also looser than the rule, which asks for cost per incremental outcome, and vendor tests with a control and cost figures do exist: Lumen's CPG case had a control campaign and a 40% lower CPM, and IAS and Kimberly-Clark reported ROAS. As worded, the row is contradicted by chapter 3's own table. | ATT-B-F18 final_wording: 'no randomised or hold-out test, by independent or academic researchers ... in the sources screened by the cutoff, so H1b is unresolved from this stream'. Its limitations: 'Vendor and commissioned optimisation tests (workstream C) may exist'. ATT-C-F26: 'no public study that compares media bought on an attention metric with a randomised or holdout control ... and reports a memory, brand or sales gain with cost per incremental outcome and uncertainty'. Its limitations: 'Lumen CPG (control arm and CPM, but outcome is attention)'. ATT-C-F13 and ATT-C-F29 cover the Lumen CPM and IAS ROAS figures. Chapter 12: 'Unresolved (ATT-B-F18) (ATT-C-F26)'. | supported with qualification | H1b row bounded to the sources screened, cost per incremental outcome, per outcome, with ATT-C-F26 added. | Holds: ATT-B-F18 is independent-only and ATT-C-F26 sets the cost-per-incremental-outcome bar; wording aligned with H1a ('independent, commissioned or vendor') and 'of buying on attention' added from the registered H1b population. | chapters/14-appendix.md |
| I1-106 | Supported, not at scale: live optimisation and guarantees exist; no offer discloses volume against a market denominator (ATT-E-F23) | The row states live optimisation and guarantees as established facts. The corrected claim says all of this rests on vendor or seller statements, and that caveat was the point of the verifier's correction. The guarantee that satisfies the H2 rule, the Wall Street Journal's, ran in October to November 2024, so the present tense 'guarantees exist' leaves out that date. H2 is also a US test (US offers, US spend), and the row never says so. | ATT-E-F23 final_wording: 'Contractual-style guarantees were live in the US (WSJ, October to November 2024, described by Adweek as a promotion), and attention is marketed as a live optimisation input in named buying platforms ...; all of these rest on vendor or seller statements ... no offer discloses US spend, impressions or buyers against a denominator.' Chapter 6 line 73: 'All of it rests on vendor or seller statements'. hypotheses.json H2 rule_supported requires 'At least one live (not announced) US offer'. | supported with qualification | Inserted 'all on vendor or seller statements' into the R3-12 H2 row. | The vendor-statement caveat holds and was added; the proposed rewrite, with the dated WSJ guarantee and US detail, was not used because R3-12 fixes the row wording as 'guarantees exist' and chapter 6 carries the detail. | chapters/14-appendix.md |
| I1-107 | Unresolved: no public source links the two (ATT-D-F28) | This is an absence claim with no bound. The marked claim limits it to the filings, accounts and releases screened by 28 September 2026. Its limitations also say private pricing and contracts are out of reach. | ATT-D-F28 final_wording: 'full-text reads of the DV and IAS 10-Ks, DV's Q2 2026 10-Q and its September 2026 preliminary merger proxy, UK registry accounts for Lumen and DAIVID, and vendor releases for Adelaide, Amplified and Neurons found no filing ... in the sources screened by 28 September 2026.' Chapter 5 line 101: 'That holds for the sources screened by 28 September 2026'. | supported with qualification | H4 row: 'in the filings, accounts and releases screened by 28 September 2026, no source links the two'. | ATT-D-F28 bounds the absence to filings, accounts and releases screened by 28 September 2026. | chapters/14-appendix.md |
| I1-108 | Almost every campaign number is a model estimate; people are observed in reference panels, one in-home TV panel and participant tests | The row carries no claim marker. It also leaves out the device-signal index category. Chapter 1 and ATT-C-F31 code DoubleVerify's Authentic Attention, the only accredited metric, as an index of device signals (a candidate proxy under the Q2 rule), not as a model estimate. 'One in-home TV panel' is ambiguous, because Amplified's reference panel also uses an in-home TV micro-camera. The point intended is that TVision's is the only in-home panel whose observations are the campaign measurement. | ATT-C-F31 final_wording: 'DV Authentic Attention ... is an index of device and on-screen signals ... which the Q2 rule codes as a candidate proxy, not a validated model estimate; direct observation of people appears only in panel work (TVision's in-home panel, Amplified's participant panels, dentsu's eye-tracking panels)'. Chapter 1 standfirst: 'almost every attention number sold for a live campaign is a model estimate or an index of device signals'. ATT-G-F02: 'In CTV, TVision is the only cohort A product whose attention data come from direct observation of people'. Chapter 1 line 46 (ATT-AMPLIFIED-F10): Amplified 'collects facial images from device cameras and an in-home TV micro-camera'. | supported with qualification | Q2 row: 'model estimate or an index of device signals (ATT-C-F31); people are observed in reference panels, TVision's in-home TV panel and creative tests (ATT-G-F02) (ATT-KANTAR-F02)'. | The row had no marker and left out the device-signal index; replaced 'one in-home TV panel' with TVision's and 'participant tests' with 'creative tests', matching chapter 1's standfirst. | chapters/14-appendix.md |
| I1-109 | Every company and product screened, with its owner, group and status. Scored products are tagged; the rest are listed with the reason in the product register. | Neither sentence matches the data. (1) The universe table (51 rows) is not every company and product screened. SeenThis Attention, Impact Media's Guaranteed Attention Suite, Nativo, the Wall Street Journal guarantee, Viant's attention-adjusted CPM and Adelaide High Attention Markets are in the product register and the text, but none appears in vendor-universe.json. The instrument's own note calls the table 'a reasoned sample, not a census'. (2) 'The rest are listed with the reason in the product register' is wrong for 11 unscored universe rows: System1, CreativeX, Ipsos, Sundogs, CubeAI, GlassView, KERV Interactive, Mobian, Veritonic, Nielsen and Comscore. None of them has a product-register record, and their reason sits only in the universe 'note' field, which the rendered table does not show. (3) The universe instrument's note says how the sample was assembled and what was left out 'is stated in the method chapter', but chapter 13 says neither. This needs fixing in the instrument text or chapter 13. | data/vendor-universe.json: 51 rows, 20 scored, 11 with an empty matrix_id; a case-insensitive grep finds no SeenThis, Impact Media, Nativo or Wall Street Journal row. data/product-register.json (44 records) includes 'SeenThis Attention', 'Impact Media Guaranteed Attention Suite', 'Nativo attention-measured inventory', 'WSJ Election AU performance guarantee', 'Viant attention-adjusted CPM' and 'Adelaide High Attention Markets'. Chapter 13 line 15 names SeenThis and Impact Media as register additions. The build/instruments.mjs universe columns omit 'note', and its note text is 'A reasoned sample, not a census ...'. downloads.json describes vendor-universe as 'the reasoned vendor sample'. Stats: universe 51, products_registered 44. | supported with qualification | Universe intro now: [universe] rows, reasoned sample not a census, tagged products in the matrix, register reasons including offers not in the table, other rows' reasons in the note field of data/vendor-universe.csv. | Holds: vendor-universe.json has 51 rows and is a reasoned sample; SeenThis, Impact Media, Nativo, WSJ and Viant sit in the register but not in the table; the 11 name-only rows (and Magnite, Rita) have no register record, and their reason is in the CSV note field. Wording was kept to what the data show. Out of scope for this file: the instrument note says the method chapter explains how the sample was assembled and what was left out, and chapter 13 does not. | chapters/14-appendix.md |
| I1-110 | 7 wording, precision and labelling issues in 14-appendix | Integrity pass findings rated P2: 14-appendix-07, 14-appendix-08, 14-appendix-09, 14-appendix-10, 14-appendix-11, 14-appendix-12, 14-appendix-13 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 7 passages reworded for precision; no verdict changed | Precision and consistency. | chapters/14-appendix.md |
| I1-111 | His firm runs two agents on the AdCP list. | Ambiguous and incomplete interest statement. It follows 'The author works at Samba, a TV data firm', so readers will take 'his firm' to be Samba, but the agents belong to No Fluff Advisory, the author's own firm. It also leaves out that one of those agents is among the three that passed the fake-vendor test that section 10 of this page reports ('A test with a fake attention firm passed'). | Chapter 13 Interests: 'No Fluff Advisory, the author's firm, operates two agents listed in the AgenticAdvertising.org registry; one, a sales agent, is among those cited in chapter 10 as passing a test-mode conformance storyboard.' ATT-I-F07 final_wording: three registered sales agents '(Addiscovery, Purrsonality Seller and No Fluff Advisory's sales agent, operated by the author's firm) passing the accountability storyboard'. ATT-I-F04 limitations: 'Two listed agents belong to the author's firm'. fig-protocol-states caveat says the same. | supported with qualification | Interests now read 'His own firm, No Fluff Advisory, runs two agents on the AdCP list. One of them is among those that passed the test with a fake attention firm.' (dropped the typed 'three'). | Holds: chapter 13 and ATT-I-F07 say No Fluff Advisory, not Samba, runs the two agents and one passed the placeholder-vendor storyboard. | plain.md |
| I1-112 | It did not pass on TV screens. | 'Did not pass' implies DoubleVerify's CTV attention was audited and failed. The record shows only that it is not accredited and that no DoubleVerify entry was under review; the only denial on record is for pre-bid attention (March 2024). | ATT-DOUBLEVERIFY-F04 final_wording: 'DoubleVerify's CTV attention is not MRC-accredited: the April 2026 letter accredits CTV only for impressions, video viewability, completion and quartile metrics ... and on 29 September 2026 MRC's digital and television under-review lists carry no DoubleVerify entry.' Chapter 2 table: 'Not accredited ...; no entry under review'. Denial appears only for pre-bid (ATT-DOUBLEVERIFY-F02). | supported with qualification | 'It did not pass on TV screens' became 'It has no pass for TV screens, and the MRC's lists showed no review of that under way.' | Holds: ATT-DOUBLEVERIFY-F04 shows CTV attention is not accredited and no DV entry is under review, not a failed audit; its limitation says an audit could be off the public lists. | plain.md |
| I1-113 | A true trading unit would pass four tests. It would pass the MRC check. It would set a price or a promise. It would state a fix if missed. And someone else would check it. No score passes all four. | Misstates Q1's registered four conditions. The registered rule is (1) accreditation for the environment traded, (2) a billing or guarantee use with a stated threshold and remedy, (3) independent audit, (4) disclosed adoption. The page drops disclosed adoption, drops the stated threshold, and splits the remedy out as a separate fourth test. | hypotheses.json Q1 comparator: 'accreditation for the environment traded, use as a billing unit or guarantee with stated threshold and remedy, independent audit, and disclosed adoption'. Chapter 6 ('Coding H2 and Q1') and chapter 12 restate the same four. ATT-E-F24. | supported with qualification | Four tests restated to match Q1; 'No score passes all four' bounded to 'No score found passes all four'. | Holds: hypotheses.json Q1 and ATT-E-F24 name accreditation for the environment, threshold and remedy, independent audit and disclosed adoption. | plain.md |
| I1-114 | In every one, the firm that sells the score also checks if the promise was kept. No one else checks. | Two problems. (a) 'No one else checks' is an unbounded absence claim. The paper finds only that no independent audit was found, and notes that contracts may hold audit clauses no one can see. (b) 'Every one' covers all nine offers, but Impact Media's suite names no metric vendor, so it cannot be said that the seller of its score checks it. The paper's claim covers the guarantees whose metric vendor is known. | ATT-E-F22 final_wording: 'In every case the metric's own vendor measures delivery: this review found no independent third-party audit of guarantee delivery'; limitations: 'Contracts may contain audit clauses not visible in public sources.' Chapter 6 table: Impact Media 'Metric vendor: Not stated'. figure-data.json guarantees: Impact Media suite 'not stated'. | supported with qualification | Replaced with 'Where an offer says who checks, it is the firm that sells the score. No outside check was found.' | Holds (ATT-E-F22 new final wording: measurer stated only for some offers, A+E and Impact Media not stated; no independent audit found); superseded by the R3-03 wording. | plain.md |
| I1-115 | No study says what the extra cost of buying on attention was. | Unbounded absence claim. The paper bounds it to the studies screened. | ATT-C-F29 final_wording: 'No study screened discloses measurement fees, and none reports a campaign's media premium for attention-bought inventory'. Chapter 4 inshort uses the same bounded wording. | supported with qualification | 'No study says' became 'No study the paper checked says'. | Holds: ATT-C-F29 is bounded to studies screened. | plain.md |
| I1-116 | ### 8. It works best on screens with a web page or video {#channels} | 'Works best' says the measurement performs better on these screens. The paper codes only coverage (vendors documenting live measurement) and says outright that coverage does not mean the measurement is validated or accredited there. | ATT-G-F01 limitations: 'Coverage means the vendor documents live measurement, not that it has been validated or accredited in that channel.' Chapter 8: 'It does not mean the measurement has been validated or accredited in that channel.' Abstract finding 8 heading: 'Coverage follows the screen.' | supported with qualification | Heading 8 now 'It is measured most on screens with a web page or video'. | Holds: ATT-G-F01 limitation says coverage is not validation or accreditation, so 'works best' overclaims. | plain.md |
| I1-117 | 15 wording, precision and labelling issues in plain | Integrity pass findings rated P2: plain-07, plain-08, plain-09, plain-10, plain-11, plain-12, plain-13, plain-14, plain-15, plain-16, plain-17, plain-18, plain-19, plain-20, plain-21 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 15 passages reworded for precision; no verdict changed | Precision and consistency. | plain.md |
| AJ-01 | Lumen SPOTLIGHT / outcome_validation: v 2, grade D | Blind second scoring pass on the band-boundary rows: a separate AI scorer, without the first scores, scored this cell differently by two or more points or as n/e against a score. | The first pass excluded Simonov et al. by analogy with Lumen’s media model; SPOTLIGHT sells the eye-tracking measurement the study tested, so independent validation against memory (anchor 4) is met. [51, 578] | supported with qualification | Lumen SPOTLIGHT / outcome_validation: v 4, grade B | Adjudication changes a first-pass score only for a factual, anchor-application or convention error, after re-reading the deciding source. | data/scores.json (lumen: matrix-extra row Lumen SPOTLIGHT); data/second-pass.json |
| AJ-02 | Lumen SPOTLIGHT / assurance: n/e | Blind second scoring pass on the band-boundary rows: a separate AI scorer, without the first scores, scored this cell differently by two or more points or as n/e against a score. | Lumen’s company-level ‘validated and audited’ statement covers the technology SPOTLIGHT uses but names an audit outside its scope: self-attestation (anchor 1), as Tobii’s academic-validation statement is scored. [20, 8] | supported with qualification | Lumen SPOTLIGHT / assurance: v 1, grade D | Adjudication changes a first-pass score only for a factual, anchor-application or convention error, after re-reading the deciding source. | data/scores.json (lumen: matrix-extra row Lumen SPOTLIGHT); data/second-pass.json |
| I2-001 | The author works at Samba, co-leads AgenticAdvertising.org's Signals & Measurement working group, and his firm runs two agents in its registry. | This short interests statement leaves out the one documented commercial link between the author's employer and a scored vendor: Adelaide, which is scored and profiled, names Samba as a tune-in data provider for its CTV model. It also repeats the 'his firm' ambiguity that I1-111 fixed in the plain edition. Coming straight after 'works at Samba', 'his firm' reads as Samba's agents. The sentence also omits that one of the agents passed the test-mode storyboard reported in chapter 10. Chapter 13 has the full statement, but the abstract is the version most readers see. | ch13 Interests: 'Adelaide, which is scored and profiled, names Samba among the tune-in attribution providers whose data train its CTV model'; ATT-E-F21; ATT-ADELAIDE-F07; corrections I1-099 (missing link) and I1-111 ('readers will take his firm to be Samba'); ATT-I-F07 (No Fluff Advisory's agent passed the storyboard). | contradicted | 'The author works at Samba, co-leads AgenticAdvertising.org's Signals & Measurement working group, and his firm runs two agents in its registry; the full statement of interests is in [chapter 13](#method).' -> 'The author works at Samba, whose tune-in data Adelaide, a scored and profiled vendor, names among the inputs that train its CTV model (ATT-ADELAIDE-F07). He co-leads AgenticAdvertising.org's Signals & Measurement working group. His own firm, No Fluff Advisory, runs two agents in that organisation's registry, and one of them passed the test-mode storyboard reported in [chapter 10](#agents) (ATT-I-F07). The full statement of interests, including past employers, is in [chapter 13](#method).' | Holds. ATT-ADELAIDE-F07 and ATT-E-F21 document that Adelaide (scored and profiled) names Samba among the tune-in providers that train its CTV AU. Chapter 13 carries the link (correction I1-099), but the abstract left it out. 'His firm' straight after 'works at Samba' repeated the ambiguity fixed in plain.md (I1-111). ATT-I-F07 shows No Fluff Advisory's sales agent passing the accountability storyboard with the placeholder vendor. I split the replacement into short sentences and added claim markers ATT-ADELAIDE-F07 and ATT-I-F07. I changed 'its registry' to 'that organisation's registry' so the referent is clear. I did not attach ATT-I-F04, because its final_wording gives the registry total and does not say the firm owns two agents. The two-agent fact stays an author disclosure, as it is in the original and in chapter 13. Chapter 13 lists past employers, so 'including past employers' holds. | 00-abstract |
| I2-002 | The author works at Samba, a TV data firm. | The plain edition's interests statement still leaves out that Adelaide, a scored and profiled vendor, names Samba as a data provider for its model. I1-099 added this link to chapter 13, but the change was never carried into the plain edition. | ch13 Interests; ATT-E-F21 final_wording: 'Adelaide names it among tune-in attribution providers used to calibrate AU'; corrections I1-099. | contradicted | Interests paragraph: 'The author works at Samba, a TV data firm. Adelaide, one of the firms scored here, names Samba as one source of the data it uses to train its TV score. The author also co-leads a working group on AdCP...' | Holds. Chapter 13 Interests and ATT-ADELAIDE-F07 say Adelaide's CTV model is trained partly on tune-in attribution data from Samba (corrections I1-099). The plain edition left this out. I said 'data' instead of 'viewing data' to stay within the source. I also changed 'He co-leads' to 'The author also co-leads' so the pronoun does not seem to point at Adelaide. | plain |
| I2-003 | Viant, a DSP, completed its acquisition of TVision, DV's only documented CTV attention data source, on 5 May 2026 without addressing third-party access. | Wrong completion date. Viant completed the TVision deal on 1 May 2026. 5 May is the date of Viant's closing press release. Every other mention in the paper, the TVision profile header and the events figure all say 1 May. The underlying claim's wording has the same error. | ATT-U-F08 passage (Viant 10-Q): 'On May 1, 2026, we completed our acquisition of TVision'; fig-consolidation event 2026-05-01; ch5 table '1 May 2026 \| Viant completes its purchase of TVision'. ATT-DOUBLEVERIFY-F13 final_wording repeats '5 May 2026' and needs the same fix. | contradicted | threats[0].t: '... on 5 May 2026 without addressing third-party access.' -> 'Viant, a DSP, completed its purchase of TVision, DV's only documented CTV attention data source, on 1 May 2026; its closing release and Q2 10-Q omit third-party access.' src += ATT-D-S02 (Viant Q2 2026 10-Q). | Holds. Viant's Q2 2026 10-Q (ATT-D-S02; ATT-D-F14, ATT-U-F09) dates completion to 1 May 2026. The register lists the 5 May release (ATT-E-S24 = ATT-DOUBLEVERIFY-S53) with event_date '2026-05-01 (completion per 10-Q)'. ATT-DOUBLEVERIFY-F13's final_wording already says 1 May, so the ledger needs no change. The suggested 'its filings do not address third-party access' was wider than the evidence. ATT-U-F08 covers only the closing release and the Q2 2026 10-Q, so the absence is limited to those two documents. ATT-D-S02 was added to src so the date has a source. | evidence/swot/doubleverify.json |
| I2-004 | Ownership change: the pending acquisition by Nielsen's parent (expected by end of Q4 2026, a forecast) would take DV private | Out of date and inconsistent. The end-of-Q4-2026 forecast comes from the original release, which a corrected release of 7 August 2026 replaced. Chapter 5, the DV SWOT and the claim ledger all give a target of closing by Q1 2027. | ATT-DOUBLEVERIFY-F17: 'DV targets closing by the first quarter of 2027, a forecast from the corrected release of 7 August 2026, repeated in the 10-Q, PREM14A and DEFM14A'; ch5 'DV targets closing by the first quarter of 2027'; DV SWOT 'closing is targeted by Q1 2027'. | contradicted | risks[0]: '...(closing targeted by the first quarter of 2027, a forecast from the corrected release of 7 August 2026)...' -> '...(closing targeted by the first quarter of 2027, a forecast from DV's corrected release of 7 August 2026)...' | Holds against ATT-DOUBLEVERIFY-F17 (DV targets closing by Q1 2027, a forecast from the corrected release of 7 August 2026), ch5 and the DV SWOT. The working file already carried the Q1 2027 wording, so the end-of-Q4-2026 text the reviewer quoted survived only in the stale built edition (dist). I finished the replacement by naming DV as the issuer of the corrected release. The cited sources ATT-V-DOUBLEVERIFY-S01 (corrected 8-K/A Exhibit 99.1) and S02 (DEFM14A) exist in the register and support the forecast. Rebuild dist to clear the stale text. | data/profiles.json (DoubleVerify, Risks) |
| I2-005 | Accreditation risk: about 23 months from MRC pre-assessment (October 2024) to a pre-audit review (May 2026) with no audit begun | The arithmetic is wrong and makes Adelaide look slower than it was. October 2024 to May 2026 is about 19 months, and chapter 2 says so. The 23 months in the evidence file runs from pre-assessment to 'Audit not begun' as read at 29 September 2026, which is a different interval. | ch2: 'MRC told Digiday in October 2024 that Adelaide was in pre-assessment. The pre-audit review came about 19 months later.'; ATT-ADELAIDE-F01; ATT-A-F14; evidence/profiles/adelaide.json ('about 23 months after it confirmed pre-assessment' for the audit-not-begun state). | contradicted | None this pass. Risks[0] already reads: "Accreditation risk: about 19 months from MRC pre-assessment (October 2024) to a pre-audit review (May 2026), with no audit begun at 29 September 2026; guarantee and billing uses rest on an unaccredited metric (ATT-ADELAIDE-S01, S36)." | Holds. October 2024 (Digiday, ATT-A-S13 / alias ATT-ADELAIDE-S36, 2024-10-29) to May 2026 (MRC 'Pre-audit reviewed 5/2026', ATT-A-S10 / alias ATT-ADELAIDE-S01) is about 19 months, matching ATT-ADELAIDE-F03 final_wording ('about 19 months later (derived), with the audit not begun at 29 September 2026') and chapter 2. The 23 months came from evidence/profiles/adelaide.json, which measures to the 29 September 2026 read date. data/profiles.json already carried the corrected text from an earlier pass, so no further edit was needed. Its wording ('with no audit begun at 29 September 2026') means the same as the proposed replacement ('and no audit begun by 29 September 2026') and stays closer to the F03 wording. Both cited source IDs resolve in the register, and the file parses. scripts/check.mjs reports no profile errors (its one FAIL, hero stat 'corrections logged' 161 vs 163, is outside this target). Outside this target: evidence/profiles/adelaide.json lines 184 and 287 still say 23 months, and dist/attention-measurement-2026.md line 1185 shows the old text until the next render. | data/profiles.json (Adelaide, Risks) |
| I2-006 | Its 27 January 2026 release says it 'secures a minimum attention score' on Adelaide's AU for a campaign across all screens, with one performance report. | This attributes to Vevo's release a statement it does not make. The Vevo SWOT, chapter 2 and chapter 6 all say the release never names the unit of the minimum, and that reading it as AU is the paper's inference. | Vevo SWOT weakness: 'The release never names the unit of the guaranteed minimum; that it is Adelaide AU is a strong inference, not a stated term'; ch6: 'It does not state the unit of the minimum; reading it as AU is an inference from the partnership'; ch2 likewise. | contradicted | what_sold: replaced "Its 27 January 2026 release says it 'secures a minimum attention score' on Adelaide's AU for a campaign across all screens, with one performance report." with "Its 27 January 2026 release, announcing a partnership with Adelaide, says it 'secures a minimum attention score' for a campaign across all screens, with one performance report. The release does not name the unit of the minimum; reading it as Adelaide's AU is this paper's inference." JSON parses; no tokens in profiles.json; scripts/check.mjs shows no failure tied to this file (its one FAIL, hero stat 'corrections logged' 161 vs 163, comes from corrections-ledger edits elsewhere). | Finding holds. Ledger ATT-VEVO-F05 says the release does not state the unit of the minimum and reading it as AU is by implication; ATT-VEVO-F04 quotes the release without a unit; ch6 and the Vevo SWOT agree. The suggested 'issued with Adelaide' was not used: the source register lists the release (ATT-C-S34) as issued by Vevo LLC, titled 'Vevo Partners With Adelaide to Launch Attention Guaranteed', so the fix says 'announcing a partnership with Adelaide' (matching F04 wording) and splits the inference into its own short sentence. | data/profiles.json (Vevo, What is sold) |
| I2-007 | IAS's June 2026 CESP answer most plausibly described its attention metrics as under MRC review, contrary to MRC's January record, which risks buyer trust. | This imputes a misstatement to IAS on the strength of the paper's own reading of which column a footnote marker belongs to in a third-party grid. It then adds a reputational verdict ('risks buyer trust'). The chapters never make this claim, and no other vendor's CESP answers are treated this way. | ATT-IAS-F17: 'Because Xpln.ai carries its own triple-asterisk marker, the **Under review note most plausibly attaches to IAS' (a layout inference); ch2 cites CESP only for XPLN's 'no' answer. | contradicted | threats[3].t: "IAS's June 2026 CESP answer most plausibly described its attention metrics as under MRC review, contrary to MRC's January record, which risks buyer trust." -> "CESP's June 2026 grid appears, by footnote layout, to mark IAS's MRC answer 'Under review', conflicting with MRC's January 2026 removal of IAS Attention Measurement from in-process status." | Finding holds. Ledger ATT-IAS-F17 final_wording rests the IAS attribution of the CESP '**Under review' note on a layout inference ('most plausibly attaches'). The CESP grid relays vendor answers, but no source shows IAS made a statement to buyers. 'Risks buyer trust' is an unsupported reputational verdict, and no chapter makes this claim. One correction to the finding's evidence: ch2 also cites CESP for the yes audit answers of Adelaide, DoubleVerify and IAS, not only for XPLN's no. The suggested replacement runs 39 words, over the 28-word SWOT limit in scripts/check.mjs, and reads as a bare record with no threat. I compressed it to 28 words. It keeps the hedge on the layout reading, takes out the imputed misstatement and the trust verdict, and states the conflict between the two public records, which ATT-IAS-F17 and ATT-IAS-F01 support. Sources are unchanged (ATT-IAS-S34 resolves to ATT-A-S14, the CESP grid; ATT-IAS-S01 resolves to ATT-A-S07, the MRC January to March 2026 decisions report), and both resolve through register aliases. Basis stays inference. check.mjs reports no SWOT warnings. Its only failure is an unrelated hero-stat count (161 vs 163). | evidence/swot/ias.json |
| I2-008 | SeenThis sells its June 2026 attention offer as "independently verified" by Lumen, whose metric the offer uses | The quotation attributed to SeenThis has no support in the cited claims. Neither ATT-E-F22 nor ATT-R2-F01 contains 'independently verified' in its final wording or passages. A WebFetch of the SeenThis release returned only metadata. The paper uses the quote to criticise SeenThis's use of 'independent'. | ATT-R2-F01 passages quote 'Available through attentive CPM...', 'available globally', 'An attentive outcome threshold is agreed before launch...', 'We run more. At no extra cost.', with no 'independently verified'; ATT-E-F22 is about who measures delivery. | contradicted | Deleted from the independence paragraph: 'SeenThis sells its June 2026 attention offer as "independently verified" by Lumen, whose metric the offer uses (ATT-E-F22) (ATT-R2-F01).' | Holds. No support passage in ATT-R2-F01 (ATT-NFA-S01 release body, ATT-NFA-S02 product page) or ATT-E-F22 contains 'independently verified'. The phrase appears only in ATT-R2-F01's limitations field, with no locator or quoted passage. The Lumen verify file says SeenThis 2026 used the label 'independent' but quotes no SeenThis passage. A ledger claim cannot be added from this target, so the sentence is deleted. The paragraph still makes its point with the supported Lumen, dentsu, Outbrain, Nativo and Adelaide examples. ATT-R2-F01 is still cited in chapters 6 and 12. Out of scope here: chapter 6 table row 63 says 'None found (verified by Lumen)', which rests on the same unsupported wording and should be checked by whoever edits chapter 6. | 04-vendor-evidence |
| I2-009 | diverge from MRC's record and can mislead buyers about what is accredited. | Saying DV's descriptions 'can mislead buyers' imputes misleading conduct. The record does not support it: DV's product-page qualifier ('only direct indexed impressions') is narrower than MRC's scope, and the 2023 date may be an announcement date. Chapter 2 and chapter 12 describe the gap neutrally ('For scope, the letter is the record'). | ATT-U-F18: 'MRC's own record is broader than that qualifier'; ATT-DOUBLEVERIFY-F22; ch12: 'DoubleVerify's own pages describe its scope differently from MRC's record.' | contradicted | risks[2]: '...diverge from MRC's record and can mislead buyers about what is accredited.' -> '...diverge from MRC's record, which is broader than the product-page qualifier and dates first accreditation to December 2022; buyers should read scope from MRC's letter (ATT-D-S08, ATT-U-S38, ATT-DOUBLEVERIFY-S41, ATT-A-S08, ATT-A-S05).' | Holds. ATT-DOUBLEVERIFY-F22 and ATT-U-F18 record the divergence but do not support the claim that DV misleads buyers. Ch2 ('For scope, the letter is the record') and ch12 describe the gap neutrally. MRC's record is broader than the product-page qualifier (U-F18) and dates accreditation to December 2022 (F22, ATT-A-S08). I applied the replacement and added source IDs so the line is traceable: the 10-K, the product page, the November 2024 release, the MRC Oct-Dec 2022 updates and the MRC April 2026 letter (ATT-A-S05). All of them exist in the source register. The text has no em dash. | data/profiles.json (DoubleVerify, Risks) |
| I2-010 | a Crossref author query shows no peer-reviewed journal article by the founder after 2013 (S34). | This makes an absence claim about a named individual's academic record from a single Crossref author query. Such queries miss name variants and co-authored work. The founder's publications are also not evidence about the product's validation, and no other vendor's founder is checked this way, so Amplified is held to a different standard. The reputational risk is high and the evidential value is nil. | ATT-AMPLIFIED-F14 (bounded to 'a Crossref author query returns ... only from 2012-2013'); no equivalent check for Adelaide, Lumen, TVision, xpln.ai or other founders in any profile. | contradicted | Replaced "No independent, commissioned-with-published-method, or replicated validation of the in-flight metric was found; a Crossref author query shows no peer-reviewed journal article by the founder after 2013 (S34)." with "No independent, commissioned-with-published-method, or replicated validation of the in-flight metric was found in the sources screened." Required follow-up (not done here): make the identical replacement in evidence/profiles/amplified.json capabilities.outcome_validation.observed, then rerun node build/build-matrix.mjs. | Finding holds. ATT-AMPLIFIED-F14 is bounded to what one Crossref author query returned (journal articles from 2012-2013, plus two Palgrave books and a 2024 book chapter), so the cell over-reads it as "no peer-reviewed journal article after 2013". A founder publication record is not evidence about the in-flight metric, and no other profile runs a founder check, so the clause applied an uneven standard. Dropped it. Added "in the sources screened" to bound the absence claim, matching the other profiles' Outcome validation cells. node scripts/check.mjs shows no marker or source failure (its one FAIL, the hero stat for corrections logged, 161 vs 163, is outside this file). NOTE: data/profiles.json is generated by build/build-matrix.mjs from evidence/profiles/amplified.json, so the next build will overwrite this edit unless the same change goes into evidence/profiles/amplified.json (capabilities.outcome_validation.observed, line ~298). The verify note evidence/verify/amplified.json (finding capabilities.outcome_validation, line ~209) quotes the removed clause and needs the same correction. This run was limited to data/profiles.json, so those two edits are deferred to the owner of the evidence files. | data/profiles.json (Amplified, Outcome validation cell) |
| I2-011 | Nielsen, which distributes AU through its Outcomes Marketplace, agreed in August 2026 to acquire DoubleVerify, the only vendor MRC lists as accredited for attention metrics. | Wrong party. DV agreed to be acquired by Neptune BidCo US Inc., the parent of Nielsen Company (US), an affiliate of Elliott and Brookfield funds. It was not Nielsen. Everywhere else the paper says 'the parent of Nielsen'. | ch5: 'The buyer is Neptune BidCo US Inc., the parent of Nielsen Company (US), LLC'; ch5 table; DV profile header 'acquisition by Nielsen's parent pending'. | contradicted | Threat now reads: "Nielsen's Outcomes Marketplace carries AU; Nielsen's parent agreed in August 2026 to buy DoubleVerify, sole holder of MRC attention accreditation, a deal pending at 29 September 2026." Added src ATT-D-S12 (8-K naming Neptune BidCo) and ATT-DOUBLEVERIFY-S13 (EDGAR submissions index, pending status per ATT-DOUBLEVERIFY-F17); kept ATT-V-ADELAIDE-S11, ATT-V-ADELAIDE-S12, ATT-ADELAIDE-S02. | Holds. The legal buyer is Neptune BidCo US Inc., the parent of Nielsen Company (US) (ATT-D-F12, ATT-DOUBLEVERIFY-F16, 8-K ATT-D-S12), and the rest of the paper says 'the parent of Nielsen' (ch0, ch5, DV profile and SWOT). Nielsen's own release headline does say 'Nielsen to Acquire DoubleVerify' (ATT-V-ADELAIDE-S12), so the old line followed the press-release shorthand rather than inventing a party, but it broke the paper's convention. Pending status at 29 September 2026 rests on ATT-DOUBLEVERIFY-F17 (and ATT-D-F12). Reworded to stay within the 28-word SWOT line limit. | evidence/swot/adelaide.json |
| I2-012 | while its outcome validation, accreditation and finances remain weak. | 'Finances remain weak' is the paper's own judgement on a private company's solvency. Lumen's directors state that no material uncertainty exists on going concern, and the accounts show no turnover from which to judge. A summary line like this is the most quotable sentence in the profile and carries commercial harm. | ch5: 'The directors state that no material uncertainty exists on going concern ... The accounts are unaudited and show no turnover, so none of this can be read as a change in demand'; ATT-LUMEN-F05/F06. | contradicted | summary: 'while its outcome validation, accreditation and finances remain weak.' -> 'while its outcome validation and accreditation remain weak. Its unaudited accounts show two loan covenant breaches at 31 December 2025, and the directors report no material going-concern uncertainty.' | Holds. 'Finances remain weak' was the paper's own solvency judgement. ATT-LUMEN-F05 records the directors' view that no material uncertainty exists on going concern, and ATT-LUMEN-F03 shows no turnover, so the record does not support it. The suggested replacement was not used as written: its 'widening loss' repeats the unqualified wording that I1-035 already corrected in chapter 5, where the rise is almost entirely a non-cash share-based charge by inference (ATT-LUMEN-F03/F04). The summary now gives only the record facts (unaudited accounts, two covenant breaches at 31 December 2025, directors' going-concern view per ATT-LUMEN-F05 / ATT-LUMEN-S06, which the Threats line cites). It is split into two short sentences. | evidence/swot/lumen.json |
| I2-013 | AU activation is live in several DSPs, including DV360, where Google's own Help Center documents the packaged 'Maximize Adelaide AU' custom-bidding algorithm from mid-July 2026. | The DV360 status conflicts across the paper. The SWOT gives a go-live date (mid-July 2026) that chapter 6 says is undocumented. The Adelaide profile's activation cell says the packaged integration is 'announced, not confirmed live'. Claim ATT-E-F12 says Google's announcement still read 'Expected mid July' on 28 September and that general availability is unconfirmed. ATT-ADELAIDE-F15 says it is live per the Help Center. The two claims need reconciling, and then one reading needs applying everywhere. | ch6: 'By 29 September 2026 its Help Center documented live steps for the Maximize Adelaide AU algorithm ... the go-live date is not documented'; profile Activation cell: 'so the packaged integration is announced, not confirmed live'; ATT-E-F12 vs ATT-ADELAIDE-F15. | supported with qualification | Same sentence change as I2-science-09; cross-file reconciliation listed in reason as deferred to the profile and claim-ledger owners. | Holds. The paper's settled reading is ATT-ADELAIDE-F15 (review_outcome 'contradicted', correction found the live operating page that ATT-E-F12's review did not), R2-01 and I1-041: live by 29 September 2026 per Google's Help Center, announced for mid-July 2026, go-live date undocumented. Chapter 6 (lines 8, 19, 76) and chapter 7 already use it. Applied that reading here (same rewrite as I2-science-09). 'Marketed as live' matches chapter 6 and the profile note that TTD and Amazon DSP live status rests on vendor or partner releases. Out of scope for this file, deferred: (1) data/profiles.json Adelaide activation cell 'observed' (line 73) still says the DV360 packaged integration is 'announced, not confirmed live'; change to 'Google's DV360 Help Center documented live activation steps for the Maximize Adelaide AU algorithm by 29 September 2026 (ATT-V-ADELAIDE-S03); announced for mid-July 2026; go-live date not documented'. (2) data/profiles.json Adelaide timeline event (lines 568-570) says 'general availability not confirmed at 29 September 2026'; same change. (3) ATT-E-F12 final_wording ('General availability is therefore unconfirmed by Google's own documentation') conflicts with ATT-ADELAIDE-F15; mark E-F12 superseded by F15 or append that the operating page documents live steps by 29 September 2026 with the go-live date undocumented. | evidence/swot/adelaide.json |
| I2-014 | Ownership uncertainty: Bain Capital and WPP split at the freeze not stated; WPP carries Kantar at nil; reports of break-up plans are unverified. | This contradicts chapter 5 and the ledger, which state the split from Kantar's audited 2025 accounts. Repeating 'reports of break-up plans' that the paper itself calls unverified is also a risky insinuation about a named company. | ATT-KANTAR-F22: 'the split at the freeze is stated in Kantar's audited Annual Report and Accounts 2025: Bain Capital indirectly 60%, WPP affiliates 40% ... no change reported in the H1 2026 results'; ch5: 'its audited 2025 accounts put its ownership at 60% Bain Capital and 40% WPP affiliates'. | supported with qualification | risks[1] (data/profiles.json line 2774), before: "Ownership change: Kantar's audited 2025 accounts put Bain Capital at 60% and WPP affiliates at 40%; WPP carries its stake at nil; reports of break-up plans are unverified (ATT-V-KANTAR-S01, ATT-KANTAR-S08)." After: "Ownership: Kantar's audited 2025 accounts put ownership at 60% Bain Capital and 40% WPP affiliates, with no change reported in H1 2026; WPP carries its stake at nil (ATT-V-KANTAR-S01, ATT-V-KANTAR-S02, ATT-KANTAR-S08)." Still needed to survive a rebuild: replace risks[1] in evidence/profiles/kantar.json (line 325) with the same 'after' text. | The finding holds. ATT-KANTAR-F22 final_wording gives the split from Kantar's audited Annual Report and Accounts 2025 (ATT-V-KANTAR-S01) as Bain Capital indirectly 60% and WPP affiliates 40%, with no change reported in the H1 2026 results (ATT-V-KANTAR-S02). Chapter 5 states the same split. The nil carrying value rests on WPP's FY2025 20-F (ATT-KANTAR-S08). When this editor read data/profiles.json, risks[1] had already been partly corrected to 'Ownership change: ... 60% ... 40%; WPP carries its stake at nil; reports of break-up plans are unverified (ATT-V-KANTAR-S01, ATT-KANTAR-S08)'. It still repeated the unverified break-up reports and used an 'Ownership change' label that no source supports. The entity status_note already says those reports are not relied on. The edit applies the reviewer's replacement and keeps parenthetical source ids, adding ATT-V-KANTAR-S02 for the H1 2026 no-change statement. Validation: data/profiles.json parses; all three source ids exist in source-register.json; scripts/check.mjs reports no profile failures. Its single FAIL is an unrelated hero-stat count, 'corrections logged' 161 vs 163. Caveat for the orchestrator: data/profiles.json is regenerated by build/build-matrix.mjs from evidence/profiles/kantar.json, and no data/profile-corrections.json overlay exists. evidence/profiles/kantar.json line 325 (risks[1]) still reads 'Ownership uncertainty: Bain Capital and WPP split at the freeze not stated; WPP carries Kantar at nil; reports of break-up plans are unverified.' The next matrix build will revert this edit unless the same text goes into that source record. Per instructions, that file was not edited. | data/profiles.json (Kantar, Risks) |
| I2-015 | the peer-reviewed MSI study found attention explained most variance in brand search lift, in-sample, without holdout. | 'Most variance' reads as a majority of variance, but attention's R-squared was .116, and the source says attention explains 'the most variation' of the four measures. With quarter-hour fixed effects the gap nearly closes. Chapter 3 says 'slightly more'. The cited item is also an MSI working paper; it became peer-reviewed only on publication in Marketing Science. The wording overstates in TVision's favour compared with how other vendors' evidence is treated. | ATT-TVISION-F18 passage: 'Attention explains the most variation in search lift'; R2 .116 vs .082, .070, .035 (.180 vs .176 with FE); ch3: 'attention explained slightly more brand-search lift ... With quarter-hour fixed effects the gap nearly closes'. | supported with qualification | Same edit as I2-science-10: strengths[3].t now reads 'Independent academics have used TVision panel data; an MSI working paper found attention explained slightly more brand-search lift than tuning, in-sample, and fixed effects nearly closed the gap.' | Holds. The figures were read from MSI Working Paper 21-132 (ATT-B-S06). The Marketing Science version (ATT-V-TVISION-S10) was read only as an abstract and its tables were not compared, so 'peer-reviewed MSI study' runs the two documents together. The fix names the working paper and uses 'slightly more' plus the fixed-effects qualifier. The suggestion's 'later published in Marketing Science' and the R-squared pair were left out to stay within the 28-word line limit. ATT-V-TVISION-S10 stays in src as the journal record of the same academic use of TVision data. 'Than tuning' holds because attention beat both tuning measures, .082 and .035. | evidence/swot/tvision.json |
| I2-016 | Two rest on something other than Adelaide’s AU: SeenThis on Lumen’s measure and A+E’s guarantee on TVision panel data. | The caption attributes Impact Media's suite to Adelaide by implication. Its metric vendor is 'not stated' in the figure data and chapter 6. Vevo's unit is also only inferred. As written, the caption overstates how many offers rest on AU. | data/figure-data.json#guarantees: Impact Media row 'not stated'; Vevo 'Adelaide*'; ch6 table: Impact Media 'Metric vendor: Not stated'. | supported with qualification | Same edit location as I2-buyer-11: replace the sentence 'Two rest on something other than Adelaide’s AU: SeenThis on Lumen’s measure and A+E’s guarantee on TVision panel data.' with 'SeenThis rests on Lumen’s measure, A+E’s guarantee on TVision panel data and Impact Media’s suite on a measure its release does not name; the rest use Adelaide’s AU, Vevo’s by inference.' | Finding holds. Impact Media's metric vendor is 'not stated' in figure-data.json and chapter 6 (ATT-R2-F02: the release does not say how 'seen' is measured), so 'Two rest on something other than Adelaide's AU' implies the suite rests on AU. Vevo's cell is 'Adelaide*' and ATT-VEVO-F05 says AU is an inference. The suggested wording changes the typed count from two to three; a count-free version is used instead so no new typed count of the dataset is added. Target is a figure, so no code edited. | fig-guarantees |
| I2-017 | In every guarantee found, the metric's vendor measures delivery; published remedies are extra delivery or conditional billing. | The decision tool, which sits in data/decision-tool.json and renders in chapter 12, still carries the wording that review R3 corrected. Vevo, Duolingo, Teads and A+E do not state who measures delivery. The chapter text says so, and the same chapter's prose contradicts this table cell. | ch6: 'Where a guarantee discloses who measures delivery, it is the vendor whose metric the offer uses ... For Vevo, Duolingo and Teads, Adelaide's role is implied but not stated, and for A+E's guarantee on TVision data the measurer is not stated'; ATT-E-F22; ch13 notes R3 'corrected how the paper described who measures guarantee delivery'. | supported with qualification | none | The quoted sentence is not in chapters/12-next.md. The chapter prose already uses the R3-corrected wording in the standfirst, the in-short and the section on who measures delivery, which say 'Where a guarantee says who measures delivery, it is the metric's own vendor; the others do not say' and name Vevo, Duolingo, Teads and A+E. The old sentence appears only in the stale built copies (dist/data/decision-tool.json, dist/attention-measurement-2026.md, build/page-data.json, all dated 29-30 Sep). The source data/decision-tool.json line 38 already reads 'Where a guarantee names who measures delivery, it is the metric's own vendor; Vevo, Duolingo, Teads and A+E do not say.' That matches ATT-E-F22. A rebuild picks up the fix. The data file is outside this target, and nothing in 12-next.md needed to change. | 12-next |
| I2-018 | In CESP's grid of 10 June 2026, it answered no to whether its attention solution had been audited by an independent third party | The CESP grid is reported selectively. The paper cites xpln.ai's 'no', which counts against the one vendor actually in MRC audit. It omits that Adelaide, DoubleVerify and IAS answered 'yes' to the same question, although only DV holds an accreditation. Readers get a worse impression of xpln.ai than the same source supports. | ATT-IAS-F17 passage (CESP row 'Has your Attention solution been audited by an independent third party?': Adelaide Yes, Adloox-Peer39 No, DoubleVerify Yes, IAS Yes, Xpln.ai No); ATT-XPLN-F19. | supported with qualification | Ch2 XPLN paragraph: after '...audited by an independent third party (ATT-XPLN-F19).' added 'Adelaide, DoubleVerify and IAS answered yes, and of these only DoubleVerify holds an MRC attention accreditation (ATT-U-F17) (ATT-A-F12) (ATT-A-S14).' | The finding holds. The CESP Audit row (ATT-U-F17 and ATT-IAS-F17 passages, source ATT-A-S14) reads Adelaide Yes, Adloox-Peer39 No, DoubleVerify Yes, IAS Yes, Xpln.ai No. Citing only xpln.ai's 'no' left out that the self-reported 'yes' answers came from vendors with no MRC attention accreditation except DoubleVerify (ATT-A-F12). Split into two sentences for voice. Used 'of these' rather than 'of the four' to avoid a count. Cited the grid directly, because no claim's final_wording itemises the Adelaide and IAS 'yes' answers. | 02-standards |
| I2-019 | What it lacks is public evidence that its numbers predict outcomes better than time in view, and any independent audit of whether the guarantees are met. | This absence claim is unbounded in the most visible places: the hero summary and dist subtitle. The OG description ('no independent audit of attention guarantees') and llms_txt have the same problem. The chapters bound it to what this review found in the sources screened. | ch6: 'this review found no independent audit of delivery'; ch13: 'Every absence claim is bounded to what was searched'. | supported with qualification | hero_summary second sentence now reads 'What it lacks, in the public record screened, is evidence that its numbers predict outcomes better than time in view, and any independent audit of whether its guarantees are met.' site.og.description and site.llms_txt now carry 'in the sources screened' before both absence clauses. | Holds. The hero summary, which is also the dist subtitle and the page description, made an unbounded absence claim. Chapter 6, chapter 12 and ATT-E-F22 bound it ('This review found no independent third-party audit of guarantee delivery in the sources screened'), and chapter 13 bounds every absence claim to what was searched. The OG description and llms_txt carried the same unbounded wording. jsonld.abstract was already bounded. The synopsis is bounded by its own 'on the public record' framing. | paper.config.json |
| I2-020 | Outcome framing conflicts with the IAB and MRC guidelines (to which IAS staff contributed), which say attention 'should not be considered or used as a measure of outcomes' and exclude unvalidated predictive models from scope. | IAS is held to a different standard. Adelaide's AU is framed the same way ('a placement's probability of attention and subsequent impact'), and Adelaide staff also contributed to the guidelines, yet Adelaide's profile carries no 'conflicts with the guidelines' verdict. Chapter 2 also says the guidelines allow use 'in conjunction with measures of outcomes', and that MRC has not judged either score. | ch2: 'Adelaide describes AU as "a placement's probability of attention and subsequent impact", and IAS says Quality Attention predicts ... The guidelines say attention should not be used as a measure of outcomes, while allowing its use "in conjunction with measures of outcomes" ... MRC has not judged IAS's score ... Adelaide's audit has not begun'; ATT-IAS-F18. | supported with qualification | risks[1] before: "Outcome framing conflicts with the IAB and MRC guidelines (to which IAS staff contributed), which say attention 'should not be considered or used as a measure of outcomes' and exclude unvalidated predictive models from scope." After: "Outcome framing: the IAB and MRC guidelines (to which IAS staff contributed) say attention 'should not be considered or used as a measure of outcomes' and place unvalidated predictive models outside their scope, while IAS markets Quality Attention as predicting business results (ATT-A-S01, ATT-A-S22). MRC has not assessed the score against those clauses because IAS left the process (ATT-A-S07). Adelaide's AU carries a similar outcome framing (ATT-ADELAIDE-S12)." | Finding holds. ATT-A-F17 names Adelaide staff among the guideline contributors as well as IAS staff. ATT-A-F18 and ATT-ADELAIDE-F04 show AU framed as 'a placement's probability of attention and subsequent impact'. Chapter 2 says MRC has not judged IAS's score against the outcome clauses (ATT-IAS-F18) and allows use 'in conjunction with measures of outcomes'. So 'conflicts with' was a verdict the paper does not reach, and Adelaide's profile carries no such verdict. Applied the proposed meaning but split it into short sentences and added source IDs for each fact: ATT-A-S01 (guidelines, contributors and scope), ATT-A-S22 (IAS's 'predict ... business result' page), ATT-A-S07 (MRC January-March 2026 decisions report, the canonical ID behind the ATT-IAS-S01 alias) and ATT-ADELAIDE-S12 (AU definition). All markers resolve in source-register.json. Pipeline note: data/profiles.json is written by build/build-matrix.mjs from evidence/profiles/ias.json and data/profile-overlay.json, and no data/profile-corrections.json exists. A rebuild will drop this edit, along with earlier direct edits such as IAS risks[0] and limits[2]. Carry the same text into evidence/profiles/ias.json risks[1], or a profile-corrections 'set' entry ias risks.1, before the next build-matrix run. | data/profiles.json (IAS, Risks) |
| I2-021 | Self-referential chain: IAS supplies the inputs, trains the score, sells the optimisation and reports the outcome, with no independent link documented (PPC Land makes the same criticism). | The same chain fits DoubleVerify: DV supplies the signals, builds the indices, sells Scibids optimisation and publishes the outcome cases, all of them DV-published. DV's profile carries no such risk, so IAS alone is singled out. | DV profile: 'Activation and optimisation from the same vendor: Universal Attention Segments and Scibids attention optimisation'; DV Outcome validation: 'All outcome evidence found reaches the public through DoubleVerify'; DV Performance evidence: Scibids '63% increase in attention levels' (vendor assertion). | supported with qualification | risks[5] before: "Self-referential chain: IAS supplies the inputs, trains the score, sells the optimisation and reports the outcome, with no independent link documented (PPC Land makes the same criticism)." After: "Self-referential chain: IAS supplies the inputs, trains the score, sells the optimisation and reports the outcome, with no independent link documented (PPC Land makes the same criticism). DoubleVerify sells a similar chain (its own signals and indices, Scibids optimisation, outcome cases published by DV), though MRC accreditation covers its post-bid display and video reporting (ATT-E-S28, ATT-C-S18, ATT-A-S05)." | Finding mostly holds. The DV profile documents a similar single-vendor chain: DV's own device signals and indices, Universal Attention Segments and Scibids optimisation (ATT-E-S28), and outcome evidence that reaches the public only through DV (outcome_validation, e.g. ATT-C-S18). The proposed 'same structure' overstates the match in two ways. DV's indices are not described as trained on outcomes, and DV's post-bid measurement carries MRC accreditation (ATT-A-S05), an independent link that IAS lacks. The comparator therefore says 'similar chain', qualifies it with the accreditation, and cites sources. All markers resolve. The PPC Land clause already in the risk is unchanged and out of scope. Pipeline note: same as I2-vendor-20. Carry this text into evidence/profiles/ias.json risks[5], or a profile-corrections entry ias risks.5, or a rebuild will drop it. | data/profiles.json (IAS, Risks) |
| I2-022 | selection toward favourable cases is likely but cannot be checked (inference). | 'Likely' imputes cherry-picking to Adelaide alone. Other vendors' case material (DV, IAS, Lumen, Amplified) is equally selected and gets no such inference. The evidence supports only that the selection method is undisclosed. | ch4 treats DV, IAS and Amplified cases as vendor assertions without a selection inference; Adelaide's Outcomes Guide releases disclose no case-selection method ([12, 13, 14]). | supported with qualification | 'The lift comparators are not stated, "37% cost savings" is undefined, and selection toward favourable cases is likely but cannot be checked (inference).' -> 'The lift comparators are not stated and "37% cost savings" is undefined. The releases do not say how cases were chosen, so selection toward favourable cases cannot be ruled out (inference), as for any vendor's published cases.' | Holds. Chapter 4 makes no selection inference for the DV, IAS, Lumen or Amplified cases, and the evidence shows only that the selection method is undisclosed (ATT-C-F05; releases ATT-C-S05 to S07). The fix keeps it marked as inference and bounds it to what the releases say. The long sentence is split in two to keep the voice. | 04-vendor-evidence |
| I2-023 | Earlier employers include Verve Group (Moments.AI), Hearts & Science and Lotame. | Correction I1-099 left two disclosure questions to the author, and the statement does not record a decision on either. First, whether Samba's AgenticAdvertising.org membership belongs here (AAO maintains AdCP, assessed in chapter 10). Second, whether a Teads link belongs here: the native-advertising paper disclosed one, and Teads appears in this paper's trading table and in the Lumen and Adelaide routes. The disclosure should be consistent across the series. | corrections-ledger I1-099: 'The native paper's interests statement disclosed the equivalent link (Samba TV and Teads) ... The author may also want to decide whether Samba's AgenticAdvertising.org membership and the Teads link (Teads is in the trading table here) belong in this statement too'. | supported with qualification | Inserted after '...passing a test-mode conformance storyboard.': 'The companion native-advertising paper disclosed two further links, which this paper's own sources do not cover: AgenticAdvertising.org listed Samba among its members, and Teads named Samba among its measurement-partner integrations. Teads is not scored here; its pilot and beta guarantees sit in the trading table in [chapter 6](#trading) (ATT-E-F07).' | Holds. I1-099 left two questions to the author, and the statement recorded no decision on either. The companion native-advertising paper's interests statement (native-2026/chapters/00-abstract.md) disclosed both links: Samba listed among AdCP's members, and Teads naming Samba among its measurement-partner integrations. Its sources are native J-F08 (AdCP member logos) and the Teads profile's measurement_support. This paper's register has no source for either link. I checked the source register: Samba appears only in the Adelaide and E-stream sources, and the AAO sources are registry and compliance endpoints. So I added the two links as a cross-reference to the companion paper's published disclosure, not as facts verified here, and said in the text that this paper's own sources do not cover them. Teads is in the product register as 'listed', 'Trading table (not scored)'. I anchored its role here with ATT-E-F07 (pilot and beta guarantees on Adelaide AU), the claim behind its trading-table row. The replacement was empty. I did not add any other disclosure; anything beyond the Samba-Teads commercial link stays the author's decision, as the native ledger also recorded. Validated: 3 claim markers and 23 stat tokens resolve, anchor #trading exists, no em dash added. The plain.md interests passage may need the same sentence for edition consistency (outside this target). | 13-method |
| I2-024 | Group loss rose from £1.13m (FY2024) to £3.76m (FY2025, as filed), £2.72m of it share-based payment; turnover is not filed. | An inference is tagged as record. The accounts show £2.72m of share-based payments credited to reserves. That all of it was charged to the FY2025 loss is the paper's inference, and chapter 5 flags it as one. The profile's Risks line ('Rising losses (record): ... with £2.72m of it share-based payment') has the same problem. | ATT-LUMEN-F03: 'share-based payments credited to reserves £2,720,974'; ATT-LUMEN-F04: 'if the whole £2,720,974 credit to reserves was charged to profit or loss in the year...'; ch5: 'If that whole credit was charged in the year ... by this paper's inference'. | supported with qualification | weaknesses[4].t: 'Group loss rose from £1.13m (FY2024) to £3.76m (FY2025, as filed), £2.72m of it share-based payment; turnover is not filed.' -> 'Group loss rose from £1.13m (FY2024) to £3.76m (FY2025, as filed); by inference, a £2.72m non-cash share-based payment charge explains almost all the rise. Turnover is not filed.' | Holds. ATT-LUMEN-F03 records £2,720,974 of share-based payments credited to reserves. That the whole amount was charged to the FY2025 loss is the inference in ATT-LUMEN-F04 (medium confidence), and the line had stated it as record. Rewritten so the record figures stand alone and the inference is labelled in the text, matching chapter 5's 'almost entirely ... by this paper's inference'. 'Almost all the rise' replaces the suggested 'most of the rise' because the share-based charge grew by £2.64m against a £2.63m rise in the loss (F03 prior-year £78,592). Kept at 28 words with source ATT-LUMEN-S04 (alias of ATT-D-S27, which carries F03/F04). Out of scope: the profiles.json Risks line ('with £2.72m of it share-based payment') has the same defect and needs the same fix in that file. | evidence/swot/lumen.json |
| I2-025 | Lumen's filed accounts show two loan covenant breaches and a wider loss, which by this paper's inference came almost entirely from a non-cash share-based charge | This chapter summary states adverse facts about a private company without the date they refer to, and without the directors' going-concern statement that the chapter body includes. 'Almost entirely' also fits the widening of the loss, not the loss itself: share-based payments are about 72% of it. | ch5 body: breaches in 'amended accounts for the three months to 31 December 2025 ... The directors state that no material uncertainty exists on going concern'; ATT-LUMEN-F04 ('about 72% of the FY2025 group loss'). | supported with qualification | Summary paragraph: replaced "Lumen's filed accounts show two loan covenant breaches and a wider loss, which by this paper's inference came almost entirely from a non-cash share-based charge (ATT-LUMEN-F03) (ATT-LUMEN-F04) (ATT-LUMEN-F05)." with "Lumen's unaudited accounts show two loan covenant breaches at 31 December 2025, and the directors state that no material uncertainty exists on going concern (ATT-LUMEN-F05). Its loss widened in the year to September 2025, almost entirely through a non-cash share-based charge by this paper's inference (ATT-LUMEN-F03) (ATT-LUMEN-F04)." | Holds. ATT-LUMEN-F05 dates both breaches to the 31 December 2025 reporting date and records the directors' no-material-uncertainty view. ATT-LUMEN-F03/F04 put share-based payments at about 72% of the FY2025 loss. The increase in the share-based charge (2,642,382) roughly equals the increase in the loss (2,627,034), so 'almost entirely' fits the widening, not the loss itself. Kept the reviewer's meaning but split it into two shorter sentences, so each claim marker sits on the fact it supports. | 05-market |
| I2-026 | No advertiser, campaign, impression count or spend under the guarantee is disclosed, and no product documentation page for the offer exists. | This is an unbounded absence claim, tagged vendor_doc. The Vevo profile's Limits bounds the same point to the pages screened. | Vevo profile Limits: 'No product documentation page exists in the pages screened'; ch13: 'Every absence claim is bounded to what was searched'. | supported with qualification | weaknesses[4].t: '...and no product documentation page for the offer exists.' -> '...and no product documentation page for the offer was found in the pages screened.' | Finding holds. The SWOT weakness stated an unbounded absence ('no product documentation page for the offer exists'), while the Vevo profile bounds the same point to the pages screened (limits: 'No product documentation page exists in the pages screened'; what_sold: 'no product documentation page for the offer was found'). Chapter 13 requires every absence claim to be bounded to what was searched. Source ids ATT-VEVO-S01, S03 and S05 all exist in the source register; basis and src left unchanged. | evidence/swot/vevo.json |
| I2-027 | about 23 months from MRC pre-assessment (October 2024) to a pre-audit review (May 2026) | Wrong interval in the Adelaide profile's Risks field. October 2024 to May 2026 is about 19 months, which is what chapter 2 says. | ATT-ADELAIDE-F03 final_wording: 'MRC's record shows the pre-audit reviewed in May 2026, about 19 months later (derived)'. Chapter 2: 'The pre-audit review came about 19 months later.' The same text is in evidence/profiles/adelaide.json. | contradicted | See I2-science-01. | Duplicate of I2-science-01, applied in the same edit. | data/profiles.json |
| I2-028 | on 5 May 2026 without addressing third-party access | The DoubleVerify SWOT threat dates Viant's completion of the TVision deal to 5 May 2026. Everywhere else the paper says 1 May 2026. 5 May is the date of Viant's closing press release, not the completion date. | ATT-D-F14: 'completed the deal on 1 May 2026'. ATT-U-F09: 'Viant's 10-Q records completion ... on 1 May 2026'. ATT-E-F17 calls 5 May the 'closing release'. The abstract, chapters 5 and 8 and the TVision profile header all give 1 May 2026. | contradicted | Covered by the threats[0] edit under I2-vendor-03 ('5 May 2026' -> '1 May 2026'). | Same error as I2-vendor-03, and it holds for the same reasons. The I2-vendor-03 edit fixes it: the threat now reads 1 May 2026, as the 10-Q, the abstract, chapters 5 and 8 and the TVision profile do. | evidence/swot/doubleverify.json |
| I2-029 | Viant completed its acquisition of TVision on 5 May 2026 without addressing third-party access. | The final_wording of ATT-DOUBLEVERIFY-F13 gives the wrong completion date. The claim is cited in chapters 1, 2 and 8, so the sentence prints in the published claims table in chapter 13, where it contradicts the abstract and chapter 5. | ATT-D-F14 and ATT-U-F09 (Viant 10-Q) give completion on 1 May 2026. The 5 May release is the closing announcement (ATT-E-F17). | contradicted | final_wording: 'Viant completed its acquisition of TVision on 5 May 2026 without addressing third-party access.' -> 'Viant completed its acquisition of TVision on 1 May 2026, and its closing release of 5 May 2026 does not address third-party access.' source_ids += ATT-D-S02; review_note gets an I2 note. | Holds. Viant's 10-Q (ATT-D-S02, carried by ATT-D-F14 and ATT-U-F09) gives completion on 1 May 2026. 5 May 2026 is the date of the closing release (ATT-E-S24, whose register event_date reads '2026-05-01 (completion per 10-Q)'). The sentence was reworded to tie 'does not address third-party access' to the closing release, the document actually read, which bounds the absence claim. ATT-D-S02 was added to source_ids to support the 1 May date. Carry-over: needs the same change in evidence/research (doubleverify profile) or the overlay to survive build:all. | data/claim-ledger.json |
| I2-030 | (expected by end of Q4 2026, a forecast) | The DoubleVerify profile's Risks field carries a stale closing forecast. Chapter 5 and the DV SWOT threat both say closing is targeted by Q1 2027, set in the corrected release of 7 August 2026. | ATT-DOUBLEVERIFY-F17: 'DV targets closing by the first quarter of 2027, a forecast from the corrected release of 7 August 2026, repeated in the 10-Q, PREM14A and DEFM14A.' Chapter 5 and the DV SWOT ('closing is targeted by Q1 2027') agree. | contradicted | See I2-science-02. | Duplicate of I2-science-02, applied in the same edit. | data/profiles.json |
| I2-031 | Bain Capital and WPP split at the freeze not stated | The Kantar profile's Risks field says the ownership split is not stated. Chapter 5 and the Kantar SWOT both state it (60% Bain Capital, 40% WPP affiliates) from Kantar's audited 2025 accounts. The product register's Kantar 'owner' fields also say '(split at the freeze not stated)'. | ATT-KANTAR-F22: 'the split at the freeze is stated in Kantar's audited Annual Report and Accounts 2025: Bain Capital indirectly 60%, WPP affiliates 40%'. Chapter 5: 'its audited 2025 accounts put its ownership at 60% Bain Capital and 40% WPP affiliates'. Kantar SWOT: 'WPP carries its 40% stake at nil'. data/product-register.json has the stale wording three times. | contradicted | kantar.risks[1] -> "Ownership change: Kantar's audited 2025 accounts put Bain Capital at 60% and WPP affiliates at 40%; WPP carries its stake at nil; reports of break-up plans are unverified (ATT-V-KANTAR-S01, ATT-KANTAR-S08)." kantar.entity.owner_or_parent: 'The percentage split at the freeze is not stated in the sources read.' -> "Kantar's audited Annual Report and Accounts 2025 states the split at the freeze: Bain Capital indirectly 60%, WPP affiliates 40% (ATT-V-KANTAR-S01)." | Holds. ATT-KANTAR-F22 final_wording says the split at the freeze is stated in Kantar's audited Annual Report and Accounts 2025: Bain Capital indirectly 60%, WPP affiliates 40%. That source is ATT-V-KANTAR-S01 in the register. Chapter 5 and the Kantar SWOT agree. The wording differs slightly from the proposal, and source ids were added. The Kantar entity.owner_or_parent in this file had the same stale sentence and was corrected. data/product-register.json (3 rows) still carries the stale wording; that is outside this target. | data/profiles.json |
| I2-032 | giving it a trading role no other specialist unit shows in these sources | The Adelaide profile's Differentiation field is stale. Since the SeenThis addition (R2), the paper documents another specialist's metric with a trading role: SeenThis Attention guarantees and prices media on Lumen's measure. The sentence is a comparative claim about named competitors and is now false. | Chapter 6 trading table: 'SeenThis Attention (aCPM, CPAS) \| SeenThis \| Lumen \| ... Yes: extra delivery at no cost'. Chapter 6: 'Two models bill on attention. SeenThis prices on aCPM or CPAS.' Appendix Q1: an unaccredited metric 'prices media, as Lumen's does in SeenThis Attention'. | contradicted | See I2-science-11. | Holds: the 'no other specialist unit' sentence is false since SeenThis Attention (Lumen) was added. Fixed by the I2-science-11 rewrite, which names both exceptions (SeenThis/Lumen and A+E/TVision). The proposed 'widest trading role of any specialist unit' was not adopted: it is an unsupported comparative, and it leaves out A+E/TVision. | data/profiles.json |
| I2-033 | so the packaged integration is announced, not confirmed live | The Activation row of the Adelaide profile table (printed in chapter 11) says the DV360 algorithm is only announced. That contradicts chapter 6, chapter 5, the matrix cell and the corrected claim, which all say Google's Help Center documents it as live. It understates a named vendor's live status. | ATT-ADELAIDE-F15 (corrected): 'by 29 September 2026 Google's DV360 Help Center documents the live activation steps for the Maximize Adelaide AU algorithm ... The packaged integration is live per platform documentation; only the exact go-live date is undocumented.' The capability-matrix activation 'why' for Adelaide: 'Only DV360 is documented by the platform itself (Google Help Center activation steps...)'. Chapter 6: 'DV360, whose Help Center documents a live Adelaide algorithm with no go-live date'. | contradicted | See I2-science-08. | Duplicate of I2-science-08, applied in the same edit. | data/profiles.json |
| I2-034 | Xaxis Switzerland (ATT-F-F17) and dentsu in the UK in 2022 (ATT-F-F11) | The chapter summary reads as if Xaxis Switzerland announced optimisation in 2022. The cited claim, the chapter's own table and its body text all say February 2023. | ATT-F-F17: 'was reported in February 2023 to use Adelaide data in Xaxis Switzerland for a customised attention algorithm'. Chapter 7 table: 'optimisation (Xaxis Switzerland, reported 2023)'. Body text: 'Xaxis Switzerland ... was reported in February 2023'. | supported with qualification | Same edit as I2-science-23. | Holds, and it is the same sentence as I2-science-23. The I2-science-23 edit fixes it: Xaxis Switzerland is now 'reported in February 2023', which matches F17, the chapter table and the body text. I did not use this finding's single-sentence wording, which mixes 'announced' and 'reported' in one list. | 07-buyers |
| I2-035 | the one buyer-side guarantee outside chapter 6's table, A+E's on TVision data, states no threshold or remedy, so under the rule's unresolved clause it would code "unresolved", as chapter 6 codes seller guarantees with unpublished terms | Left over from an earlier draft. Chapter 6's trading table and its Q1 table now include A+E's guarantee and already code it 'Unresolved', so it is not 'outside chapter 6's table' and the conditional 'would code' no longer applies. | Chapter 6 trading table row: 'Attention guarantee on a Budweiser buy \| A+E, for dentsu and AB InBev \| TVision'. Q1 table row: 'TVision data: A+E inventory for Budweiser ... \| Unresolved'. Chapter 6 count: 'six code "diagnostic" and twelve "unresolved"' includes it. | supported with qualification | Q1 paragraph: 'In this paper's reading, the one buyer-side guarantee outside chapter 6's table, A+E's on TVision data, states no threshold or remedy, so under the rule's unresolved clause it would code "unresolved", as chapter 6 codes seller guarantees with unpublished terms (ATT-F-F14).' -> 'The only executed guarantee found in buyer-side sources, A+E's on TVision data, states no threshold or remedy, and chapter 6 codes it "unresolved" under the rule's unresolved clause (ATT-F-F14).' | Holds. Chapter 6's trading table (A+E for dentsu and AB InBev, TVision) and its Q1 table ('TVision data: A+E inventory for Budweiser ... \| Unresolved') already include A+E's guarantee and code it, so 'outside chapter 6's table' and 'would code' are out of date. Q1's rule_unresolved in hypotheses.json covers unreadable trading terms. I changed 'the one buyer-side guarantee' to 'The only executed guarantee found in buyer-side sources', the summary paragraph's wording. Kargo's guarantee to MediaCom was also made to a buyer, but it comes from a vendor case study. I dropped 'In this paper's reading' because the coding is now attributed to chapter 6. | 07-buyers |
| I2-036 | In every guarantee found, the metric's vendor measures delivery; published remedies are extra delivery or conditional billing. | This is the stale 'every guarantee measured by the vendor' wording that the third outside review corrected. Vevo, Duolingo, Teads and A+E do not state who measures. 'Conditional billing' also contradicts chapter 12's own 'Every published remedy is more media' and the figure data, which code Impact Media's remedy as n/a. | Chapter 12: 'Where a guarantee says who measures delivery, it is the vendor whose metric defines it ... Vevo, Duolingo, Teads and A+E do not state who measures'. ATT-E-F22. figure-data guarantees: remedy 'yes' only for WSJ, Kargo and SeenThis; Impact Media 'n/a'. Chapter 12: 'Every published remedy is more media.' | supported with qualification | None in this pass. The equivalent fix is already in place: the 'fails' field of the 'Is an attention guarantee enforceable?' row, with src extended to ATT-E-S07, ATT-E-S10, ATT-E-S02 and ATT-F-S25. | Holds, and it is already fixed in the working copy. The 12-next editor applied the same fix earlier in this pass as I2-science-04 (log applied-12-next__data_decision-tool_json_.json). The quoted sentence is no longer in the file. The current 'fails' text in the guarantee row reads 'Where a guarantee names who measures delivery, it is the metric's own vendor; Vevo, Duolingo, Teads and A+E do not say. Every published remedy is extra delivery.' This means the same as the proposed replacement. It matches ATT-E-F22, chapter 12 ('Every published remedy is more media') and the figure data, where Impact Media's remedy is coded n/a. The stale wording survives only in dist/attention-measurement-2026.md, which is a build output and will refresh on rebuild. | data/decision-tool.json |
| I2-037 | three state a remedy, each extra delivery or conditional billing. Two rest on something other than Adelaide’s AU: SeenThis on Lumen’s measure and A+E’s guarantee on TVision panel data. | The caption disagrees with its own data and with the chapter 6 table. The three remedies (WSJ, Kargo, SeenThis) are all extra delivery; Impact Media's remedy is coded n/a, not conditional billing. Three offers, not two, rest on something other than AU: Impact Media's metric vendor is 'not stated'. | data/figure-data.json#guarantees: Impact Media suite cells ['not stated','yes','n/a',...]. Remedy 'yes' rows: WSJ, Kargo, SeenThis. Chapter 6 table: 'Impact Media \| Not stated \| ... \| Not applicable: bills only when seen'. Chapter 12: 'Every published remedy is more media.' | supported with qualification | In scripts/paper-figures.mjs line 121, set the fig-guarantees caption to: 'Every attention guarantee or attention-priced offer found, by what its public terms state. Two publish a numeric threshold (Kargo’s campaign, at an average of 25 AU, and Impact Media’s seconds seen); three state a remedy, each a form of extra delivery. SeenThis rests on Lumen’s measure, A+E’s guarantee on TVision panel data and Impact Media’s suite on a measure its release does not name; the rest use Adelaide’s AU, Vevo’s by inference.' Then regenerate figures/fig-guarantees.json (and the dist/package copies) through the figure build. This one caption covers all three findings. | Finding holds on both points (the remedy wording, as in I2-science-13, and the vendor attribution, as in I2-vendor-16). The count is correct: seven rows read Adelaide, with Vevo's starred. The suggested 'Seven rest on...' still adds a typed count of the paper's own dataset, so the same meaning is written without the number. Every fact traces to the figure's listed sources: ATT-NFA-S03 for Impact Media, ATT-E-S01 for Vevo, ATT-NFA-S01/S02 for SeenThis, ATT-F-S25/S23 for A+E. Target is a figure, so no code edited. | fig-guarantees |
| I2-038 | Attention measurement now has a final standard, one accredited metric and a documented trade in guarantees whose scale nobody discloses. | The hero summary, also printed as the paper's subtitle, describes the trade as a 'trade in guarantees'. Chapter 6 finds the trade is mainly targeting and bidding, with guarantees and priced offers few, and its undisclosed scale applies to all of it. 'One accredited metric' also departs from the abstract and fig-accreditation, which say one accredited product (a suite of metrics). | Chapter 6 summary: 'attention is traded in the US mainly as a targeting and bidding input inside buying platforms; guarantees and priced offers are few'. Abstract In short: 'a documented trade whose scale is not disclosed'. fig-accreditation title: 'One attention product is accredited'. Chapter 2 lists several accredited metrics (display and video suite plus V4). | supported with qualification | hero_summary first sentence now reads 'Attention measurement now has a final standard, one accredited attention product and a documented trade, mostly as a bidding input, whose scale is not disclosed.' 'one (MRC-)accredited attention metric' became 'one (MRC-)accredited attention product' in og.description, llms_txt, publications.synopsis and jsonld.abstract. | Holds. Chapter 6's standfirst says attention is traded 'mainly as a targeting and bidding input inside buying platforms; guarantees and priced offers are few', so 'trade in guarantees' misdescribes it. ATT-A-F12, the abstract, chapter 2 and fig-accreditation all say one accredited attention product. MRC accredits a suite of metrics (display and video Authentic Attention plus the V4 metrics), so 'one accredited metric' is inaccurate. Modified wording: 'mostly as a bidding input' in place of 'mostly in bidding', which follows the abstract's 'marketed as a bidding input'. 'whose scale nobody discloses' became 'whose scale is not disclosed', the abstract's wording, so it is no longer phrased as a universal. For consistency, the same metric-to-product fix went into site.og.description, site.llms_txt, site.publications.synopsis and site.jsonld.abstract. | paper.config.json |
| I2-039 | None was found using attention as the unit it pays for, and no deal terms were found in public. | Overstated. Chapter 7 documents one agency guarantee with a public threshold and remedy (Kargo for MediaCom, 2022), drawn from the vendor's case study. What chapter 7 says is that no buyer-side source states a contract term. | Chapter 7 summary: 'the one guarantee an agency received with a public threshold and remedy, Kargo's for MediaCom in 2022, comes from the metric vendor's case study'. 'No buyer-side source screened states who pays or a contract term.' | supported with qualification | 'None was found using attention as the unit it pays for, and no deal terms were found in public.' -> 'None was found using attention as the unit it pays for. No deal terms were found in buyers' own sources. In 2022 a seller made one agency, MediaCom, a promise with a set level and a fix if missed. The source for it is the scoring firm's own case study.' | Holds. Chapter 7 says no buyer-side source states a contract term. It also documents Kargo's 2022 guarantee to MediaCom, with a public threshold (25 AU) and a remedy, and that record comes from Adelaide's case study (ATT-E-F03, ATT-F-F28). 'No deal terms were found in public' also clashed with plain section 6, which reports offers with levels and fixes. I split the replacement into short sentences and bounded the absence to buyers' own sources. | plain |
| I2-040 | The paper's searches found no case against an attention firm. | Wrong as written. The paper documents Nielsen's patent actions against TVision, which are cases against an attention firm. What the searches found none of is privacy or biometric-data actions. | Chapter 9: 'this review found no public privacy or biometric-data regulator action or lawsuit naming a seed attention-measurement vendor ... Nielsen has brought patent actions against TVision since 2021'. Chapter 5: 'Viant took on Nielsen's pending patent litigation against TVision'. | supported with qualification | 'The paper's searches found no case against an attention firm.' -> 'The paper's searches found no privacy lawsuit or regulator action against an attention firm.' | Holds. Chapter 9 bounds the finding to privacy or biometric-data regulator action or lawsuit (ATT-H-F30), and documents Nielsen's patent actions against TVision (ATT-TVISION-F21). I added 'regulator action' to match chapter 9's scope. The next sentence, 'Not every firm was searched by name', stays as the bound. | plain |
| I2-041 | Their changes were checked against sources before being applied and are logged in the corrections ledger. | Process disclosure that does not match the ledger. The ledger holds 117 entries from the integrity pass (I1) and none from the adversarial pass, so the 161 total in the abstract, hero stats and chapter 13 does not include it. Either log the adversarial-pass changes and recompute the counts, or say the pass made no logged changes. | data/corrections-ledger.json id prefixes: R2 6, R3 19, I1 117, SC 19 (total 161). No entry's review_pass names an adversarial pass. stats.json corrections = 161. | supported with qualification | None in chapters/13-method.md. Required elsewhere: log the review2 decisions in data/corrections-ledger.json under an adversarial-pass review_pass label, then rerun stats.mjs and the build. | Holds as a point-in-time mismatch. data/corrections-ledger.json has R2 6, R3 19, I1 117, SC 19 and now AJ 2 (163 entries). No entry's review_pass names the adversarial pass, which is this review round (I2 findings). Its decision logs exist in review2/applied-*.json, but no one has written them into the ledger yet. Round 1 set the pattern: log-r1.mjs wrote the integrity-pass decision logs into the ledger after the editors finished. The sentence becomes true once the lead author does the same for review2, with a review_pass label such as 'Internal review pass 2 (adversarial)'. That is a data-file edit outside this target. Saying the pass 'made no logged changes' would be false, because its changes are being applied now. Chapter 13 has no typed counts. Every count is a [corrections] or [corrections_scores] token, so the counts recompute when stats.mjs reruns. Note for the lead: stats.json still says corrections 161 against 163 in the ledger, and check.mjs fails 'hero stat corrections logged says 161 but data gives 163'. Separately, the AJ-01/AJ-02 entries (blind second scoring pass, data/second-pass.json) contradict this chapter's line 'No row was scored twice independently, so no inter-rater statistic exists'. The In short line 'Those reviews and a consistency audit of the scores produced the ... entries' also leaves that pass out. Neither is part of this finding, so I left both to the lead author. | 13-method |
| I2-042 | Accreditation risk: about 23 months from MRC pre-assessment (October 2024) to a pre-audit review (May 2026) with no audit begun | Wrong arithmetic in the Adelaide profile's Risks, rendered in chapter 11. October 2024 to May 2026 is about 19 months, not 23. | ATT-ADELAIDE-F03 final_wording: 'MRC's record shows the pre-audit reviewed in May 2026, about 19 months later (derived)'. Chapter 2 says: 'The pre-audit review came about 19 months later.' | contradicted | adelaide.risks[0]: 'about 23 months from MRC pre-assessment (October 2024) to a pre-audit review (May 2026) with no audit begun;' -> 'about 19 months from MRC pre-assessment (October 2024) to a pre-audit review (May 2026), with no audit begun at 29 September 2026;' (citation (ATT-ADELAIDE-S01, S36) kept; these alias to ATT-A-S10 and ATT-A-S13). | Holds. October 2024 to May 2026 is 19 months. ATT-ADELAIDE-F03 final_wording and its derivation say about 19 months, and chapter 2 agrees. F03 also supports 'audit not begun at 29 September 2026'. | data/profiles.json |
| I2-043 | Ownership change: the pending acquisition by Nielsen's parent (expected by end of Q4 2026, a forecast) | The DoubleVerify profile Risks bullet, rendered in chapter 11, gives a closing forecast that has been superseded. It contradicts chapter 5 and the DV SWOT. | ATT-DOUBLEVERIFY-F17 final_wording: 'DV targets closing by the first quarter of 2027, a forecast from the corrected release of 7 August 2026'. Chapter 5: 'closing targeted by the first quarter of 2027'. evidence/swot/doubleverify.json: 'closing is targeted by Q1 2027'. | contradicted | doubleverify.risks[0] -> "Ownership change: the pending acquisition by Nielsen's parent (closing targeted by the first quarter of 2027, a forecast from the corrected release of 7 August 2026) would take DV private and end SEC reporting, removing the main public financial series; product strategy under Nielsen is not documented (ATT-V-DOUBLEVERIFY-S01, S02)." status_note and entity.status_note: first two sentences -> 'At 29 September 2026 the merger was pending: DV filed its definitive proxy (DEFM14A) that day for a special meeting on 29 October 2026, the HSR waiting period expired on 21 September 2026, and EDGAR shows no closing 8-K. DV targets closing by the first quarter of 2027 (a forecast from the corrected release of 7 August 2026), subject to stockholder approval and specified foreign antitrust or investment clearances; ...' (rest unchanged). timeline 2026-09-29: 'merger pending, no definitive proxy' -> 'merger pending; definitive proxy (DEFM14A) filed for a 29 October 2026 special meeting', src + ATT-V-DOUBLEVERIFY-S02. | Holds. ATT-DOUBLEVERIFY-F17 is 'contradicted' and its correction says DV targets closing by the first quarter of 2027, a forecast from the corrected release of 7 August 2026. Chapter 5 and the DV SWOT agree. The same stale forecast and a stale 'no definitive proxy' also appeared in the DV status_note, entity.status_note and the 2026-09-29 timeline row, so those were corrected too (same claim; DEFM14A filed 29 September 2026 for a 29 October meeting; HSR expired 21 September 2026). | data/profiles.json |
| I2-044 | At 29 September 2026 the deal was pending, with a preliminary proxy (PREM14A, 11 September 2026) and no definitive proxy filed. | The final_wording of ATT-D-F12, shown in the chapter 13 claims table, says no definitive proxy had been filed at 29 September. This contradicts ATT-DOUBLEVERIFY-F17 and chapter 5, which record the definitive proxy filed that day. | ATT-DOUBLEVERIFY-F17 review note: 'EDGAR lists a DEFM14A filed 29 September 2026'. Its final_wording: 'DV filed its definitive proxy (DEFM14A) on 29 September 2026 for a special meeting on 29 October 2026, and the HSR waiting period expired on 21 September 2026.' | contradicted | final_wording last sentence now: 'At 29 September 2026 the deal was pending. DV filed its definitive proxy (DEFM14A) that day for a special meeting on 29 October 2026. The HSR waiting period had expired on 21 September 2026, and DV targets closing by the first quarter of 2027.' source_ids += ATT-V-DOUBLEVERIFY-S02; support and passage += two DEFM14A quotes; limitations: closing forecast now 'by the first quarter of 2027, per the corrected release of 7 August 2026 and the DEFM14A'; review_note gets an I2 note. | Holds. ATT-D-F12 final_wording said no definitive proxy had been filed at 29 September 2026. The DEFM14A (ATT-V-DOUBLEVERIFY-S02, filed 29 September 2026) sets a special meeting for 29 October 2026, records HSR expiry on 21 September 2026 and says DV is 'currently targeting to consummate the Merger by the first quarter of 2027', matching ATT-DOUBLEVERIFY-F17. Applied with the same meaning, split into short sentences, and closing attributed to DV as a target. ATT-D-F12 cited no source for these facts, so ATT-V-DOUBLEVERIFY-S02 was added to source_ids with two verbatim support passages. The limitations field still gave 'expected closing (end of Q4 2026)', which now contradicted the new wording, so it was updated to the Q1 2027 target. Carry-over: build:all regenerates the ledger from evidence files plus data/claims-overlay.json, so this edit needs a claims-overlay final_wording entry for ATT-D-F12 (and the source in evidence/research/D.json) to survive a full rebuild. | data/claim-ledger.json |
| I2-045 | In every guarantee found, the metric's vendor measures delivery; published remedies are extra delivery or conditional billing. | The decision-tool row brings back a statement that correction R3-03 removed. Four guarantees do not say who measures delivery. Impact Media's conditional billing is coded as no remedy, so 'conditional billing' is not a published remedy. | Corrections ledger R3-03: 'In every guarantee the metric's own vendor measures delivery is unsupported where the measurer is not disclosed.' Chapter 6: 'For Vevo, Duolingo and Teads, Adelaide's role is implied but not stated, and for A+E's guarantee on TVision data the measurer is not stated.' The chapter 6 table codes Impact Media's remedy 'Not applicable'. Chapter 12: 'Every published remedy is more media.' | contradicted | Row 'Is an attention guarantee enforceable?', field 'fails': replaced 'In every guarantee found, the metric's vendor measures delivery; published remedies are extra delivery or conditional billing.' with 'Where a guarantee names who measures delivery, it is the metric's own vendor; Vevo, Duolingo, Teads and A+E do not say. Every published remedy is extra delivery.' Added sources for the offers now named to src: ATT-E-S07 (TripleLift), ATT-E-S10 (Duolingo), ATT-E-S02 (Teads), ATT-F-S25 (A+E). Vevo was already cited as ATT-E-S01, an alias of ATT-C-S34. check.mjs reports no unknown sources. | Holds. ATT-E-F22 final_wording says the measurer is disclosed only for WSJ, Kargo, TripleLift (Adelaide) and SeenThis (Lumen). For Vevo, Duolingo and Teads it is implied but not stated, and for A+E it is not stated. Correction R3-03 removed the 'every guarantee' wording. Chapter 6 codes Impact Media's remedy 'Not applicable: bills only when seen', so conditional billing is not a published remedy. The published remedies (Kargo added-value PMP impressions, ATT-E-F03; WSJ makegood of added-value impressions, ATT-E-F02; SeenThis extra delivery at no cost, ATT-R2-F01) are all extra delivery, which matches chapter 12's 'Every published remedy is more media'. | 12-next (data/decision-tool.json) |
| I2-046 | The rule needs one live US offer in any of the three trading classes. The Journal's guarantee was live in the US from October to November 2024, though Adweek called it a promotion and its threshold is not public (ATT-E-F02). A+E's guarantee ran in the US in 2020 and 2021, on terms not disclosed (ATT-F-F14). Attention is marketed as a live optimisation input in named buying platforms: The Trade Desk and Amazon DSP, whose releases state no geography (ATT-E-F11), and DV360, whose Help Center documents a live Adelaide algorithm with no go-live date {{claim:ATT-ADELAIDE-F15} | The H2 rule needs a live (not announced) US offer: an optimisation input in a named buying platform, a billing unit, or a contractual guarantee. The paragraph rests the verdict on items that each fail a limb. The Journal's guarantee is not documented as US-only or as contractual (Adweek called it a promotion). A+E's terms are undisclosed and it may have guaranteed presence. The Trade Desk and Amazon DSP releases state no geography. The evidence that meets every limb is in the ledger and chapter 7 but is not cited here. The verdict can stand; its stated basis does not. | ATT-E-F02: 'the sources say advertisers, not US advertisers ... Adweek described it as a promotion'. ATT-E-F03: Kargo sources 'do not say contractual'. ATT-E-F10: Yahoo DSP AU segments with Verizon, 'an optimisation use in a named US buying platform'. ATT-ADELAIDE-F28: iFIT used Adelaide segments in The Trade Desk. Chapter 7: JLR North America described AU custom bidding in DV360 (ATT-V-ADELAIDE-S04). hypotheses.json H2 rule_supported. | supported with qualification | Coding H2 and Q1, paragraph 1: replaced the Journal/A+E/TTD-Amazon-DV360 basis with: 'The rule needs one live US offer in any of the three trading classes, and optimisation in named buying platforms meets it. Yahoo launched AU pre-bid segments in its DSP with Verizon as brand launch partner (ATT-E-F10), and Adelaide said in June 2026 that the targeting is available there (ATT-ADELAIDE-F14). iFIT bought Adelaide segments in The Trade Desk in 2025 (ATT-ADELAIDE-F28). JLR North America described custom bidding on AU in DV360 in May 2025 (ATT-V-ADELAIDE-S04), and DV360's Help Center now documents a live Adelaide algorithm with no go-live date (ATT-ADELAIDE-F15). The Trade Desk and Amazon DSP releases themselves state no geography (ATT-E-F11). Guarantees do not carry the verdict on their own. The Journal's ran from October to November 2024, but the sources do not say its advertisers were in the US, and Adweek called it a promotion (ATT-E-F02). A+E's ran in the US in 2020 and 2021 on terms not disclosed, and may have guaranteed presence rather than attention (ATT-F-F14) (ATT-TVISION-F23).' Also: 'All of it rests on vendor or seller statements' -> 'Apart from JLR's account to a trade publication, all of it rests on vendor or seller statements (ATT-E-F23).' | Holds. The H2 rule (hypotheses.json) needs a live US offer: optimisation in a named buying platform, a billing unit or a contractual guarantee. The cited basis each fails a limb. ATT-E-F02 final_wording says the sources say 'advertisers', not 'US advertisers', and Adweek called it a promotion. ATT-F-F14 discloses no terms, and ATT-TVISION-F23 says presence or attention is ambiguous. ATT-E-F11 states no geography. Stronger evidence that meets every limb exists. ATT-E-F10 ('an optimisation use in a named US buying platform', Verizon launch partner) and ATT-ADELAIDE-F14 (Adelaide June 2026 status). ATT-ADELAIDE-F28 (iFIT, scope_geo US, 2025, Adelaide segments in The Trade Desk). ATT-V-ADELAIDE-S04 (JLR North America, AdExchanger May 2025, AU custom bidding in DV360, already cited in chapter 7). Changes from the proposed replacement: split into short sentences; added 'in 2025' (ATT-ADELAIDE-F28 scope_time) and 'in May 2025' (S04 pub_date); worded the Yahoo status as 'Adelaide said in June 2026 that the targeting is available there', matching F14, which has no Yahoo-side confirmation; added ATT-TVISION-F23 to back the presence-or-attention clause, which ATT-F-F14 alone does not state. 'All of it rests on vendor or seller statements' became 'Apart from JLR's account to a trade publication, all of it ...', because JLR is a buyer account in independent trade press and the old sentence would have been false. iFIT appeared in The Trade Desk's own publication, so it is still a seller statement. The verdict, the scale sentences and the currency-grade sentence are unchanged. plain.md was not touched (single-file scope) and may need a matching edit if it restates the H2 basis. | 06-trading |
| I2-047 | Contractual-style guarantees were live in the US (WSJ, October to November 2024, described by Adweek as a promotion), and attention is marketed as a live optimisation input in named buying platforms (The Trade Desk since October 2024, Amazon DSP since June 2026, DoubleVerify segments and Scibids optimisation per DV | ATT-E-F23 is cited three times in the chapter 6 H2 verdict and shown in the claims table, but its final_wording was never updated after correction I1-041. It calls the WSJ promotion 'contractual-style' and US. It also counts DV segments, which name no platform, and Viant's marketplace, which was in testing, as live optimisation in named platforms. The corrected chapter text rejects both. | Corrections ledger I1-041 says DV 'names no platform, so DV segments and Scibids optimisation do not meet the named buying platform limb', and Viant called its targeting 'testing'. ATT-E-F02: sources say 'advertisers', not 'US advertisers'. ATT-E-F17: Viant 'testing' in August 2026. | supported with qualification | Replaced 'Contractual-style guarantees were live in the US (...); all of these rest on vendor or seller statements.' with: 'It is marketed as a live optimisation input in named buying platforms. Yahoo DSP, a US platform, launched Adelaide AU segments in June 2023 with Verizon as brand launch partner, and Adelaide said in June 2026 that the targeting is available there. The Trade Desk has offered AU tiers since October 2024, and its own publication reported iFIT’s use of Adelaide segments there in October 2025. Amazon DSP has offered AU tiers since June 2026; neither release states geography. DV360’s Help Center documents a live Adelaide algorithm with no go-live date, and JLR North America described AU custom bidding in DV360 in May 2025. Guarantees ran but do not carry the verdict on their own: the Wall Street Journal’s from October to November 2024, for advertisers the sources do not place in the US, which Adweek described as a promotion, and A+E’s for Budweiser in the US in 2020 and 2021, on terms not disclosed. DoubleVerify’s segments and Scibids optimisation name no platform, and Viant described its TVision pre-bid attention targeting, offered through the Havas marketplace, as in testing in August 2026. Apart from JLR’s account to a trade publication, all of these rest on vendor or seller statements.' The opening sentence and the closing sentences on thresholds, scale and currency grade are unchanged. source_ids += ATT-V-ADELAIDE-S04; review_note gets an I2 trace to ATT-E-F10, ATT-ADELAIDE-F14, ATT-E-F11, ATT-ADELAIDE-F28, ATT-ADELAIDE-F15, ATT-E-F02, ATT-F-F14, ATT-E-F15 and ATT-E-F17. | Holds. The ATT-E-F23 final_wording predated correction I1-041. It called the WSJ guarantee 'contractual-style' and 'live in the US', although ATT-E-F02 says the sources say 'advertisers', not 'US advertisers', and Adweek called it a promotion. It also counted DV segments and Scibids optimisation as optimisation in named platforms, though ATT-E-F15 says DV names no DSPs, and it counted Viant's DSP although ATT-E-F17 records 'Began testing'. Chapter 6 line 8 cites ATT-E-F23 for DV360, which the old wording did not mention. The replacement was rewritten to match the current chapter 6 H2 paragraph. Two parts of the proposal were left out. 'Used by US buyers' is dropped because no source places iFIT or Verizon in the US; only Yahoo DSP (ATT-E-F10: 'named US buying platform') and JLR North America are marked US. 'None is documented as contractual' is dropped because no claim makes that absence claim as worded. The chapter 6 exception for JLR's trade-press account was added to the vendor-statement caveat. ATT-V-ADELAIDE-S04 (JLR) was added to source_ids, since no other claim carries it; the review note traces each other element to its claim. Carry-over: data/claims-overlay.json still holds the old ATT-E-F23 final_wording and will overwrite this on build:all unless it is updated to the same text. | data/claim-ledger.json |
| I2-048 | This chapter's screen for independent and academic work covered the Journal of Advertising and the Journal of Advertising Research through Crossref title filters, ARF reports and logged web searches, but not the International Journal of Advertising | The screen found two independent 2025 JAR articles on attention in CTV and linear TV that use TVision panel data, one funded through a JAR and TVision data release, and did not read them. Chapter 3 codes H1a and says no study makes the held-out comparison, but never mentions them. The only mention is a passing line in chapter 12. The absence claims in the abstract (finding 3), chapter 3 and the appendix verdict row are therefore stronger than the screen supports. | ATT-TVISION-F24: 'Crossref records two 2025 Journal of Advertising Research articles on attention in CTV and linear TV: Bhan, Rallabandi and Lee ... and Lee and Choi (funder field JARxTVision Data Release and Call for Ideas). The articles were not read.' Chapter 12: 'which this paper did not read'. | supported with qualification | After the screen sentence (which keeps ATT-B-F01), added: 'This review also found two 2025 Journal of Advertising Research articles on attention in CTV and linear TV that draw on TVision panel data, but did not read them (ATT-TVISION-F24). Whether either compares attention with time in view on held-out data is not known, so every H1a code below is bounded by that gap.' | Holds. ATT-TVISION-F24 records two 2025 JAR articles on attention in CTV and linear TV that the paper did not read. Its review note quotes both abstracts as using TVision panel data. Chapter 3 coded H1a without mentioning them. The proposed replacement was one very long sentence, so the same meaning went in as two added sentences after the screen sentence. The text says 'this review found' rather than 'the screen found' because F24 comes from the TVision stream, not stream B's logged screen. The abstract, chapter 12 and the appendix are outside this target. | 03-validation |
| I2-049 | Google's own announcement of an off-the-shelf algorithm reads 'Expected mid July', S19, so the packaged integration is announced, not confirmed live | The Adelaide Activation dimension (chapter 11 profile table) says the DV360 algorithm is not confirmed live. Chapters 6 and 7 and the Adelaide SWOT say Google's Help Center documented it as live by 29 September, based on the corrected claim. The timeline row at profiles.json line 570 has the same stale wording. | ATT-ADELAIDE-F15 (contradicted, then corrected): 'by 29 September 2026 Google's DV360 Help Center documents the live activation steps for the Maximize Adelaide AU algorithm ... The packaged integration is live per platform documentation; only the exact go-live date is undocumented.' | supported with qualification | adelaide.capabilities.activation.observed: "Google's own announcement of an off-the-shelf algorithm reads 'Expected mid July', S19, so the packaged integration is announced, not confirmed live)" -> "Google announced an off-the-shelf algorithm for mid-July, S19, and by 29 September 2026 its DV360 Help Center documented live activation steps for the 'Maximize Adelaide AU' algorithm, with an additional media fee and no documented go-live date, ATT-V-ADELAIDE-S03)"; src + ATT-V-ADELAIDE-S03. Timeline row (date "2026 (page undated; 'Expected mid July')"): event_type -> 'integration announced, then live (platform documentation)'; 'general availability not confirmed at 29 September 2026' -> "by 29 September 2026 the Help Center documents live activation steps for 'Maximize Adelaide AU', with an additional media fee and no documented go-live date"; src + ATT-V-ADELAIDE-S03. | Holds. ATT-ADELAIDE-F15 is 'contradicted'. Its correction says that by 29 September 2026 Google's DV360 Help Center documents live activation steps for 'Maximize Adelaide AU', with an additional media fee, and only the go-live date is undocumented. Chapters 4, 6 and 7, the matrix cell and the Adelaide SWOT agree. The timeline row had the same stale wording and was corrected too. Source ATT-V-ADELAIDE-S03 (the Help Center page, grade C) is in the register. | data/profiles.json |
| I2-050 | where Google's own Help Center documents the packaged 'Maximize Adelaide AU' custom-bidding algorithm from mid-July 2026. | The Adelaide SWOT Strength gives a go-live date of mid-July 2026. Correction R2-01 removed that date because Google's documentation does not give one. | Corrections ledger R2-01: 'the go-live date is undocumented'. ATT-ADELAIDE-F15: 'only the exact go-live date is undocumented'. Chapter 6: 'the go-live date is not documented'. | supported with qualification | Strength 1 now reads: "AU is marketed as live in several DSPs; Google's DV360 Help Center documented live steps for the 'Maximize Adelaide AU' algorithm by 29 September 2026; go-live date undocumented." Added src ATT-E-S16 (Google announcement page, 'Expected mid July') to support the undocumented go-live date. | Holds. R2-01 removed the mid-July 2026 go-live date: Google's announcement page says only 'Expected mid July' and the operating page (ATT-V-ADELAIDE-S03) documents 'Maximize Adelaide AU' with no date. ATT-ADELAIDE-F15 final_wording: live per platform documentation by 29 September 2026, 'only the exact go-live date is undocumented'. Chapter 6 line 19 says the same. Merged with I2-vendor-13 and I2-buyer-20 into one rewrite of the same sentence, within 28 words. | evidence/swot/adelaide.json |
| I2-051 | the peer-reviewed MSI study found attention explained most variance in brand search lift, in-sample, without holdout. | This reads R-squared as dominance. Attention explained 11.6% of variance and was only slightly above tuning (8.2%) in single-measure regressions; with quarter-hour fixed effects the gap nearly closes. 'Most variance' can be read as a majority, and the SWOT drops both qualifiers. | ATT-B-F06: attention 'explained slightly more brand-search lift (R-squared .116) than exposure tuning (.082), presence (.070) ... in separate single-measure regressions without a hold-out'. Chapter 3: '.180 for attention against .176 for exposure tuning' with QH fixed effects. | supported with qualification | strengths[3].t: 'the peer-reviewed MSI study found attention explained most variance in brand search lift, in-sample, without holdout.' -> 'an MSI working paper found attention explained slightly more brand-search lift than tuning, in-sample, and fixed effects nearly closed the gap.' (28 words; src unchanged) | Holds. ATT-B-F06 and ATT-TVISION-F18 (source ATT-B-S06, alias ATT-TVISION-S38) give R-squared .116 for attention against .082 exposure tuning, .070 presence and .035 OTS tuning in single-measure regressions, and .180 against .176 with quarter-hour fixed effects. 'Most variance' can read as a majority of variance and the line dropped the fixed-effects qualifier. The suggested replacement runs to about 45 words, well over the 28-word SWOT line limit enforced by scripts/check.mjs. The fix keeps the meaning ('slightly more', in-sample, gap nearly closes with fixed effects) and leaves the R-squared figures to chapter 3, which carries them under ATT-B-F06. | evidence/swot/tvision.json |
| I2-052 | AU is the metric in every AU guarantee found (WSJ, Vevo, Duolingo and others via the AU Ecosystem), giving it a trading role no other specialist unit shows in these sources | The first clause is circular. The second is contradicted by the paper's own trading table: SeenThis guarantees and prices media on Lumen's measure, and A+E's guarantee rested on TVision data. The claim favours Adelaide over named competitors. | Chapter 6 table: SeenThis Attention (metric vendor Lumen, guarantee and price) and A+E (TVision). fig-guarantees: 'Two rest on something other than Adelaide's AU: SeenThis on Lumen's measure and A+E's guarantee on TVision panel data.' | supported with qualification | adelaide.differentiation[2] -> "AU defines most of the guarantees found (WSJ, Kargo, Duolingo and others via the AU Ecosystem, and Vevo's by inference); the exceptions are SeenThis, which guarantees and prices on Lumen's measure, and A+E's 2020-21 guarantee on TVision data (ATT-ADELAIDE-S26, S27, S38; ATT-E-S08; ATT-NFA-S01; ATT-F-S25)." | Holds. 'AU is the metric in every AU guarantee' is circular, and chapter 6 shows SeenThis guaranteeing and pricing on Lumen's measure (ATT-R2-F01) and A+E's guarantee resting on TVision data (ATT-F-F14). Chapter 6 and ATT-E-F22 say 'Most guarantees found are defined on Adelaide's AU; the exceptions are SeenThis and A+E.' Changed from the proposed text: Vevo's guarantee does not state its unit, so it is marked 'by inference' (chapter 2, ATT-VEVO-F05). Citations added for Kargo, SeenThis and A+E. The I2-buyer-06 wording 'widest trading role' was not used because it is an unsupported comparative. | data/profiles.json |
| I2-053 | The closest studies use the same data they learn from, and they don't agree. | This misstates why H1a is unresolved. The main paper says the in-sample studies are 'not directly comparable with one another', not that they conflict. Conflict between comparable studies is a separate unresolved clause in the H1a rule. The sentence also blurs into the next one, about ARF vendor agreement. | Chapter 3: 'The closest comparisons are in-sample and not directly comparable with one another.' hypotheses.json H1a rule_unresolved lists 'only in-sample associations' and 'studies of comparable design conflict' as distinct grounds. | supported with qualification | 'The closest studies use the same data they learn from, and they don't agree.' -> 'The closest studies test on the same data they learn from, and they can't be compared with each other.' | Holds. Chapter 3 and abstract finding 3 say the closest comparisons are 'in-sample and not directly comparable with one another'. The H1a rule_unresolved lists in-sample-only and comparable-design conflict as separate grounds. 'They don't agree' misstated the reason. Merged with duplicate I2-buyer-25. 'Test on' states the in-sample point more exactly than 'use'. | plain |
| I2-054 | discloses about five of seven method elements at summary level for the in-flight product | The Amplified SWOT credits about five of seven method elements. The scored method-disclosure cell for the same in-flight product counts 'about three of seven at summary level' from public pages. The two counts use different bases and the text does not reconcile them. | Amplified Campaign Performance method disclosure (chapter 11 profile table): 'That is about three of seven at summary level, no data.' | supported with qualification | None to evidence/swot/amplified.json. Fix needed in data/profiles.json (amplified.capabilities.method_disclosure.observed): replace 'That is about three of seven at summary level, no data.' with the verifier's correction, e.g. 'Counting Amplified's 2023 ARF self-report (ATT-V-AMPLIFIED-S01), about five of seven elements are published at summary level: construct; signals and model inputs; reference labels; panel size and recruitment; model type. A held-out design is described without statistics; no recalibration cadence or data are published.' Then add ATT-V-AMPLIFIED-S01 to that cell's src. | The SWOT's 'about five of seven' already matches the scored cell. capability-matrix.json Amplified Campaign Performance method_disclosure is v 3, and its 'why' reads 'about five of seven elements are published at summary level' with the 2023 ARF self-report (ATT-V-AMPLIFIED-S01). The finding's premise that the scored cell counts three is wrong. The 'about three of seven' text is the profile's pre-verification count (profiles.json capabilities.method_disclosure.observed, rendered in the profile table). evidence/verify/amplified.json marks that count 'contradicted': it notes that public vendor-supplied material, including the ARF compendium, gives about five, and that the privacy policy also adds tag-signal detail. Adding 'about three of seven on its current public pages, which the scored cell counts' would bring a contradicted count into the SWOT and misstate which count the score uses. The unreconciled count is real, but it lives in profiles.json, outside this target. | evidence/swot/amplified.json |
| I2-055 | 2 wording, precision and consistency issues in evidence/swot/adelaide.json | Adversarial pass findings rated P2: I2-vendor-27, I2-buyer-20 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 2 passages reworded; no verdict or score changed | Precision and consistency. | evidence/swot/adelaide.json |
| I2-056 | 1 wording, precision and consistency issues in evidence/swot/amplified.json | Adversarial pass findings rated P2: I2-vendor-28 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | evidence/swot/amplified.json |
| I2-057 | 1 wording, precision and consistency issues in evidence/swot/lumen.json | Adversarial pass findings rated P2: I2-vendor-29 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | evidence/swot/lumen.json |
| I2-058 | 1 wording, precision and consistency issues in data/profiles.json (DoubleVerify, Differentiation and Production scale) | Adversarial pass findings rated P2: I2-vendor-30 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | data/profiles.json (DoubleVerify, Differentiation and Production scale) |
| I2-059 | 4 wording, precision and consistency issues in 03-validation | Adversarial pass findings rated P2: I2-vendor-31, I2-buyer-26, I2-science-19, I2-science-20 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 4 passages reworded; no verdict or score changed | Precision and consistency. | 03-validation |
| I2-060 | 1 wording, precision and consistency issues in data/profiles.json (Kantar, Limits) | Adversarial pass findings rated P2: I2-vendor-32 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | data/profiles.json (Kantar, Limits) |
| I2-061 | 1 wording, precision and consistency issues in data/capability-matrix.json (Kantar assurance vs eye square) | Adversarial pass findings rated P2: I2-vendor-33 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | data/capability-matrix.json (Kantar assurance vs eye square) |
| I2-062 | 2 wording, precision and consistency issues in 02-standards | Adversarial pass findings rated P2: I2-vendor-34, I2-science-22 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 2 passages reworded; no verdict or score changed | Precision and consistency. | 02-standards |
| I2-063 | 1 wording, precision and consistency issues in 05-market | Adversarial pass findings rated P2: I2-vendor-35 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | 05-market |
| I2-064 | 2 wording, precision and consistency issues in fig-accreditation | Adversarial pass findings rated P2: I2-vendor-36, I2-buyer-19 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 2 passages reworded; no verdict or score changed | Precision and consistency. | fig-accreditation |
| I2-065 | 4 wording, precision and consistency issues in plain | Adversarial pass findings rated P2: I2-vendor-37, I2-buyer-24, I2-buyer-25, I2-science-24 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 4 passages reworded; no verdict or score changed | Precision and consistency. | plain |
| I2-066 | 1 wording, precision and consistency issues in fig-channel-coverage | Adversarial pass findings rated P2: I2-buyer-16 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | fig-channel-coverage |
| I2-067 | 1 wording, precision and consistency issues in fig-four-layers | Adversarial pass findings rated P2: I2-buyer-17 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | fig-four-layers |
| I2-068 | 1 wording, precision and consistency issues in fig-evidence-ladder | Adversarial pass findings rated P2: I2-buyer-18 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | fig-evidence-ladder |
| I2-069 | 2 wording, precision and consistency issues in data/decision-tool.json | Adversarial pass findings rated P2: I2-buyer-21, I2-buyer-22 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 2 passages reworded; no verdict or score changed | Precision and consistency. | data/decision-tool.json |
| I2-070 | 1 wording, precision and consistency issues in 12-next | Adversarial pass findings rated P2: I2-buyer-23 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | 12-next |
| I2-071 | 2 wording, precision and consistency issues in evidence/profiles/adelaide.json | Adversarial pass findings rated P2: I2-buyer-27, I2-science-27 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 2 passages reworded; no verdict or score changed | Precision and consistency. | evidence/profiles/adelaide.json |
| I2-072 | 5 wording, precision and consistency issues in data/profiles.json | Adversarial pass findings rated P2: I2-buyer-28, I2-buyer-29, I2-buyer-30, I2-buyer-31, I2-science-26 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 5 passages reworded; no verdict or score changed | Precision and consistency. | data/profiles.json |
| I2-073 | 1 wording, precision and consistency issues in data/claim-ledger.json | Adversarial pass findings rated P2: I2-buyer-32 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | data/claim-ledger.json |
| I2-074 | 3 wording, precision and consistency issues in 00-abstract | Adversarial pass findings rated P2: I2-buyer-33, I2-science-16, I2-science-30 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 3 passages reworded; no verdict or score changed | Precision and consistency. | 00-abstract |
| I2-075 | 1 wording, precision and consistency issues in fig-guarantees | Adversarial pass findings rated P2: I2-science-13 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | fig-guarantees |
| I2-076 | 1 wording, precision and consistency issues in 12-next (data/decision-tool.json) | Adversarial pass findings rated P2: I2-science-14 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | 12-next (data/decision-tool.json) |
| I2-077 | 1 wording, precision and consistency issues in 11-fq | Adversarial pass findings rated P2: I2-science-15 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | 11-fq |
| I2-078 | 2 wording, precision and consistency issues in 08-channels | Adversarial pass findings rated P2: I2-science-17, I2-science-29 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 2 passages reworded; no verdict or score changed | Precision and consistency. | 08-channels |
| I2-079 | 1 wording, precision and consistency issues in 04-vendor-evidence | Adversarial pass findings rated P2: I2-science-18 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | 04-vendor-evidence |
| I2-080 | 1 wording, precision and consistency issues in 06-trading | Adversarial pass findings rated P2: I2-science-21 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | 06-trading |
| I2-081 | 1 wording, precision and consistency issues in 07-buyers | Adversarial pass findings rated P2: I2-science-23 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | 07-buyers |
| I2-082 | 1 wording, precision and consistency issues in evidence/swot/doubleverify.json | Adversarial pass findings rated P2: I2-science-25 | Each finding checked against the claim ledger, sources or stats before editing. | supported with qualification | 1 passages reworded; no verdict or score changed | Precision and consistency. | evidence/swot/doubleverify.json |
| CO-01 | Bands named Front rank, Strong, Mid and Narrow | Method changes adopted across the series after the audio paper’s fourth and fifth reviews. | Audio paper commits 2a1d9ea9 and ae109d53; the change was applied and rebuilt here. | structural | Bands renamed Most documented, Well documented, Partly documented and Least documented; thresholds unchanged; bands_rule records the old names | Keep the three papers’ methods and labels consistent. | data/scoring-rules.json; chapters 0, 11, 13 |
| CO-02 | One scoring pass with a consistency audit; no second scorer | Method changes adopted across the series after the audio paper’s fourth and fifth reviews. | Audio paper commits 2a1d9ea9 and ae109d53; the change was applied and rebuilt here. | structural | Blind second scoring pass on the six rows nearest a band line (66 cells): 78.8% exact, 98.4% within one, quadratic-weighted kappa 0.88, no band moved; two material disagreements adjudicated (AJ-01, AJ-02); both passes published as data/second-pass.json and .csv | Keep the three papers’ methods and labels consistent. | data/second-pass.json; build-matrix; make-csv; datasets; chapters 11 and 13 |
| CO-03 | Band sensitivity shown only for missing evidence (±) | Method changes adopted across the series after the audio paper’s fourth and fifth reviews. | Audio paper commits 2a1d9ea9 and ae109d53; the change was applied and rebuilt here. | structural | Provisional markers: ± for bands that move when missing evidence is filled and † for bands a blind second scorer would move, in the matrix, the quotient figure and the CSV | Keep the three papers’ methods and labels consistent. | build/instruments.mjs; scripts/make-figures.mjs; build/make-csv.mjs; scripts/stats.mjs |
| CO-04 | Research cutoff 2 October 2026 with a freeze recheck on that day | Method changes adopted across the series after the audio paper’s fourth and fifth reviews. | Audio paper commits 2a1d9ea9 and ae109d53; the change was applied and rebuilt here. | structural | Cutoff set to the publication date, 1 October 2026, at the author’s decision, without a separate freeze recheck; status-sensitive facts stand as read on 28 to 30 September; registered rules unchanged | Keep the three papers’ methods and labels consistent. | paper.config.json; chapters 0, 2, 10, 13; fig-accreditation caveat |
| CO-05 | No buyer brief at the top of the abstract; headline not bounded | Method changes adopted across the series after the audio paper’s fourth and fifth reviews. | Audio paper commits 2a1d9ea9 and ae109d53; the change was applied and rebuilt here. | structural | Buyer brief added to the abstract, with the line that the paper audits what is documented in public and does not test whether buying on attention works; grade caps and anchor thresholds described as assumptions; ‘before publication’ added to the no-vendor line | Keep the three papers’ methods and labels consistent. | chapters 0 and 13 |
| CO-06 | Matrix print band labels in capitals at 6.5pt | Method changes adopted across the series after the audio paper’s fourth and fifth reviews. | Audio paper commits 2a1d9ea9 and ae109d53; the change was applied and rebuilt here. | structural | Matrix print labels at 7pt without capitals, so the matrix page meets the 7pt print floor | Keep the three papers’ methods and labels consistent. | build/paper.template.html |
| FC-01 | Adelaide profile risk and SWOT line said that in AU guarantees Adelaide both supplies the metric and measures delivery. | The line goes further than chapter 6 and ATT-E-F22, which say the measurer is Adelaide only where an offer discloses it. | ATT-E-F22: measurer disclosed for the WSJ, Kargo and TripleLift offers (ATT-E-S07, S08, S12); implied but not stated for Vevo, Duolingo and Teads. The line cited the Duolingo, Vevo and AU-hub pages, which do not name a measurer. [94, 93, 91] | supported as stated | Profile risk and SWOT line now say: where an AU guarantee discloses who measures delivery (the Journal, Kargo, TripleLift) it is Adelaide, which also supplies the metric; for Vevo, Duolingo and Teads the measurer is not stated; no independent audit found. Sources ATT-E-S07, S08, S12 added. | Verified against the primary source or the paper’s own data before any change. | Adelaide profile (risks); Adelaide SWOT |
| FC-02 | Source buttons and SWOT references that used alias ids (profile-local ids mapped to a canonical register id) did not resolve in the page. | Profile, SWOT and figure source buttons with alias ids opened nothing, and SWOT references were dropped. | data/source-aliases.json maps profile-local ids to canonical ids; the page data kept the alias form. | supported as stated | Page data, instruments, figure source buttons and the client source lookup now canonicalise alias ids; the alias map ships with the page. | Verified against the primary source or the paper’s own data before any change. | build/make-page-data.mjs, build/render.mjs, build/instruments.mjs, build/paper.template.html |
| FC-03 | The TVision panel-size claim’s limitation and the TVision profile said panel size appears only in trade reporting and that no 2026 primary statement was found. | TVision’s own 2026 report states the panel size, and the claim’s corrected wording already said so. | TVision 2026 TV Attention & Creative Impact Report, Methodology & Metrics (p. 21): ‘an opt-in panel of 5,000 homes across the United States and weighted to US Census’ (ATT-V-TVISION-S01). [292, 415] | supported as stated | Claim limitation and confidence note rewritten, with the report and the recruitment site added as sources; profile dependency, method-validity, disclosure tally and panel limit now cite the report (disclosure element 4 counted as published, matching the scored cell). | Verified against the primary source or the paper’s own data before any change. | ATT-TVISION-F07; TVision profile |
| FC-04 | Chapter 11, the scoring rules, the pipeline and quotient figures and the footer said single scorer, one scoring pass or no independent second scorer. | Chapter 13 and data/second-pass.json report a blind second AI scoring pass on six rows. | data/second-pass.json: 6 rows, 66 cells; chapter 13 method text. | supported as stated | Wording now says one initial scoring pass, a consistency audit and blind rescoring of the six rows nearest a band line by a separate AI scorer; every other row scored once; no human second scorer. | Verified against the primary source or the paper’s own data before any change. | chapter 11; data/scoring-rules.json; fig-pipeline (boxes, footnote and description); fig-quotient; footer |
| FC-05 | Chapter 12, the appendix, nine claim limitations and eight profile status notes said the status is rechecked, or must be rechecked, at the 2 October freeze; the claim on studies said sources were screened by 2 October 2026. | The paper is published on 1 October 2026 without a freeze recheck, as chapter 13 states. | Chapter 13: the author chose to publish on 1 October 2026 without a separate freeze recheck. | supported as stated | All now say the status was not rechecked before the 1 October 2026 cutoff (claims via the overlay, profiles via profile corrections); ATT-B-F01 and the plain edition give 1 October 2026 as the end of screening. Verifier review notes keep their dated wording. | Verified against the primary source or the paper’s own data before any change. | chapter 12; chapter 14; plain; claims ATT-A-F12, ATT-I-F01, ATT-I-F17, ATT-ADELAIDE-F01, ATT-AMPLIFIED-F23, ATT-DOUBLEVERIFY-F04, ATT-DOUBLEVERIFY-F17, ATT-IAS-F03, ATT-TVISION-F10, ATT-B-F01; eight profile status notes |
| FC-06 | The deletion test paragraph did not say what is held fixed. | Readers cannot tell whether the field mean moves when a cell is deleted. | build/quotient.mjs: deletion recomputes the row’s k, admission and eligibility with F frozen. | supported as stated | Chapter 11 now says every other row and the field mean are held fixed. | Verified against the primary source or the paper’s own data before any change. | chapter 11 |
| FC-07 | Lumen profile limit and tension line said LAMP results cannot show value beyond time in view. | Too strong: the sources show the comparisons reviewed do not establish it, not that it is impossible. | ATT-LUMEN-S20 (PwC review): the published model uses viewable time as an input; no held-out comparison against time in view is published. [8] | supported with qualification | Now: the published comparisons reviewed here do not establish incremental predictive value beyond time in view; the Havas study limitation (ATT-LUMEN-F26) now says the study does not establish value beyond time in view. | Verified against the primary source or the paper’s own data before any change. | Lumen profile (limits); ATT-LUMEN-F26 limitation |
| FC-08 | Old band names (front rank, Mid) survived in chapter 11, a count read as ‘1 products’, and Markdown tables whose evidence column held only claim markers exported empty. | Leftover labels and an empty column in the Markdown edition. | Chapter 11 text; dist Markdown accreditation (chapter 2) and market-event (chapter 5) tables. | supported as stated | Band names updated; the count sentence reworded; the Markdown edition now renders each claim marker as its source numbers. | Verified against the primary source or the paper’s own data before any change. | chapter 11; build/render.mjs (Markdown edition) |
| PUB-01 | The paper said it was published on 1 October 2026, its cutoff date, and that statuses were not rechecked before publication on that date. | Publication slipped to 6 October 2026, so those statements would no longer be true. | Publication date of the site release. | structural | Publication and modified dates set to 6 October 2026; the research cutoff stays 1 October 2026; the abstract, chapters 2, 12, 13 and 14, nine claim notes and eight profile status notes now tie status-sensitive facts to the cutoff, not to the publication date. Nothing read after the cutoff was added. | Keep every stated date true. | paper.config.json; abstract; chapters 2, 12, 13, 14; claims overlay; profile corrections |



## Appendix: verdicts, vendor universe and downloads


*The registered hypotheses and how each was coded, the reasoned sample of companies and products screened, and the published data files behind the paper.*

### How each registered test was coded

Each hypothesis and question was registered on 28 September 2026 with its decision rules, before any research agent ran. The verdicts below apply those rules to the evidence read by the cutoff; unresolved means the public sources screened do not settle it, not that the hypothesis is refuted. The codes are dated to screening on 28 and 29 September 2026 and were not rechecked before the 1 October 2026 cutoff.

| Test | Statement (short) | Verdict | Where |
|---|---|---|---|
| H1a | Attention adds out-of-sample predictive value beyond viewability and time in view | Unresolved for memory, brand and sales: in the sources screened, no independent, commissioned or vendor study makes a held-out comparison against time in view [46, 51, 52] [8, 10, 71, 64] | [Chapter 3](#validation) |
| H1b | Buying on attention improves outcomes at comparable cost | Unresolved for memory, brand and sales: in the sources screened, no independent, commissioned or vendor test of buying on attention against a randomised or holdout control reports an outcome gain with cost per incremental outcome [51, 4, 5] [69, 65, 11, 10] | [Chapter 3](#validation) |
| H2 | Attention is used in trading, not only planning | Supported; scale undocumented: live optimisation and guarantees exist, all on vendor or seller statements; no offer discloses volume against a market denominator [91, 92, 334, 93, 25] | [Chapter 6](#trading) |
| H3 | MRC accredits attention metrics against the November 2025 guidelines | Contradicted as of 28 and 29 September 2026, on a narrow basis: the standards listed in the letter for the only attention accreditation found, DoubleVerify's, do not include the guidelines; whether they informed the audit is not public [42, 2, 425, 563] | [Chapter 2](#standards) |
| H4 | Verification bundling changes specialists' demand, pricing or revenue | Unresolved: in the filings, accounts and releases screened by 28 September 2026, no source links the two [47, 231, 285, 232] | [Chapter 5](#market) |
| H5 | Attention signals are carried and used in agent-driven buying | Unresolved at 28 and 29 September 2026: no named attention vendor's signal is documented beyond schema representation in AdCP v3.1.24; the seller test passes use a placeholder vendor and simulated delivery [142, 141, 151, 33] | [Chapter 10](#agents) |
| Q1 | Currency or diagnostic | No metric-environment pair codes currency or partial currency; the rest code diagnostic or unresolved [2] [63, 222]. Where terms can be read, an unaccredited metric codes diagnostic even when it prices media, as Lumen's does in SeenThis Attention [2, 48, 3] [99, 100] | [Chapter 6](#trading) |
| Q2 | What each product measures | Almost every campaign number is a model estimate or an index of device signals [10, 63, 73, 74, 9]; people are observed in reference panels, TVision's in-home TV panel and creative tests [45, 106, 220, 110] [249] | [Chapter 1](#defining) |
| Q3 | Where measurement is live and where it is thin | Display, online video, CTV and social are covered; audio and DOOH are thin; in-game is study only; retail media, an unscored attribute, is thin [105, 119, 42, 45, 115, 110, 113] [113, 29] [162, 260] | [Chapter 8](#channels) |
| Q4 | How the market consolidated | Events found from 2017 to the cutoff, classified by type and dated where a source allows (inferred events are undated); the list is not a census, and no source found attributes any event to a failure of attention metrics [87, 80, 82, 83, 465, 86] | [Chapter 5](#market) |

### The vendor universe

The 51 companies and products in the vendor sample, a reasoned sample rather than a census, with owner, group and status. Products in the matrix are tagged. The product register gives the inclusion reason for every registered product, including trading offers not in this table, such as SeenThis Attention and Impact Media's suite. For the other rows, including name-only candidates, the reason is in the note field of data/vendor-universe.csv.

**Vendor universe**

| Company / product | Legal entity / owner | HQ | Scope group | Formats | Buying routes | Main geographies | Status at cutoff | Src |
|---|---|---|---|---|---|---|---|---|
| Adelaide | not established (workstream D) | New York, US (Adelaide Metrics Inc.) | specialist | display; OLV; social; CTV; linear TV; podcasts (MRC pre-audit scope adds cinema and DOOH) | DSP pre-bid and custom bidding; Nielsen Outcomes Marketplace; dashboard; Custom AU | US entity; US availability documented | live | [336] |
| Amplified | not established (workstream D) | Adelaide, South Australia (offices Adelaide, London, New York, Sydney) | specialist | TV; cinema; BVOD; web; social; OOH; audio (site lists) | SaaS platform; 'API-ready' claim; integrations service | global; New York office listed; US sale not otherwise documented | live | [347] |
| Amplified | not established (workstream D) | Adelaide, South Australia | specialist | digital; programmatic; CTV (in-flight verification) | SaaS platform; in-flight verification 'no tracking tags required' | global; US not documented | live | [347] |
| Amplified | not established (workstream D) | Adelaide, South Australia | specialist | video creative by platform and format | attentionai-api v1; CSV download | global; US not documented | live | [429] |
| Lumen Research | not established (workstream D) | London, UK (Lumen Research Ltd, registered office) | specialist | display; video; social (tags and DSP plug-in); bespoke studies across TV, BVOD, print, OOH, POS | LAMP measurement tag; LAMP DSP plug-in (Attention Bidding); self-service planning platform; data used in IAS Quality Attention | global; US not documented in pages opened | live | [392] |
| Lumen Research | not established (workstream D) | London, UK | specialist | social; digital; in-game; TV/VOD; OOH; POS; print; packaging | reports; heatmaps; surveys | global; US not documented | live | [279] |
| TVision | Viant Technology Inc. (acquisition completed 2026-05-01) | not established (Viant release dateline Irvine, CA) | media_owned | linear TV; CTV | Ad Scoreboard; CTV Campaign Measurement; Audience and CTV Activation; Viant DSP integration; data licensing; calibration data | US (US clients in testimonials; panel country not stated on home page) | live | [411] |
| DoubleVerify | DoubleVerify Holdings, Inc. (listed; workstream D) | not established in sources opened | verification | display; video; social; CTV (site channels; accreditation stated for desktop and mobile only) | Universal Attention Segments; Scibids AI; DV snapshot | global | live | [209] |
| Integral Ad Science | Novacap (per brief; not re-verified here) | not established in sources opened | verification | IAS site channels: CTV and video; open web; social; mobile in-app; digital audio (not attention-specific) | IAS suite; Lumen eye-tracking data inside the metric | global | live | [73] |
| xpln.ai | not established | not stated (French CESP grid lists it) | specialist | display; video; social; walled gardens; CTV (vendor claim) | dashboards | not stated; MRC audit (US accreditor) in process for mobile web and in-app | live | [122] |
| Adloox - Peer39 | not established | not established | verification | not established | not established | France (CESP grid); US not documented | unconfirmed | [3] |
| GumGum | not established (workstream D) | Santa Monica, California | media_owned | not stated for Attentive Lift | On Device brand-lift API | Australia (launch); US not documented | live | [458] |
| GumGum | not established (workstream D) | Santa Monica, California | historical | n/a | gumgum.com/attention redirects to /results | n/a | retired | [459] |
| Kantar | not established (workstream D) | not established (release dateline London) | research | digital; TV; audio; print; outdoor ads | Kantar Marketplace self-serve or serviced; dashboard | US list price from $4,500 | live | [246] |
| Kantar | not established (workstream D) | not established | research | YouTube; Facebook; Instagram; Twitter; TikTok; YouKu; custom contexts incl. audio, Netflix, Spotify, LinkedIn | serviced only; dashboard; PowerPoint export | US list price from $13,950 | live | [248] |
| Kantar | not established (workstream D) | not established | research | digital; TV; static; creator content | Kantar Marketplace self-serve or serviced; dashboards | US list price from $500 | live | [247] |
| Adverteyes | not established (executive chairman is the Realeyes CEO) | not established | specialist | video; social and creator content | API; Chrome extension; MCP (claimed); Nielsen Outcomes Marketplace | global (90-country database); US not stated | live | [460] |
| Neurons | not established | not established | research | image and video ads | self-serve platform | not stated | live | [579] |
| Tobii | Tobii (product sold on tobii.com; workstream D) | not established | research | ads; websites; packaging; video | Tobii Sticky API; Alchemer and Qualtrics integration | global panel; US not stated | live | [455] |
| iMotions | not established (Smart Eye per brief) | not established | research | video creative | iMotions Online; Affectiva facial coding API and SDK sold separately | global markets; US not stated | live | [580] |
| DAIVID | not established | not established | research | retail media creative; video; social; display | Creative Data Feed API | not stated | live | [581] |
| eye square | eye square GmbH (legal form); owner not established | not established | research | video; social media ads; e-commerce | reports | not stated | live | [582] |
| Brainsights | not established | not established (Canadian billboard product) | research | video creative (EEG); OOH billboards (SI Maps) | reports; SI Maps | Canada; US not documented | live | [422] |
| Dragonfly AI | not established | London, UK (per vendor page) | research | creative incl. packaging and digital shelf | Connect creative data API; Studio3 platform | not stated | live | [583] |
| Viomba | not established | not established | specialist | display; social; video (vendor claim) | MCP server (claimed) | not stated | unconfirmed | [562] |
| System1 | not established | not established | research | not established | AECP partner of Adelaide | not established | unconfirmed | [176] |
| CreativeX | not established | not established | research | not established | AECP partner of Adelaide | not established | unconfirmed | [176] |
| Ipsos | not established | not established | research | not established | AECP partner of Adelaide | not established | unconfirmed | [176] |
| Sundogs | not established | not established | research | not established | AECP partner of Adelaide | not established | unconfirmed | [176] |
| CubeAI | not established | not established | specialist | not established | not established | not established | unconfirmed | [41] |
| GlassView | not established | not established | trading | not established | not established | not established | unconfirmed | [41] |
| KERV Interactive | not established | not established | trading | not established | not established | not established | unconfirmed | [41] |
| Mobian | not established | not established | specialist | not established | not established | not established | unconfirmed | [41] |
| Veritonic | not established | not established | research | audio (per audio paper) | not established | not established | unconfirmed | [41] |
| Nielsen | not established | not established | research | cross-platform (Nielsen ONE) | Outcomes Marketplace carries Adelaide AU and Adverteyes | US | live | [331] |
| Comscore | not established | not established | research | not established | attention metrics in deal IDs (headline only) | not established | unconfirmed | [438] |
| EDO | not established | not established | outcome_comparator | TV (convergent TV) | Amazon DSP; The Trade Desk; Google DV360 | US | live | [584] |
| Upwave | not established | not established | outcome_comparator | not stated | not stated | US | live | [585] |
| Vevo | not established | New York (release dateline) | trading | CTV; mobile; desktop | Adelaide AU; one performance report | US dateline | live | [78] |
| Teads | Teads Holding Co. (per register; not re-verified here) | not established | trading | omnichannel (not specified in the post) | not specified | US (NewFronts) | live | [95] |
| Duolingo | not established | not established | trading | mobile in-app (OLV) | Adelaide AU guarantees and AU reporting | not stated | live | [97] |
| Adlook | not established | not established | trading | not stated | DSP AU-CPM buying model with Adelaide | US dateline | unconfirmed | [98] |
| TripleLift | not established | not established | trading | not established here | not established here | not established here | unconfirmed | [41] |
| Ogury | not established | not established | trading | not established here | not established here | not established here | unconfirmed | [41] |
| Kargo | not established | not established | trading | not established here | not established here | not established here | unconfirmed | [41] |
| Magnite | not established | not established | trading | not established here | not established here | not established here | unconfirmed | [41] |
| Yahoo | not established | not established | trading | not established here | not established here | not established here | unconfirmed | [41] |
| Rita | Adelaide (all-stock acquisition announced 2024-08-19) | not established (EU panel) | historical | n/a | data feeds into AU models | EU | retired | [341] |
| Realeyes | not established | not established | historical | n/a | n/a | n/a | retired | [461] |
| SPARK Neuro | not established | not established | historical | n/a | n/a | n/a | retired | [451] |
| Oracle (Moat) | Oracle Corporation | not established | historical | n/a | n/a | n/a | retired | [586] |

### Downloads

- `attention-measurement-2026.pdf` — the print edition: A4, with page numbers and chapter bookmarks
- `attention-measurement-2026.md` — the reading edition in Markdown, with numbered references
- `data/source-register.csv` — every source with publication, event and access dates, type, geography, funding, access limits and reliability notes
- `data/claim-ledger.csv` — every material claim with sources, passage locator, scope, evidence class, review outcome, confidence and final wording
- `data/capability-matrix-wide.csv` — vendors × dimensions, scores only (0–5, n/e, n/a)
- `data/capability-matrix-long.csv` — one row per cell with availability, evidence grade, date, rationale and sources
- `data/capability-dimensions.csv` — the dimension definitions and anchors
- `data/fq.csv` — Focus Quotient per vendor: group means, coverage, band, the reweightings and every dimension score
- `data/fq-method.csv` — FQ groups, weights, shrinkage, field mean, band thresholds, and the reweighting, shrinkage and missing-evidence tests
- `data/fq-missing-evidence.csv` — every scored cell turned into n/e one at a time, with the FQ before and after
- `data/swot.csv` — every SWOT line for the profiled vendors with its basis (record, vendor documentation, vendor claim, inference) and sources
- `data/deep-profiles.csv` — the deep-profile summaries: what is sold, buyers, earnings, dependencies, fit, limits
- `data/deep-profile-records.csv` — every profiled capability, performance claim, financial figure and timeline event with sources
- `data/study-evidence.csv` — studies with design, sample, sponsor and limits
- `data/corrections-ledger.csv` — every change made after verification and challenge passes, with implementation and verification
- `data/vendor-universe.csv` — the reasoned vendor sample with scope group, owner, formats, routes and status
- `data/agentic-map.csv` — AI and agentic capabilities with status labels and observability
- `data/rules.csv` — rules and governance instruments by jurisdiction
- `data/decision-tool.csv` — the decision tool rows
- `data/product-register.csv` — the product register: every scored or listed product with its purchased service, operating entity, owner, cohort, inclusion reason, aliases and US availability
- `data/hypotheses.csv` — the hypotheses registered before extraction, with the rules for supported, contradicted and unresolved
- `data/figure-data.json` — the canonical dataset behind every data figure
- `data/derived-checks.json` — every derived number with its formula and inputs
- `data/stats.json` — every count the prose and figures quote about the research itself, computed from the ledgers
- `data/second-pass.csv` — the blind second scoring pass: one row per re-scored cell with both passes’ scores, grades and reasons, and the adjudication of each material disagreement
- `data/second-pass.json` — the same cells with both passes’ sources, each re-scored row’s band under both passes, and the agreement statistics’ inputs
- `data/scoring-rules.json` — dimension definitions, anchors, cell rules and the FQ definition, fixed before scoring



## References

The 725 sources cited in this paper, numbered in order of first citation. The full register of 985 sources opened during the research, with access dates, access limits and reliability notes, is in data/source-register.csv.

[1] Adelaide vs. Lumen Research: How Our Attention Measurement Approaches Differ. Adelaide Metrics Inc. (Kristin Rose), 2026-09-28. https://www.adelaidemetrics.com/blog/adelaide-vs-lumen-research-attention-measurement-comparison Accessed 2026-09-29.

[2] Accredited Services: Digital. Media Rating Council, undated (live page). https://mediaratingcouncil.org/accreditation/digital Accessed 2026-09-28.

[3] CESP Attention Grid (Published audit and accreditation). CESP (France), 2026-06-10. https://www.cesp.org/wp-content/uploads/Attention-grid-2026-10-06-26-Published-Audit-Accreditation.pdf Accessed 2026-09-28.

[4] ARF Attention Measurement Validation Initiative Phase 2 (2nd Edition). Advertising Research Foundation (ARF), 2024-06. https://thearf-org-unified-admin.s3.amazonaws.com/ARFReports/ARF_AttentionMeasurementValidationInitiativePhase2_2E_2024_06.pdf Accessed 2026-09-28.

[5] ARF Releases Large-Scale Study of Attention Metrics in Live Campaigns (Phase 3 release). Advertising Research Foundation (ARF), 2026-06-09. https://thearf.org/wp-content/uploads/2026/06/Attention-Phase-3-Press-Release-June-2026_FINAL_CLEAN.pdf Accessed 2026-09-28.

[6] Predicting Advertising Success Beyond Traditional Measures: New Insights from Neurophysiological Methods and Market Response Modeling. Venkatraman, Dimoka, Pavlou, Vo, Hampton, Bollinger, Hershfield, Ishihara, Winer (Temple, Duke, UCLA, NYU), 2015-08. https://web.stanford.edu/~knutson/nfc/venkatraman15.pdf Accessed 2026-09-28.

[7] Best Measures of Attention To Creative Tactics in TV Advertising. Bellman, Nenycz-Thiel, Kennedy, Hartnett, Varan (Journal of Advertising Research), 2019-01-24. https://doi.org/10.2501/JAR-2019-002 Accessed 2026-09-28. Access: abstract only.

[8] Lumen / PwC Attention Methodology & Case Study Review. PricewaterhouseCoopers LLP for Lumen Research, 2023-03. https://2871934.fs1.hubspotusercontent-na1.net/hubfs/2871934/Lumen%20_%20PwC%20Attention%20Methodology%20&%20Case%20Study%20Review%202023.pdf Accessed 2026-09-28.

[9] dentsu Attention Economy Phase 2 (2021). dentsu (with Lumen Research, TVision and Amplified Intelligence), 2021. https://assets-eu-01.kc-usercontent.com/27bd3334-62dd-01a3-d049-720ae980f906/887ba048-d8e4-4b28-9836-9b0ce45fa9e4/dentsu%20Attention%20Economy%20Phase%202%202021.pdf Accessed 2026-09-28.

[10] What Do 9,000 Brand Lift Studies Teach Us About Attention & Memory?. Havas Media Network with Lumen Research and Brand Metrics, 2024-09. https://havasmedianetwork.com/wp-content/uploads/2025/02/lumen-brand-metrics-havas-attentionwhitepaper.pdf Accessed 2026-09-28.

[11] Kimberly-Clark Drives Sales Lift with IAS Quality Attention. Integral Ad Science, 2025-09-29. https://integralads.com/insider/kimberly-clark-drives-sales-lift-with-ias-quality-attention/ Accessed 2026-09-28.

[12] Visa Collaborates With Adelaide to Increase Media Quality Transparency & Enhance Brand Impact on the Debit Awareness Campaign in Germany. Adelaide Metrics Inc., 2023-08-01. https://www.adelaidemetrics.com/blog/visa-collaborates-with-adelaide-to-increase-media-quality-transparency-enhance-brand-impact-on-the-debit-awareness-campaign-in-germany Accessed 2026-09-29.

[13] Adelaide Unveils 2024 Outcomes Guide. Adelaide Metrics Inc., 2024-01-18. https://www.adelaidemetrics.com/blog/adelaide-unveils-2024-outcomes-guide-proving-the-connection-between-attention-metrics-and-brand-outcomes-in-over-40-case-studies Accessed 2026-09-28. Access: partial extract.

[14] Adelaide Launches 2025 Outcomes Guide: Collection of Case Studies Demonstrates Average 41% Upper-Funnel and 55% Lower-Funnel Lift From Attention Metrics. Adelaide Metrics Inc., 2025-01-21. https://www.adelaidemetrics.com/blog/adelaide-launches-2025-outcomes-guide-collection-of-case-studies-demonstrates-average-41-upper-funnel-and-55-lower-funnel-lift-from-attention-metrics Accessed 2026-09-28. Access: partial extract.

[15] Adelaide Releases 2026 Outcomes Guide Showing How Attention Drives Business Outcomes Across 60 Campaigns. Adelaide Metrics Inc., 2026-01-20. https://www.adelaidemetrics.com/blog/adelaide-releases-2026-outcomes-guide Accessed 2026-09-28. Access: partial extract.

[16] Study: Contextual Targeting Generates 2.1X More Ad Attention. Wurl, 2025-07-31. https://www.wurl.com/wurlnews/study-contextual-targeting-generates-2-1x-more-ad-attention/ Accessed 2026-09-28. Access: partial extract.

[17] Study: Contextual Targeting Generates 2.1X More Ad Attention. The Measure (Jon Lafayette), 2025-07-31. https://www.themeasure.net/contextual-targeting-generates-2-1-times-more-attention-for-ads-new-study-finds/ Accessed 2026-09-28.

[18] Wurl and TVision Unveil New Report Linking Emotion Alignment to Higher Ad Attention on Streaming TV. Wurl and TVision (Business Wire release, syndicated by Ad Tech Daily), 2025-07-31. https://adtechdaily.com/2025/08/29/wurl-and-tvision-unveil-new-report-linking-emotion-alignment-to-higher-ad-attention-on-streaming-tv/ Accessed 2026-09-28.

[19] In Context with TVision: The efficacy of scene-level contextual targeting. Wurl, 2026-02-02. https://www.wurl.com/blog/in-context-with-tvision-the-efficacy-of-scene-level-contextual-targeting/ Accessed 2026-09-28.

[20] Our Technology - Lumen Research. Lumen Research, 2021-04-12. https://lumen-research.com/attention-technology/ Accessed 2026-09-28.

[21] Native ads vs social ads study. Outbrain, 2019-11-25. https://www.outbrain.com/blog/native-ads-vs-social-ads-study/ Accessed 2026-09-28.

[22] Adelaide Releases Independent Review by MediaSense Validating AU's Effectiveness & Predictive Power. ExchangeWire (Adelaide release), 2025-07-07. https://www.exchangewire.com/blog/2025/07/07/adelaide-releases-independent-review-by-mediasense-validating-aus-effectiveness-predictive-power/ Accessed 2026-09-28.

[23] Attention 2025 (Newsworks, Lumen and Peter Field report). Newsworks with Lumen Research and Peter Field, 2025-09-25. https://www2.newsworks.org.uk/attentionreport2025 Accessed 2026-09-29.

[24] Dentsu and A+E Team to Guarantee Attention by Tracking Eye Movement. MediaVillage (Charlene Weisler), unknown. https://www.mediavillage.com/article/dentsu-and-ae-team-to-guarantee-attention-by-tracking-eye-movement/ Accessed 2026-09-28. Access: abstract only.

[25] Google's Move Toward Attention Metrics Could Be A Viewability Killer, Says Jaguar Land Rover. AdExchanger, 2025-05-20. https://www.adexchanger.com/marketers/googles-move-toward-attention-metrics-could-be-a-viewability-killer-says-jaguar-land-rover/ Accessed 2026-09-29.

[26] Attention-Based PMPs Help Mars And Nespresso Be More Seen And More Green. AdExchanger (Anthony Vargas), 2023-05-02. https://www.adexchanger.com/data-exchanges/attention-based-pmps-helps-mars-and-nespresso-be-more-seen-and-more-green/ Accessed 2026-09-28.

[27] The 2026 Industry Pulse Report (IAS Insider page). Integral Ad Science, 2025-12. https://integralads.com/insider/the-2026-industry-pulse-report/ Accessed 2026-09-28. Access: partial extract.

[28] Radio, Podcasts, and Music Streaming—An Electroencephalography and Physiological Analysis of Listeners' Attitude, Attention, Memory, and Engagement (Brain Sciences 14(4):330). ARN Neurolab (Australian Radio Network) authors; MDPI Brain Sciences, 2024-03-29. https://pmc.ncbi.nlm.nih.gov/articles/PMC11047838/ Accessed 2026-09-28.

[29] The Ultimate Guide to Measuring Intrinsic In-Game Ads. Anzu, 2024-08-22. https://www.anzu.io/blog/in-game-measurement-guide Accessed 2026-09-28.

[30] Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), OJ L series 2024/1689. European Parliament and Council of the European Union, 2024-07-12. http://publications.europa.eu/resource/celex/32024R1689 Accessed 2026-09-28.

[31] AdCP Accountability (performance standards, measurement terms, makegood_policy). AgenticAdvertising.org, 2026-09-23. https://github.com/adcontextprotocol/adcp/blob/v3.1.24/docs/media-buy/advanced-topics/accountability.mdx Accessed 2026-09-29.

[32] AdCP issue #4644 'Vendor-bound optimization goals: bind buyer→seller→vendor end-to-end for vendor-attested metrics' and PR #4668. AdCP GitHub issue tracker, 2026-05-17. https://github.com/adcontextprotocol/adcp/issues/4644 Accessed 2026-09-28.

[33] AdCP compliance storyboard media_buy_seller/vendor_metric_optimization_flow.yaml. AgenticAdvertising.org, 2026-09-23. https://github.com/adcontextprotocol/adcp/blob/v3.1.24/static/compliance/source/protocols/media-buy/scenarios/vendor_metric_optimization_flow.yaml Accessed 2026-09-28.

[34] IAB Tech Lab Agent & MCP Server Registry (browse page). IAB Tech Lab, 2026-09-29. https://registry.iabtechlab.ai/ Accessed 2026-09-29.

[35] AAO brand resolution records and brand.json checks for attention vendors. AgenticAdvertising.org; vendor domains, 2026-09-29. https://agenticadvertising.org/api/brands/resolve?domain=doubleverify.com Accessed 2026-09-29.

[36] OpenProposal — Object Model Specification (AAMP 3.0, draft for public comment), release v3.0-draft-1. IAB Tech Lab, 2026-09-24. https://github.com/IABTechLab/OpenProposal/blob/v3.0-draft-1/spec/openproposal-3.0.md Accessed 2026-09-28.

[37] PubMatic AgenticOS product page. PubMatic, undated. https://pubmatic.com/agenticos/ Accessed 2026-09-28.

[38] PubMatic Launches Decision Fabric on AgenticOS, Giving Partner Decision Models a Native Environment Inside the Programmatic Supply Path. PubMatic, 2026-06-01. https://pubmatic.com/news/pubmatic-launches-decision-fabric-on-agenticos/ Accessed 2026-09-28.

[39] DAIVID Creative Analyzer API (OpenAPI 3.1 specification and ReDoc page), version v1. DAIVID, undated (live at access). https://api.daivid.co/v1/openapi.json Accessed 2026-09-29.

[40] TripleLift Brings Agentic Intelligence to Cannes as Advertising Moves from Delivery to Impact. TripleLift (PR Newswire), 2026-06-18. https://www.prnewswire.com/news-releases/triplelift-brings-agentic-intelligence-to-cannes-as-advertising-moves-from-delivery-to-impact-302804320.html Accessed 2026-09-29.

[41] IAB and MRC Attention Measurement Guidelines, FINAL Version 1.0. Interactive Advertising Bureau (IAB) and Media Rating Council (MRC), 2025-11. https://www.iab.com/wp-content/uploads/2025/11/IAB_MRC_Attention_Measurement_Guidelines_November_2025.pdf Accessed 2026-09-28.

[42] MRC letter to DoubleVerify, Inc. (GI041326 DoubleVerify Accreditation Letter), with Encl. 1 Specific Accredited Statistics. Media Rating Council (George W. Ivie), 2026-04-13. https://mediaratingcouncil.org/sites/default/files/Accreditation%20Letters/DoubleVerify/GI041326_DoubleVerify%20Accreditation%20Letter.pdf Accessed 2026-09-28.

[43] What Are Attention Metrics? Definition & How They Work. Adelaide, undated. https://www.adelaidemetrics.com/attention-resources/the-fundamentals-of-attention-metrics Accessed 2026-09-28.

[44] Attention Technology: What Advertisers Need to Know. Lumen Research, undated. https://lumen-research.com/blog/attention-technology-advertising/ Accessed 2026-09-28.

[45] TVision Methodology Overview. TVision, March 10 (year not shown). https://www.tvisioninsights.com/resources/tvision-methodology-overview Accessed 2026-09-28.

[46] Measuring Gaining and Holding Attention to Social Media Ads with Viewport Logging: A Validation Study Using Mobile Eye-Tracking. Bruns, Kopka, Borgmann, Prior, Langner (University of Wuppertal), 2025-07-21. https://www.tandfonline.com/doi/full/10.1080/00913367.2025.2524186 Accessed 2026-09-28.

[47] DoubleVerify Holdings, Inc. Form 10-K for fiscal year 2025. DoubleVerify Holdings, Inc. (SEC filing), 2026-02-26. https://www.sec.gov/Archives/edgar/data/1819928/000110465926020499/dv-20251231x10k.htm Accessed 2026-09-28. Access: partial extract.

[48] Services Under Review: Digital. Media Rating Council, undated (live page; entries dated up to 9/2026). https://mediaratingcouncil.org/accreditation/services-under-review/digital Accessed 2026-09-28.

[49] Media Rating Council Accreditation Decisions, January 2026 through March 2026. Media Rating Council, 2026-04-17. https://mediaratingcouncil.org/sites/default/files/News/MRC%20Accreditation%20Status%20Quarterly%20Update_Q1%202026.pdf Accessed 2026-09-28.

[50] Services Under Review: Television. Media Rating Council, undated (live page). https://mediaratingcouncil.org/accreditation/services-under-review/television Accessed 2026-09-28.

[51] Attention Spillovers from News to Ads: Evidence from an Eye-Tracking Experiment. Simonov (Columbia), Valletti (Imperial College London), Veiga (Imperial College London), 2024 (online); Journal of Marketing Research 62(2): 294-315, 2025. https://assets.tina.io/a371bbc3-e9fe-472a-a32b-f55a8990b0bc/Simonov%20et%20al.%20Attention%20Spillovers%20from%20News%20to%20Ads-%20Evidence%20from%20an%20Eye-Tracking%20Experiment%20%20.pdf Accessed 2026-09-28.

[52] How Viewer Tuning, Presence and Attention Respond to Ad Content and Predict Brand Search Lift (MSI Working Paper 21-132). McGranaghan (Delaware), Liaukonyte (Cornell), Wilbur (UC San Diego), 2021-09. https://www.msi.org/wp-content/uploads/2021/09/MSI_Report_21-132.pdf Accessed 2026-09-28.

[53] Where Vision Meets Memory: An Eye-Tracking Study of In-App Ads in Mobile Sports Games with Mixed Visual-Quantitative Analytics. Büyükakgül, Yüce, Katırcı (Eskisehir Technical University), 2025-12-10. https://pmc.ncbi.nlm.nih.gov/articles/PMC12733859/ Accessed 2026-09-28.

[54] Consumer Neuroscience-Based Metrics Predict Recall, Liking and Viewing Rates in Online Advertising. Guixeres, Bigné, Ausín Azofra, Alcañiz Raya, Colomer Granero, Fuentes Hurtado, Naranjo Ornedo (Universitat Politècnica de València; Universitat de València), 2017-10-31. https://pmc.ncbi.nlm.nih.gov/articles/PMC5671759/ Accessed 2026-09-28.

[55] Eye Fixations on Advertisements and Memory for Brands: A Model and Findings. Wedel, Pieters (IDEAS/RePEc record), 2000. https://ideas.repec.org/a/inm/ormksc/v19y2000i4p297-312.html Accessed 2026-09-28. Access: abstract only.

[56] Sales Effects of Attention to Feature Advertisements: A Bayesian Mediation Analysis. Zhang, Wedel, Pieters (Tilburg University research portal record), 2009-10. https://research.tilburguniversity.edu/en/publications/sale-effects-of-attention-to-feature-advertisements-a-bayesian-me/ Accessed 2026-09-28. Access: abstract only.

[57] Attention Trajectories Capture Utility Accumulation and Predict Brand Choice. Martinovici, Pieters, Erdem (Journal of Marketing Research), 2022 (online); 2023. https://journals.sagepub.com/doi/10.1177/00222437221141052 Accessed 2026-09-28. Access: abstract only.

[58] Online webcam-based eye tracking in cognitive science: A first look. Semmelmann, Weigelt (Ruhr-University Bochum), 2017-06-07. https://link.springer.com/content/pdf/10.3758/s13428-017-0913-7.pdf Accessed 2026-09-28.

[59] Comparison of Webcam and Remote Eye Tracking. Wisiecka, Krejtz, Krejtz, Sromek, Cellary, Lewandowska (ETRA 2022), 2022. https://dl.acm.org/doi/10.1145/3517031.3529615 Accessed 2026-09-28. Access: abstract only.

[60] Putting Things into Context: A Meta-Analysis of Media Context Effects on Attitudinal Outcomes. Kwon, Nyilasy, King, Reid (Journal of Advertising), 2021-03-30. https://www.tandfonline.com/doi/full/10.1080/00913367.2021.1895920 Accessed 2026-09-28. Access: abstract only.

[61] Multimodal Physiological Responses Predict Advertisement Effectiveness. Han, Ni, Amon, Necaise, Han, Lee (Journal of Advertising 54(5)), 2025. https://doi.org/10.1080/00913367.2025.2524185 Accessed 2026-09-28. Access: abstract only.

[62] News UK and Lumen research shows editorial context drives advertising attention. Newsworks (publishing a News UK / Lumen Research release), 2022-03-22. https://newsworks.org.uk/news-and-opinion/news-uk-and-lumen-research-shows-importance-of-context-and-relevance/ Accessed 2026-09-28.

[63] New Adelaide Research: Nativo Inventory Delivers 40% Higher AU Scores Than Industry Benchmarks. Nativo (PR Newswire release), 2025-08-26. https://www.prnewswire.com/news-releases/new-adelaide-research-nativo-inventory-delivers-40-higher-au-scores-than-industry-benchmarks-302537887.html Accessed 2026-09-28.

[64] Adelaide Insights: AU & Brand Impact Correlations. Adelaide Metrics Inc. (Kaitlin Nizolek), 2020-03-18. https://www.adelaidemetrics.com/blog/au-and-brand-impact-correlations Accessed 2026-09-28.

[65] Lucid & Adelaide Prove Attention Metrics Drive Better Brand Outcomes Than Traditional Metrics. Adelaide Metrics Inc., 2022-03-16. https://www.adelaidemetrics.com/blog/lucid-adelaide-prove-attention-metrics-drive-better-brand-outcomes-than-traditional-metrics Accessed 2026-09-28.

[66] Who's Auditing the Metrics? MediaSense Just Did - and AU Passed. Adelaide Metrics Inc., 2025-07-01. https://www.adelaidemetrics.com/blog/whos-auditing-the-metrics-mediasense-just-did--and-au-passed Accessed 2026-09-28. Access: partial extract.

[67] iFIT boosts brand lift 61% by capitalizing on attention with Adelaide, not cheap reach. The Current (owned and operated by The Trade Desk Inc.), 2025-10-03. https://www.thecurrent.com/marketing-strategy-ifit-brand-lift-capitalizing-attention-adelaide Accessed 2026-09-28.

[68] Unveiling the Connection: Attention & Outcomes. A Brand Lift Meta-Analysis. Teads (Phil Summer) and Lumen (Erik Bogsnes), 2023-07. https://iabeurope.eu/wp-content/uploads/2023/07/Teads-x-Lumen-Attention-Whitepaper.pdf Accessed 2026-09-28.

[69] CPG Brand Drives 2X Higher Ad Exposure and 40% Lower CPM with Attention Bidding. Lumen Research, 2024-03-04. https://lumen-research.com/blog/cpg-brand-attention-bidding/ Accessed 2026-09-28.

[70] High-attention media delivers greater profits for advertisers. Newsworks (study by Newsworks, Lumen and Peter Field), 2025-09-25. https://newsworks.org.uk/news-and-opinion/high-attention-media-delivers-greater-profits-for-advertisers/ Accessed 2026-09-28.

[71] Powering Media Quality & Effectiveness: Case Study DV Authentic Attention (Mondelez). DoubleVerify, 2021. https://arf-research-dox.s3.amazonaws.com/Supplementary+Materials/DoubleVerify/DV_CaseStudy_AuthenticAttention_Mondelez_Whitepaper_2021.pdf Accessed 2026-09-28.

[72] Club Med Drove Stronger Brand Awareness and Brand Favorability with High-Attention Impressions. DoubleVerify, not stated. https://doubleverify.com/lp/case-study/web/verify/club-med-drove-stronger-brand-awareness-and-brand-favorability-with-high-attention-impressions-during-the-thats-lesprit-libre-campaign Accessed 2026-09-28. Access: partial extract.

[73] Attention Metrics & Measurement Solutions. Integral Ad Science, undated. https://integralads.com/solutions/attention/ Accessed 2026-09-28.

[74] TVision and Upwave Unveil 'The Impact of Attention on Brand Lift' Report. TVision and Upwave, 2022-07-18. https://www.tvisioninsights.com/resources/press-release-tvision-upwave Accessed 2026-09-28.

[75] Groundbreaking research from VCCP Media and Amplified proves brands with distinctive assets can drive results faster. Amplified (formerly Amplified Intelligence) and VCCP Media, 2025-05. https://www.amplified.co/insight/vccp-research-amplified-distinctive-assets Accessed 2026-09-28. Access: partial extract.

[76] Audience Group Sees 74% Brand Recall Lift with Attention. Amplified, not stated. https://www.amplified.co/case-studies/audience-group Accessed 2026-09-28.

[77] Spotify India - Measuring Attention That Moves the Needle. Amplified, not stated. https://www.amplified.co/case-studies/spotify-india-measuring-attention-that-moves-the-needle Accessed 2026-09-28.

[78] Vevo Partners With Adelaide to Launch 'Attention Guaranteed'. Vevo LLC, 2026-01-27. https://www.hq.vevo.com/press/vevo-partners-with-adelaide-to-launch-attention-guaranteed Accessed 2026-09-28.

[79] Tobii Acquires Sticky, A Leading Solution Provider for Market Research Studies Based on Webcam Eye Tracking. Tobii (press release via PR Newswire UK), 2017-05-17. https://www.prnewswire.co.uk/news-releases/tobii-acquires-sticky-a-leading-solution-provider-for-market-research-studies-based-on-webcam-eye-622743784.html Accessed 2026-09-28.

[80] GumGum Acquires Attention Measurement And Optimization Platform Playground xyz. GumGum (press release via GlobeNewswire), 2021-12-16. https://www.globenewswire.com/news-release/2021/12/16/2354020/0/en/GumGum-Acquires-Attention-Measurement-And-Optimization-Platform-Playground-xyz.html Accessed 2026-09-28.

[81] Adelaide Secures $1.4 Million Seed Expansion Led by Aperiam Ventures to Advance Attention Measurement Activation. Adelaide (company blog), 2024-08-19. https://www.adelaidemetrics.com/blog/adelaide-secures-1-4-million-seed-expansion-led-by-aperiamventures-to-advance-attention-measurement-activation Accessed 2026-09-28.

[82] Viant Technology Inc. Form 10-Q for the quarter ended June 30, 2026 (Note 6, TVision acquisition). Viant Technology Inc. (SEC filing), 2026-08-10. https://www.sec.gov/Archives/edgar/data/1828791/000182879126000072/dsp-20260630.htm Accessed 2026-09-28.

[83] Integral Ad Science Holding Corp. Form 8-K: completion of merger with Igloo Group Acquisition Company (Novacap). Integral Ad Science Holding Corp. (SEC filing), 2025-12-23. https://www.sec.gov/Archives/edgar/data/1842718/000114036125046475/ef20061707_8k.htm Accessed 2026-09-28.

[84] DoubleVerify Holdings, Inc. Form 8-K (Items 1.01, 5.02, 9.01): Agreement and Plan of Merger with Neptune BidCo US Inc. and Wallace Merger Sub Inc.. DoubleVerify Holdings, Inc. (SEC filing), 2026-08-10. https://www.sec.gov/Archives/edgar/data/1819928/000110465926092934/tm2621904d5_8k.htm Accessed 2026-09-28.

[85] Contextual Advertising Leader Introduces The GumGum Platform and Mindset Graph, Ushering in a New Era of Digital Advertising. GumGum (press release page), 2024-10-04. https://gumgum.com/press-releases/gumgum-platform-mindset-graph Accessed 2026-09-28.

[86] Realeyes Spins Out Adverteyes, a Creative Intelligence Company. Adverteyes / Realeyes (PRNewswire-PRWeb), 2025-11-18. https://www.prweb.com/releases/realeyes-spins-out-adverteyes-a-creative-intelligence-company-302618807.html Accessed 2026-09-29.

[87] Oracle Advertising EOL Frequently Asked Questions, Version 3.0. Oracle Corporation, 2024-08-12. https://www.oracle.com/a/ocom/docs/oracle-advertising-end-of-life-faqs.pdf Accessed 2026-09-28.

[88] Oracle Advertising EOL Frequently Asked Questions, Version 1 (copy hosted by Adweek). Oracle Corporation (document); hosted at static-prod.adweek.com, 2024-06-17. https://static-prod.adweek.com/wp-content/uploads/2024/06/oracle-advertising-end-of-life-faqs.pdf Accessed 2026-09-28.

[89] Media Buying Briefing: Attention metrics vie for the industry's, well, attention. Will it take root?. Digiday (Michael Bürgi), 2021-05-10. https://digiday.com/marketing/media-buying-briefing-attention-metrics-vie-for-the-industrys-well-attention-will-it-take-root/ Accessed 2026-09-28.

[90] The Wall Street Journal Offers Advertisers a Performance Guarantee Timed to the Election. Mark Stenberg, Adweek, 2024-09-05. https://www.adweek.com/media/wall-street-journal-adelaide-election-guarantee-advertisers/ Accessed 2026-09-28. Access: partial extract.

[91] WSJ promising ad guarantee during election. Talking Biz News, 2024-09-09. https://talkingbiznews.com/media-news/wsj-promising-ad-guarantee-during-election/ Accessed 2026-09-28.

[92] WSJ Finds People Paid More Attention to Ads During the Election (The WSJ Delivered On Its Election-Time Performance Guarantee). Mark Stenberg, Adweek, 2025-01-28. https://www.adweek.com/media/wall-street-journal-election-performance-guarantee-adelaide/ Accessed 2026-09-28. Access: partial extract.

[93] Case Study: Tech | EssenceMediacom & Kargo Guarantee AU. Adelaide, 2023-01-17. https://www.adelaidemetrics.com/case-studies/mediacom-kargo-guaranteed-au Accessed 2026-09-28.

[94] TripleLift Partners with Adelaide Launching First-Ever Attention-Based Buying Guarantee for CTV. TripleLift (PR Newswire), 2023-09-13. https://www.prnewswire.com/news-releases/triplelift-partners-with-adelaide-launching-first-ever-attention-based-buying-guarantee-for-ctv-301926477.html Accessed 2026-09-28.

[95] Why Attention Is No Longer Just a Metric—It's a Guarantee. Teads, 2025-05-07. https://www.teads.com/blog/guaranteeing-attention-connecting-outcomes/6467/ Accessed 2026-09-28.

[96] Teads Holding Co. Form 10-K for fiscal year ended 31 December 2025. Teads Holding Co. (CIK 0001454938), 2026-03-16. https://www.sec.gov/Archives/edgar/data/1454938/000145493826000015/tead-20251231.htm Accessed 2026-09-28.

[97] Duolingo Launches Attention Guarantees for Advertisers, Powered by Adelaide. Adelaide (Sam Gonzalez-Freedson), 2026-09-15. https://www.adelaidemetrics.com/blog/duolingo-launches-attention-guarantees-for-advertisers-powered-by-adelaide Accessed 2026-09-28.

[98] Adlook and Adelaide Unveil New AU-CPM Buying Model Enabling Advertisers to Transact on Attention Metrics. Adelaide (Jordan Weiers), 2023-04-12. https://www.adelaidemetrics.com/blog/adlook-and-adelaide-unveil-new-au-cpm-buying-model-enabling-advertisers-to-transact-on-attention-metrics Accessed 2026-09-28.

[99] SeenThis launches SeenThis Attention for digital video. SeenThis, 2026-06-30. https://seenthis.co/resources/news-article/seenthis-launches-seenthis-attention-for-digital-video Accessed 2026-09-29.

[100] SeenThis Attention (product page). SeenThis, undated (live page). https://attention.seenthis.co/ Accessed 2026-09-29.

[101] Guaranteed Attention Suite by Impact Media: Revolutionizing the Digital Advertising Industry with Media Products that Guarantee Human Attention. Impact Media (PR Newswire), 2024-04-18. https://www.prnewswire.com/news-releases/guaranteed-attention-suite-by-impact-media-revolutionizing-the-digital-advertising-industry-with-media-products-that-guarantee-human-attention-302120145.html Accessed 2026-09-29.

[102] 2023 Outlook Survey: Ad Spend, Opportunities, and Strategies for Growth. IAB, 2022-11. https://www.iab.com/wp-content/uploads/2022/11/IAB_2023_Outlook_Survey.pdf Accessed 2026-09-28.

[103] 2024 Outlook: A Snapshot into Ad Spend, Opportunities, and Strategies for Growth. IAB, 2023-11. https://www.iab.com/wp-content/uploads/2023/11/IAB_2024-Outlook_Survey.pdf Accessed 2026-09-28.

[104] 2025 Outlook Study (January 16, 2025). IAB, 2025-01-16. https://www.iab.com/wp-content/uploads/2025/01/IAB_Outlook_-Study_January_16_2025_v2.pdf Accessed 2026-09-29.

[105] Adelaide Attention Glossary. Adelaide, undated (live page). https://www.adelaidemetrics.com/adelaide-attention-glossary Accessed 2026-09-28.

[106] Beyond Impressions and GRPs: How To Measure Media Quality Across CTV and Linear TV With AU. Adelaide, 2026-07-29. https://www.adelaidemetrics.com/blog/ctv-linear-tv-attention-measurement Accessed 2026-09-28.

[107] Magellan AI and Adelaide Partner to Boost Podcast Advertising Effectiveness with Advanced Attention Measurement. Adelaide / Magellan AI, 2024-01-10. https://www.adelaidemetrics.com/blog/magellan-ai-and-adelaide-partner-to-boost-podcast-advertising-effectiveness-with-advanced-attention-measurement Accessed 2026-09-28.

[108] Adelaide Expands Adelaide for Publishers with Advanced Audio Capabilities, Welcoming Major Audio Partners. Adelaide, 2024-05-09. https://www.adelaidemetrics.com/blog/adelaide-expands-adelaide-for-publishers-with-advanced-audio-capabilities-welcoming-major-audio-partners Accessed 2026-09-28.

[109] Attention Metrics 101 | Media Quality Measurement. Adelaide, 2023-12-04. https://www.adelaidemetrics.com/attention-resources/revolutionizing-media-quality-measurement-with-attention-metrics Accessed 2026-09-28.

[110] Lumen launches new attention measurement solution for advertisers with Netflix. Lumen Research, 2026-03-05. https://lumen-research.com/press-releases/lumen-partners-with-netflix/ Accessed 2026-09-28.

[111] Lumen Research brings attention measurement to Netflix ads in five European markets. PPC Land, 2026-03 (article says the announcement is dated 5 March 2026). https://ppc.land/lumen-research-brings-attention-measurement-to-netflix-ads-in-five-european-markets/ Accessed 2026-09-28.

[112] Dentsu Partners With Lumen Research for First-Of-Its-Kind Audio Attention Study. dentsu / Lumen Research (release republished on Lumen's site), 2023-08-01. https://lumen-research.com/blog/lumen-audio-attention-economy/ Accessed 2026-09-28.

[113] Dentsu and Lumen Release Gaming & Advertising Attention Report 2024. dentsu / Lumen Research (release on Lumen's site), 2024-01-18 (release dateline 16 January 2024). https://lumen-research.com/blog/dentsu-and-lumen-release-in-game-advertising-report-with-twitch-ads-anzu-and-activision-blizzard/ Accessed 2026-09-28.

[114] 'The Attention Dividend' – premium large format DOOH commands 5x more attention than digital channels. Ocean Outdoor (with Lumen Research), 2025-06-11. https://oceanoutdoor.com/blog/the-attention-dividend-premium-large-format-dooh-commands-5x-more-attention-than-digital-channels Accessed 2026-09-28.

[115] IAS Introduces New Quality Attention Optimization and Social Attention Measurement. Integral Ad Science, 2024-12-17. https://integralads.com/news/quality-attention-optimization-social-attention-measurement/ Accessed 2026-09-28.

[116] DoubleVerify Debuts First-of-Its-Kind Attention Measurement for Social, Launching with Snap. DoubleVerify, 2025-06-30. https://doubleverify.com/company/newsroom/doubleverify-debuts-first-of-its-kind-attention-measurement-for-social-launching-with-snap Accessed 2026-09-28.

[117] DoubleVerify Expands DV Authentic Attention to TikTok as the First Badged TikTok Marketing Partner for Impression-Level Attention Insights. DoubleVerify, 2025-12-10. https://doubleverify.com/company/newsroom/doubleverify-expands-dv-authentic-attention-to-tiktok-as-the-first-badged-tiktok-marketing-partner-for-impression-level-attention-insights Accessed 2026-09-28.

[118] IAS Launches First-to-Market AI-Powered Social Attention Measurement for Snap. Integral Ad Science, 2025-06-30. https://integralads.com/news/snap-first-to-market-ai-powered-social-attention-measurement/ Accessed 2026-09-28.

[119] Boost Ad Performance with Attention Metrics (Campaign Strategy). Amplified, undated (live page). https://www.amplified.co/platform/campaign-strategy Accessed 2026-09-28.

[120] Attention for Audio. Amplified, undated (live page). https://www.amplified.co/attention-for-audio Accessed 2026-09-28.

[121] CESP Attention Grid (Published Methodology). CESP (Centre d'Etude des Supports de Publicite), 2026-06-12. https://www.cesp.org/wp-content/uploads/Attention-grid-2026-10-06-26-Published-Methodology.pdf Accessed 2026-09-29.

[122] xpln.ai (home page). xpln.ai, undated. https://xpln.ai/ Accessed 2026-09-28.

[123] Regulation (EU) 2026/1744 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 (Digital Omnibus on AI), OJ L series 2026/1744. European Parliament and Council of the European Union, 2026-07-24. http://publications.europa.eu/resource/celex/32026R1744 Accessed 2026-09-28.

[124] Commission Guidelines on prohibited artificial intelligence practices established by Regulation (EU) 2024/1689 (AI Act), C(2025) 5052 final. European Commission, 2025-07-29. https://ai-act-service-desk.ec.europa.eu/sites/default/files/2025-08/guidelines_on_prohibited_artificial_intelligence_practices_established_by_regulation_eu_20241689_ai_act_english_ied3r5nwo50xggpcfmwckm3nuc_112367-1.PDF Accessed 2026-09-28.

[125] Regulation (EU) 2016/679 (General Data Protection Regulation), OJ L 119, 4.5.2016. European Parliament and Council of the European Union, 2016-05-04. http://publications.europa.eu/resource/celex/32016R0679 Accessed 2026-09-28.

[126] Guidelines 2/2023 on Technical Scope of Art. 5(3) of ePrivacy Directive, Version 2.0. European Data Protection Board, 2024-10-07. https://www.edpb.europa.eu/system/files/2024-10/edpb_guidelines_202302_technical_scope_art_53_eprivacydirective_v2_en_0.pdf Accessed 2026-09-28.

[127] Biometric data guidance: Biometric recognition - Key data protection concepts. Information Commissioner's Office (UK), undated page. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/lawful-basis/biometric-data-guidance-biometric-recognition/key-data-protection-concepts/ Accessed 2026-09-28.

[128] Biometric Information Privacy Act, 740 ILCS 14. Illinois General Assembly, 2008 (P.A. 95-994, eff. 10-3-08). https://www.ilga.gov/Legislation/ILCS/Articles?ActID=3004&ChapterID=57 Accessed 2026-09-28.

[129] H.B. No. 149 (enrolled), Texas Responsible Artificial Intelligence Governance Act, amending Business & Commerce Code Section 503.001. Texas Legislature (89th Legislature, Regular Session), 2025. https://capitol.texas.gov/tlodocs/89R/billtext/html/HB00149F.htm Accessed 2026-09-28.

[130] California Civil Code Section 1798.140 (CCPA definitions). California Legislature, as amended by Stats. 2025, Ch. 67, Sec. 27 (AB 1170). https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?lawCode=CIV&sectionNum=1798.140. Accessed 2026-09-28.

[131] Text of Regulations (CCPA Updates, Cyber, Risk, ADMT, and Insurance Regulations), approved text. California Privacy Protection Agency, 2025-09-22 (OAL approval, per ATT-H-S34). https://cppa.ca.gov/regulations/pdf/ccpa_updates_cyber_risk_admt_appr_text.pdf Accessed 2026-09-28.

[132] RCW 19.373.010 Definitions (Washington My Health My Data Act). Washington State Legislature, 2023 (2023 c 191 s 3). https://app.leg.wa.gov/RCW/default.aspx?cite=19.373.010 Accessed 2026-09-28.

[133] RCW 19.375.010 Definitions (Washington biometric identifiers). Washington State Legislature, 2017 (2017 c 299 s 3). https://app.leg.wa.gov/RCW/default.aspx?cite=19.375.010 Accessed 2026-09-28.

[134] House Bill 24-1130, signed act, 'Concerning protecting the privacy of an individual's biometric data'. Colorado General Assembly, 2024. https://leg.colorado.gov/bill_files/42280/download Accessed 2026-09-28.

[135] Policy Statement of the Federal Trade Commission on Biometric Information and Section 5 of the Federal Trade Commission Act. Federal Trade Commission, 2023-05-18 (date from ATT-H-S17). https://www.ftc.gov/system/files/ftc_gov/pdf/p225402biometricpolicystatement.pdf Accessed 2026-09-28.

[136] PIPEDA Findings #2020-004: Joint investigation of The Cadillac Fairview Corporation Limited. Office of the Privacy Commissioner of Canada, OIPC Alberta, OIPC British Columbia, 2020-10-28. https://www.priv.gc.ca/en/opc-actions-and-decisions/investigations/investigations-into-businesses/2020/pipeda-2020-004/ Accessed 2026-09-28.

[137] Sentença, Ação Civil Pública nº 1090663-42.2018.8.26.0100, Idec v. Concessionária da Linha 4 do Metrô de São Paulo S.A. (ViaQuatro), 37ª Vara Cível, Foro Central Cível, TJSP. Tribunal de Justiça do Estado de São Paulo (Judge Patrícia Martins Conceição); copy hosted by Idec, 2021-05-07. https://idec.org.br/sites/default/files/doc_5._sentenca_viaquatro.pdf Accessed 2026-09-28.

[138] AdCP schema core/vendor-metric-value.json. AgenticAdvertising.org, 2026-09-23. https://github.com/adcontextprotocol/adcp/blob/v3.1.24/static/schemas/source/core/vendor-metric-value.json Accessed 2026-09-28.

[139] AdCP schema core/reporting-capabilities.json. AgenticAdvertising.org, 2026-09-23. https://github.com/adcontextprotocol/adcp/blob/v3.1.24/static/schemas/source/core/reporting-capabilities.json Accessed 2026-09-28.

[140] AdCP schema core/forecast-vendor-metric-value.json. AgenticAdvertising.org, 2026-09-23. https://github.com/adcontextprotocol/adcp/blob/v3.1.24/static/schemas/source/core/forecast-vendor-metric-value.json Accessed 2026-09-28.

[141] AgenticAdvertising.org agent registry API snapshot. AgenticAdvertising.org, 2026-09-28. https://agenticadvertising.org/api/registry/agents Accessed 2026-09-28.

[142] AdCP schema core/product.json. AgenticAdvertising.org, 2026-09-23. https://github.com/adcontextprotocol/adcp/blob/v3.1.24/static/schemas/source/core/product.json Accessed 2026-09-28.

[143] AdCP release notes (docs/reference/release-notes.mdx), sections Version 3.1.0: 'Delivery reporting — reach_window and viewability.viewed_seconds (#4618)' and 'Optimization goals — kind: vendor_metric (#4668)'. AgenticAdvertising.org, 2026-09-23. https://github.com/adcontextprotocol/adcp/blob/v3.1.24/docs/reference/release-notes.mdx Accessed 2026-09-28.

[144] AdCP schema enums/performance-standard-metric.json. AgenticAdvertising.org, 2026-09-23. https://github.com/adcontextprotocol/adcp/blob/v3.1.24/static/schemas/source/enums/performance-standard-metric.json Accessed 2026-09-28.

[145] AdCP schema core/performance-standard.json. AgenticAdvertising.org, 2026-09-23. https://github.com/adcontextprotocol/adcp/blob/v3.1.24/static/schemas/source/core/performance-standard.json Accessed 2026-09-28.

[146] AdCP issue #4579 'delivery-metrics.json missing fields for attention/duration optimization goals' and close comment. AdCP GitHub issue tracker (opened by yvassiliadis; closed with note by bokelley), 2026-05-15. https://github.com/adcontextprotocol/adcp/issues/4579 Accessed 2026-09-28.

[147] GitHub records for AAMP child repositories: release lists and code search for 'attention' in the IABTechLab organisation. IAB Tech Lab (GitHub), 2026-09-28. https://github.com/IABTechLab Accessed 2026-09-28.

[148] From Brief to Buy: How Butler/Till Used DV Neura to Embed Verification Into Agentic Workflows. DoubleVerify (Ben Cann, SVP Product Management), 2026-07-24. https://doubleverify.com/blog/omnichannel/optimize/from-brief-to-buy-how-butler/till-used-dv-neura-to-embed-verification-into-agentic-workflows Accessed 2026-09-28.

[149] What's new in AdCP 3.1 (docs/reference/whats-new-in-3-1.mdx). AgenticAdvertising.org, 2026-09-23. https://github.com/adcontextprotocol/adcp/blob/v3.1.24/docs/reference/whats-new-in-3-1.mdx Accessed 2026-09-28.

[150] AdCP schema files at v3.2.0-rc.7 (core/product.json, core/vendor-metric-value.json, core/reporting-capabilities.json, enums/indicator-type.json) and whats-new-in-3-2.mdx. AgenticAdvertising.org, 2026-09-27. https://github.com/adcontextprotocol/adcp/tree/v3.2.0-rc.7/static/schemas/source Accessed 2026-09-28.

[151] prebid/salesagent docs/test-obligations/storyboard-issue-map.yaml ('Storyboard -> GitHub issue map (AdCP 3.1.1)'). Prebid.org (prebid/salesagent, 'A media sales agent that implements the AdCP Media Buy protocol'), 2026-09-17. https://github.com/prebid/salesagent/blob/main/docs/test-obligations/storyboard-issue-map.yaml Accessed 2026-09-28.

[152] prebid/salesagent tests/bdd/dormant_scenarios.txt and tests/bdd/features/BR-UC-004-deliver-media-buy-metrics.feature. Prebid.org, 2026-09-17. https://github.com/prebid/salesagent/blob/main/tests/bdd/dormant_scenarios.txt Accessed 2026-09-28.

[153] Agentic Audiences Embedding Taxonomy (specs/v1.0/embedding-taxonomy.md). IAB Tech Lab (IABTechLab/agentic-audiences), 2026 (file date not retrieved; repository last pushed 2026-07-27). https://github.com/IABTechLab/agentic-audiences/blob/main/specs/v1.0/embedding-taxonomy.md Accessed 2026-09-28.

[154] IABTechLab/buyer-agent v2.4.2, src/ad_buyer/tools/reporting/stats_retrieval.py. IAB Tech Lab, 2026-08-18. https://github.com/IABTechLab/buyer-agent/blob/v2.4.2/src/ad_buyer/tools/reporting/stats_retrieval.py Accessed 2026-09-28.

[155] IAS launches 'IAS Agent' enabling marketers to surface deep campaign insights in real-time and increase ad performance. Integral Ad Science, 2025-12-16. https://integralads.com/news/ias-launches-ias-agent/ Accessed 2026-09-28.

[156] Omnicom First to Use Amplified Data via API in 2022 (case study). Amplified, undated. https://www.amplified.co/case-studies/omg-attention-planning-data-api Accessed 2026-09-28.

[157] Official MCP Registry API search results. Model Context Protocol registry (registry.modelcontextprotocol.io), 2026-09-28. https://registry.modelcontextprotocol.io/v0/servers?search=attention Accessed 2026-09-28.

[158] Third-party custom bidding algorithms (section 'Optimizing towards Adelaide's AU'). Google (Display & Video 360 Help Center), undated (live page; Adelaide section absent from the 9 June 2026 Wayback copy). https://support.google.com/displayvideo/answer/14210738?hl=en Accessed 2026-09-29.

[159] Adelaide AU in Custom Bidding (Display & Video 360 announcements). Google (Display & Video 360 Help Center), 2026 (undated page; copyright 2026). https://support.google.com/displayvideo/answer/17312629?hl=en Accessed 2026-09-28.

[160] Google puts Adelaide attention scores into DV360 bidding from mid-July. PPC Land, 2026-07-23. https://ppc.land/google-puts-adelaide-attention-scores-into-dv360-bidding-from-mid-july/ Accessed 2026-09-29.

[161] Buyers Can Now Target High-Attention Inventory In The Trade Desk. Anthony Vargas, AdExchanger, 2024-10-08. https://www.adexchanger.com/technology/buyers-can-now-target-high-attention-inventory-in-the-trade-desk/ Accessed 2026-09-28.

[162] Adelaide Brings Attention-Based Pre-Bid Segments to Amazon DSP, Helping Advertisers Reduce Waste and Prioritize High-Quality Inventory. Adelaide (Kaitlin Nizolek), 2026-06-09. https://www.adelaidemetrics.com/blog/attention-based-pre-bid-targeting-now-in-amazon-dsp Accessed 2026-09-28.

[163] Amazon DSP Adds Adelaide's Pre-Bid Attention Targeting. AdExchanger, 2026-06-10. https://www.adexchanger.com/technology/amazon-dsp-adds-adelaides-pre-bid-attention-targeting/ Accessed 2026-09-29.

[164] PubMatic & Adelaide Unveil First-to-Market Integration to Reinvent Attention-Based Media Quality Targeting. Adelaide Metrics Inc., 2024-12-11. https://www.adelaidemetrics.com/blog/pubmatic-adelaide-unveil-first-to-market-integration-to-reinvent-attention-based-media-quality-targeting Accessed 2026-09-29.

[165] Comscore and Adelaide Expand Access to High-Quality, Attention-Optimized Curated Deals across PubMatic's platform. Comscore, 2025-06-10. https://www.comscore.com/Insights/Press-Releases/2025/6/Comscore-and-Adelaide-Expand-Access Accessed 2026-09-28.

[166] Adelaide Joins Index Marketplaces, Bringing Attention-Based Media Quality To Sell-Side Curation. Adelaide (Sarah Powell), 2026-01-07. https://www.adelaidemetrics.com/blog/adelaide-joins-index-marketplaces-bringing-attention-based-media-quality-to-sell-side-curation Accessed 2026-09-28.

[167] Adelaide Debuts New AU Ecosystem Hub for Quality-Driven Advertising. Adelaide (Sam Gonzalez-Freedson), 2025-10-23. https://www.adelaidemetrics.com/blog/announcing-au-ecosystem-hub Accessed 2026-09-28.

[168] Beyond Viewability and Duration: How To Measure Media Quality Across Display and OLV With AU. Adelaide Metrics Inc. (Kristin Rose, Kaitlin Nizolek), 2026-07-30. https://www.adelaidemetrics.com/blog/display-olv-attention-measurement Accessed 2026-09-29.

[169] The Evolution of Gaze Detection: How Adelaide Is Raising the Standard. Adelaide Metrics Inc. (Claire Browne), 2025-03-10. https://www.adelaidemetrics.com/blog/the-evolution-of-gaze-detection-how-adelaide-is-raising-the-standard Accessed 2026-09-29.

[170] Adelaide's Attention Unit 'follows data science best practice', according to audit. The Media Leader (Jack Benjamin), 2025-07-02. https://uk.themedialeader.com/adelaides-attention-unit-follows-data-science-best-practice-according-to-audit/ Accessed 2026-09-29.

[171] MediaSense Audit: Validating the Predictive Power of AU (report landing page). Adelaide Metrics Inc., undated. https://www.adelaidemetrics.com/free-attention-guides/mediasense-au-effectiveness-and-transparency-review Accessed 2026-09-29. Access: partial extract.

[172] Form D, Notice of Exempt Offering of Securities: Adelaide Metrics Inc. (accession 0001937524-22-000001). Adelaide Metrics Inc. (SEC EDGAR filing), 2022-07-12. https://www.sec.gov/Archives/edgar/data/1937524/000193752422000001/primary_doc.xml Accessed 2026-09-29.

[173] Form D, Notice of Exempt Offering of Securities: Adelaide Metrics Inc. (accession 0001937524-24-000001). Adelaide Metrics Inc. (SEC EDGAR filing), 2024-08-01. https://www.sec.gov/Archives/edgar/data/1937524/000193752424000001/primary_doc.xml Accessed 2026-09-29.

[174] The Legacy of a Measurement Business is the Incentives it Creates ($7M seed announcement). Adelaide Metrics Inc. (Marc Guldimann), 2022-06-02 (structured data; no visible date). https://www.adelaidemetrics.com/blog/the-legacy-of-a-measurement-business-is-the-incentives-it-creates Accessed 2026-09-29.

[175] Introducing Custom AU. Adelaide (Marc Guldimann), 2025-11-05. https://www.adelaidemetrics.com/blog/introducing-custom-au Accessed 2026-09-28.

[176] Adelaide Introduces the AU Ecosystem for Creative Partners to Align Creative and Media Quality. Adelaide (Jordan Weiers), 2026-07-08. https://www.adelaidemetrics.com/blog/adelaide-au-ecosystem-for-creative-partners-to-align-creative-and-media-quality Accessed 2026-09-29.

[177] Services Under Review: Digital (Wayback copies 2 November 2024 and 9 June 2026). Media Rating Council (archived by the Internet Archive), 2026-06-09 (capture). http://web.archive.org/web/20260609000330/https://www.mediaratingcouncil.org/accreditation/services-under-review/digital Accessed 2026-09-29.

[178] Mobkoi and xpln.ai expand mobile attention deal, now 1bn impressions deep. PPC Land, 2026-06-21. https://ppc.land/mobkoi-and-xpln-ai-expand-mobile-attention-deal-now-1bn-impressions-deep/ Accessed 2026-09-29.

[179] Nielsen and Adelaide Launch the Industry's First Unified View of Reach and Media Attention. Nielsen, 2025-10-07. https://www.nielsen.com/news-center/2025/nielsen-and-adelaide-launch-the-industrys-first-unified-view-of-reach-and-media-attention/ Accessed 2026-09-29.

[180] Nielsen to Acquire DoubleVerify, Creating a Leading, Independent Media Intelligence Company. Nielsen, 2026-08-06. https://www.nielsen.com/news-center/2026/nielsen-to-acquire-doubleverify-creating-a-leading-independent-media-intelligence/ Accessed 2026-09-29.

[181] EDGAR submissions index, DoubleVerify Holdings, Inc. (CIK 0001819928). US Securities and Exchange Commission, live index. https://data.sec.gov/submissions/CIK0001819928.json Accessed 2026-09-29.

[182] Microsoft is contacting clients telling them it will close its DSP. Digiday, 2025-05-14. https://digiday.com/media-buying/microsoft-advertising-is-closing-the-xandr-dsp/ Accessed 2026-09-29.

[183] Adelaide Partners with Magnite, PubMatic, TRUSTX, Sovrn, Xandr, Google Display & Video 360, and Other Supply-side Partners to Launch the Industry's First High Attention Marketplaces Worldwide. Adelaide (Kaitlin Nizolek), 2023-03-07. https://www.adelaidemetrics.com/blog/adelaide-partners-with-magnite-pubmatic-trustx-sovrn-xandr-google-display-amp-video-360-and-other-supply-side-partners-to-launch-the-industrys-first-high-attention-marketplaces-worldwide Accessed 2026-09-28.

[184] Privacy Policy. Amplified (Amplified Intelligence), 2026-06. https://www.amplified.co/privacy-policy Accessed 2026-09-28.

[185] QMS: Proving OOH is more than a passing glance. Amplified, undated. https://www.amplified.co/case-studies/qms-ooh-attention-study Accessed 2026-09-29.

[186] In-Flight Attention Measurement for Campaigns (Campaign Performance). Amplified, undated (live page). https://www.amplified.co/platform/measurement Accessed 2026-09-29.

[187] Amplified vs Adelaide: Real human measurement vs modelled attention. Amplified, undated (live page). https://www.amplified.co/compare/adelaide Accessed 2026-09-29.

[188] Amplified’s Reach Curves Boost Smartly’s Attention Tools. Amplified, 2025-06-03. https://www.amplified.co/insight/smartly-partner-with-amplified Accessed 2026-09-29.

[189] ARF Attention Measurement Validation Initiative Phase 1 (compendium), company profile 'Amplified Intelligence Technologies', PDF pp. 39-44. Advertising Research Foundation (profile content supplied by the company through ARF's provider survey), 2023-09. https://thearf-org-unified-admin.s3.amazonaws.com/ARFReports/ARF_AttentionMeasurementValidationInitiativePhase1_2023_09.pdf Accessed 2026-09-29.

[190] IAB Europe TCF Global Vendor List (v3), vendor 1142. IAB Europe, 2026-09-24. https://vendor-list.consensu.org/v3/vendor-list.json Accessed 2026-09-29.

[191] Investing in the future of advertising (Investor centre). Amplified, undated (live page). https://www.amplified.co/investor Accessed 2026-09-29.

[192] Amplified Launches Ground-Breaking Predictive AI to Transform Creative and Media Performance. Amplified, 2026-05-20. https://www.amplified.co/insight/amplified-launches-attentionai Accessed 2026-09-29.

[193] Amplified Wrapped: From Thesis to Industry Standard (company timeline). Amplified, undated (live page; covers 2017 to 2026). https://www.amplified.co/timeline Accessed 2026-09-29.

[194] Yahoo & Amplified Bring Attention to Programmatic Buying. Amplified (media release with Yahoo quotes), 2023-10-03. https://www.amplified.co/insight/yahoo-amplified-partnership Accessed 2026-09-29.

[195] Amplified partners with Chalice Custom Algorithms. Amplified (release with Chalice quotes), 2023-10-09 (dateline; page date 10 October 2023). https://www.amplified.co/insight/chalice-partnership Accessed 2026-09-29.

[196] AttentionAI integration: put predictive attention data where your media decisions already happen. Amplified, undated (live page). https://www.amplified.co/attentionai-integration Accessed 2026-09-29.

[197] Connect secures exclusive African rights to AttentionAI, new predictive Amplified attention product. Amplified, 2026-07-14. https://www.amplified.co/insight/connect-secures-exclusive-african-rights-to-attentionai Accessed 2026-09-29.

[198] DOOH Attention Study Shows Brand and Sales Impact. Amplified, 2026-08-14 (page updated 2026-09-18). https://www.amplified.co/insight/amplified-qms-dooh-research Accessed 2026-09-29.

[199] Screenvision Proves Cinema Attention Drives Brand Impact. Amplified, undated. https://www.amplified.co/case-studies/screenvision-cinema-attention-study Accessed 2026-09-29.

[200] Australian brands are losing 63 cents in every social media dollar to low-attention formats. Amplified (with Snap Inc. ANZ quotes), 2026-09-24 (updated 2026-09-25). https://www.amplified.co/insight/australian-wastage-low-attention-formats Accessed 2026-09-29.

[201] Amplified appoints Anneliese Urquhart as CEO. Amplified, 2024-12-10 (dateline; page date 9 December 2024). https://www.amplified.co/insight/anneliese-urquhart-ceo-announcement Accessed 2026-09-29.

[202] Connected TV home screen ads achieve more attention and less wastage versus other media channels, new study. Amplified (with Samsung Ads and OMG Australia), 2025-11-19. https://www.amplified.co/insight/ctv-samsung-attention-study Accessed 2026-09-29.

[203] CreativeX’s New Datalink Integrates Amplified Attention Data to Revolutionise Creative Data Connectivity. Amplified (republished CreativeX release), 2025-12-04. https://www.amplified.co/insight/creativexs-new-datalink-integrates-amplified-attention-data-to-revolutionise-creative-data-connectivity Accessed 2026-09-29.

[204] Amplified Raises $3M in Oversubscribed Round Led by TEN13. Amplified, 2021-09-30. https://www.amplified.co/insight/global-attention-platform-amplified-raises-3m Accessed 2026-09-29.

[205] Media Rating Council Accreditation Decisions, April 2026 through June 2026. Media Rating Council, 2026-07-13. https://mediaratingcouncil.org/sites/default/files/News/MRC%20Accreditation%20Status%20Quarterly%20Update_Q2%2026%207.14.26.pdf Accessed 2026-09-28.

[206] DoubleVerify fourth quarter and full year 2025 results (Exhibit 99.1 to 8-K). DoubleVerify (press release furnished on Form 8-K), 2026-02-26. https://www.sec.gov/Archives/edgar/data/1819928/000110465926020508/dv-20260226xex99d1.htm Accessed 2026-09-28.

[207] DoubleVerify Holdings, Inc. Form 10-Q for the quarter ended June 30, 2026. DoubleVerify Holdings, Inc. (SEC filing), 2026-08-07. https://www.sec.gov/Archives/edgar/data/1819928/000110465926092375/dv-20260630x10q.htm Accessed 2026-09-28.

[208] DV Quarterly Benchmarks Q1 2026 (Quality and Attention Benchmark Report). DoubleVerify, 2026-05 (PDF created 2026-05-27). https://doubleverify.com/hubfs/content/DV_QualityandAttentionBenchmark_Q1_2026.pdf Accessed 2026-09-29.

[209] DV Authentic Attention - Exposure and Engagement. DoubleVerify, undated. https://doubleverify.com/products/advertisers/verify/authentic-attention Accessed 2026-09-28.

[210] Media Rating Council Accreditation Updates, January 2024 through March 2024. Media Rating Council, 2024-04-24. https://mediaratingcouncil.org/sites/default/files/News/050624_MRC_Accreditation_Status_Communication_FINAL.pdf Accessed 2026-09-29.

[211] DV Quarterly Benchmarks Q1 2025 (Quality and Attention Benchmark Report). DoubleVerify, 2025 (Q1 edition; data January-February 2025). https://s3.amazonaws.com/media.mediapost.com/uploads/DV_QualityandAttentionBenchmark_Q1_2025.pdf Accessed 2026-09-29.

[212] DoubleVerify's Guide to Attention Measurement (Attention Matters). DoubleVerify, 2023-02 (upload path; document copyright 2022). https://doubleverify.com/wp-content/uploads/2023/02/DV_Guide_AttentionMatters.pdf Accessed 2026-09-29.

[213] DoubleVerify and Scibids Unveil Industry-First AI-Powered Attention Solution for Advertisers: DV Algorithmic Optimizer. DoubleVerify, 2023-06-19. https://doubleverify.com/company/newsroom/doubleverify-and-scibids-unveil-industry-first-ai-powered-attention-solution-for-advertisers-dv-algorithmic-optimizer Accessed 2026-09-28.

[214] DoubleVerify Unveils Attention Offering for Publishers, Raising the Bar for Performance Optimization and Transparency. DoubleVerify, 2024-11-06. https://doubleverify.com/company/newsroom/doubleverify-unveils-attention-offering-for-publishers-raising-the-bar-for-performance-optimization-and-transparency Accessed 2026-09-29.

[215] Media Rating Council Accreditation Updates, October 2022 through December 2022. Media Rating Council, 2023-01-24. https://www.mediaratingcouncil.org/sites/default/files/News/Accreditation-Matters/012423%20MRC%20Accreditation%20Status%20Communication.pdf Accessed 2026-09-28.

[216] Nielsen to Acquire DoubleVerify, Creating a Leading, Independent Media Intelligence Platform (corrected Exhibit 99.1 to Form 8-K/A). Nielsen and DoubleVerify Holdings, Inc., 2026-08-07. https://www.sec.gov/Archives/edgar/data/1819928/000110465926092753/tm2621904d8_ex99-1.htm Accessed 2026-09-29.

[217] DoubleVerify Introduces DV Neura, the Dynamic AI Engine Powering the Future of Media Quality and Effectiveness. DoubleVerify, 2026-06-17. https://doubleverify.com/company/newsroom/doubleverify-introduces-dv-neura-the-dynamic-ai-engine-powering-the-future-of-media-quality-and-effectiveness Accessed 2026-09-29.

[218] DV Neura MCP endpoint, unauthenticated probe and OAuth protected-resource metadata. DoubleVerify (endpoint); observation by this review, live endpoint. https://mcp.doubleverify.com/mcp Accessed 2026-09-29. Access: partial extract.

[219] Viant Technology Closes Acquisition of TVision Insights. Viant Technology, 2026-05-05. https://www.viantinc.com/company/news/press-releases/viant-technology-closes-acquisition-of-tvision-insights/ Accessed 2026-09-28.

[220] DoubleVerify Partners With TVision to Expand Industry-Leading Attention Measurement in CTV. DoubleVerify, 2023-06-13. https://doubleverify.com/company/newsroom/doubleverify-partners-with-tvision-to-expand-industry-leading-attention-measurement-in-ctv Accessed 2026-09-28.

[221] DoubleVerify Holdings, Inc. Definitive Proxy Statement (DEFM14A) for the Special Meeting on the Merger. DoubleVerify Holdings, Inc., 2026-09-29. https://www.sec.gov/Archives/edgar/data/1819928/000110465926111962/tm2621904-12_defm14a.htm Accessed 2026-09-29.

[222] DoubleVerify Launches Programmatic Pre-Bid Attention Optimization Segments To Maximize Campaign Performance. DoubleVerify, 2023-05-04. https://doubleverify.com/company/newsroom/doubleverify-launches-programmatic-pre-bid-attention-optimization-segments-to-maximize-campaign-performance Accessed 2026-09-28.

[223] Infillion Confirms That It's Buying MediaMath's Assets. AdExchanger, 2023-08-24. https://www.adexchanger.com/programmatic/infillion-confirms-that-its-buying-mediamaths-assets/ Accessed 2026-09-29.

[224] DoubleVerify, IAS develop attention measurement solutions for Snap. Marketing Dive, 2025-07-02. https://www.marketingdive.com/news/doubleverify-ias-develop-attention-measurement-solutions-for-snap/752190/ Accessed 2026-09-29.

[225] Integral Ad Science: Dimensions and metrics. Funnel (Funnel Knowledge Base), undated live page. https://help.funnel.io/en/articles/8704186-what-data-can-i-get-from-integral-ad-science-ias Accessed 2026-09-29.

[226] IAS Extends Quality Attention Product to Publishers to Drive Yield and Maximize Revenue. Integral Ad Science, 2024-10-03. https://integralads.com/news/ias-extends-quality-attention-publishers-platforms/ Accessed 2026-09-29.

[227] How Stellantis Is Turning Attention Into a Competitive Advantage. Integral Ad Science, 2026-06-04 (modified 2026-08-05). https://integralads.com/insider/how-stellantis-is-turning-attention-into-a-competitive-advantage/ Accessed 2026-09-29. Access: partial extract.

[228] Toyota Spain Drives Stronger Conversions and Higher Attention with IAS Quality Attention. Integral Ad Science, 2025-11-03. https://integralads.com/insider/toyota-spain-drives-stronger-conversions-and-higher-attention/ Accessed 2026-09-29.

[229] IAS Announces First Attention Product to Unify Media Quality and Eye Tracking. Integral Ad Science, 2024-01-04. https://integralads.com/news/ias-announces-quality-attention-measurement-product-2/ Accessed 2026-09-29.

[230] Quality Attention Drives Brand Lift for Pernod Ricard (IAS and INTAGE). Integral Ad Science (IAS Japan), 2025-09-05. https://integralads.com/apac/insider/quality-attention-drives-brand-lift-for-pernod-ricard/ Accessed 2026-09-29. Access: partial extract.

[231] Integral Ad Science Holding Corp. Form 10-K for fiscal year 2024. Integral Ad Science Holding Corp. (SEC filing), 2025-02-28. https://www.sec.gov/Archives/edgar/data/1842718/000184271825000023/ias-20241231.htm Accessed 2026-09-28. Access: partial extract.

[232] IAS fourth quarter and full year 2024 financial results (Exhibit 99.1 to 8-K). Integral Ad Science (press release furnished on Form 8-K), 2025-02-28. https://www.sec.gov/Archives/edgar/data/1842718/000184271825000022/q424earningsrelease.htm Accessed 2026-09-28.

[233] Services Under Review: Digital (archived snapshot, 16 September 2025). Media Rating Council (via Internet Archive), 2025-09-16 (snapshot). https://web.archive.org/web/20250916044540/https://www.mediaratingcouncil.org/accreditation/services-under-review/digital Accessed 2026-09-29.

[234] Services Under Review: Digital (archived snapshot, 20 January 2026). Media Rating Council (via Internet Archive), 2026-01-20 (snapshot). https://web.archive.org/web/20260120190827/https://www.mediaratingcouncil.org/accreditation/services-under-review/digital Accessed 2026-09-29.

[235] Explaining Quality Attention. PPC Land, 2026-09-12. https://ppc.land/ias-quality-attention/ Accessed 2026-09-29.

[236] IAS Partners with Lumen Research to Further Bolster Attention Measurement. Integral Ad Science, 2023-04-30 (page footer; modified 2023-07-31). https://integralads.com/news/ias-lumen-attention-partnership/ Accessed 2026-09-29.

[237] IAS Announces Mobile In-App Support for First Attention Product to Unify Media Quality and Eye Tracking. Integral Ad Science, 2024-07-30. https://integralads.com/news/ias-expands-mobile-in-app-support-quality-attention-eye-tracking-media-quality/ Accessed 2026-09-29.

[238] IAS Announces Quality Attention Measurement Product. Integral Ad Science, 2023-06-15. https://integralads.com/news/ias-announces-quality-attention-measurement-product/ Accessed 2026-09-29.

[239] From media quality to revenue optimization: IAS and Mastercard are turning media signals into real sales performance. Integral Ad Science, 2026-03-26. https://integralads.com/insider/ias-and-mastercard-turn-media-quality-into-performance-engine/ Accessed 2026-09-29.

[240] Papyrus: Inside a Mobile Scheme Built to Fake Clicks, Scrolls, and Attention. Integral Ad Science (IAS Threat Lab), 2026-08-06. https://integralads.com/insider/papyrus-inside-a-mobile-scheme-built-to-fake-clicks-scrolls-and-attention/ Accessed 2026-09-29.

[241] IAS and Mastercard turn media quality into live sales signals. PPC Land, 2026-03-31. https://ppc.land/ias-and-mastercard-turn-media-quality-into-live-sales-signals/ Accessed 2026-09-29.

[242] How to connect Integral Ad Science (IAS). Funnel (Funnel Knowledge Base), 2026-08-05. https://help.funnel.io/en/articles/8943138-how-to-connect-integral-ad-science-ias Accessed 2026-09-29.

[243] DoubleVerify Debuts First-of-Its-Kind Attention Measurement for Social, Launching with Snap. DoubleVerify, 2025-06-30. https://ir.doubleverify.com/news-events/press-releases/press-releases-details/2025/DoubleVerify-Debuts-First-of-Its-Kind-Attention-Measurement-for-Social-Launching-with-Snap/default.aspx Accessed 2026-09-29.

[244] Integral Ad Science Holding Corp. Form 15-12G. Integral Ad Science Holding Corp. (SEC filing), 2026-01-02. https://www.sec.gov/Archives/edgar/data/1842718/000114036126000141/ef20061704_15-12g.htm Accessed 2026-09-29.

[245] EDGAR submissions record, Integral Ad Science Holding Corp. (CIK 0001842718). U.S. Securities and Exchange Commission (EDGAR), live record (latest filing listed 2026-02-12). https://data.sec.gov/submissions/CIK0001842718.json Accessed 2026-09-29.

[246] Ad testing (LINK+), Kantar Marketplace. Kantar, undated (footer copyright 2026). https://www.kantar.com/marketplace/solutions/ad-testing-and-development/ad-testing Accessed 2026-09-28.

[247] AI-powered ad testing (LINK AI), Kantar Marketplace. Kantar, undated (footer copyright 2026). https://www.kantar.com/marketplace/solutions/ad-testing-and-development/ai-powered-ad-testing Accessed 2026-09-28.

[248] Digital media optimisation (Context Lab), Kantar Marketplace. Kantar, undated (footer copyright 2026). https://www.kantar.com/marketplace/solutions/media-planning-and-effectiveness/digital-media-optimisation Accessed 2026-09-28.

[249] Kantar's new attention measurement framework helps brands cut through the noise. Kantar, 2023-11-14. https://www.kantar.com/press-center/kantars-new-attention-measurement-framework-helps-brands-cut-through-the-noise Accessed 2026-09-28.

[250] Beyond viewability: The role of attention in creative effectiveness. Kantar (Ecem Erdem, Global Creative Thought Leadership Manager), 2023-11-21. https://www.kantar.com/north-america/inspiration/agile-market-research/the-role-of-attention-in-creative-effectiveness Accessed 2026-09-29.

[251] Kantar APIs. Kantar, undated (live at access). https://www.kantar.com/campaigns/api Accessed 2026-09-29.

[252] Datalink Partner Spotlight: Kantar. CreativeX (Isobel Bruce), 2025-12-15. https://www.creativex.com/blog/datalink-partner-spotlight-kantar Accessed 2026-09-29.

[253] Kantar Annual Report and Accounts 2025 (Kantar Global Holdings S.à r.l.). Kantar Global Holdings S.à r.l., 2026-03-31. https://www.kantar.com/investor-relations/financial-announcement/-/media/E8C198AFCA004736AEE16E8E060F2D56.ashx Accessed 2026-09-29.

[254] Marketing Metric Certification – Compare Metrics with Confidence (MASB catalog of certified metrics). Marketing Accountability Standards Board (MASB), undated (live list; latest entry Jun 2026). https://themasb.org/marketing-metric-certification/ Accessed 2026-09-29.

[255] MASB Members and Participants. Marketing Accountability Standards Board (MASB), undated (live at access). https://themasb.org/membership/members/ Accessed 2026-09-29.

[256] Kantar introduces LINK+; the next generation of ad-testing. Kantar (press release), 2023-06-06. https://www.kantar.com/press-center/kantar-introduces-link-the-next-generation-of-ad-testing Accessed 2026-09-29.

[257] Accredited Services and Services Under Review (digital, television, radio, out-of-home, multimedia, outside the US). Media Rating Council (MRC), undated (live at access). https://mediaratingcouncil.org/accreditation Accessed 2026-09-29.

[258] From playback to payback: How attention impacts media and creative effectiveness. Kantar (Duncan Southgate, Global Creative & Media Lead), 2024-07-15. https://www.kantar.com/inspiration/advertising-media/how-attention-impacts-media-and-creative-effectiveness Accessed 2026-09-29.

[259] Attention beyond views for creative effectiveness. Kantar (Ecem Erdem, Global Creative Thought Leadership Manager), 2024-02-28. https://www.kantar.com/inspiration/advertising-media/attention-beyond-views-for-creative-effectiveness Accessed 2026-09-29.

[260] Kantar broadens LINK creative effectiveness solutions to include more media formats. Kantar, 2025-05-06. https://www.kantar.com/press-center/kantar-broadens-link-creative-effectiveness-solutions-to-include-more-media-formats Accessed 2026-09-28.

[261] Creative Testing and Optimization (Kantar North America). Kantar, undated (live at access). https://www.kantar.com/north-america/solutions/decision-intelligence/creative Accessed 2026-09-29.

[262] Can ad testing really predict sales impact?. Kantar (Duncan Southgate, Global Creative & Media Lead), 2022-09-25. https://www.kantar.com/inspiration/advertising-media/can-ad-testing-really-predict-sales-impact Accessed 2026-09-29.

[263] Shaping the future with efficiency and agility (Kantar Marketplace North America Agile Playbook, July 2025). Kantar, 2025-07. https://sites.kantar.com/marketing/KMP/2025/Kantar-Marketplace-NA_Agile_Playbook-July_2025.pdf Accessed 2026-09-29.

[264] Kantar's LINK suite closes the creator content effectiveness gap. Kantar (press release), 2026-06-02 (dateline 'LONDON: 2 JUNE'; page header dated 18 June 2026). https://www.kantar.com/press-center/kantars-link-suite-closes-the-creator-content-effectiveness-gap Accessed 2026-09-29.

[265] Kantar Completes Sale of Kantar Media to H.I.G. Capital. Kantar Global Holdings S.a r.l. (press release), 2025-08-04. https://www.kantar.com/press-center/kantar-completes-sale-of-kantar-media-to-hig-capital Accessed 2026-09-28.

[266] Affectiva and Kantar Renew Long-Term Partnership to Advance Attention and Emotional Insights in Advertising and Media. Affectiva, a Smart Eye company (press release via Business Wire), 2024-10-03. https://www.affectiva.com/news-item/affectiva-and-kantar-renew-long-term-partnership-to-advance-attention-and-emotional-insights-in-advertising-and-media/ Accessed 2026-09-29.

[267] Kantar Global Surveys Privacy Notice. Kantar, 2026-05. https://www.kantar.com/-/media/C19FDF6B4B4540F8B31067E2DFA8EE84.ashx Accessed 2026-09-29.

[268] Affectiva and the ARF Attention Initiative: Validation, and the Key Issue. Affectiva / Smart Eye (blog, go.smarteye.se), 2024-06 (inferred: undated page; embedded screenshot filename dated 2024-06-24). https://go.smarteye.se/affectiva-and-the-arf-attention-initiative-validation-and-the-key-issue Accessed 2026-09-29.

[269] WPP plc Annual Report on Form 20-F for fiscal year 2025. WPP plc (SEC filing), 2026-03-19. https://www.sec.gov/Archives/edgar/data/806968/000162828026019643/wpp-20251231.htm Accessed 2026-09-29.

[270] EP3861424B1: Collecting of points of interest on web-pages by eye-tracking. European Patent Office (via Google Patents), 2023-11-29. https://patents.google.com/patent/EP3861424B1/en Accessed 2026-09-29.

[271] Eye-tracking panel. Lumen Research, 2021-06-21. https://lumen-research.com/eye-tracking-panel/ Accessed 2026-09-29.

[272] Lumen Research is Granted Patent in the EU for Eye-Tracking Data Collection Platform. Lumen Research, 2024-02-28. https://lumen-research.com/blog/lumen-eye-tracking-platform/ Accessed 2026-09-29.

[273] IAS and Lumen Announce Launch of IAS Quality Attention. Integral Ad Science (PR Newswire release hosted by Lumen), 2024-01-04. https://lumen-research.com/blog/ias-lumen-quality-attention/ Accessed 2026-09-29.

[274] Unlock Deeper Attention Insights on Snapchat. Integral Ad Science, 2025-06-30. https://integralads.com/insider/unlock-deeper-attention-insights-snapchat/ Accessed 2026-09-29.

[275] Lumen Announces Native Billing Integration for Attention Metrics with The Trade Desk. Lumen Research, 2024-03-25. https://lumen-research.com/blog/lumen-announces-native-billing-integration-for-attention-metrics-with-the-trade-desk/ Accessed 2026-09-29.

[276] Teads and Lumen Launch First-to-Market Native Attention Integration. Teads (PR Newswire release on Lumen's site), 2023-06-14. https://lumen-research.com/blog/teads-and-lumen-launch-first-to-market-native-attention-integration-in-teads-ad-manager/ Accessed 2026-09-29.

[277] Teads and Lumen Research bring proven attention science to CTV for omnichannel impact. Teads (release on Lumen's site), 2026-05-06. https://lumen-research.com/white-papers/tv/teads-and-lumen-research-bring-proven-attention-science-to-ctv-for-omnichannel-impact/ Accessed 2026-09-29.

[278] Omnicom Media Group Expands Its Partnership with Lumen, Further Advancing Attention-Based Planning Capabilities Across Its Regions. Omnicom Media Group (PR Newswire APAC), 2024-05-24. https://www.prnewswire.com/apac/news-releases/omnicom-media-group-expands-its-partnership-with-lumen-further-advancing-attention-based-planning-capabilities-across-its-regions-302154918.html Accessed 2026-09-28.

[279] Creative Consulting v2 (Lumen SPOTLIGHT). Lumen Research, 2021-04-15. https://lumen-research.com/creative-consulting-v2/ Accessed 2026-09-28.

[280] Find out more spot light (SPOTLIGHT respondent experience and outputs). Lumen Research, 2021-04-12. https://lumen-research.com/find-out-more-spot-light/ Accessed 2026-09-29.

[281] Attention prediction model. Lumen Research, 2021-06-21. https://lumen-research.com/attention-prediction-model/ Accessed 2026-09-29.

[282] LAMP measurement tag. Lumen Research, 2021-04-10. https://lumen-research.com/lamp-measurement-tag/ Accessed 2026-09-29.

[283] Lumen Releases Report from PwC on the ROI of Attention Technology. Lumen Research, 2023-04-27. https://lumen-research.com/blog/lumen-releases-report-from-pwc-on-the-roi-of-attention-technology/ Accessed 2026-09-29.

[284] Accredited Services: Television; Accredited Services: Outside US. Media Rating Council, live page. https://mediaratingcouncil.org/accreditation/outside-us Accessed 2026-09-29.

[285] Lumen Research Limited, group of companies' accounts for the year ended 30 September 2025 (iXBRL). Lumen Research Limited (filed at UK Companies House), 2025-12-15. https://find-and-update.company-information.service.gov.uk/company/08682432/filing-history/MzQ5Mzk4MzU4NmFkaXF6a2N4/document?format=xhtml&download=1 Accessed 2026-09-28.

[286] Teads and Lumen Research Bring Proven Attention Science to CTV for Omnichannel Impact. Teads, 2026-05-06. https://www.teads.com/blog/teads-and-lumen-research-bring-proven-attention-science-to-ctv-for-omnichannel-impact/10475/ Accessed 2026-09-28.

[287] Lumen Research partners with Cint to strengthen the link between attention and brand outcomes. Lumen Research, 2026-04-20. https://lumen-research.com/press-releases/lumen-research-partners-with-cint-to-strengthen-the-link-between-attention-and-brand-outcomes/ Accessed 2026-09-29.

[288] Lumen Research Limited, amended unaudited financial statements for the period ended 31 December 2025 (PDF). Lumen Research Ltd (filed at UK Companies House), 2026-06-27. https://find-and-update.company-information.service.gov.uk/company/08682432/filing-history/MzUyNzI4MjYzOWFkaXF6a2N4/document?format=pdf&download=0 Accessed 2026-09-29.

[289] Nexxen Partners with Lumen and TVision to Deliver Its First High Attention CTV Solution. Nexxen (release on Lumen's site), 2023-08-08. https://lumen-research.com/blog/nexxen-lumen-ctv-attention-pmp/ Accessed 2026-09-29.

[290] LG Ad Solutions Partners with Lumen and TVision. LG Ad Solutions (release on Lumen's site), 2024-02-02. https://lumen-research.com/blog/lg-ad-solutions-partners-with-lumen-and-tvision/ Accessed 2026-09-29.

[291] Our Panel. TVision Insights, Inc., undated page. https://www.tvisioninsights.com/our-panel Accessed 2026-09-28.

[292] 2026 TV Attention & Creative Impact Report (PDF). TVision Insights, 2026-04-26 (release date of the report per TVision's post). https://info.tvisioninsights.com/hubfs/TVision-creative-report2026-final.pdf Accessed 2026-09-29.

[293] VideoAmp and TVision Expand License Agreement to Include CTV Co-viewing Data. TVision Insights, February 11 (year not shown; 2026 per the trade copy ATT-TVISION-S20). https://www.tvisioninsights.com/resources/videoamp-and-tvision-expand-license-agreement-to-include-ctv-co-viewing-data Accessed 2026-09-29.

[294] Global Partnership Between Adelaide and TVision Furthers Commitment to Omnichannel Attention Measurement and Enables Programmatic Activation. Adelaide, 2023-01-24. https://www.adelaidemetrics.com/blog/global-partnership-between-adelaide-and-tvision-furthers-commitment-to-omnichannel-attention-measurement-and-enables-programmatic-activation Accessed 2026-09-29.

[295] Lumen Research and TVision Announce Global Partnership. Lumen Research, 2023-04-21. https://lumen-research.com/blog/tvision-lumen-ctv-attention/ Accessed 2026-09-29.

[296] xpln.ai partners with TVision. Advanced Television, 2026-04-23 (from URL path). https://www.advanced-television.com/2026/04/23/xpln-ai-partners-with-tvision/ Accessed 2026-09-29.

[297] Data Licensing for TV Measurement. TVision Insights, undated (live page). https://www.tvisioninsights.com/data-licensing Accessed 2026-09-29.

[298] TVision's Total View Offers Linear, CTV Person-Level Measurement. Next TV (Broadcasting+Cable), 2023-04-17. https://www.nexttv.com/news/tvisions-total-view-offers-linear-ctv-person-level-measurement Accessed 2026-09-29.

[299] How Viewer Tuning, Presence, and Attention Respond to Ad Content and Predict Brand Search Lift (Marketing Science 41(5):873-895). Matthew McGranaghan, Jura Liaukonyte, Kenneth C. Wilbur, 2022-09. https://doi.org/10.1287/mksc.2021.1344 Accessed 2026-09-29. Access: abstract only.

[300] The Impact of Ad Repetition on Viewer Attention in Connected and Linear Television (Crossref metadata). Hyung Sup (Zack) Bhan, Kalyan Rallabandi and Seoungwoo Lee; Journal of Advertising Research, 2025-09-23 (published online). https://api.crossref.org/works/10.1080/00218499.2025.2538940 Accessed 2026-09-29. Access: abstract only.

[301] How Co-Viewing Influences Attention to TV Ads and Programs Across Connected TV (CTV) and Linear TV (Crossref metadata). Mi Hyun Lee and Jaewon Royce Choi; Journal of Advertising Research, 2025-12-18 (published online). https://api.crossref.org/works/10.1080/00218499.2025.2540660 Accessed 2026-09-29. Access: abstract only.

[302] Accredited Services: Television. Media Rating Council, undated (live list). https://mediaratingcouncil.org/accreditation/television Accessed 2026-09-29.

[303] Accredited Services: Multimedia. Media Rating Council, undated (live list). https://mediaratingcouncil.org/accreditation/multimedia Accessed 2026-09-29.

[304] Viant Technology Q2 2026 Earnings Call: Complete Transcript. Benzinga (transcript of Viant Technology's call of 10 August 2026), 2026-09-21. https://www.benzinga.com/news/26/09/61887625/viant-technology-q2-2026-earnings-call-complete-transcript Accessed 2026-09-29.

[305] Future of TV Briefing: OpenX adds attention targeting for CTV ads backed by TVision. Digiday, 2026-03-11. https://digiday.com/future-of-tv/future-of-tv-briefing-openx-adds-attention-targeting-for-ctv-ads-backed-by-tvision/ Accessed 2026-09-29.

[306] US Privacy Policy (TVision panel). TVision Insights, Inc., 2024-09-09 (updated; effective date 2018-12-18). https://www.mytvpanel.com/us-privacy-policy Accessed 2026-09-29.

[307] TVision Achieves ISO 27001 Certification Validating Compliance with Information Security Safeguards. TVision Insights, May 13 (year not shown; 2024 or 2025, since TVision's 13 June 2025 statement cites the certification). https://www.tvisioninsights.com/resources/iso-27001-cert Accessed 2026-09-29.

[308] Viant Technology Inc. Form 8-K, Exhibit 99.1: Viant Announces Agreement to Acquire TVision. Viant Technology Inc. (CIK 0001828791), 2026-04-15. https://www.sec.gov/Archives/edgar/data/1828791/000182879126000029/dsp-20260414xexx991.htm Accessed 2026-09-28.

[309] Viant Technology Inc. Form 8-K, Exhibit 99.1: Second Quarter 2026 results. Viant Technology Inc., 2026-08-10. https://www.sec.gov/Archives/edgar/data/1828791/000182879126000068/dsp-20260810xexx991.htm Accessed 2026-09-28.

[310] Havas & Viant Bring the Attention Marketplace to Life for Brands That Demand More. Havas Media Network, 2026-06-24. https://havasmedianetwork.com/news/havas-viant-bring-the-attention-marketplace-to-life-for-brands-that-demand-more/ Accessed 2026-09-29.

[311] TVision's Ad Scoreboard 2.0 Delivers Person-Level Attention, Presence and Co-viewing Data to Brands for Campaign Planning, Measurement and Optimization. TVision Insights, August 25 (year not shown; 2026 inferred from the boilerplate 'TVision, a Viant Technology company'). https://www.tvisioninsights.com/resources/tvisions-ad-scoreboard-2.0 Accessed 2026-09-29.

[312] The Nielsen Company (US), LLC v. TVision Insights, Inc., No. 2025-1371 (Fed. Cir.). United States Court of Appeals for the Federal Circuit (Dyk, Reyna, Bissoon), 2026-08-14. https://www.cafc.uscourts.gov/opinions-orders/25-1371.OPINION.8-14-2026_2739381.pdf Accessed 2026-09-29.

[313] TVision's Statement on the IAB/MRC 'Attention Taskforce' Draft Attention Measurement Guidelines. TVision Insights, June 13 (2025 inferred: the draft was released 12 May 2025 for comment to 12 July 2025). https://www.tvisioninsights.com/resources/tvisions-statement-on-the-iab/mrc-attention-taskforce-draft-attention-measurement-guidelines Accessed 2026-09-29.

[314] Viant Technology Announces First Quarter 2026 Financial Results (8-K Exhibit 99.1). Viant Technology Inc., 2026-05-11. https://www.sec.gov/Archives/edgar/data/1828791/000182879126000034/dsp-20260511xexx991.htm Accessed 2026-09-29.

[315] OpenX and TVision Launch First Supply-Side Attention Targeting Solution for CTV, Enabling Real-Time Activation on High-Engagement Inventory. OpenX Technologies (Business Wire), 2026-03-11. https://www.openx.com/press-releases/openx-and-tvision-launch-first-supply-side-attention-targeting-solution-for-ctv-enabling-real-time-activation-on-high-engagement-inventory/ Accessed 2026-09-28.

[316] Nielsen Loses At Fed. Circ., Fights Verdict In Patent Row. Law360 (Adam Lidgett), 2026-08-14. https://www.law360.com/articles/2513377/nielsen-loses-at-fed-circ-fights-verdict-in-patent-row Accessed 2026-09-29. Access: partial extract.

[317] Vevo Unveils New Ad Product Suite ‘Vevo Evolve’ for Advanced Targeting and Optimization. Vevo LLC, 2025-04-24. https://www.hq.vevo.com/press/vevo-unveils-new-ad-product-suite-vevo-evolve-for-advanced-targeting-and-optimization Accessed 2026-09-29.

[318] For Advertisers. Vevo LLC, undated live page. https://www.hq.vevo.com/for-advertisers Accessed 2026-09-29.

[319] A Vevo & Channel 4 study – Unlocking the Power of Premium. egta (business case text in Vevo's voice), 2025-04-17. https://www.egta.com/unlocking-the-power-of-premium/ Accessed 2026-09-29.

[320] Unlocking the Power of Premium (study deck, 2025). Vevo, Channel 4 and Amplified Intelligence (hosted by egta), 2025-04. https://www.egta.com/wp-content/uploads/2025/04/Unlocking-the-Power-of-Premium-2025.pdf Accessed 2026-09-29.

[321] Attention Please: Unlocking the Power of Premium on YouTube. Helen Myers, Advertising Week, 2025-04-02. https://advertisingweek.com/attention-please-unlocking-the-power-of-premium-on-youtube/ Accessed 2026-09-29.

[322] Vevo Names DISQO as Preferred Brand Outcomes Measurement Partner in the U.S.. Vevo LLC, 2026-05-20. https://www.hq.vevo.com/press/vevo-names-disqo-as-preferred-brand-outcomes-measurement-partner-in-the-u-s Accessed 2026-09-29.

[323] Vevo Partners With AudienceProject for Campaign Measurement. Vevo LLC, 2025-02-25. https://www.hq.vevo.com/press/vevo-partners-with-audienceproject-for-campaign-measurement Accessed 2026-09-29.

[324] Vevo Refines 2026 Upfront Strategy to Focus on Custom Solutions, Artist-Led Programming, and Flexible Buying Options Across Screens and Platforms. Vevo LLC, 2026-04-22 (dateline; page datePublished 2026-04-23). https://www.hq.vevo.com/press/vevo-refines-2026-upfront-strategy-to-focus-on-custom-solutions-artist-led-programming-and-flexibl Accessed 2026-09-29.

[325] The Guardian Launches Attention-Based Packages and Joins Adelaide's AU Ecosystem. Adelaide Metrics Inc., 2026-06-11. https://www.adelaidemetrics.com/blog/the-guardian-joins-au-ecosystem Accessed 2026-09-29.

[326] Vevo to Shutter Its Video Apps and Website in Coming Weeks. Marc Schneider, Billboard, 2018-05-24. https://www.billboard.com/pro/vevo-video-apps-website-shutting-down/ Accessed 2026-09-29.

[327] How We Use Data at Adelaide. Adelaide Metrics Inc. (Perry Papadopoulos), 2023-11-28 (structured data; text refers to the August 2024 Rita acquisition, so the page was updated later). https://www.adelaidemetrics.com/blog/how-we-use-data-at-adelaide Accessed 2026-09-29.

[328] Brands Achieve Significant Gains in Media Quality and Performance with Adelaide's Attention Pre-Bid Segments. Adelaide Metrics Inc. (Jordan Weiers), 2024-10-08. https://www.adelaidemetrics.com/blog/brands-achieve-significant-gains-in-media-quality-and-performance-with-adelaides-attention-pre-bid-segments Accessed 2026-09-29.

[329] Adelaide Adds Facebook, Instagram, and YouTube to AU Measurement Suite. Adelaide Metrics Inc. (Marc Guldimann), 2020-09-01. https://www.adelaidemetrics.com/blog/adelaide-adds-facebook-instagram-and-youtube-to-attention-unit-measurement-suite Accessed 2026-09-29.

[330] NewFronts, New Signals: Google Announces AU-Powered Custom Bidding in DV360. Adelaide (Sarah Powell), 2025-05-05. https://www.adelaidemetrics.com/blog/google-unveils-adelaide-attention-au-integration-for-dv360-custom-bidding Accessed 2026-09-28.

[331] Nielsen and Adelaide Launch the Industry's First Unified View of Reach and Media Attention. Adelaide (Kaitlin Nizolek), 2025-10-07. https://www.adelaidemetrics.com/blog/nielsen-adelaide-reach-attention-integration Accessed 2026-09-28.

[332] Adelaide's Upgraded Dashboard Delivers More Control, Efficiency, & Transparency. Adelaide Metrics Inc., 2025-02-27. https://www.adelaidemetrics.com/blog/adelaides-upgraded-dashboard-delivers-more-control-efficiency-transparency Accessed 2026-09-29.

[333] Attention firm Adelaide penetrates holding company media agencies with new tool. Digiday (Michael Bürgi), 2024-03-12. https://digiday.com/media-buying/attention-firm-adelaide-penetrates-holding-company-media-agencies-with-new-tool/ Accessed 2026-09-28.

[334] Yahoo Advertising Launches High Attention Pre-Bid Segments, Powered by Adelaide. Yahoo Inc., 2023-06-13. https://www.yahooinc.com/press/yahoo-advertising-launches-high-attention-pre-bid-segments-powered-by-adelaide Accessed 2026-09-28.

[335] Adelaide Announces AFP, Adelaide for Publishers, Including Free Attention Audits. The New York Times Joins As Launch Client.. Adelaide Metrics Inc., 2024-04-05 (dateline; structured data 2024-04-04). https://www.adelaidemetrics.com/blog/adelaide-announces-afp-adelaide-for-publishers-including-free-attention-audits-the-new-york-times-joins-as-launch-client Accessed 2026-09-29.

[336] Adelaide — Attention Metrics for Digital Media (homepage). Adelaide, undated (live page). https://www.adelaidemetrics.com/ Accessed 2026-09-28.

[337] Breaking Ground: Adelaide & Kantar Forge Game-Changing Partnership to Connect Attention Metrics with LIFT and LIFT+ Solutions. Adelaide Metrics Inc., 2024-05-01. https://www.adelaidemetrics.com/blog/breaking-ground-adelaide-kantar-forge-game-changing-partnership-to-connect-attention-metrics-with-lift-and-lift-solutions Accessed 2026-09-29.

[338] Unlocking video's full potential: an attention-based analysis of YouTube's premium ad solutions. Adelaide Metrics Inc. (study commissioned by Google), 2025-04-29. https://www.adelaidemetrics.com/case-studies-youtube-premium-attention-analysis Accessed 2026-09-29.

[339] Adelaide becomes the first attention metrics firm to enter the MRC accreditation process. Digiday (Michael Bürgi), 2024-10-29. https://digiday.com/media-buying/adelaide-becomes-the-first-attention-metrics-firm-to-enter-the-mrc-accreditation-process/ Accessed 2026-09-28.

[340] AU Ratings Privacy Policy. Adelaide Metrics Inc., 2024-09-13. https://www.adelaidemetrics.com/privacy-au-ratings Accessed 2026-09-28.

[341] Adelaide Expands Leadership in Media Quality Measurement with Strategic Acquisition of Rita. Adelaide (Marc Guldimann), 2024-08-19. https://www.adelaidemetrics.com/blog/adelaide-expands-leadership-in-media-quality-measurement-with-strategic-acquisition-of-rita Accessed 2026-09-28.

[342] Driving Responsible and Sustainable Brand Growth: Study by MAGNA Media Trials and Adlook Demonstrates the Power of Pre-Bid Optimization Across Multiple Metrics. MAGNA (IPG Mediabrands) and Adlook, 2023-07-19. https://magnaglobal.com/the-power-of-pre-bid-optimization-across-multiple-metrics/ Accessed 2026-09-29.

[343] Media Buying Briefing: Omnicom Media Group takes the next step in turning attention metrics into a valuable KPI for brands. Digiday (Michael Bürgi), 2021-11-22. https://digiday.com/media/media-buying-briefing-omnicom-media-group-takes-the-next-step-in-turning-attention-metrics-into-a-valuable-kpi-for-brands/ Accessed 2026-09-28.

[344] Our Capabilities (navigation 'Compare' list). Amplified, undated. https://www.amplified.co/our-capabilities Accessed 2026-09-28.

[345] Crossref REST API works query, author 'Nelson-Field' (60 rows). Crossref, 2026-09-29 (query date). https://api.crossref.org/works?query.author=Nelson-Field&rows=60&select=DOI,title,container-title,issued,author,type Accessed 2026-09-29.

[346] Why Does The Attention-Memory Threshold Matter?. Amplified, 2022-06-27. https://www.amplified.co/insight/why-does-the-attention-memory-threshold-matter Accessed 2026-09-29.

[347] Attention Metrics (using attention; solutions list). Amplified, undated. https://www.amplified.co/using-attention Accessed 2026-09-28.

[348] amplified.co sitemap.xml. Amplified, live file. https://www.amplified.co/sitemap.xml Accessed 2026-09-29.

[349] Attention by Amplified: Smarter Advertising (Platform). Amplified, undated. https://www.amplified.co/platform Accessed 2026-09-28.

[350] Seamless Attention Integrations (Integrations service). Amplified, undated (live page). https://www.amplified.co/services/integrations Accessed 2026-09-29.

[351] Amplified's attention platform now includes attention benchmark data. Amplified, 2023-12-05 (dateline; page date 4 December 2023). https://www.amplified.co/insight/attention-benchmark-data-now-available Accessed 2026-09-29.

[352] THE ICONIC Uses Live Attention in ‘Got You Looking’. Amplified, 2024-09-26. https://www.amplified.co/insight/the-iconic-love-media-attentionprove-live-attention-insights Accessed 2026-09-29.

[353] Our technology. Amplified, undated (live page). https://www.amplified.co/our-technology Accessed 2026-09-29.

[354] Alex Khan joins to lead US expansion as EVP, Global Partnerships. Amplified, 2022-11-07. https://www.amplified.co/insight/alex-khan-to-lead-us-expansion Accessed 2026-09-29.

[355] Attention Standardisation Is Here – Adopt Attention Before You're Left Behind. Amplified, undated (only a 23 July 2026 stamp shown, which is the site-wide update date on other posts; content refers to the draft guidelines, so it was likely written between May and November 2025, inference). https://www.amplified.co/insight/attention-standardisation-is-here Accessed 2026-09-29.

[356] The Attention Economy: A Category Blueprint (book page and table of contents). Karen Nelson-Field (Palgrave Macmillan Singapore / Springer Nature), 2024-10-27 (hardcover; eBook 2024-10-26; softcover 2025-10-27). https://link.springer.com/book/10.1007/978-981-97-0084-4 Accessed 2026-09-29. Access: partial extract.

[357] The Attention Economy and How Media Works: Simple Truths for Marketers (book page and table of contents). Karen Nelson-Field (Palgrave Macmillan Singapore / Springer Nature), 2020-01-05 (hardcover; eBook 2020-01-04). https://link.springer.com/book/10.1007/978-981-15-1540-8 Accessed 2026-09-29. Access: partial extract.

[358] Bespoke Attention Research (Research service). Amplified, undated (live page). https://www.amplified.co/services/research Accessed 2026-09-29.

[359] Pinterest x Amplified: Reframing Passive Attention as a Brand Power Play. Amplified, undated. https://www.amplified.co/case-studies/pinterest-x-amplified-case-study Accessed 2026-09-29.

[360] Havas Increased Lower-Funnel Outcomes by 83% with DV's Attention Measurement and Optimization. DoubleVerify, undated. https://doubleverify.com/lp/case-study/web/verify/havas-increased-lower-funnel-outcomes-by-83-with-dvs-attention-measurement-and-optimization Accessed 2026-09-29. Access: partial extract.

[361] DoubleVerify and Attain's Joint Solution Empowered Jack in the Box to Bridge the Gap Between Impression-Level Attention Measurement and Real-Time Sales Outcomes. DoubleVerify, undated. https://doubleverify.com/lp/case-study/web/verify/doubleverify-and-attains-joint-solution-empowered-jack-in-the-box-to-bridge-the-gap-between-attention-measurement-and-real-time-sales-outcomes Accessed 2026-09-29. Access: partial extract.

[362] DoubleVerify Launches DV Authentic Streaming TV at CES 2026. DoubleVerify, 2026-01-06. https://doubleverify.com/company/newsroom/doubleverify-launches-dv-authentic-streaming-tv-at-ces-2026 Accessed 2026-09-29.

[363] DV Scibids AI: Maximize Campaign ROI Without the Manual Lift (Optimize page). DoubleVerify, undated (live page). https://doubleverify.com/products/advertisers/optimize Accessed 2026-09-29.

[364] Channels / CTV: Drive Performance on Streaming Campaigns. DoubleVerify, undated (live page). https://doubleverify.com/channels/ctv Accessed 2026-09-29.

[365] Supercharge Your CTV Media Strategy with Actionable Attention Insights (product fact sheet landing page). DoubleVerify, undated. https://doubleverify.com/lp/product-fact-sheet/ctv/verify/supercharge-your-ctv-media-strategy-with-actionable-attention-insights Accessed 2026-09-29. Access: partial extract.

[366] DV Solutions for Commerce Media. DoubleVerify, undated (live page). https://doubleverify.com/channels/commerce-media Accessed 2026-09-29.

[367] DoubleVerify Research Reveals Retail Media is an Opportunity for Safe Engagement Despite Some Viewability Challenges. DoubleVerify, 2024-05-21. https://doubleverify.com/company/newsroom/doubleverify-research-reveals-retail-media-is-an-opportunity-for-safe-engagement-despite-some-viewability-challenges Accessed 2026-09-29.

[368] DoubleVerify Expands Global Measurement Capabilities To Include Amazon Custom Audiences. DoubleVerify, 2023-10-31. https://doubleverify.com/company/newsroom/doubleverify-expands-global-measurement-capabilities-to-include-amazon-custom-audiences Accessed 2026-09-29.

[369] DV Neura MCP Server (APIs.io listing). APIs.io (API Evangelist), undated. https://apis.io/apis/doubleverify/dv-neura-mcp-server/ Accessed 2026-09-29.

[370] DoubleVerify Launches First-of-its-Kind Global Attention Benchmark Report. DoubleVerify, 2023-12-19. https://doubleverify.com/company/newsroom/doubleverify-launches-first-of-its-kind-global-attention-benchmark-report Accessed 2026-09-29.

[371] DoubleVerify's 8th Annual Global Insights Report Highlights the Rising Importance of Attention Metrics. DoubleVerify, 2024-06-18. https://doubleverify.com/company/newsroom/doubleverifys-8th-annual-global-insights-report-highlights-the-rising-importance-of-attention-metrics Accessed 2026-09-28.

[372] DoubleVerify Solutions Privacy Notice. DoubleVerify Inc., 2025-10-09. https://doubleverify.com/privacy-notice Accessed 2026-09-28.

[373] DoubleVerify and Attain Launch Pioneering Joint Solution that Links Attention Metrics to Real-Time Sales Outcomes. DoubleVerify, 2023-10-18. https://doubleverify.com/company/newsroom/doubleverify-and-attain-launch-pioneering-joint-solution-that-links-attention-metrics-to-real-time-sales-outcomes Accessed 2026-09-29.

[374] Drive Proven Results with Quality Attention (IAS Insider). Integral Ad Science, 2024-01-04. https://integralads.com/insider/quality-attention-measurement/ Accessed 2026-09-29. Access: partial extract.

[375] Taking Action on Attention: Volume II. Integral Ad Science, 2024-07-11. https://integralads.com/insider/taking-action-on-attention-volume-ii/ Accessed 2026-09-29. Access: partial extract.

[376] HP Drives Attention and Results with IAS and Lumen Research. Integral Ad Science, 2023-10-18. https://integralads.com/insider/hp-attention-ias-lumen-case-study/ Accessed 2026-09-29. Access: partial extract.

[377] Quality Attention Boosts Brand Lift for Leading Beverage Company (IAS and Upwave). Integral Ad Science, 2024-05-06. https://integralads.com/insider/upwave-cpg-brand-attention-case-study/ Accessed 2026-09-29. Access: partial extract.

[378] Quality Attention Drives Sales Lift for Major CPG Brand (IAS and NCSolutions). Integral Ad Science, 2024-05-06. https://integralads.com/insider/ncsolutions-cpg-brand-attention-case-study/ Accessed 2026-09-29. Access: partial extract.

[379] Connected TV (CTV) Measurement Solutions. Integral Ad Science, modified 2026-05-11. https://integralads.com/solutions/ctv-video/ Accessed 2026-09-29.

[380] RESEARCH: The Mind on The Stream (IAS and Neuro-Insight). Integral Ad Science, 2024-08-21. https://integralads.com/insider/mind-on-stream-ctv-advertising-research/ Accessed 2026-09-29. Access: partial extract.

[381] IAS and Anzu Partner to Provide Media Quality Measurement for In-Game Advertising Environments. Integral Ad Science, 2022-07-28. https://integralads.com/news/ias-anzu-partner-for-in-game-advertising-environments/ Accessed 2026-09-29.

[382] IAS Enhances Signal UI with Unified View and Attention Metrics. Integral Ad Science, 2022-04-05 (structured data; modified 2026-01-29). https://integralads.com/insider/ias-enhances-signal-unified-view-attention-metrics/ Accessed 2026-09-29.

[383] Integral Ad Science Holding Corp. Form 10-Q for the quarter ended 30 September 2025. Integral Ad Science Holding Corp. (SEC filing), 2025-11-04. https://www.sec.gov/Archives/edgar/data/1842718/000184271825000109/ias-20250930.htm Accessed 2026-09-29.

[384] IAS Receives ISO 27001 Certification for Information Security Management. Integral Ad Science, 2024-05-22. https://integralads.com/news/ias-achieves-iso-certification-27001-information-security/ Accessed 2026-09-28.

[385] IAS Technology Solutions and Reporting Platform Privacy Policy (policy IASLG-17). Integral Ad Science, most recent revision 01.12.2023 (date format as printed). https://go.integralads.com/rs/469-VBI-606/images/IAS_Technical_Solutions_and_Platform_Privacy.pdf Accessed 2026-09-29.

[386] IAS Website Privacy Policy. Integral Ad Science, updated 2026-03-09 (modified 2026-04-30). https://integralads.com/ias-privacy-data-management/policies/privacy-policy/ Accessed 2026-09-29.

[387] Express your emotions: Why emotional response is important in advertising. Kantar (Duncan Southgate, Global Creative & Media Lead), 2022-11-29. https://www.kantar.com/inspiration/advertising-media/express-your-emotions-why-emotional-response-is-important-in-advertising Accessed 2026-09-29.

[388] Affectiva Announces New Emotion AI Attention Metric with Smart Eye's Sophisticated Eye Tracking Technology. Affectiva (Smart Eye company), press release, 2023-09-12. https://www.affectiva.com/news-item/affectiva-announces-new-emotion-ai-attention-metric-with-smart-eyes-sophisticated-eye-tracking-technology/ Accessed 2026-09-29.

[389] Kantar Creative Effectiveness Awards India 2026. Kantar, 2026 (undated page). https://www.kantar.com/campaigns/apac/kantar-creative-effectiveness-awards-india-2026 Accessed 2026-09-29.

[390] MMAP Endorsements. Marketing Accountability Standards Board (MASB), undated. https://themasb.org/mmap-metric-catalog/endorsements/ Accessed 2026-09-29.

[391] Can AI really decode creative effectiveness?. Kantar (Duncan Southgate, Global Creative & Media Lead), 2026-08-19. https://www.kantar.com/inspiration/agile-market-research/can-ai-really-decode-creative-effectiveness Accessed 2026-09-29.

[392] Media Buying - Lumen Research. Lumen Research, 2021-03-15 (dateModified 2024-04-01). https://lumen-research.com/media-buying/ Accessed 2026-09-28.

[393] Eye-tracking (How does the eye-tracking work?). Lumen Research Ltd, undated page. https://lumen-research.com/eye-tracking/ Accessed 2026-09-28.

[394] Our Technology (The Lumen Attention Technology Suite). Lumen Research, 2024-03-29. https://lumen-research.com/our-technology-v2/ Accessed 2026-09-29.

[395] Snapchat & Lumen Research Release Global Rich Media Attention Measurement Solution. Lumen Research, 2024-03-18. https://lumen-research.com/blog/snapchat-lumen-research-release-global-rich-media-attention-measurement-integration/ Accessed 2026-09-29.

[396] LAMP DSP plug in. Lumen Research, 2021-04-11. https://lumen-research.com/lamp-dsp-plug-in/ Accessed 2026-09-29.

[397] Lumen Launches Attentive Private Marketplaces with The MediaGrid. Lumen Research, 2023-05-31. https://lumen-research.com/blog/lumen-launches-attentive-private-marketplace-apmp-mediagrid-criteo/ Accessed 2026-09-29.

[398] Avocet & Lumen Research eye U.S. expansion following merger. Digiday, 2022-11-15. https://digiday.com/media-buying/avocet-lumen-research-eye-u-s-expansion-following-merger/ Accessed 2026-09-29.

[399] Silverpush and Epitaph release attention study with Lumen Research to prove performance of YouTube and context. Lumen Research, 2026-02-17. https://lumen-research.com/press-releases/silverpush-and-epitaph-release-attention-study-with-lumen-research-to-prove-performance-of-youtube-and-context/ Accessed 2026-09-29.

[400] Programmatic Attention Bidding: Lower Costs, Higher Rewards. Lumen Research, 2024-03-04. https://lumen-research.com/blog/programmatic-attention-bidding-lower-costs-higher-rewards/ Accessed 2026-09-29.

[401] Lumen Research Privacy Policy. Lumen Research Ltd, 2026-06. https://lumen-research.com/privacy/ Accessed 2026-09-29.

[402] Editorial context drives enhanced advertising attention. Mediashotz (syndicated copy of the News UK / Lumen release), 2022-03. https://mediashotz.co.uk/editorial-context-drives-enhanced-advertising-attention/ Accessed 2026-09-28.

[403] Our Technology. TVision Insights, undated (live page). https://www.tvisioninsights.com/our-technology Accessed 2026-09-29.

[404] ISpot Leads $16M VC Round In Panel Provider TVision To Measure Co-Viewing. AdExchanger, 2022-11-15. https://www.adexchanger.com/digital-tv/ispot-leads-16m-vc-round-in-panel-provider-tvision-to-measure-co-viewing/ Accessed 2026-09-29.

[405] xpln.ai and TVision team up to scale CTV attention data across channels. PPC Land (Luis Rijo), 2026-04-24. https://ppc.land/xpln-ai-and-tvision-team-up-to-scale-ctv-attention-data-across-channels/ Accessed 2026-09-29.

[406] CIMM and TVision Explore Potential for Age-Based Bias in Active Meter Audience Measurement Panels, Compared to Passive Meter Panels. TVision Insights (with CIMM), October 23 (year not shown). https://www.tvisioninsights.com/resources/tvision-active-passive Accessed 2026-09-29.

[407] Digital Activation. TVision Insights, undated (live page). https://www.tvisioninsights.com/digital-activation Accessed 2026-09-29.

[408] CTV Campaign Measurement. TVision Insights, undated (live page). https://www.tvisioninsights.com/ctv-campaign-measurement Accessed 2026-09-29.

[409] TV Planning & Measurement. TVision Insights, undated (live page). https://www.tvisioninsights.com/tv-planning-measurement Accessed 2026-09-29.

[410] CTV Activation. TVision Insights, undated (live page). https://www.tvisioninsights.com/ctv-activation Accessed 2026-09-29.

[411] See How People Really Watch TV (TVision home page). TVision Insights, undated (promotes 'The September 2026 State of Streaming'). https://www.tvisioninsights.com/ Accessed 2026-09-28.

[412] Calibration Data. TVision Insights, undated (live page). https://www.tvisioninsights.com/calibration-data Accessed 2026-09-29.

[413] TVision's New TV Attention and Creative Impact Report. TVision Insights, April 26 (year not shown; the report is titled 2026 and uses July-December 2025 data). https://www.tvisioninsights.com/resources/tvisions-new-tv-attention-and-creative-impact-report Accessed 2026-09-29.

[414] Website Privacy Policy. TVision Insights, Inc., 2022-08-25 (last updated). https://www.tvisioninsights.com/privacy-policy Accessed 2026-09-29.

[415] TVision - Get Paid to Watch TV (panel recruitment site). TVision Insights, undated (footer (c) 2023). https://www.mytvpanel.com/ Accessed 2026-09-29.

[416] Consumers Are 5X More Likely to View Brands Advertising Within Non-Premium Content as Low Quality. Vevo LLC (with Channel 4), 2023-09-28. https://www.hq.vevo.com/press/consumers-are-5x-more-likely-to-view-brands-advertising-within-non-premium-content-as-low-quality Accessed 2026-09-29.

[417] Vevo Launches Nostalgia-Based Buying Capability. Vevo LLC, 2026-07-09. https://www.hq.vevo.com/press/vevo-launches-nostalgia-based-buying-capability Accessed 2026-09-29.

[418] Zellmer v. Meta Platforms, Inc., No. 22-16925 (9th Cir.), Opinion. U.S. Court of Appeals for the Ninth Circuit (Judge R. Nelson), 2024-06-17. https://cdn.ca9.uscourts.gov/datastore/opinions/2024/06/17/22-16925.pdf Accessed 2026-09-28.

[419] Threats to the validity of eye-movement research in psychology. Orquin (Aarhus University), Holmqvist (North-West University; Masaryk University; Regensburg University), 2017-12-07. https://link.springer.com/content/pdf/10.3758/s13428-017-0998-z.pdf Accessed 2026-09-28.

[420] About - Brainsights. Brainsights, undated. https://www.brainsights.com/company/about Accessed 2026-09-28.

[421] Creative Optimization. Brainsights, undated. https://www.brainsights.com/solutions/creative-optimization Accessed 2026-09-28.

[422] Brainsights (home page). Brainsights, undated. https://www.brainsights.com/ Accessed 2026-09-28.

[423] ARF Attention Validation Measurement Initiative Reveals Majority of Attention Measurement Companies Focus on Both Creative and Media (Phase 1 release). Advertising Research Foundation (ARF), 2023-10-03. https://thearf.org/wp-content/uploads/2023/10/ARF-Attention-Validation_Phase-1-Findings-Release-FINAL.pdf Accessed 2026-09-28.

[424] Creative optimisation with AttentionAI (target of the /services/creative-testing redirect). Amplified, undated (live page). https://www.amplified.co/creative-optimisation-with-attentionai Accessed 2026-09-29.

[425] Attention Measurement: The Industry Framework for Measuring Attention. IAB, 2025-11-11. https://www.iab.com/guidelines/attention/ Accessed 2026-09-28.

[426] New State of CTV Advertising Report from TVision and DoubleVerify Shows Viewer Attention to CTV Advertising is on the Rise. TVision Insights, 2024-05-30. https://www.tvisioninsights.com/resources/press-release-ctv-q1-2024-dv Accessed 2026-09-29.

[427] UK Companies House: LUMEN RESEARCH LTD (08682432), overview, persons with significant control, filing history. UK Companies House, 2026-09-25. https://find-and-update.company-information.service.gov.uk/company/08682432 Accessed 2026-09-28.

[428] About Us - Lumen Research. Lumen Research, undated (article modified 2026-02-24). https://lumen-research.com/about-us/ Accessed 2026-09-28.

[429] AttentionAI. Amplified, undated. https://www.amplified.co/attentionai Accessed 2026-09-28.

[430] DRAFT IAB / MRC Attention Measurement Guidelines, Public Comment Version. IAB and MRC, 2025-05-12. https://www.iab.com/wp-content/uploads/2025/05/IAB_MRC_Attention_Measurement_Guidelines_for_Public_Comment_May_2025.pdf Accessed 2026-09-28.

[431] Guide to Attention in Digital Advertising (updated). IAB Europe, 2025-10. https://iabeurope.eu/wp-content/uploads/Guide-to-Attention.pdf Accessed 2026-09-28.

[432] Updated Guide to Attention in Digital Advertising (web page). IAB Europe, undated. https://iabeurope.eu/knowledge_hub/updated-guide-to-attention-in-digital-advertising/ Accessed 2026-09-28.

[433] DoubleVerify Continues Leadership in Attention Metrics, Earning MRC Accreditation for Attention Measurement Solution. DoubleVerify, 2023-01-04. https://doubleverify.com/company/newsroom/doubleverify-continues-leadership-in-attention-metrics-earning-mrc-accreditation-for-attention-measurement-solution Accessed 2026-09-28.

[434] Attention in Context: What Real Campaigns Reveal About Media Performance (ARF resource page). Advertising Research Foundation (ARF), 2026-06-04. https://thearf.org/category/ua_resource/attention-in-context-what-real-campaigns-reveal-about-media-performance/ Accessed 2026-09-28. Access: partial extract.

[435] Viant Technology Inc. Form 8-K (Items 2.02, 3.02, 8.01, 9.01): Agreement and Plan of Merger with TVision. Viant Technology Inc. (SEC filing), 2026-04-15. https://www.sec.gov/Archives/edgar/data/1828791/000182879126000029/dsp-20260414.htm Accessed 2026-09-28.

[436] EDGAR full-text search index (efts.sec.gov/LATEST/search-index): query log. U.S. SEC, 2026-09-28. https://efts.sec.gov/LATEST/search-index Accessed 2026-09-28.

[437] IAS to be Acquired by Novacap for $1.9 Billion to Further Support Its Strategic Goals (Exhibit 99.1 to 8-K). Integral Ad Science (press release furnished on Form 8-K), 2025-09-24. https://www.sec.gov/Archives/edgar/data/1842718/000119312525214050/d938971dex991.htm Accessed 2026-09-28.

[438] Attention Measurement Playbook for Marketers. CIMM and IAB, 2025-11. https://www.iab.com/wp-content/uploads/2025/11/CIMM_IAB_Attention_Measurement_Playbook_for_Marketers_November_2025.pdf Accessed 2026-09-28.

[439] New Study Helps Quantify the Impact of TV Ad Attention on Key Sales and Marketing Metrics. TVision Insights, 2021 (inferred: page says the academic study was 'published earlier this year'; no date on page). https://www.tvisioninsights.com/resources/attention-impacts-sales-metrics Accessed 2026-09-28.

[440] How emotion impacts CTV ad attention (report landing page). Wurl, 2025. https://try.wurl.com/reports/emotion-boosts-ctv-attention-report Accessed 2026-09-28. Access: partial extract.

[441] Haleon Says Attention Is A Good Measure Of Media Quality, But It's Too Soon To Make It A Buying Standard. AdExchanger (Anthony Vargas), 2024-02-22. https://www.adexchanger.com/measurement/haleon-says-attention-is-a-good-measure-of-media-quality-but-its-too-soon-to-make-it-a-buying-standard/ Accessed 2026-09-28.

[442] Contextual Ad Pioneer, GumGum, Merges Playground xyz's APAC Media Business, Creating a Powerhouse Ad Suite for the Mindset Era. GumGum (press release page), 2024-04. https://gumgum.com/press-releases/playgroundxyz-apac-integration Accessed 2026-09-28.

[443] GumGum results page (redirect target of gumgum.com/attention) and redirect observations. GumGum (web page); redirects observed by this review. https://gumgum.com/results Accessed 2026-09-28.

[444] SEC XBRL companyfacts: DoubleVerify Holdings, Inc. (CIK 0001819928), RevenueFromContractWithCustomerExcludingAssessedTax. U.S. SEC (structured data from the issuer's 10-K and 10-Q filings), 2026-08-07. https://data.sec.gov/api/xbrl/companyfacts/CIK0001819928.json Accessed 2026-09-28.

[445] TVision Insights Inc. Form D (new notice, Rule 506(b)). TVision Insights Inc. (SEC filing), 2022-11-23. https://www.sec.gov/Archives/edgar/data/1686186/000168618622000001/xslFormDX01/primary_doc.xml Accessed 2026-09-28.

[446] Life360 to Acquire Nativo, Accelerating Growth and Expanding Its Advertising Platform (Exhibit 99.3 to 8-K). Life360 (press release furnished on Form 8-K), 2025-11-10. https://www.sec.gov/Archives/edgar/data/1581760/000158176025000225/nexuspressrelease_11725x.htm Accessed 2026-09-28.

[447] Life360, Inc. Form 10-Q for the quarter ended March 31, 2026 (Note 7, Business Combinations). Life360, Inc. (SEC filing), 2026-05-11. https://www.sec.gov/Archives/edgar/data/1581760/000158176026000078/lifx-20260331.htm Accessed 2026-09-28.

[448] Outbrain Inc. Form 8-K (Items 5.03, 7.01, 8.01, 9.01): corporate name change to Teads Holding Co.. Outbrain Inc. / Teads Holding Co. (SEC filing), 2025-06-09. https://www.sec.gov/Archives/edgar/data/1454938/000095015725000478/form8-k.htm Accessed 2026-09-28.

[449] UK Companies House company search results for 'adverteyes'. UK Companies House, 2026-09-28. https://find-and-update.company-information.service.gov.uk/search/companies?q=adverteyes Accessed 2026-09-28.

[450] UK Companies House: REALEYES DATA SERVICES LIMITED (06286989), overview, persons with significant control, filing history. UK Companies House, 2026-08-13. https://find-and-update.company-information.service.gov.uk/company/06286989 Accessed 2026-09-28.

[451] SPARK Neuro homepage. SPARK Neuro (web page). https://sparkneuro.com/ Accessed 2026-09-28.

[452] Smart Eye Annual Report 2025. Smart Eye AB, 2026-04. https://smarteye.se/wp-content/uploads/2026/04/Smart-Eye-Annual-Report-2025.pdf Accessed 2026-09-28.

[453] Lumen Research Limited, group of companies' accounts for the period ended 31 December 2025 (iXBRL). Lumen Research Limited (filed at UK Companies House), 2026-03-10. https://find-and-update.company-information.service.gov.uk/company/08682432/filing-history/MzUwOTMyMDI1MWFkaXF6a2N4/document?format=xhtml&download=1 Accessed 2026-09-28.

[454] UK Companies House: DAIVID LIMITED (12583663), overview, persons with significant control, filing history. UK Companies House, 2026-05-15. https://find-and-update.company-information.service.gov.uk/company/12583663 Accessed 2026-09-28.

[455] Tobii Sticky: Online Eye Tracking Market Research Platform (product page). Tobii (web page). https://www.tobii.com/products/software/consumer-research-software/sticky Accessed 2026-09-28.

[456] About Amplified: Attention Economy Leaders. Amplified (web page). https://www.amplified.co/about-us Accessed 2026-09-28.

[457] Neuroscience startup Neurons Inc raises EUR 6M as one of the largest seed rounds in Denmark. Tech Funding News (reporting), 2022-08-25. https://techfundingnews.com/neuroscience-startup-neurons-inc-secures-e6m-in-one-of-the-largest-global-seed-rounds/ Accessed 2026-09-28.

[458] GumGum partners with On Device to power new Attentive Lift solution. GumGum, 2025-12-03. https://gumgum.com/press-releases/on-device-attentive-lift-au Accessed 2026-09-28.

[459] Campaign Results (target of the gumgum.com/attention redirect). GumGum, undated. https://gumgum.com/attention Accessed 2026-09-28.

[460] Realeyes Spins Out Adverteyes, a Creative Intelligence Company. Adverteyes (Colin Pye), 2025-11-26. https://adverteyes.ai/realeyes-spins-out-adverteyes-a-creative-intelligence-company/ Accessed 2026-09-28.

[461] Home - Realeyes. Realeyes, undated. https://www.realeyesit.com/ Accessed 2026-09-28.

[462] ABN Lookup: current details for ABN 21 636 344 360. Australian Business Register (Australian Government), live record (ABN last updated 25 Sep 2019). https://abr.business.gov.au/ABN/View?abn=21636344360 Accessed 2026-09-29.

[463] Amplified Reveals New Brand Identity for Growth. Amplified, 2025-04-29. https://www.amplified.co/insight/amplified-unveils-new-brand-identity Accessed 2026-09-29.

[464] Amplified acquires attention verification company AttentvAds. Amplified, 2022-12-13. https://www.amplified.co/insight/acquisition-of-attentvads Accessed 2026-09-29.

[465] Nielsen to Acquire DoubleVerify, Creating a Leading, Independent Media Intelligence Platform (Exhibit 99.1 to 8-K). DoubleVerify / Nielsen (joint press release furnished on Form 8-K), 2026-08-06. https://www.sec.gov/Archives/edgar/data/1819928/000110465926092121/tm2621904d1_ex99-1.htm Accessed 2026-09-28.

[466] DoubleVerify Holdings, Inc. preliminary proxy statement (PREM14A) for the merger. DoubleVerify Holdings, Inc., 2026-09-11. https://www.sec.gov/Archives/edgar/data/1819928/000110465926107126/tm2621904-10_prem14a.htm Accessed 2026-09-29.

[467] SEC XBRL companyfacts: Integral Ad Science Holding Corp. (CIK 0001842718), RevenueFromContractWithCustomerExcludingAssessedTax. U.S. SEC (structured data from the issuer's 10-K and 10-Q filings), 2025-11-04. https://data.sec.gov/api/xbrl/companyfacts/CIK0001842718.json Accessed 2026-09-28.

[468] Investment by Bain Capital Private Equity values Kantar at c.$4.0bn; WPP leverage target to be met a year ahead of plan; c.£1.0bn shareholder return planned. Bain Capital / WPP (party release), 2019-07-12. https://www.baincapital.com/news/investment-bain-capital-private-equity-values-kantar-c40bn-wpp-leverage-target-be-met-year Accessed 2026-09-29.

[469] LUMEN RESEARCH LTD, filing history. UK Companies House, live record. https://find-and-update.company-information.service.gov.uk/company/08682432/filing-history Accessed 2026-09-29.

[470] Viant Technology Inc. Prospectus Supplement (Form 424B5). Viant Technology Inc. (SEC filing), 2026-09-16. https://www.sec.gov/Archives/edgar/data/1828791/000162828026062297/viant-prosuppx424b5.htm Accessed 2026-09-29.

[471] Teads Holding Co. Form 10-Q for quarter ended 30 June 2026. Teads Holding Co., 2026-08-07. https://www.sec.gov/Archives/edgar/data/1454938/000145493826000054/tead-20260630.htm Accessed 2026-09-28.

[472] EDGAR full-text search: 'attention-adjusted', filings dated 2026-01-01 to 2026-09-29. U.S. Securities and Exchange Commission (EDGAR full-text search), 2026-09-29 (query date). https://efts.sec.gov/LATEST/search-index?q=%22attention-adjusted%22&dateRange=custom&startdt=2026-01-01&enddt=2026-09-29 Accessed 2026-09-29.

[473] Newsroom. Vevo LLC, live listing (latest item dated Jul 28, 2026). https://www.hq.vevo.com/newsroom Accessed 2026-09-29.

[474] SEC EDGAR full-text search results (queries logged in 'searches'). US Securities and Exchange Commission (efts.sec.gov), 2026-09-28 (query date). https://efts.sec.gov/LATEST/search-index?q=%22attention-adjusted%22&dateRange=custom&startdt=2026-01-01&enddt=2026-09-28 Accessed 2026-09-28.

[475] The ARF Attention Report Shows Promise and Opportunity for Attention Metric Adoption. TVision (Alex Banks, SVP Brand & Agency Partnerships), 2022-05-03. https://www.tvisioninsights.com/resources/arf-attention-report Accessed 2026-09-28.

[476] Media Buying Briefing: Holding companies and attention metrics, a KPI or a currency?. Digiday (Michael Bürgi), 2023-02-27. https://digiday.com/media-buying/media-buying-briefing-holding-companies-and-attention-metrics-a-kpi-or-a-currency/ Accessed 2026-09-28.

[477] IPG Mediabrands Partnership With Adelaide Activates New Era in Media Quality Management. Adelaide (Kaitlin Nizolek), 2024-04-25. https://www.adelaidemetrics.com/blog/ipg-mediabrands-partnership-with-adelaide-activates-new-era-in-media-quality-management Accessed 2026-09-28.

[478] WPP Media launches as fully integrated, AI-powered media company. WPP / WPP Media, 2025-05-28. https://www.wppmedia.com/news/announcing-wpp-media Accessed 2026-09-28.

[479] Omnicom Group Inc. Annual Report on Form 10-K for fiscal year 2025. Omnicom Group Inc. (SEC EDGAR, accession 0000029989-26-000006), 2026-02-20. https://www.sec.gov/Archives/edgar/data/29989/000002998926000006/omc-20251231.htm Accessed 2026-09-28.

[480] Omnicom Completes Acquisition of Interpublic, Forming the World's Leading Marketing and Sales Company. Omnicom Group Inc., 2025-11-26. https://www.omc.com/newsroom/omnicom-completes-acquisition-of-interpublic-forming-the-worlds-leading-marketing-and-sales-company-built-for-intelligent-growth-in-the-next-era/ Accessed 2026-09-28.

[481] Havas announces successful listing on Euronext Amsterdam, marking a new era of growth and innovation. Havas N.V., 2024-12-16. https://www.havas.com/press_release/havas-announces-successful-listing-on-euronext-amsterdam-marking-a-new-era-of-growth-and-innovation/ Accessed 2026-09-28.

[482] Dentsu drops sale of international arm but is open to deals in local markets. Campaign US (Nikita Mishra), 2026-02-13. https://www.campaignlive.com/article/dentsu-drops-sale-international-arm-open-deals-local-markets/1948417 Accessed 2026-09-28.

[483] Havas Media Group Becomes First Agency Network to Implement Attention Data into Planning Tools on a Global Scale. Havas Media Group release (republished by Marketing Communication News), 2022-11-17. https://marcommnews.com/havas-media-group-becomes-first-agency-network-to-implement-attention-data-into-planning-tools-on-a-global-scale/ Accessed 2026-09-28.

[484] Havas Unveils First Custom Attention Metric in Partnership with Adelaide (Meaningful Ad Unit). Adelaide (blog/press release), 2023-06-19. https://www.adelaidemetrics.com/blog/havas-media-network-partners-with-adelaide-in-north-america-to-build-industry-first-meaningful-ad-unit-mau Accessed 2026-09-28.

[485] Link Between Attention and Drivers of Brand Outcomes Revealed in New Research. Havas Media Network, 2024-09-16. https://havasmedianetwork.com/news/link-between-attention-and-drivers-of-brand-outcomes-revealed-in-new-research/ Accessed 2026-09-28. Access: partial extract.

[486] Dentsu UK&I launches real-time trading and measurement of effective attention in UK advertising industry first. dentsu UK&I, 2022-04-07. https://www.dentsu.com/uk/en/media-and-investors/dentsu-uk-and-i-launches-real-time-trading-and-measurement-of-effective-attention-in-uk-advertising-industry-first Accessed 2026-09-28.

[487] The Attention Economy | dentsu. dentsu (US site), 2024-01-12. https://www.dentsu.com/us/en/attention-economy Accessed 2026-09-28.

[488] Dentsu planning tool unifies ATL and BTL. MediaCat UK (Elliot Wright), 2026-03-19. https://mediacat.uk/dentsu-planning-tool-unifies-atl-and-btl/ Accessed 2026-09-28.

[489] How OMD's attention planning tool transforms campaign strategy. Adweek (Olivia Morley), unknown. https://www.adweek.com/agencies/how-omds-attention-planning-tool-transforms-campaign-strategy/ Accessed 2026-09-28. Access: partial extract.

[490] Introducing Open Intelligence from WPP Media. WPP Media, 2025-06-03. https://www.wppmedia.com/news/introducing-open-intelligence Accessed 2026-09-28.

[491] Publicis Groupe 2025 Universal Registration Document. Publicis Groupe S.A., 2026-04 (file name indicates 30 Apr 2026; inferred). https://documents.publicisgroupe.com/urd2025/PBS2025_URD_EN_MEL26_04_30.pdf?v2= Accessed 2026-09-28.

[492] Why Mars is more interested in measuring attention, not impressions. Digiday (Seb Joseph), 2020-03-12. https://digiday.com/marketing/mars-interested-measuring-attention-not-impressions/ Accessed 2026-09-28.

[493] Media Experts Redefine What Quality Means in AI World Dominated by Social Media and Digital Video, According to New IAS Industry Pulse Report. Integral Ad Science, 2025-12-08. https://integralads.com/news/2026-ias-industry-pulse-report/ Accessed 2026-09-28.

[494] Survey: The Attention Metrics Landscape in 2024. Lumen Research (blog), 2024 (exact date not shown). https://lumen-research.com/blog/survey-the-attention-metrics-landscape-in-2024/ Accessed 2026-09-28.

[495] Listing Metrics and KPIs in DOOH Advertising. Quividi, 2022-03-08. https://quividi.com/metrics-and-kpis-in-digital-out-of-home-advertising/ Accessed 2026-09-28.

[496] xpln.ai Launches in the US with CRO, Gina Cavallo, to Capture Demand for Next Gen Attention Solutions. ExchangeWire PressBox (xpln.ai release), 2026-01-29. https://www.exchangewire.com/blog/2026/01/29/xpln-ai-launches-in-the-us-with-cro-gina-cavallo-to-capture-demand-for-next-gen-attention-solutions/ Accessed 2026-09-29.

[497] Privacy and Cookies Policy. Adverteyes Inc., 2026-02-02 (effective date); sitemap lastmod 2026-06-19. https://adverteyes.ai/privacy-cookies/ Accessed 2026-09-29.

[498] Methodology. Adverteyes, undated; sitemap lastmod 2026-07-21. https://adverteyes.ai/company/methodology/ Accessed 2026-09-29.

[499] Human Measurement. Adverteyes, undated; sitemap lastmod 2026-07-20. https://adverteyes.ai/human-measurement/ Accessed 2026-09-29.

[500] Research Participant Privacy Policy, Online Research Platform (InContext Studies), Version 1.0, March 2026. eye square GmbH, 2026-03. https://www.eye-square.com/en/software-privacy-policy/ Accessed 2026-09-29.

[501] Privacy Policy Eye Tracking (Online Eye-Tracking Specific Privacy Policy). eye square GmbH, 2024-01-15. https://www.eye-square.com/en/privacy-policy-eye-tracking/ Accessed 2026-09-29.

[502] Privacy Statement | eye square. eye square GmbH, 2021-08-17. https://www.eye-square.com/en/privacy-statement/ Accessed 2026-09-29.

[503] eye square | SEAL - InContext Eye & Emotion Tracking. eye square GmbH, 2025-08-21. https://www.eye-square.com/en/incontext-eye-tracking/ Accessed 2026-09-29.

[504] eye square Privacy Policy (last updated Sept 2025). eye square GmbH, 2025-09. https://www.eye-square.com/en/privacy-policy/ Accessed 2026-09-29.

[505] Attorney General Ken Paxton Secures $1.4 Billion Settlement with Meta Over Its Unauthorized Capture of Personal Biometric Data. Office of the Attorney General of Texas, 2024-07-30. https://www.texasattorneygeneral.gov/news/releases/attorney-general-ken-paxton-secures-14-billion-settlement-meta-over-its-unauthorized-capture Accessed 2026-09-28.

[506] FTC Legal Library: Policy Statements (listing). Federal Trade Commission, listing as displayed on access date. https://www.ftc.gov/legal-library/browse/policy-statements Accessed 2026-09-28.

[507] Roberts v. Cooler Screens Incorporated, Class Action Complaint, Case No. 2022CH01824, Circuit Court of Cook County, Chancery Division. Plaintiff's counsel (filed with Cook County Circuit Clerk); copy hosted by Courthouse News Service, 2022-03-03. https://www.courthousenews.com/wp-content/uploads/2022/03/roberts-v-cooler-screens-complaint-cook-county.pdf Accessed 2026-09-28.

[508] Decision NAIH-85-3/2022 (earlier case number NAIH-7350/2021) concerning Budapest Bank Zrt.. Nemzeti Adatvédelmi és Információszabadság Hatóság (Hungarian National Authority for Data Protection and Freedom of Information), 2022-02-08. https://naih.hu/hatarozatok-vegzesek?download=517%3Amesterseges-intelligencia-alkalmazasanak-adatvedelmi-kerdesei Accessed 2026-09-28.

[509] Justiça multa ViaQuatro em R$ 500 mil por biometria facial sem consentimento. Consumidor Moderno (Matheus Ferreira), 2023-05-16. https://consumidormoderno.com.br/justica-multa-viaquatro-em-r-500-mil-por-biometria-facial-sem-consentimento/ Accessed 2026-09-28.

[510] Privacy Policy. iMotions A/S (Smart Eye group), 2022-06-22. https://imotions.com/privacy-policy/ Accessed 2026-09-28.

[511] Tobii Eye Tracking Data Transparency Policy. Tobii AB, undated page. https://www.tobii.com/company/tobii-eye-tracking-data-transparency-policy Accessed 2026-09-28.

[512] Company (About, Leadership, FAQs). Adverteyes, undated; sitemap lastmod 2026-09-04. https://adverteyes.ai/company/ Accessed 2026-09-29.

[513] Media Analytics - iMotions (iMotions Online study type page). iMotions A/S, 2026-06-26. https://imotions.com/products/imotions-online/media-analytics/ Accessed 2026-09-29.

[514] Facial Coding for Ad Testing - iMotions. iMotions A/S, 2025-03-05. https://imotions.com/solutions/facial-coding-for-ad-testing/ Accessed 2026-09-29.

[515] iMotions Online Study Type: Media Analytics (one-pager PDF). iMotions A/S, 2026-08-13. https://imotions.com/wp-content/uploads/brochures/Online%20-%20iMotions_Online_Media_Analytics_one-pager.pdf Accessed 2026-09-29.

[516] iMotions' stance on the AI Act. iMotions A/S, 2025-02-13. https://imotions.com/imotions-stance-on-the-ai-act/ Accessed 2026-09-29.

[517] Monitoring Viewer Attention During Online Ads (arXiv:2504.06237v1; presented at the ECCV 2024 Workshops). Mina Bishay, Graham Page, Waleed Emad, Mohammad Mavadati (Affectiva, A Smart Eye Company), 2025-04-08. https://arxiv.org/abs/2504.06237 Accessed 2026-09-29.

[518] Tobii Annual and Sustainability Report 2025. Tobii AB (publ), 2026 (FY2025 report; exact publication date not captured). https://storage.mfn.se/dec5df2f-b324-4b88-83c7-d4d5ee9207dd/tobii-annual-and-sustainability-report-2025.pdf Accessed 2026-09-29.

[519] A Comparison of the Performance of Webcam vs. Infrared Eye Tracking Technology. Liz Burton, William Albert, Mark Flynn (Proceedings of the Human Factors and Ergonomics Society Annual Meeting 58, pp. 1437-1441), 2014-09. https://doi.org/10.1177/1541931214581300 Accessed 2026-09-29. Access: abstract only.

[520] White Paper: Sticky by Tobii Data protection and privacy. Tobii (Tobii Connect knowledge base), 2026-03-23. https://connect.tobii.com/s/article/White-Paper-Sticky-by-Tobii-Data-protection?language=en_US Accessed 2026-09-29.

[521] Tobii Sticky's Analytics overview. Tobii (Tobii Connect knowledge base), 2026-03-27. https://connect.tobii.com/s/article/Sticky-analytics-interpretation-results-metrics?language=en_US Accessed 2026-09-29.

[522] Privacy Policy (Solutions Privacy Notice). XPLN.AI SAS, undated (no effective date in the retrieved text). https://xpln.ai/privacy Accessed 2026-09-29.

[523] xpln.ai Documentation: '1. What is xpln.ai ?'. xpln.ai, 2022-11-07 (Notion last-edited time). https://xpln.notion.site/7cadfdbe54ac49d7aa23093bae5221cf Accessed 2026-09-29.

[524] Austrian Standards certification database: certificate S 001201, eye square GmbH, ISO 20252:2019. Austrian Standards International, 2026-09-29. https://www.austrian-standards.at/en/products-solutions/apply-standards/certification/certification-database Accessed 2026-09-29.

[525] ISO 20252 | Certification | eye square (page and certificate image ISOCert.png). eye square GmbH, 2021-11-25. https://www.eye-square.com/en/iso-certification/ Accessed 2026-09-29.

[526] About Us | Implicit Market Research Berlin | eye square. eye square GmbH, 2023-05-22. https://www.eye-square.com/en/company/ Accessed 2026-09-29.

[527] Legal Information | eye square. eye square GmbH, 2015-05-28. https://eye-square.com/en/legal-information/ Accessed 2026-09-29.

[528] CCPA Updates, Cybersecurity Audits, Risk Assessments, Automated Decisionmaking Technology (ADMT), and Insurance Regulations (rulemaking page). California Privacy Protection Agency, 2025-09-22 (approval stated on page). https://cppa.ca.gov/regulations/ccpa_updates.html Accessed 2026-09-28.

[529] Adverteyes API documentation: Models. Adverteyes, undated. https://api.adverteyes.ai/Models.html Accessed 2026-09-29.

[530] Adverteyes API documentation: Legacy Metrics (Metric Dictionary). Adverteyes, undated. https://api.adverteyes.ai/Metrics.html Accessed 2026-09-29.

[531] Adverteyes API documentation: Insights Overview. Adverteyes, undated. https://api.adverteyes.ai/Insights.html Accessed 2026-09-29.

[532] Adverteyes API documentation: Attention and Emotions. Adverteyes, undated. https://api.adverteyes.ai/AttentionAndEmotions.html Accessed 2026-09-29.

[533] Adverteyes API documentation: Authentication. Adverteyes, undated. https://api.adverteyes.ai/Authentication.html Accessed 2026-09-29.

[534] Adverteyes API documentation: Changelog. Adverteyes, 2026-08-24 (latest entry). https://api.adverteyes.ai/Changelog.html Accessed 2026-09-29.

[535] API product page ('Pipe Creative Intelligence Into Your Own Stack'). Adverteyes, undated; sitemap lastmod 2026-07-20. https://adverteyes.ai/product/api/ Accessed 2026-09-29.

[536] Clinch Taps AI To Help Marketers Predict Which Ads Will Actually Work.. Inside Radio, 2025-12-29. https://www.insideradio.com/free/clinch-taps-ai-to-help-marketers-predict-which-ads-will-actually-work/article_f70d21fd-c872-41f2-9be7-dcf1a736cbdf.html Accessed 2026-09-29.

[537] Powering Predictive Creative Performance Within Flight Control. Adverteyes, 2026-01-06. https://adverteyes.ai/resource/powering-predictive-creative-performance-within-flight-control/ Accessed 2026-09-29.

[538] Adverteyes + CreativeX Datalink: Predicting Sales from Creative. Colin Pye, Adverteyes, 2026-02-17. https://adverteyes.ai/partnership-with-creativex-datalink/ Accessed 2026-09-29.

[539] Vidmob Integrates Adverteyes into its Creative Data Platform. Colin Pye, Adverteyes (originally Realeyes), 2024-06-06. https://adverteyes.ai/vidmob-integrates-adverteyes-into-its-creative-data-platform/ Accessed 2026-09-29.

[540] Nielsen Launches Outcomes Marketplace, Debuts Ad Attention Metrics from Realeyes to Expand Beyond Reach and Frequency. Nielsen, 2025-07-31. https://www.nielsen.com/news-center/2025/nielsen-launches-outcomes-marketplace-debuts-ad-attention-metrics-realeyes-to-expand-beyond-reach-and-frequency/ Accessed 2026-09-29.

[541] MCP product page ('Ask Adverteyes From Inside Your Own AI Tools'). Adverteyes, undated; sitemap lastmod 2026-08-11. https://adverteyes.ai/product/mcp/ Accessed 2026-09-29.

[542] Adverteyes MCP: Creative Intelligence, On Demand. Max Kalehoff, Adverteyes, 2026-07-21. https://adverteyes.ai/adverteyes-mcp-creative-intelligence-on-demand/ Accessed 2026-09-29.

[543] Welcome to the Ask Adverteyes MCP Server. Adverteyes, undated. https://mcp.adverteyes.ai/ Accessed 2026-09-29.

[544] As AI Ad Production Surges, DAIVID Launches New Tool To Help Advertisers Maintain Effectiveness Of Gen AI Creative. DAIVID (company blog release), 2025-06-24. https://blog.daivid.co/2025/06/24/as-ai-ad-production-surges-daivid-launches-new-tool-to-help-advertisers-maintain-effectiveness-of-gen-ai-creative/ Accessed 2026-09-29.

[545] Daivid launches gen AI ad analysis tool. Research Live (published by MRS), 2025-06-27. https://www.research-live.com/article/news/daivid-launches-gen-ai-ad-analysis-tool/id/5140413 Accessed 2026-09-29.

[546] DAIVID Brings Creative Intelligence To Retail Media Networks With AI-Powered Tool. DAIVID (company blog release), 2025-07-24. https://blog.daivid.co/2025/07/24/daivid-brings-creative-intelligence-to-retail-media-networks-with-ai-powered-tool/ Accessed 2026-09-29.

[547] DAIVID Launches New AI-Powered Solution That Enables Marketers To Benchmark Ad Effectiveness Against Competitors in Real Time. DAIVID (company blog release), 2025-08-15. https://blog.daivid.co/2025/08/15/daivid-launches-new-ai-powered-solution-that-enables-marketers-to-benchmark-ad-effectiveness-against-competitors-in-real-time/ Accessed 2026-09-29.

[548] DAIVID - AI-Powered Creative Intelligence Platform (website: home, Products & Use Cases, Real world alignment - 80%, Global, About pages). DAIVID, 2026-04-13. https://daivid.co/ Accessed 2026-09-29.

[549] DAIVID Launches New Tool to Help Advertisers Maintain Effectiveness of Gen AI Creative. ExchangeWire PressBox (DAIVID release), 2025-06-24. https://www.exchangewire.com/blog/2025/06/24/daivid-launches-new-tool-to-help-advertisers-maintain-effectiveness-of-gen-ai-creative/ Accessed 2026-09-29.

[550] Creative Data Provider DAIVID Gives Charities Free Access To AI Ad Testing Tool To Boost Campaign Impact On International Charity Day. DAIVID (company blog release), 2025-09-02. https://blog.daivid.co/2025/09/02/creative-data-provider-daivid-gives-charities-free-access-to-ai-ad-testing-tool-to-boost-campaign-impact-on-international-charity-day/ Accessed 2026-09-29.

[551] AdCP registry documentation (docs/registry/index.mdx). AgenticAdvertising.org, 2026-09-23. https://github.com/adcontextprotocol/adcp/blob/v3.1.24/docs/registry/index.mdx Accessed 2026-09-28.

[552] IABTechLab/AAMP README ('AAMP - Agentic Advertising Management Protocols'). IAB Tech Lab, 2026-09-22. https://github.com/IABTechLab/AAMP Accessed 2026-09-28.

[553] IAB Tech Lab Releases AAMP 2.3, Bringing Enterprise-Grade Infrastructure and Privacy Diligence to Agentic Advertising. IAB Tech Lab (PR Newswire), 2026-07-30. https://www.prnewswire.com/news-releases/iab-tech-lab-releases-aamp-2-3--bringing-enterprise-grade-infrastructure-and-privacy-diligence-to-agentic-advertising-302838652.html Accessed 2026-09-28.

[554] IAB Tech Lab Introduces AAMP 3.0. IAB UK, 2026-09-22. https://www.iabuk.com/news-article/iab-tech-lab-introduces-aamp-30 Accessed 2026-09-28.

[555] IAS Agent product page. Integral Ad Science, undated. https://integralads.com/ias-agent/ Accessed 2026-09-28.

[556] AdCP measurement taxonomy (docs/measurement/taxonomy.mdx). AgenticAdvertising.org, 2026-09-23. https://github.com/adcontextprotocol/adcp/blob/v3.1.24/docs/measurement/taxonomy.mdx Accessed 2026-09-28.

[557] (Author's work) The Metric That Was Never About You (slug which-metric-breaks-first). Evgeny Popov, No Fluff Advisory, 2026-07-10. https://nofluffadvisory.com/essays/which-metric-breaks-first/ Accessed 2026-09-28.

[558] Official MCP Registry API search results for 'amplified' and 'attentionai'. Model Context Protocol registry (registry.modelcontextprotocol.io), 2026-09-29 (query date). https://registry.modelcontextprotocol.io/v0/servers?search=amplified Accessed 2026-09-29.

[559] DV Neura: The Cognitive Engine for Modern Advertising (product page). DoubleVerify, undated (live page). https://doubleverify.com/products/neura Accessed 2026-09-29.

[560] DV Launches Ask Neura, Expanding Natural-Language Access to Media Intelligence. DoubleVerify, 2026-09-17. https://doubleverify.com/company/newsroom/dv-launches-ask-neura-expanding-natural-language-access-to-media-intelligence Accessed 2026-09-29.

[561] AdCP optimization and reporting (docs/media-buy/media-buys/optimization-reporting.mdx), 'Worked example: Adelaide attention end-to-end'. AgenticAdvertising.org, 2026-09-23. https://github.com/adcontextprotocol/adcp/blob/v3.1.24/docs/media-buy/media-buys/optimization-reporting.mdx Accessed 2026-09-28.

[562] Viomba MCP - Human Attention Intelligence for AI Agents (home page meta tags). Viomba, undated. https://www.viomba.com/ Accessed 2026-09-28. Access: partial extract.

[563] Standards and Guidelines. Media Rating Council, undated (live page). https://mediaratingcouncil.org/standards-and-guidelines Accessed 2026-09-28.

[564] Understanding Attention Metrics. Amplified, 2025-01-07. https://www.amplified.co/insight/blog-understanding-attention-metrics Accessed 2026-09-28.

[565] DoubleVerify Reports Second Quarter 2026 Financial Results (Exhibit 99.1). DoubleVerify Holdings, Inc., 2026-08-06. https://www.sec.gov/Archives/edgar/data/0001819928/000110465926092119/tm2621904d3_ex99-1.htm Accessed 2026-09-29.

[566] NewFronts: Teads Talks Tying Attention to Campaign Outcomes. Jon Lafayette, Broadcasting+Cable / Next TV, 2023-05-03. https://www.nexttv.com/news/newfronts-teads-talks-tying-attention-to-campaign-outcomes Accessed 2026-09-28.

[567] adcontextprotocol/adcp GitHub releases (tag list with publication times). AgenticAdvertising.org (AdCP maintainers), 2026-09-27. https://github.com/adcontextprotocol/adcp/releases Accessed 2026-09-28.

[568] DoubleVerify Expands DV Authentic AdVantage to Meta and TikTok, an AI-Powered Solution to Optimize Media Quality and Performance. DoubleVerify, 2026-06-22. https://doubleverify.com/company/newsroom/doubleverify-expands-dv-authentic-advantage-to-meta-and-tiktok-an-ai-powered-solution-to-optimize-media-quality-and-performance Accessed 2026-09-28.

[569] Sightly Partners with Adelaide to Enhance Attention-based Media Quality Measurement on Youtube. Adelaide / Sightly, 2024-10-07. https://www.adelaidemetrics.com/blog/sightly-partners-with-adelaide-to-enhance-attention-based-media-quality-measurement-on-youtube Accessed 2026-09-28.

[570] Quarterly Measurement Update. Adelaide (Sam Gonzalez-Freedson), 2022-02-11. https://www.adelaidemetrics.com/blog/quarterly-measurement-update Accessed 2026-09-28.

[571] xpln.ai Documentation: '3. Implementing xpln.ai Measurement'. xpln.ai, 2026-08-24 (Notion last-edited time). https://xpln.notion.site/860f9462233c4df58859f3d86f81662a Accessed 2026-09-29.

[572] AdCP compliance storyboard vendor_metric_accountability.yaml. AgenticAdvertising.org, 2026-09-23. https://github.com/adcontextprotocol/adcp/blob/v3.1.24/static/compliance/source/protocols/media-buy/scenarios/vendor_metric_accountability.yaml Accessed 2026-09-28.

[573] Creative + Media Quality: 5.3x Stronger Brand Lift. Adverteyes, 2026-05-01. https://adverteyes.ai/resource/creative-media-quality-brand-lift/ Accessed 2026-09-29. Access: partial extract.

[574] Neurons Attention Prediction (knowledge base). Neurons, undated (footer 2026). https://knowledge.neuronsinc.com/how-neurons-attention-prediction Accessed 2026-09-29.

[575] The making of our #attentionAI model. Neurons, 2021-10-20. https://www.neuronsinc.com/insights/the-making-of-our-attentionai-model Accessed 2026-09-29.

[576] Data Privacy policy (version 1.0.1). Neurons Inc, 2019-09-18 (last update stated on page). https://www.neuronsinc.com/legal/privacy-policy Accessed 2026-09-29.

[577] Neurons AI Tech Paper (PDF, 26 pages). Neurons, undated (content 'as of 2025'; cites DeepGaze MSDB, ICCV 2025, leaderboard figures). https://drive.google.com/file/d/1tFO3REKlOiqufhxCIrYLogQlhv72DBEG/view?usp=sharing Accessed 2026-09-29.

[578] Focus Attention: Lumen x YOC Study 2024. YOC with Lumen Research, 2024. https://yoc.com/api/media/file/YOC_Lumen-Study_Report2024.pdf Accessed 2026-09-29.

[579] Neurons | See Through the Eyes of Your Audience (home page). Neurons, undated. https://www.neuronsinc.com/ Accessed 2026-09-28.

[580] Media Analytics - iMotions. iMotions, undated. https://imotions.com/products/media-analytics/ Accessed 2026-09-28.

[581] DAIVID rolls out AI-powered creative data feed API for retail media. Mi3 (staff writer; AI-assisted rewrite of a DAIVID announcement), 2025-07-25. https://www.mi-3.com.au/25-07-2025/daivid-rolls-out-ai-powered-creative-data-feed-api-retail-media Accessed 2026-09-29.

[582] eye square (English home page). eye square GmbH, 2015-06-22. https://www.eye-square.com/en/ Accessed 2026-09-29.

[583] Dragonfly AI Creative Testing & Performance Insights (home page). Dragonfly AI, undated. https://dragonflyai.co/ Accessed 2026-09-29.

[584] EDO | The TV Outcomes Company (home page). EDO, undated. https://www.edo.com/ Accessed 2026-09-28.

[585] Upwave (home page). Upwave, undated. https://www.upwave.com/ Accessed 2026-09-28.

[586] Oracle Advertising (page request). Oracle Corporation. https://www.oracle.com/advertising/ Accessed 2026-09-28. Access: partial extract.

[587] Tremor International Group Rebrands as Nexxen. Nexxen (formerly Tremor International), 2023-06-12. https://nexxen.com/tremor-international-group-rebrands-as-nexxen/ Accessed 2026-09-29. Access: partial extract.

[588] DoubleVerify developer documentation portal (login and request-access page). DoubleVerify (observation by this verifier), live page. https://developer.doubleverify.com/ Accessed 2026-09-29. Access: partial extract.

[589] IAS and Snap Announce Expansion of Measurement Partnership to Validate Ad Effectiveness in Sponsored Snaps. Integral Ad Science, 2025-11-04. https://integralads.com/news/ias-and-snap-measurement-expansion-sponsored-snaps/ Accessed 2026-09-29.

[590] Criteo and Lumen Announce Two PMP Case Studies. Lumen Research, 2023-10-10. https://lumen-research.com/blog/criteo-and-lumen-announce-two-new-high-attention-pmp-case-studies/ Accessed 2026-09-29.

[591] Co-op ‘buys attention’ with Lumen. Lumen Research, 2019-10-30. https://lumen-research.com/blog/co-op-buys-attention-with-lumen/ Accessed 2026-09-29.

[592] Mesure de l'attention: Ogury s'associe a xpln.ai. ExchangeWire PressBox (Ogury release, French), 2023-10-02. https://www.exchangewire.com/blog/2023/10/02/mesure-de-lattention-ogury-sassocie-a-xpln-ai/ Accessed 2026-09-29.

[593] Beauty: More Attention, Higher Ad Recall (case study). xpln.ai, undated. https://xpln.ai/cases/beauty Accessed 2026-09-29.

[594] Beverages: More Attention, Better Brand Association (case study). xpln.ai, undated. https://xpln.ai/cases/beverages Accessed 2026-09-29.

[595] Financial Services: More Attention, More Qualified Visits (case study). xpln.ai, undated. https://xpln.ai/cases/financial-services Accessed 2026-09-29.

[596] Travel: More Attention, More Customers (case study). xpln.ai, undated. https://xpln.ai/cases/travel Accessed 2026-09-29.

[597] xpln.ai Documentation: '2. Getting Started with Measurement'. xpln.ai, 2024-09-13 (Notion last-edited time). https://xpln.notion.site/d6aa28500c7140a38d5826acbef4b47c Accessed 2026-09-29.

[598] xpln.ai & Comcast Advertising Introduces "Focus" an Attention-Based Media Package for Advertisers & Agencies. ExchangeWire PressBox (Comcast Advertising / xpln.ai release), 2025-09-09. https://www.exchangewire.com/blog/2025/09/09/xpln-ai-comcast-advertising-introduces-focus-an-attention-based-media-package-for-advertisers-agencies/ Accessed 2026-09-29.

[599] xpln.ai Enables Attention-Driven Programmatic Buying via Index Exchange. Index Exchange (newsroom; joint release with xpln.ai), 2026-02-12. https://www.indexexchange.com/press/press-releases/xpln-ai-enables-attention-driven-programmatic-buying-via-index-exchange/ Accessed 2026-09-29.

[600] Index Exchange embeds AI attention signals into SSP for pre-bid targeting. PPC Land (Luis Rijo), 2026-02-12. https://ppc.land/index-exchange-embeds-ai-attention-signals-into-ssp-for-pre-bid-targeting/ Accessed 2026-09-29.

[601] Mobkoi Accelerates Attention Measurement by Strengthening Its Global Partnership with xpln.ai. EIN Presswire (Mobkoi / xpln.ai release), republished by National Law Review, 2026-06-17. https://natlawreview.com/press-releases/mobkoi-accelerates-attention-measurement-strengthening-its-global Accessed 2026-09-29.

[602] Seedtag & xpln.ai Expand Strategic Partnership to Advance Attention-Based Advertising Performance. ExchangeWire PressBox (xpln.ai / Seedtag release), 2026-06-03. https://www.exchangewire.com/blog/2026/06/03/seedtag-xpln-ai-expand-strategic-partnership-to-advance-attention-based-advertising-performance/ Accessed 2026-09-29.

[603] CESP Attention Grid (Published, combined 8-page grid). CESP (Centre d'Etude des Supports de Publicite), 2026-06-12 (pp. 1-4) and 2026-06-10 (pp. 5-8). https://www.cesp.org/wp-content/uploads/Attention-grid-2026-10-06-26-Published.pdf Accessed 2026-09-29.

[604] Welcome to Adverteyes API (API documentation home). Adverteyes, undated; changelog latest entry 1.28.0, 2026-08-24. https://api.adverteyes.ai/ Accessed 2026-09-29.

[605] Adverteyes API documentation: Countries. Adverteyes, undated. https://api.adverteyes.ai/Countries.html Accessed 2026-09-29.

[606] Adverteyes API documentation: Heatmaps. Adverteyes, undated. https://api.adverteyes.ai/Heatmaps.html Accessed 2026-09-29.

[607] Adverteyes API documentation: Input Parameters. Adverteyes, undated. https://api.adverteyes.ai/InputParameters.html Accessed 2026-09-29.

[608] Prove Creative Performance: Adelaide + Adverteyes Creative Attention Measurement. Colin Pye, Adverteyes, 2026-03-02. https://adverteyes.ai/adelaide-reports-which-ads-drive-performance/ Accessed 2026-09-29.

[609] In-Store Scenes in Ads Drives 3% Sales Impact for US Retailers. Colin Pye, Adverteyes (originally Realeyes), 2025-02-20. https://adverteyes.ai/in-store-scenes-in-ads-drives-3-sales-impact-for-us-retailers/ Accessed 2026-09-29.

[610] Reminder, Test Your Ad Creative. Max Kalehoff, Adverteyes, 2026-03-18. https://adverteyes.ai/real-audience-ad-testing/ Accessed 2026-09-29.

[611] Adverteyes Attention Tech Featured in ARF Attention Validation. Colin Pye, Adverteyes (originally Realeyes), 2024-06-10. https://adverteyes.ai/adverteyes-attention-tech-featured-in-arf-attention-validation/ Accessed 2026-09-29.

[612] Did it beat the benchmark? How index metrics turn attention data into creative diagnosis. Amplified, 2026-08-28. https://www.amplified.co/insight/blog-beat-the-benchmark Accessed 2026-09-29.

[613] Test your ad with AttentionAI (free test form). Amplified, undated (live page). https://www.amplified.co/free-ad-test-attentionai Accessed 2026-09-29.

[614] Public DNS lookups for api.attentionai.amplified.co (Cloudflare DoH and Google DNS). Cloudflare (1.1.1.1) and Google Public DNS, 2026-09-29 (query date). https://cloudflare-dns.com/dns-query?name=api.attentionai.amplified.co&type=A Accessed 2026-09-29.

[615] Ads From Mutual of Omaha, Chevrolet and The Farmer's Dog Top Ranking Of December's Most Effective New U.S. TV Campaigns. DAIVID (company blog release, with Kinetiq), 2026-01-08. https://blog.daivid.co/2026/01/08/ads-from-mutual-of-omaha-chevrolet-and-the-farmers-dog-top-ranking-of-decembers-most-effective-new-u-s-tv-campaigns/ Accessed 2026-09-29.

[616] Kinetiq and DAIVID Launch Joint Monthly U.S. Report Showcasing the Top-Performing TV Ads. Kinetiq (joint release with DAIVID), 2025-11-06. https://kinetiq.tv/2025/11/06/kinetiq-and-daivid-launch-joint-monthly-u-s-report-showcasing-the-top-performing-tv-ads/ Accessed 2026-09-29.

[617] Creative Data Provider DAIVID Expands Into Japanese Ad Market Through Partnership With Video Research. DAIVID via PRNewswire-PRWeb, 2025-07-31. https://www.prweb.com/releases/creative-data-provider-daivid-expands-into-japanese-ad-market-through-partnership-with-video-research-302515838.html Accessed 2026-09-29.

[618] AdGreetz and DAIVID Partnership Brings AI-Generated, AI-Optimised Ad Campaigns to SMBs Worldwide. DAIVID (company blog release), 2025-09-25. https://blog.daivid.co/2025/09/25/adgreetz-and-daivid-partnership-brings-ai-generated-ai-optimised-ad-campaigns-to-smbs-worldwide/ Accessed 2026-09-29.

[619] DAIVID Joins Adelaide's New AU Ecosystem For Creative Partners, Connecting Creative Quality Directly To Media Decisions. DAIVID (company blog release), 2026-07-09. https://blog.daivid.co/2026/07/09/daivid-joins-adelaides-new-au-ecosystem-for-creative-partners-connecting-creative-quality-directly-to-media-decisions/ Accessed 2026-09-29.

[620] Datalink Partner Spotlight: DAIVID. CreativeX (Elissa Ha), 2026-03-05. https://www.creativex.com/blog/datalink-partner-spotlight-daivid Accessed 2026-09-29.

[621] DAIVID Launches New Solution That Analyses Thousands of Ads at Once to Connect Creative Performance to Business Results. ExchangeWire PressBox (DAIVID release; also on blog.daivid.co 2026-09-21), 2026-09-21. https://www.exchangewire.com/blog/2026/09/21/daivid-launches-new-solution-that-analyses-thousands-of-ads-at-once-to-connect-creative-performance-to-business-results/ Accessed 2026-09-29.

[622] DAIVID Launches AI-Powered Solution That Predicts Attention, Emotions, Memory and Brand Impact of Ads. DAIVID (company blog release), 2024-10-29. https://blog.daivid.co/2024/10/29/daivid-launches-ai-powered-solution-that-predicts-attention-emotions-memory-and-brand-impact-of-ads/ Accessed 2026-09-29.

[623] Sky Media Partners With DAIVID On Launch Of New AI-powered Creative Advertising Toolkit To Help SMEs Get On TV. DAIVID (company blog release), 2026-01-28. https://blog.daivid.co/2026/01/28/sky-media-partners-with-daivid-on-launch-of-new-ai-powered-creative-advertising-toolkit-to-help-smes-get-on-tv/ Accessed 2026-09-29.

[624] Former Ace Metrix CEO Peter Daboll Joins Global Creative Effectiveness Platform DAIVID As Head Of U.S. DAIVID (company blog release), 2025-01-07. https://blog.daivid.co/2025/01/07/former-ace-metrix-ceo-peter-daboll-joins-global-creative-effectiveness-platform-daivid-as-head-of-u-s/ Accessed 2026-09-29.

[625] DAIVID Chosen For CIMM's Inaugural Startup Program Supporting the Future of Media Measurement. DAIVID (company blog release), 2026-02-05. https://blog.daivid.co/2026/02/05/daivid-chosen-for-cimms-inaugural-startup-program-supporting-the-future-of-media-measurement/ Accessed 2026-09-29.

[626] CIMM Announces Inaugural Startup Program Members. Coalition for Innovative Media Measurement (CIMM), a subsidiary of the ARF, 2026-02-03. https://cimm-us.org/cimm-announces-inaugural-startup-program-members/ Accessed 2026-09-29.

[627] DAIVID platform Cookie Policy (text embedded in app.daivid.co/en). DAIVID, 2026-09. https://app.daivid.co/en Accessed 2026-09-29.

[628] How Attention Works (Dragonfly AI Academy, The Science). Dragonfly AI (James Harvey), 2025-10-30. https://academy.dragonflyai.co/en/articles/3276493-how-attention-works Accessed 2026-09-29.

[629] Dragonfly AI FAQs: Creative Testing Questions & Platform Guide. Dragonfly AI, undated (current at access). https://dragonflyai.co/faqs Accessed 2026-09-29.

[630] The Science Behind Our Biological Algorithm. Dragonfly AI, undated. https://dragonflyai.co/our-science Accessed 2026-09-29.

[631] The science behind Dragonfly AI: a London AI success story. Dragonfly AI (Priyanka Jethwa), 2026-09-02. https://dragonflyai.co/resources/blog/the-science-behind-dragonfly-ai-a-london-ai-success-story Accessed 2026-09-29.

[632] MIT Saliency Benchmark: MIT300 results. MIT Saliency Benchmark (Bylinskii, Judd, Borji, Itti, Durand, Oliva, Torralba), undated (page notes performance numbers prior to 25 September 2014 and later additions). http://saliency.mit.edu/results_mit300.html Accessed 2026-09-29.

[633] MIT/Tuebingen Saliency Benchmark: Leaderboard MIT300. MIT/Tuebingen Saliency Benchmark, undated (current at access). https://saliency.tuebingen.ai/results.html Accessed 2026-09-29.

[634] 7% growth in client toothpaste (case study). Dragonfly AI, undated. https://dragonflyai.co/resources/case-studies/growth-in-client-toothpaste Accessed 2026-09-29.

[635] Turning hero images into revenue drivers (case study). Dragonfly AI, undated. https://dragonflyai.co/resources/case-studies/turning-hero-images-into-revenue-drivers Accessed 2026-09-29.

[636] Go Beyond Attention: Discover Emotion and Memory (product page). Dragonfly AI, undated. https://dragonflyai.co/emotion-memory Accessed 2026-09-29.

[637] Dragonfly AI launches new Emotion and Memory metrics to lift creative effectiveness by up to 41%. Dragonfly AI, 2026-02-09. https://dragonflyai.co/resources/news/dragonfly-ai-launches-new-emotion-and-memory-metrics-to-lift-creative-effectiveness-by-up-to-30 Accessed 2026-09-29.

[638] Reports (Dragonfly AI Academy, Studio 3). Dragonfly AI (Angus Bingham), 2025-10-28. https://academy.dragonflyai.co/en/articles/10670509-reports Accessed 2026-09-29.

[639] Scene Testing (Dragonfly AI Academy, Studio 3). Dragonfly AI (Will Champion), 2026-04-22. https://academy.dragonflyai.co/en/articles/14488314-scene-testing Accessed 2026-09-29.

[640] Dragonfly API Documentation (API reference). Dragonfly AI, 2022-03-22. https://docs.api.dragonflyai.co/ Accessed 2026-09-29.

[641] Introducing Studio 3 (platform page). Dragonfly AI, undated. https://dragonflyai.co/platform Accessed 2026-09-29.

[642] Welcome to Motion (Dragonfly AI Academy, Motion Helpdesk). Dragonfly AI (Iwan Keymer), 2025-10-28. https://academy.dragonflyai.co/en/articles/4711544-welcome-to-motion Accessed 2026-09-29.

[643] Motion as a service (Dragonfly AI Academy, Motion Helpdesk). Dragonfly AI (Iwan Keymer), 2023-12-12. https://academy.dragonflyai.co/en/articles/4829679-motion-as-a-service Accessed 2026-09-29.

[644] API - Documentation (Dragonfly AI Academy, Developer API). Dragonfly AI (Iwan Keymer), 2021-07-27. https://academy.dragonflyai.co/en/articles/4867433-api-documentation Accessed 2026-09-29.

[645] Dragonfly AI Partners With CreativeX On Launch Of Datalink. Dragonfly AI (Rik Turner), 2025-12-10. https://dragonflyai.co/resources/news/dragonfly-ai-partners-with-creativex-on-launch-of-datalink Accessed 2026-09-29.

[646] Dragonfly AI secures £5m new investment from 24Haymarket, Guinness Ventures and Foresight. Dragonfly AI, 2026-01-28. https://dragonflyai.co/resources/news/dragonfly-ai-secures-5m-new-investment-from-24haymarket-guinness-ventures-and-foresight Accessed 2026-09-29.

[647] Dragonfly AI Privacy Policy. Dragonfly Technology Solutions Ltd, 2019-10-13. https://dragonflyai.co/privacy-policy Accessed 2026-09-29.

[648] Unlocking Consumer Insights: Dragonfly Teams Up with Mars at Quirk's Chicago!. Dragonfly AI, 2024-03-13. https://dragonflyai.co/resources/news/unlocking-consumer-insights-dragonfly-teams-up-with-mars-at-quirks-chicago Accessed 2026-09-29.

[649] Securing Trust and Excellence: Dragonfly AI Proudly Announces SOC 2 Accreditation. Dragonfly AI, 2023-12-04. https://dragonflyai.co/resources/news/securing-trust-and-excellence-dragonfly-ai Accessed 2026-09-29.

[650] Dragonfly AI - Trust Center - Security & Privacy. Dragonfly AI, undated. https://trust.dragonflyai.co/ Accessed 2026-09-29. Access: partial extract.

[651] Understanding attention (Dragonfly AI Academy, The Science). Dragonfly AI (Iwan Keymer), 2021-07-02. https://academy.dragonflyai.co/en/articles/4924569-understanding-attention Accessed 2026-09-29.

[652] Data protection regulations (Dragonfly AI Academy, Privacy & Security). Dragonfly AI (Iwan Keymer), 2021-04-29. https://academy.dragonflyai.co/en/articles/5192052-data-protection-regulations Accessed 2026-09-29.

[653] New InContext Eye Tracking Technology (launch news). eye square GmbH, 2024-08-29. https://www.eye-square.com/en/launch-incontext-eye-tracking-technology/ Accessed 2026-09-29.

[654] Whitepaper: The power to understand human experience and behavior (SEAL InContext Eye Tracking), 2025. eye square GmbH (hosted by succeet.de), 2025-01-31. https://www.succeet.de/fileadmin/succeet/succeet24/Blogbeitraege/Succeet_SEAL_InContext_Eye_Tracking_Whitepaper_eye_square_31_01_2025.pdf Accessed 2026-09-29.

[655] US12229992B2: Method of performing a user-specific and device-specific calibration of point of gaze estimation. Eye Square GmbH (assignee); inventors Karol Duzinkiewicz, Jan Glinko, Artur Skrzynecki, Michael Schiessl (Google Patents rendering), 2025-02-18. https://patents.google.com/patent/US12229992B2/en Accessed 2026-09-29.

[656] The Magical 2.5 Seconds: Media Attention Benchmark. eye square GmbH, 2020. https://www.eye-square.com/en/wp-content/uploads/sites/4/2020/03/MediaAttentionBenchmark-EngFullBLUECOVERS.pdf Accessed 2026-09-29. Access: partial extract.

[657] SEAL - In Context Eye Tracking. eye square GmbH, 2024-08-23. https://www.eye-square.com/en/in-context-eye-tracking/ Accessed 2026-09-29.

[658] In Context Experience Research | eye square. eye square GmbH, 2020-01-29. https://www.eye-square.com/en/in-context-experience-research/ Accessed 2026-09-29.

[659] In Context Express | Brand & Media Experience | eye square. eye square GmbH, 2017-01-25. https://www.eye-square.com/en/in-context-express/ Accessed 2026-09-29.

[660] Media Benchmark. eye square GmbH, 2024-06-05. https://www.eye-square.com/en/media-benchmark/ Accessed 2026-09-29.

[661] 2,3 Seconds of Magic - A Current Attention Analysis by OMD Germany and eye square. eye square GmbH, 2026-01-27. https://www.eye-square.com/en/attention-analysis-2-3-seconds-of-magic/ Accessed 2026-09-29.

[662] Social Media Ad Testing. eye square GmbH, 2024-06-06. https://www.eye-square.com/en/social-media-ad-testing/ Accessed 2026-09-29.

[663] Live Online Video and TVC Ad Testing. eye square GmbH, 2024-06-06. https://www.eye-square.com/en/video-ad-testing/ Accessed 2026-09-29.

[664] Technology Market Research | Implicit & Explicit | eye square. eye square GmbH, 2019-05-21. https://www.eye-square.com/en/technology/ Accessed 2026-09-29.

[665] Market Reasearch USA | eye square | Innovation in Market Research. eye square GmbH / eye square INC, 2023-05-05. https://www.eye-square.com/en/us/ Accessed 2026-09-29.

[666] The Snapchat & eye square Partnership (Discover our joint studies). eye square GmbH, 2026-07-08. https://www.eye-square.com/en/partner-overview-snapchat-eye-square/ Accessed 2026-09-29.

[667] How Snapchat Builds Brands Like No Other. Snap Inc. (Snapchat for Business blog), 2026-06-23. https://forbusiness.snapchat.com/blog/how-snapchat-builds-brands Accessed 2026-09-29.

[668] New Snapchat, OMG and eye square Study. eye square GmbH, 2025-01-29. https://www.eye-square.com/en/new-study-on-brand-authenticity-on-social-media/ Accessed 2026-09-29.

[669] The Role of Attention and Engagement in Creative Effectiveness. Affectiva / Smart Eye (blog, go.smarteye.se), undated. https://go.smarteye.se/the-role-of-attention-and-engagement-in-creative-effectiveness Accessed 2026-09-29.

[670] iMotions Online (overview with study-type comparison table). iMotions A/S, 2026-06-26. https://imotions.com/products/imotions-online/ Accessed 2026-09-29.

[671] iMotions to Fully Integrate Affectiva's Media Analytics in Order to Establish Consolidated Global Behavioral Research Unit within Smart Eye Group. iMotions A/S (press release, also distributed by PR Newswire), 2025-01-15. https://imotions.com/blog/learning/product-news/imotions-to-fully-integrate-affectivas-media-analytics/ Accessed 2026-09-29.

[672] Lifting Brand Awareness by 64% and Ad Recall by 29% with AI-Optimized Video (Teads case). Neurons, 2026-04-29 (schema datePublished, shared by every case page and probably a site republish date). https://www.neuronsinc.com/cases/teads-media-platform Accessed 2026-09-29.

[673] Neurons Memory Prediction (knowledge base). Neurons, undated (footer 2026). https://knowledge.neuronsinc.com/neurons-memory-prediction Accessed 2026-09-29.

[674] Customer Stories (cases index). Neurons, undated (footer 2026). https://www.neuronsinc.com/cases Accessed 2026-09-29.

[675] How Our Benchmarks Are Calculated (knowledge base). Neurons, undated (footer 2026). https://knowledge.neuronsinc.com/how-our-benchmarks-are-calculated Accessed 2026-09-29.

[676] Get Medium Benchmark Ranges (API reference). Neurons, 2026-06-12 (updatedAt). https://apidocs.neuronsinc.com/reference/get-media-benchmark-ranges.md Accessed 2026-09-29.

[677] Everything You Love About Neurons - Now for Video. Neurons, 2026-05-13. https://www.neuronsinc.com/insights/manual-video-aois Accessed 2026-09-29.

[678] Benchmark use case definitions (knowledge base). Neurons, undated (footer 2026). https://knowledge.neuronsinc.com/benchmark-use-case-definitions Accessed 2026-09-29.

[679] Benchmark industry definitions (knowledge base). Neurons, undated (footer 2026). https://knowledge.neuronsinc.com/benchmark-industry-definitions Accessed 2026-09-29.

[680] Neurons API Documentation (llms.txt index of guides and API reference, Neurons Customers API v2). Neurons, undated. https://apidocs.neuronsinc.com/llms.txt Accessed 2026-09-29.

[681] Introducing the Neurons API. Neurons, 2025-02-24 (updatedAt). https://apidocs.neuronsinc.com/reference/introduction.md Accessed 2026-09-29.

[682] Neurons AI Plugins (knowledge base). Neurons, undated (footer 2026). https://knowledge.neuronsinc.com/neurons-ai-plugins Accessed 2026-09-29.

[683] Neurons AI (Figma Community plugin listing). Neurons; listing hosted by Figma, undated. https://www.figma.com/community/plugin/1158782771898008510/neurons-ai Accessed 2026-09-29. Access: partial extract.

[684] Neurons Predict (Chrome Web Store listing). Neurons (developer: Neurons Inc ApS); listing hosted by Google, 2022-12-14 (last updated, version 1.0.3). https://chromewebstore.google.com/detail/neurons-predict/hoiahgbbchpeldmpconndelibfohkijh?hl=en Accessed 2026-09-29.

[685] Analyzing 1,800+ Ads in Hours to Win a Competitive 6-Agency Pitch (Mindshare Lithuania case). Neurons, 2026-04-29 (schema datePublished, shared by every case page). https://www.neuronsinc.com/cases/mindshare Accessed 2026-09-29.

[686] Brain Science Firm Neurons Opens US Office (Daily Research News Online no. 38126). MrWeb / Daily Research News Online (Nick Thomas), 2025-04-09. https://www.mrweb.com/drno/news38126.htm Accessed 2026-09-29.

[687] Predicting Percentage of Viewers & View Time on Ads. Neurons, 2023-03-23. https://www.neuronsinc.com/insights/predict-viewer-percentage-view-time-advertising Accessed 2026-09-29.

[688] US version of 2026 Neurons Terms and Conditions (Neurons AI purchase terms). Neurons Inc ApS, 2026 (per title; no date in file metadata). https://cdn.prod.website-files.com/6246d2921358dd41407dc44a/6a7da6701bb44b334393a5d1_US%20version%20of%202026-%20Neurons%20Terms%20and%20Conditions%20%5BCLEAN%5D.pdf Accessed 2026-09-29.

[689] The science behind Neurons. Neurons, undated (footer 2026). https://www.neuronsinc.com/science Accessed 2026-09-29.

[690] Neurons AI Metrics & Explanations (knowledge base). Neurons, undated (footer 2026). https://knowledge.neuronsinc.com/neurons-ai-metrics-and-explanations Accessed 2026-09-29.

[691] Ethical AI: Responsibility, Integrity, Transparency. Neurons, undated (footer 2026). https://www.neuronsinc.com/ethical-ai Accessed 2026-09-29.

[692] Legal Documents. Neurons, undated (footer 2026). https://www.neuronsinc.com/legal Accessed 2026-09-29.

[693] What are AOI metrics in Tobii Sticky?. Tobii (Tobii Connect knowledge base), undated. https://connect.tobii.com/s/article/What-are-AOI-metrics-in-Sticky?language=en_US Accessed 2026-09-29.

[694] Tobii Sticky mean validation error explained. Tobii (Tobii Connect knowledge base), undated. https://connect.tobii.com/articles/en_US/Knowledge/Mean-validation-error-explained Accessed 2026-09-29.

[695] Calibration in Tobii Sticky. Tobii (Tobii Connect knowledge base), 2026-03-23. https://connect.tobii.com/s/article/calibration-in-Sticky?language=en_US Accessed 2026-09-29.

[696] How to set up a hardware-based eye tracker with Tobii Sticky. Tobii (Tobii Connect knowledge base), undated. https://connect.tobii.com/s/article/How-to-set-up-a-hardware-based-eye-tracker-with-Tobii-Sticky?language=en_US Accessed 2026-09-29.

[697] White Paper: Tobii Sticky Emotion Analysis. Tobii (Tobii Connect knowledge base), undated. https://connect.tobii.com/s/article/White-Paper-Emotion-Analysis?language=en_US Accessed 2026-09-29.

[698] Reimagining attention: Boosting advert effect with eye tracking insights. Tobii (blog), 2021-06-16. https://www.tobii.com/blog/boosting-advert-effect-with-eye-tracking-insights Accessed 2026-09-29.

[699] Individual AOI Metrics in Tobii Sticky. Tobii (Tobii Connect knowledge base), undated. https://connect.tobii.com/s/article/Individual-AOI-Metrics-Sticky?language=en_US Accessed 2026-09-29.

[700] What is Sticky?. Tobii (Tobii Connect knowledge base), undated. https://connect.tobii.com/s/article/What-is-Sticky?language=en_US Accessed 2026-09-29.

[701] Tobii Pro Launches New Mobile Eye Tracking Solution for Marketing and Advertising Research. Tobii Pro (press release via PR Newswire, Nasdaq copy), 2021-06-08. https://www.nasdaq.com/press-release/tobii-pro-launches-new-mobile-eye-tracking-solution-for-marketing-and-advertising Accessed 2026-09-29.

[702] Using Sticky with your phone. Tobii (Tobii Connect knowledge base), undated. https://connect.tobii.com/articles/en_US/Knowledge/Using-Sticky-with-your-phone Accessed 2026-09-29.

[703] Tobii Sticky API Documentation. Tobii (Tobii Connect knowledge base), undated. https://connect.tobii.com/s/article/Sticky-API-Documentation?language=en_US Accessed 2026-09-29.

[704] Tobii Sticky JavaScript FAQs. Tobii (Tobii Connect knowledge base), undated. https://connect.tobii.com/s/article/Sticky-JavaScript-FAQs?language=en_US Accessed 2026-09-29.

[705] Which countries can I order participants from within the Tobii Sticky platform?. Tobii (Tobii Connect knowledge base), undated. https://connect.tobii.com/articles/en_US/Knowledge/Which-countries-can-I-order-participants-from-within-the-platform Accessed 2026-09-29.

[706] Tobii Sticky: Terms of service for clients with a plan agreed upon after June 21, 2022. Tobii (Tobii Technology Inc / Tobii Pro UK Holdings Ltd. / Tobii Technology K.K. / Tobii AB), 2026-03-20. https://storage.pardot.com/151161/1774039013WQ91ClOh/Tobii_Sticky___Terms_of_service_agreed_after_21_June_2022.pdf Accessed 2026-09-29.

[707] Certificate SE556613-9654: Tobii AB, ISO 9001:2015, ISO 14001:2015, ISO/IEC 27001:2022. Scandinavian Business Certification AB (certificate hosted by Tobii), 2025-04-09 (renewal). https://a.storyblok.com/f/290827487058644/999506c025/tobii-iso-certificate-2025-04.pdf Accessed 2026-09-29.

[708] Tobii Sticky – Privacy & security. Tobii AB, 2026-03-23. https://storage.pardot.com/151161/1774194547qwo9W9Hd/Tobii_Sticky___Privacy_and_Security.pdf Accessed 2026-09-29.

[709] Amplified gain CIMM membership. Amplified, 2023-01-27. https://www.amplified.co/insight/cimm-membership Accessed 2026-09-29.

[710] Dr Nelson-Field Launches Book on Attention Economy. Amplified, 2024-10-30. https://www.amplified.co/insight/karen-nelson-field-attention-economy-category-blueprint-book Accessed 2026-09-29.

[711] Amplified Launches Creative Testing AI - Self-Service Product. Amplified, 2025-08-04. https://www.amplified.co/insight/amplified-launches-self-service-creative-testing-ai-product Accessed 2026-09-29.

[712] Looking for attentionPLAN, attentionPROVE, or attentionTRACE? (new product names). Amplified, undated (live page). https://www.amplified.co/new-product-names Accessed 2026-09-29.

[713] DoubleVerify Holdings, Inc. Form 10-K for fiscal year 2022. DoubleVerify Holdings, Inc., 2023-03-01. https://www.sec.gov/Archives/edgar/data/1819928/000155837023002596/dv-20221231x10k.htm Accessed 2026-09-29.

[714] DoubleVerify First Measurement Provider to Earn MRC Accreditation for TikTok Video Viewability Reporting. DoubleVerify, 2026-04-23. https://doubleverify.com/company/newsroom/doubleverify-first-measurement-provider-to-earn-mrc-accreditation-for-tiktok-video-viewability-reporting Accessed 2026-09-29.

[715] IAS Announces Completion of Acquisition by Novacap (Exhibit 99.1 to Form 8-K). Integral Ad Science Holding Corp. (SEC filing), 2025-12-23. https://www.sec.gov/Archives/edgar/data/1842718/000114036125046475/ef20061707_ex99-1.htm Accessed 2026-09-29.

[716] Numerator and Kantar Worldpanel combine to form new global consumer data company. Kantar / Numerator (press release), 2025-01-14. https://www.kantar.com/Press-Center/Numerator-and-Kantar-Worldpanel-combine-to-form-new-global-consumer-data-company Accessed 2026-09-29.

[717] LUMEN RESEARCH LTD, officers. UK Companies House, live record. https://find-and-update.company-information.service.gov.uk/company/08682432/officers Accessed 2026-09-29.

[718] VideoAmp, TVision expand licence agreement. Advanced Television, 2026-02-12 (from URL path). https://www.advanced-television.com/2026/02/12/videoamp-tvision-expand-licence-agreement Accessed 2026-09-29.

[719] TVision Insights Enters Definitive Agreement to Be Acquired by Viant Technology. TVision Insights, 2026-04 (undated on page; announces the 14 April 2026 agreement). https://www.tvisioninsights.com/resources/tvision-insights-enters-definitive-agreement-to-be-acquired-by-viant-technology Accessed 2026-09-29.

[720] Report: Streaming audiences increasingly disengaged from ads. Advanced Television, 2026-09-25 (from URL path). https://www.advanced-television.com/2026/09/25/report-streaming-audiences-increasingly-disengaged-from-ads/ Accessed 2026-09-29.

[721] Vevo Gets Its Investor: Abu Dhabi Media Joins “Hulu for Music Videos”. Peter Kafka, AllThingsD, 2009-10-19. https://allthingsd.com/20091019/vevo-gets-its-investor-abu-dhabi-media-joins-hulu-for-music-videos/ Accessed 2026-09-29.

[722] YouTube Confirms Stake in Vevo Music-Video Venture. Todd Spangler, Variety, 2013-07-03. https://variety.com/2013/digital/news/youtube-confirms-stake-in-vevo-music-video-venture-1200504729/ Accessed 2026-09-29.

[723] New Products, Programmatic Innovation & Reach Take Center Stage at Vevo Event. Vevo LLC, 2025-05-08 (dateline; page datePublished 2025-05-09). https://www.hq.vevo.com/press/new-products-programmatic-innovation-reach-take-center-stage-at-vevo-event Accessed 2026-09-29.

[724] Adelaide News. Adelaide, live listing. https://www.adelaidemetrics.com/news Accessed 2026-09-29.

[725] Universal Music Group Annual Report 2025 (Note 13, Investments in equity affiliates). Universal Music Group N.V., 2026 (annual report for FY2025; linked from investors.universalmusic.com). https://assets.ctfassets.net/e66ejtqbaazg/7HObJrt5UDQfjjfKYN454B/201bf057190c5ed4ee185b04a34ec095/UMG_2025_Annual_Report.pdf Accessed 2026-09-29.