Proof and Story

Abstract

What this paper found

Contextual advertising spent a decade learning to read content. In 2026 it can read text, images, audio and video at industrial scale. Two things it still cannot do are prove what it read, and read a story.

This is a survey of the contextual advertising field as it stands in August 2026, built from a structured research run rather than a reading list: nine thematic research streams, four vendor-segment sweeps, ten company profiles, adversarial fact-checking on every stream, and a red-team pass over the competitive analysis. Fifty-one vendors were scored across eleven capability dimensions. Ten were profiled in depth across twelve. Every externally sourced number below was subjected to an adversarial verification pass whose job was to refute it, and the 117 claims that did not survive intact are published in the ledger at the end rather than quietly dropped.

Three findings drive everything that follows.

The first is that meaning derivation has been solved well enough to commoditise. Foundation models made page-level semantic classification cheap, and the market responded by moving the contest somewhere else. Brian O'Kelley's cost trajectory is the clearest statement of the economics: classifying the web through a brand's own eyes cost roughly $45M in 2023 and about $270K by March 2025, with under $100K projected for 2026. When the hard part gets cheap, the moat moves to whatever is still expensive. Right now that is proof and distribution.

The second is that emotion has been productised faster than it has been validated. Scene-level emotional targeting became a real product category between 2024 and 2026, with Plutchik's wheel shipped verbatim inside Wurl's BrandDiscovery and emotion sold as the core product at Seedtag. But of the five vendors scoring 4 or better on emotion, not one has had its emotional ontology independently verified, and the field's only two content-level MRC accreditations certify safety and suitability instead. On this evidence, an independently audited emotion engine is not purchasable in this market today.

The third is the one that surprised me. Narrative is close to a blank column. Within this 51-vendor public-evidence sample, 27 vendors showed no narrative or archetype capability, 18 gave insufficient evidence to score, and six showed some adjacent capability — Overtone at 5, Anoki and Scope3 at 3, Origin Media at 2, and Seedtag and Frameplay at 1. We found no publicly available product offering transactable targeting by narrative position or by brand archetype. The research primitives have existed since 2020; their commercial validity and activation layer have not. This is the largest unclaimed capability the sample shows, and unlike the other gaps in this paper it looks structural rather than competitive.

Chapter 1

Five generations of reading

Contextual targeting has been rebuilt four times since 2006. Each generation answered a different question about the same web page, and the current one stopped asking about the page at all.

The first generation matched keywords and blocked URLs. Grapeshot was its emblem and its commercial ceiling, sold to Oracle in 2018 for a reported $400M. It was also a well-documented failure mode: keyword systems put airline ads next to crash coverage and blocked the word "shot" across vaccine reporting during a pandemic. That generation is now formally over. Oracle announced the closure of its advertising business in June 2024 and shut it on 30 September 2024, taking Grapeshot and Moat with it. The business went from roughly $2B in revenue in 2022 to about $300M in FY2024, and the technology was shuttered rather than sold.

The second generation replaced keywords with shared categories. The IAB Content Taxonomy is now at version 3.1, roughly 700 categories across five tiers, carried in OpenRTB 2.6 as cattax value 9. This layer never went away and is the reason the whole market still functions: it is the common vocabulary that lets a buyer's intent survive the trip to a seller's inventory. Adoption is uneven, with many implementations still on 1.0 or 2.2, and the Tech Lab shipped an official 1.0-to-2.0 migration mapping in January 2025 to deal with exactly that.

The third generation read for meaning instead of category. Embeddings, entity extraction, sentiment scoring and knowledge graphs let a system distinguish an article that mentions a brand from an article about a brand. Peer39, Seedtag's Liz in its earlier form, IAS's semantic knowledge graph and Illuma's live-session signals all sit here, and for most of the market this is still where production lives.

The fourth generation added the other modalities. During 2025 multimodal classification became table stakes rather than a differentiator. DoubleVerify's Universal Content Intelligence runs vision-language models across video, image, audio, speech and text at hundreds of millions of items per day. IAS ships frame-level models into Meta and TikTok. GumGum's Verity reads text, image, audio and video, and holds the first MRC content-level accreditation. The frontier extended into audio through Sounder and Barometer, and into scene and object granularity through Wurl, KERV and IRIS.TV.

The fifth generation stopped classifying and started deciding. O'Kelley's "The End of Taxonomy" in March 2025 made the argument that fixed categories are simultaneously too precise and too reductive, and that cheap inference makes it viable to evaluate every piece of content through the eyes of a specific brand. Scope3, the company he runs, had shipped the product four days before he published the argument: Brand Standards launched on 13 March 2025, with Brand Stories following in June. Seedtag shipped Liz Agent in March 2026, GumGum the Mindset Agent in July 2026. These systems answer a different question from their predecessors. Not what is this content, but what should this brand do here.

There is a temptation to read this as a straight replacement sequence where each generation kills the last. That is wrong, and the error has commercial consequences. Generation 2 is not legacy. It is the settlement layer, and the reason is that buying media requires a deterministic object that two systems can agree on after the fact. An embedding is expressive and unauditable. A category ID is reductive and enforceable. The market needs both, which is why the interesting engineering in 2026 is not in either layer but in the join between them.

Chapter 2

Deriving meaning

The end-of-taxonomy argument was right about accuracy and wrong about what a taxonomy is for. The market settled the fight in 2026 by splitting the job in two.

The accuracy ceiling is the taxonomy

The strongest technical support for the anti-taxonomy position arrived in November 2025. A benchmark study evaluated ten state-of-the-art language models zero-shot against the IAB hierarchical taxonomy across 8,660 human-annotated samples. Individual models plateaued, and the paper attributes the plateau to the compression of semantically rich text into sparse categorical representations. An ensemble of models recovered up to 65% more F1 than the best single model. Read plainly: the models were not the bottleneck. The categories were.

The evidence from the field is worse than the evidence from the lab. Adalytics documented verification systems rating explicitly vile pages as low risk, a finding the vendors dispute on methodology. A January 2026 study documented wide disagreement between vendors classifying the same news articles. There is still no standardised public benchmark for advertising content classification, which means no buyer can compare two vendors' accuracy on equal terms. That absence is not an oversight. It is a commercial preference, and it is the root of the proof problem this paper keeps returning to.

The counter-argument is about settlement, not accuracy

IAB Tech Lab's response, in a position paper in April 2026 and in its CEO's warnings from December 2025, does not claim taxonomies classify better. It claims they settle better. Language models predict likely completions rather than looking up facts, so a natural-language brief reinterpreted across five agent hops drifts semantically, and nobody can audit where it drifted. An integer category ID does not drift. When a buyer and a seller disagree about what was delivered, someone has to arbitrate against a shared reference.

Both positions are correct because they are about different jobs. The resolution visible in what has shipped so far is rich derivation upstream, deterministic identifiers downstream, and a mapping layer in between that is now the contested ground. The Tech Lab's agentic framework consolidated as AAMP in February 2026, and its most telling feature is a component called Taxonomy Guardrails.

Nobody runs a large model in the bid path

This is among the most consequential engineering facts in contextual advertising, and it is routinely obscured by marketing language about real-time AI. A real-time bidding auction has a budget of roughly 300 milliseconds end to end. The slice available to a contextual data provider is about 10 milliseconds. No useful large language model responds in 10 milliseconds at web-scale economics, so in every architecture examined here the model does not sit in the bid path.

So the work moves out of the hot path. GumGum has published its architecture, and it is the clearest public description of how this actually works: contextual signals served at roughly 6 milliseconds at the 99th percentile, at about 92,000 requests per second, from Redis edge caches running on AWS Outposts. On a cache miss the system returns a 404 and queues the page for asynchronous classification. The expensive thinking happened earlier, offline, and what the auction sees is a lookup.

Every vendor whose architecture could be examined uses a variant of this. DoubleVerify classifies visual content offline with vision-language models using keyframe extraction rather than processing every frame. IRIS.TV pre-analyses video corpora and passes only an identifier in the bid stream for a decode-side join. Scope3 classifies content as it is published and enforces the resulting verdicts through curation and deal infrastructure. Distillation of large teacher models into small task-specific ones is how language-model judgment reaches production latency at all.

Understanding the cache changes how you read vendor claims. Peer39's headline strength is freshness: First Look classification, a 24-hour time-to-live, and a claim of 1.3 billion pages a day with 90% of pages new every 24 hours. The verification pass found Peer39's own properties contradicting those figures, quoting 2.5 billion daily URLs in one place and 400 billion rather than 750 billion daily bid requests in another. The throughput numbers are unreliable. The architecture argument survives them, because a 24-hour TTL plus a First Look mechanism guarantees freshness by design regardless of what the throughput actually is. That distinction between a number you cannot verify and a mechanism you can reason about is worth carrying into every vendor conversation.

The newest surface makes the point differently. OpenAI began showing contextually matched ads in ChatGPT in February 2026, using what it calls context hints rather than keywords, with self-serve arriving by mid-year. Perplexity launched advertising first, in November 2024, and reportedly walked away from it in early 2026. Conversational context is the purest form of moment reading anyone has built. Its monetisation is not settled.

Chapter 3

Deriving emotion

Emotional targeting became a real product category between 2024 and 2026. The science underneath it is weaker than the marketing, the evidence for it is almost entirely vendor-funded, and no vendor's emotional model has been independently audited.

What actually shipped

Wurl's BrandDiscovery, launched March 2024 and owned by AppLovin, scores the scene immediately before a CTV ad break against Robert Plutchik's wheel: eight primary emotions at three intensities. The score travels in the bid request. Because Wurl sits inside its own distribution and server-side ad insertion path, the classification happens with no added latency and covers live FAST content that a crawl-based competitor cannot reach at all.

Around it a category formed. Seedtag launched "neuro-contextual advertising" in June 2025, with its Liz model classifying interest, emotion and intent, and eMarketer named the category a top trend for the first half of that year. Reticle AI classifies more than five million URLs and videos daily into 19 proprietary emotional categories, distributed through Peer39's marketplace. Anoki's ContextIQ went into Magnite's SpringServe in June 2025. Zenapse licensed its "Large Emotion Model" to LG Ad Solutions in April 2025 across roughly 200 million smart TVs. Disney has been targeting scene mood through Magic Words since February 2024, extending to live programming in 2025.

Worth noting who is absent from that list. The verification incumbents still ship sentiment rather than discrete emotion. GumGum's own methodology documentation states that its model predicts sentence-level sentiment aggregated to positive, neutral or negative, and concedes that neutral is typically the highest-scoring value. DoubleVerify has no emotion product at all. That gap is the clearest wedge any emotion-led vendor has.

The science is shakier than the category

On text, the reference dataset for fine-grained emotion is Google's GoEmotions: 58,000 Reddit comments labelled across 27 emotions plus neutral. The original 2020 baseline reported about 0.46 average F1. Later literature puts fine-tuned BERT at roughly 52.75% macro-F1 and zero-shot ChatGPT at about 25.55%. Multimodal conversational emotion recognition on MELD sits in the high-40s to mid-60s depending on setup, with severe class imbalance. Speech emotion recognition scores well on acted single-corpus benchmarks and degrades sharply across corpora, languages and noise conditions.

Read those numbers against the marketing language of human-like emotional understanding. Roughly half of fine-grained emotion labels are wrong or contested in the best published conditions, partly because human annotators disagree with each other about the same text.

The evidence problem

The outcome studies are directionally consistent and almost entirely funded by the vendors whose products they validate. Wurl and TVision found emotionally aligned CTV ads captured 2.1 times more attention across 50-plus campaigns. Kantar measured Wurl's emotion and genre targeting lifting brand awareness 33% and favourability 28%. Seedtag's EEG study with Columbia's Moran Cerf, published November 2025, reports 3.5 times higher neural engagement than non-contextual placement and 30% higher than standard IAB contextual. GumGum with SPARK Neuro reported 43% higher neural engagement and 2.2 times recall.

Set against that is the largest independent synthesis available. A 2021 meta-analysis in the Journal of Advertising by Kwon, Nyilasy, King and Reid pooled 96 studies and 597 effect sizes across 31 years, with a combined sample of 139,233, and found media-ad congruence has a very weak influence on attitudes and purchase intent. Felt involvement, transportation and trust in the vehicle mattered considerably more, and arousal from context correlated weakly negatively with ad memory.

There is a deeper objection than effect size. Lisa Feldman Barrett and colleagues, reviewing the field in Psychological Science in the Public Interest in 2019, found that facial expressions fail tests of reliability and specificity: people scowl in only a minority of anger episodes and scowl frequently when not angry, and expression-to-emotion mappings vary across cultures and individuals. That critique targets biometric emotion inference rather than content classification, but part of it transfers. The emotion a scene expresses is not the emotion a viewer feels. The UK's Information Commissioner's Office warned in October 2022 that emotion analysis technologies are immature and carry systemic bias risk.

This suggests the framing that is both scientifically defensible and legally durable: these systems classify the emotion that content expresses or is likely to evoke. That is a media-planning construct, not a claim about any individual's internal state. Almost no vendor says this plainly, which leaves the plain version available to whoever wants it.

The legal line runs through biometrics

The EU AI Act's Article 5(1)(f), applicable since February 2025, prohibits AI inference of a natural person's emotions only from biometric data, and only in workplace and education settings. Biometric emotion recognition elsewhere is high-risk with transparency duties. Classifying what a scene or an article expresses involves no biometric data and no identified person, which is exactly why every vendor in this category positions content emotion as the privacy-safe path.

Nobody can buy audited emotion

Two structural gaps sit under this whole category. The first is audit. Of the five vendors in this sample scoring 4 or better on emotion, we found no published independent verification of any emotional ontology. Reticle's 19 categories and Wurl's Plutchik-derived families are proprietary and unpublished, and Seedtag's neuro-contextual layer is a positioning term rather than a documented methodology. Meanwhile the field's only two content-level MRC accreditations, held by GumGum and Zefr, certify safety and suitability. On the public evidence gathered here, an independently audited emotion engine is not something a buyer can currently purchase.

The second is interoperability. The IAB Content Taxonomy has no emotion field. IPTC is only now developing an editorial tone descriptor. So every emotion taxonomy in market is proprietary and non-portable: Plutchik's eight-by-three at Wurl, 19 categories at Reticle, five Watson emotions at Mantis, 39 at DAIVID, 83 psychographic dimensions at Zenapse, ad-hoc moods at Disney and Tubi. None of them can be exchanged, none carries provenance, and none carries activation rules. A buyer who standardises on one has standardised on one vendor.

Chapter 4

The blank column

Of 51 vendors scored on narrative and archetype capability, 27 show none and 18 give no evidence to score. The research primitives have existed since 2020. The activation layer has not.

Start with what the number means, and with a distinction the matrix insists on. A score of 0 means the record shows no such capability. n/e means the record gives no basis to score at all. Those are different claims, and collapsing them overstates the finding, so they are kept apart here. Of 51 vendors, 27 score 0 and 18 are n/e. Six score above zero: Overtone at 5 through narrative-shift primitives, Anoki at 3 for naming narrative context as a signal without formalising or exposing it, Scope3 at 3 for compiling brand values, Origin Media at 2, and Seedtag and Frameplay at 1. That is not a market with a leader and some laggards. It is a market where the column is nearly empty and nothing in it is transactable.

The research primitives exist

Computational narrative analysis is not speculative, though it has been demonstrated on books and screenplays rather than on live CTV, short video or news, and that generalisation is unproven. Reagan and colleagues derived six emotional arc shapes across 1,327 Project Gutenberg texts in 2016. Boyd, Blackburn and Pennebaker identified a consistent structure of staging, plot progression and cognitive tension across roughly 40,000 narratives in Science Advances in 2020, which maps onto Freytag's staging closely enough to be useful. The TRIPOD dataset annotates screenplay turning points, setup, complications, climax and aftermath, from multimodal video. Work through 2025 pushed this toward production quality.

Commercially, StoryFit sells script-level emotional arc and character analytics to studios for both development and marketing. So the capability exists and someone is already paying for it. It is simply being sold to the people who make the content rather than the people who buy advertising around it.

Transportation theory supplies the reason to care. A 2024 systematic review in Psychology & Marketing and a run of 2025 studies show that narrative absorption lowers resistance to persuasion. The researchers themselves note how little is understood about which story structures make advertising compelling. That is academic demand for exactly the operationalisation nobody has built.

Archetypes are the missing key

Jung's archetypes, popularised for brand strategy by Mark and Pearson, dominate brand consulting and creative tooling. No major programmatic vendor sells targeting on them. The nearest commercial analogues stop short in instructive ways. Quilt.AI's Sphere detects brand values and archetypes inside video creative, which is creative intelligence rather than media activation. Zenapse matches consumers to 16 psychographic master segments, which is persona-shaped. Reticle matches the emotion an ad evokes to placements across its 19 categories, the closest live implementation of creative-context resonance matching, claiming 31% higher attention and 21% higher recall.

So the components of an archetype-resonance stack all exist in market: emotion-scored creative, emotion-scored content, and psychographic personas. Nobody has unified them under an archetypal model. The reason to want that unification is that archetype is the brand-side key that makes the other two signals actionable. Emotion tells you what a moment feels like. Arc position tells you where in a story it sits. Archetype tells you which brand should want it.

The case against building it

An empty column is not automatically an opportunity. It is also what a market looks like when something does not work, cannot be measured, or cannot be executed. That case deserves stating properly, because a reader will raise it.

There is no measured evidence anywhere in this study that placing an ad by narrative position beats placing it by scene emotion. Every effect size in chapter 3 concerns emotional congruence, and none isolates arc position as a variable. The same is true of archetype resonance: the closest live implementation reports higher attention and recall, and it matches emotion between creative and placement rather than matching archetype. So the layer's value is inferred from adjacent results rather than demonstrated.

The execution objection is more serious than the evidence objection. Ad breaks are scheduled by pod position and duration, not by where the story has got to. In server-side ad insertion the break points are usually fixed in the manifest before any narrative analysis could influence them. Targeting a position on an arc requires either that the content owner exposes arc metadata per break, or that break placement itself becomes narrative-aware, and the second is a scheduling change rather than a targeting feature. On the open web the problem is different but real: most pages do not contain a narrative arc at all.

The honest version of the opportunity is therefore narrower than "nobody has built it". Arc position is buildable today only where content is episodic, where the owner already indexes scenes, and where break placement is not fixed upstream, which describes premium streaming and very little else. Archetype matching has no such constraint, because it operates on the creative and the content's tone rather than on timing, which makes it the more immediately buildable of the two and the one with less excuse for being absent.

Trade criticism has already named the underlying pathology. ADOTAT's March 2025 piece called out AI-washing and keyword engines with a thesaurus addiction, and landed on the line that matters here: no shared language means no scale. IRIS.TV's identifier, now covering more than 70 million videos, is the closest thing to connective tissue, and it standardises the content identifier rather than the signal payload. That distinction is the subject of chapter 6.

Chapter 5

Reach without meaning

Contextual finished its jump from web pages to every major channel between 2025 and 2026. It arrived as a set of channel-specific beachheads rather than one layer, and the channel with the most reach has the least understanding.

Each channel got there differently

CTV moved fastest. Anoki's ContextIQ, which analyses video, audio, tone, emotion and objects, went into Magnite's SpringServe in June 2025 and Index Exchange afterwards. Viant's acquisition of IRIS.TV in November 2024 turned a neutral content identifier into a DSP-owned asset, now carrying more than 1,200 taxonomy categories with genre and tone in the bid request. The verification duopoly followed: DoubleVerify launched Authentic Streaming TV at CES 2026 with IMDb-licensed program signals, and IAS launched Total TV in April 2026 across Disney, NBCUniversal, Paramount and Prime Video.

Audio is the cleanest example of cross-channel normalisation anyone has shipped. In July 2026 Comscore's Proximic launched transcript-level targeting, meaning episode-level rather than show-level, across Spotify, SiriusXM, Triton Digital, Acast and Libsyn, and described it as applying the same identity-free contextual and suitability approach it already used across display, video and CTV. One taxonomy, four channels. That is the thing everyone else claims and few can demonstrate.

Out-of-home is natively spatiotemporal, which makes it the channel that already buys moments rather than places: weather, rush hour, the end of a match, a foot-traffic spike. It also consolidated into an unexpected owner. T-Mobile closed its $600M acquisition of Vistar Media on 3 February 2025 and bought Blis for $175M, a business whose defining architecture is data structured on geography rather than identity.

In-app is where the argument breaks

In-app advertising has no page to crawl. The content is inside a binary, rendered by an SDK, and often not text at all. So the industry substituted proxies: app-store metadata, app category, coarse geography, session state. Those are real signals. None of them tells you what the content means.

The transport layer proves it. In OpenRTB 2.6, the App object marks only its identifier as recommended, and the bundle, store URL, category, section category, page category, keywords and nested content object are all optional. The Content object itself is shaped for episodic audiovisual media, with fields for episode, series, season, artist, album, ISRC, livestream and length. There is no field anywhere in the specification that describes what is on an app screen right now. The only current-view fields are optional arrays of coarse categories. What exists is a transport layer for app identity, plus a television content object borrowed into the app parent.

What fills that vacuum in practice is visible in Digital Turbine's own specification, which defines contextual app targeting as store category and subcategory, framework type, Android API level, plus behavioural fields such as impression depth, session duration, previous bundle and completion rates, plus device state including tracking-permission status, battery level, dark mode and headset connectivity. The taxonomy is the app store's, not the IAB's. No in-app text, screen content or gameplay is analysed at any point.

In-game says the quiet part out loud. Anzu's September 2025 announcement states that intrinsic in-game advertising had until then been largely limited to contextual targeting based on genre or title. Its remedy was not deeper scene understanding. It was 62,400 third-party audience segments, universal identifiers, footfall attribution and lookalike audiences. The one genuine counter-example found anywhere in this study is Odeeo's Game Emotions, which targets live game state such as a new personal high score or losing a life. It is sold direct to agencies only, is not exposed pre-bid to third-party demand, publishes no detection mechanism and carries no performance data.

The platform-policy backdrop makes this harder rather than easier, and it is the opposite of the Chrome story in chapter 7. Google's October 2025 retirement killed the Android programme alongside the browser one: Topics, Protected Audience and Attribution Reporting on Android, plus Protected App Signals, the SDK Runtime and On-Device Personalization, all listed as scheduled for phaseout with no published sunset dates as of August 2026. Topics on Android was the only platform-sanctioned mechanism for inferring meaning on device from app identity. Its removal leaves the channel with less sanctioned infrastructure for understanding than it had two years ago.

The matrix in chapter 8 makes the consequence visible. The in-app reach leaders, Ogury, InMobi, Digital Turbine, Verve and AppLovin, all score 5 on in-app. Their meaning-depth scores are 2, 2, 1, 4 and 0. The single 4 belongs to Verve, and it belongs to a web engine the company has since retired rather than to its on-device in-app product. Anzu and Frameplay get context by construction, because in intrinsic in-game advertising the ad slot is a scene object, but they stop at genre and title taxonomy. In-app emotion is entirely unclaimed: every emotion-capable vendor in the study scores no-evidence or 2-and-below on in-app.

The two architectural answers to in-app are visible in market and neither is finished. Bid-stream contextual reads what the exchange already carries, which is metadata and geography. On-device intelligence runs the model where the content actually is: Verve's ATOM does identity-free on-device cohort modelling, and Ogury sells persona targeting with no identifiers. On-device is the more interesting path because it is the only one with access to the actual content, and it is also the one that cannot easily emit a portable signal, for the same reason.

Nobody spans the whole thing

Different vendors are closest on different axes. On channel breadth with one consistent identity-free methodology, Comscore's Proximic leads, covering display, video, CTV and audio under a single taxonomy with panel-backed measurement. On owned trading infrastructure, Seedtag and GumGum. On walled-garden coverage, DoubleVerify and IAS. On physical-world spatiotemporal context, T-Mobile's combination of Vistar and Blis. None of them covers web, app, CTV, audio, DOOH, in-game and retail media together.

The unsolved layer underneath is measurement. Identity-free frequency capping does not reconcile across mobile web, in-app and CTV. Capping in CTV still runs on household IP heuristics. The honest available answers are panels, on-device cohorts and incrementality testing, which is a polite way of saying the industry currently estimates. A portable, machine-readable context object that travelled across channels would not solve frequency by itself, but it is the precondition for solving it without identity.

Chapter 6

Rails, agents and the portable signal

Agentic advertising went from thesis to shipped infrastructure between October 2025 and August 2026. Two standards camps are each building one half of a portable signal object, and neither is building the other half.

The Ad Context Protocol launched on 15 October 2025, backed by a supply-heavy consortium including Yahoo, PubMatic, Optable, Scope3, Swivel and Triton Digital, with Magnite, Kargo, Raptive, Samba TV, LG Ad Solutions and Butler/Till among launch members. It runs over Anthropic's Model Context Protocol, with an agent-to-agent transport option, and it is governed by a non-profit modelled on Prebid. Version 3.0 reached general availability on 22 April 2026, and by August 2026 the specification had grown into a full stack: natural-language briefs, media buy creation, creative sync, signal activation, a governance protocol with human escalation gates and audit logs, machine-readable brand guidelines, capability discovery, and signed requests.

Transactions followed quickly. The first agent-to-agent buy ran on LG Ads inventory on 16 October 2025. Butler/Till and PubMatic ran a CTV campaign end to end from a natural-language brief in December 2025. NBCUniversal, RPA, FreeWheel and Newton Research executed a cross-platform premium video buy including live NFL playoff inventory in January 2026. Criteo and dentsu ran a fully agent-orchestrated video campaign in May 2026.

IAB Tech Lab's response was blunt. Its CEO said the industry did not need another trade group, and answered with the Agentic RTB Framework in November 2025: containerised bidding logic co-located inside the auction's data centre, gRPC rather than HTTP, and a claimed latency reduction from 600 to 800 milliseconds down to around 100. That work was consolidated under the AAMP umbrella, named on 26 February 2026, which extends OpenRTB, AdCOM and OpenDirect rather than inventing new rails.

First, the word is already taken

Anyone writing about portable signals in 2026 has to deal with a terminology collision before making any argument at all, because the industry already uses the container vocabulary to mean close to the opposite thing.

In market, containerisation means packaging compute. Digiday's explainer from June 2026 defines a container as a self-contained package of bidding logic and the data it needs to run, deployed inside someone else's infrastructure rather than running remotely and passing data back and forth. Magnite's August 2026 position paper describes packaging applications, AI models, algorithms and business logic into portable, self-contained environments. Index Exchange shipped Index Cloud in April 2026 with Bedrock as the first containerised bidder. PubMatic's Decision Fabric followed in June. The Chalice and Equativ announcement in May 2026 is headlined around advancing portable AI models. So the compute camp has claimed both words.

The direction is not merely different. It is an explicit argument against moving signals at all. Index Exchange says enrichment now happens locally with no signal loss in transit. PubMatic's framing is that by the time a bid request reaches a DSP the signals are already compressed, filtered and delayed. OpenX puts it plainly: run your models where the impression is created and you stop paying to move the data and stop losing signal along the way.

Two camps, one object, no join

These are usually reported as a turf war. The more useful reading is that they are building complementary halves of the same thing, and the seam between them is where the value is.

ARTF containerises compute. It puts the decisioning logic next to the auction and ships dense vector embeddings that compress identity, contextual and reinforcement signals into a form fast enough for single-pass inference inside the bid. That solves the latency problem from chapter 2 by moving the thinking rather than shrinking it.

AdCP containerises governance. Machine-readable brand guidelines, signed requests, escalation gates and tamper-evident audit logs are the parts that make an automated decision defensible after the fact. What it does not do is compress meaning into something that survives an auction's time budget.

This matters beyond architecture. A signal that carries its provenance and its permitted scope is auditable, and one that does not is a number of unknown origin being used to make a decision about a person's attention. The engineering is neutral about which of those it produces. Consent and provenance treated as relevance inputs rather than as a compliance gate applied afterwards is the difference, and it has to be built into the object rather than bolted onto the process.

The gap is narrower than "nobody thought of this"

It would be wrong to present portable signal packaging as unclaimed. Both halves exist. They just exist in different specifications that do not meet.

AdCP's Signals Protocol already carries a remarkably complete governance payload on the signal object: lawful basis for consent, special-category flags, restricted attributes, policy categories, applicable countries, data-subject-rights routing with response deadlines, methodology, data sources, taxonomy mappings, refresh cadence, pricing and activation keys. That is consent, provenance, policy and activation in one place. But AdCP Signals is a discovery and licensing protocol. It moves descriptions and instructions, and the segment membership itself gets pushed into a destination platform. The rich object never travels with the impression.

The Tech Lab's Agentic Audiences, donated by LiveRamp in November 2025 and brought to version 1.0 in July 2026, does travel at impression time, carried as an encoded embedding inside the bid request. Its entire object is five fields: version, vector, dimension, model, and a type flag distinguishing identity from contextual from reinforcement signals. There is no consent field, no provenance field and no policy field.

Why the last four attempts failed

Portable, provenance-carrying signal standards have been tried repeatedly in ad tech, and the failure pattern is consistent enough to be predictive.

Seller Defined Audiences is the closest analogue and the sharpest warning. It was a portable, seller-authored audience signal riding the same bid-request carriage Agentic Audiences uses now. Publishers adopted it and buyers did not. The Tech Lab's own Benjamin Dick said in 2022 that there was no stick compelling agencies and DSPs to adopt it, and buyers objected that the construction methodology was opaque. It was rebranded Curated Audiences in December 2024. The Data Transparency Standard launched in 2019 and reached version 1.2 in April 2024, with Yahoo becoming the first DSP to adopt its labels in January 2025, a delay the Tech Lab's own CEO called a travesty. DigiTrust died in July 2020 on a funding asymmetry: publishers and brands never funded it, platforms did, and platform adoption did not keep pace. Privacy Sandbox is the largest-scale case, retired with Protected Audience running on 0.22% of Chrome page loads.

Against those, the two standards that did succeed are instructive for being unglamorous. Ads.txt reached roughly three-quarters of European news publishers and sellers.json around 61%, because both were nearly free to implement and had buy-side enforcement behind them. The pattern is not about schema quality. Every failed standard cost the party who had to populate it, benefited a party under no obligation to pay for it, and had nobody on the demand side willing to refuse a bid that lacked it.

There is one further caution specific to consent payloads. In May 2025 the Brussels Court of Appeal held that the consent string at the heart of the industry's transparency framework is itself personal data when combinable with an IP address, and that its steward is a joint controller. The portable consent object became the compliance liability rather than resolving it. Any design that packages permission into a travelling object inherits that risk and has to answer it in the design rather than in the marketing.

The design templates worth borrowing are outside advertising. Content Credentials bind assertions, a claim referencing them and a signature over the whole, with both cryptographic and perceptual bindings, verified against a trust list. Verifiable Credentials became a W3C Recommendation in May 2025 with an issuer, holder and verifier model and selective disclosure, which is exactly what a signal object needs, since a buyer should be able to verify a consent basis without receiving the underlying evidence. The IAB's own AI transparency framework already recommends Content Credentials as the machine-readable disclosure layer for advertising, but only for creative assets. Nobody has drawn that bridge to targeting signals, and the weakness to pre-empt is that manifests get stripped when content crosses platforms that do not support them.

The sceptical case is real

Adoption is genuine and tiny. MiQ's AdCP tests have been described as five-figure, several pilot buys have been measured in dollars, and a Digiday survey found only 18% of respondents thought the industry was ready to scale this. Gartner found roughly 130 of thousands of self-described agentic vendors actually deliver autonomy.

The structural objections come from serious people. Jeff Green warns that agent-to-agent direct deals recreate hundreds of thousands of ad networks. Ozone's Craig Tuck argues that outcomes-at-any-cost re-imports programmatic's margin games. Security research adds concrete failure modes: prompt injection is the top agentic risk in OWASP's 2026 listing, agents hallucinate plans and tools, and multi-agent delegation diffuses accountability. Those are the reasons AdCP's signed requests and audit logs exist, and they are also the reason the walled gardens, projected to take just over 80% of 2026 programmatic spend, may simply ship their own agents and never adopt an open protocol at all.

Scope3, the most visible agentic-native company, cut staff twice between August 2025 and February 2026 and divested Adloox to Peer39. Agentic capability shipped fastest inside incumbent platforms rather than inside the companies that named the category.

Chapter 7

The economy underneath

The forcing function everyone planned around evaporated in 2025, and the shift toward context kept moving anyway. It just moved through different plumbing than predicted.

The reversal that changed less than expected

Google announced on 22 April 2025 that it would not remove third-party cookies and would not ship a user-choice prompt. On 17 October 2025 it effectively shut down Privacy Sandbox, retiring roughly ten APIs including Topics, Protected Audience and Attribution Reporting, with code removal running from Chrome 144 in January 2026 through Chrome 150 in July 2026.

The paradox is that cookies survived in Chrome while the industry's replacement project died. Addressability did not come back. Safari and Firefox still block third-party cookies across a large share of traffic, Firefox 145 cut uniquely fingerprintable users substantially, and consent rejection degrades a majority of addressable open-web traffic regardless of what Chrome does. Contextual demand therefore repositioned rather than collapsed, moving from a cookie-replacement pitch to a performance and relevance pitch, strongest in CTV where it has become the default premium video story.

Not everyone agreed at the time. iCrossing predicted contextual investment would decrease significantly after the reversal, which is worth recording as the contrarian call that the transaction data did not support.

The market-size numbers are not worth much

Published estimates cluster between $220B and $260B for 2025 and 2026, with outliers above $300B. The Business Research Company puts 2025 at $233.89B rising to $258.32B in 2026. 360iResearch says $225.78B to $250.64B. Global Market Insights says roughly $222.8B for 2025. Earlier editions from the same publishers disagree with the current ones by more than $80B.

These figures typically count search and much of native and display as contextual, which makes them close to unusable for judging the health of the vendor category this paper is about. Any number in that range is a statement about definitions rather than about the market.

What contextual is actually worth against the alternative

The question every finance director asks is the one the vendor literature never answers: what does contextual earn relative to the identity-based targeting it replaces. There is no independent field experiment at scale. The industry bodies are occupied with cross-media measurement rather than targeting-method comparison, and every study showing contextual winning is vendor-produced.

The academic literature looks chaotic and is not. Published estimates of what publishers lose when identifiers disappear run from 4% to 70%, which both sides cite as evidence the other side is guessing. The spread is a controls problem. One study walks the same dataset of 41.8 million impressions from a 60.9% loss with no controls to an 18.3% loss with full publisher and advertiser fixed effects. Another shows the same pattern in miniature, a raw gap of 59% collapsing to 4% under proper adjustment. The defensible central estimate for the impression-level identifier premium is roughly 18 to 23%, and every figure outside that band is explained by a methodological choice. Citing 4% as fact is as vulnerable as citing 52%.

On head-to-head value, two peer-reviewed studies with different designs and different datasets reach the same answer. A 2021 Marketing Science paper ran behavioural-only and contextual-only models over the same 27.5 million impressions and found relative information gain of 12.27% for behavioural against 5.12% for contextual. A 2024 Journal of Marketing Research paper covering 3.7 billion impressions found topic-matched pages offset 42 to 44% of the revenue lost to consent refusal, with the net effect still negative. Contextual recovers a little under half.

The harm contextual actually causes

A paper arguing for contextual that does not address keyword blocklists defunding journalism is advocacy. The widely quoted $2.8B figure for over-blocking is weak evidence and should be retired: it extrapolates from 225 articles across 15 news sites on a single day in July 2019, was produced by a vendor selling an alternative to blocklists, and its stated calculation does not reconcile to its own headline number.

The replacement is considerably better. A 2026 study funded by the National Science Foundation scored 4,352 news articles across 51 domains through three major verification providers. Nearly 44% of news articles were flagged unsafe by at least one provider, and inter-rater reliability across the three ran between 0.19 and 0.53, well below the conventional acceptability floor, with 87% disagreement on hate speech specifically. Three vendors classifying identical URLs agree barely better than chance.

Industry evidence points the same way. IAS's own work with Reuters found that 54% of Reuters news URLs cleared by its contextual tool would trip a typical keyword blocklist. Reach's Mantis found 45 to 47% of Euro 2024 and Super Bowl coverage misflagged. One UK publisher was running a blocklist of 34,000 words across 22 languages. This is the strongest counterattack on any pro-contextual argument, and the honest response is that it indicts generation-one keyword systems specifically, which is also the generation this paper documents dying.

Regulation is the durable driver

Cookie policy turned out to be reversible. Privacy law has not been. In the EU, France's CNIL fined Google €325M and SHEIN €150M in September 2025 over consent failures, and a December 2025 decision extended GDPR to a non-EU adtech processor doing audience segmentation. The Brussels Court of Appeal ruled in May 2025 that TC Strings are personal data and that IAB Europe is partly a joint controller. EU AI Act obligations landed on 2 August 2026, including disclosure duties for emotion-recognition systems, which touches the category in chapter 3 directly.

In the United States, twenty comprehensive state privacy laws are in force in 2026, with Indiana, Kentucky and Rhode Island joining on 1 January. Maryland bans sensitive-data sale outright. Minnesota grants a right to an explanation of profiling. Each of these raises the cost and legal risk of identity-based targeting relative to contextual, and none of them can be reversed by a browser vendor changing its mind.

Bifurcation, not substitution

The counter-evidence deserves equal weight. Retail media grew from $58.79B to $69.33B in the US between 2025 and 2026. Clean rooms and logged-in walled-garden identity are both healthy. Google, Amazon and Meta are building proprietary agentic stacks rather than adopting open protocols, and walled gardens are projected to take just over 80% of 2026 programmatic spend.

So the honest synthesis is not that context replaces identity. It is that the two are separating by venue. Identity retreats into authenticated environments and retail media, where it works well and is legally defensible. Context becomes the operating system of the open web and CTV, where identity does not work well and is getting more expensive to defend. Anyone arguing that one wins outright is describing half the market.

Chapter 8

How many vendors, really

The commonly repeated figure of about 3,000 contextual vendors does not survive contact with any countable source. The defensible answer has two tiers and a much smaller core.

No landscape source counts anything close to 3,000 contextual-specific companies. The chiefmartec supergraphic reached 15,505 total products in May 2026 across all 49 marketing categories, up just 121 net year on year, with 1,488 added and 1,367 removed. That is a churn machine rather than a growth market, and advertising is one domain of six inside it. LUMA maintains 23 landscapes with no dedicated contextual map and publishes no per-category counts. IAB Tech Lab claims 700-plus member companies in total, meaning the entire standards-participating industry. At the narrow end, CB Insights names about 15 vendors in its contextual targeting category.

So the two-tier answer is the one to use, and it should be presented with its method rather than as a census. Roughly 500 to 1,500 companies now market a contextual feature, because essentially every SSP, DSP, verification vendor, curation platform and data provider has added one. Somewhere between 50 and 150 have contextual intelligence as their core product. The nameable, investable core is 30 to 60 companies, and this study scored 51 of them.

A quotient, and why not a sum

The matrix carries a summary column, the Contextual Quotient. Getting to it required rejecting the two obvious approaches, and the reasons are worth stating because they are the same reasons the rest of this paper distrusts vendor scorecards.

Adding the eleven scores is wrong. It counts an unscored dimension as a zero, so it penalises a company with a thin public record rather than a weak product, and 83 of the 561 cells here are unscored. It also rewards breadth over depth, which would invert the finding this chapter is about. A reach machine with shallow understanding would outrank a deep specialist, and the paper would be arguing against its own summary statistic.

Averaging the scored dimensions is the mirror error. A vendor with eight scored dimensions gets averaged over only its strong ones, so missing evidence becomes an advantage. In the first draft of this index that put a company with nine scored dimensions at the top of the field on the strength of what nobody had documented about it.

So CQ does three things instead. It groups the dimensions into Depth, Reach and Standing and takes the mean within each group, so unscored cells are skipped rather than counted. It weights the groups 45/25/30, favouring what a vendor understands over where it can act. And it shrinks each result toward the field mean in proportion to how little evidence backs it, which is the standard correction for exactly the small-sample flattery described above.

The result that survives all four weightings is worth sitting with. The top of the capability index is Verve, because Moments.AI's web depth, ATOM's on-device in-app inference and a scale-5 balance sheet combine on paper into something no competitor matches. Chapter 6 shows the two halves emit incompatible signal formats and cannot be sold as one object, and chapter 10 shows the company quietly demoted the contextual half out of its investor reporting. The strongest documented capability envelope in this field belongs to a company that took it apart. That is the paper's argument in a single row, and the author's declared interest in that company is marked in the table.

Six clusters, and what they reveal

Sorting the matrix by different columns exposes a structure that a vendor list does not. Nine vendors genuinely read content with multimodal or language-model machinery: Seedtag, GumGum, Overtone, Cognitiv, Samba TV, KERV, Wurl, Anoki and Zefr. None of them exceeds 4 on scale, and only Wurl reaches 4, as a subsidiary whose parent's attention is elsewhere.

A second cluster wins on throughput and neutrality rather than depth: Peer39, Proximic, Mantis, Weborama, Sirdata, Channel Factory and 4D. These are the category's working infrastructure. A third cluster measures whether a moment was seen without measuring what it meant: Adelaide, Lumen, TVision and Sincera. They are the best-evidenced vendors in the entire field, and none of them reads content.

A fourth cluster is reach machines wearing contextual clothing, where the context is device state, survey persona or app graph. A fifth owns the transaction point and hosts everyone else's meaning: IRIS.TV, PubMatic, Magnite, Index Exchange, Equativ and Optable. Almost every high agentic-readiness score in the field sits in that cluster. The sixth is emotion specialists, and there are five of them.

The consolidation record supports the same reading. Both verification incumbents that defined the brand-safety era left the public markets within twelve months: IAS taken private by Novacap for $1.9B at $10.30 a share, closing in December 2025, and Nielsen's $2.15B all-cash acquisition of DoubleVerify announced on 6 August 2026. Around them the deal ledger ran hot, with contextual and signal assets absorbed by telcos, retailers, agencies, DSPs and measurement companies. T-Mobile bought Vistar and Blis. Viant bought IRIS.TV and TVision. The Trade Desk bought Sincera. Experian bought Audigent. Publicis bought Lotame and then LiveRamp. Peer39 bought Adloox from Scope3. The standalone middle is vanishing, which is the subject of chapter 10.

Chapter 9

Ten players, five tiers

A closer read of the ten companies that matter most, scored across twelve dimensions, written from the vantage point of one of them. Every score below survived a hostile review that changed four of them.

The tiering is not by size. It is by what each company is structurally able to sell.

Tier 1, full-stack contextual leaders: Seedtag and GumGum. The only two combining a proprietary meaning engine, a differentiated proof layer, owned media execution and multi-market scale. Both fit the survival pattern from chapter 10: context attached to media execution rather than licensed as data alone. They are near-identical in capability and opposite in trust architecture. GumGum's derivation is MRC-accredited; Seedtag's is unpublished, and because Seedtag simultaneously owns the signal, the exchange and the supply, it is structurally harder to audit.

Tier 2, verification giants pivoting from policing to targeting: DoubleVerify and IAS. Roughly $600M to $750M of revenue each and the broadest omnichannel footprints in the market, including social walled gardens the others cannot reach. Their meaning is suitability-first, with thin emotion at IAS and none at DoubleVerify. Both are exiting public markets under pressure while their measurement cores deflate, which is exactly why both are pushing into pre-bid optimisation.

Tier 3, CTV scene-level specialists: KERV, Anoki and Wurl. The deepest video understanding in the panel and the narrowest reach. KERV has patented object-level computer vision inside Warner Bros. Discovery, NBCUniversal and Amazon. Anoki has a peer-reviewed multimodal engine live on Spectrum Reach and Allen Media. Wurl does zero-latency Plutchik emotion matching across 140 billion monthly impressions. All three are single-channel by design.

Tier 4, pre-bid data rails: Peer39. Distribution-ubiquitous inside 13-plus DSPs, best-in-class recency, and cheap. The engine underneath is taxonomy-era with no disclosed rebuild. Its June 2026 acquisition of Adloox escaped standalone data licensing into verification, which this panel's own Tier 2 shows to be deflating.

Tier 5, flankers and boutiques: Ogury and Illuma. Ogury attacks the same identity-free budgets from the audience side with survey and purchase data, and has verifiably no emotion capability at all. Illuma is a micro-cap with the panel's only licensed publisher access.

What the red team changed

The competitive analysis was reviewed by an adversarial pass instructed to attack the logic rather than the formatting. Thirteen challenges, six sustained, five requiring caveats, two overruled. Four of them moved scores, and the pattern in them is worth more than the individual corrections.

Seedtag's agentic readiness went from 4 to 3 because a demo cannot outrank shipped products. GumGum's Mindset Agent does brief-to-segment activation with The Trade Desk, and Ogury's SONA ships live connectors inside ChatGPT, Claude and Gemini. Liz Agent is impressive and sits behind a protocol hedge. That was home-team flattery, and it was the only cell in the analysis where it appeared.

DoubleVerify's in-app score went from 4 to 3, and this one changed a strategic conclusion rather than a number. The original analysis simultaneously claimed in-app was white space and scored DoubleVerify a 4 there, which cannot both be true. Resolving it in favour of the staleness argument, since those claims date to 2019 through 2021, makes the white space real and reframes the opportunity as displacing a dormant incumbent.

Peer39's real-time recency went from 5 to 4 for symmetry: Seedtag took an evidence haircut for unaudited vendor claims, and Peer39's own contradictory figures had not. Applying a standard unevenly is how an analysis loses a room.

Where this leaves Seedtag

Three threats are sharper than the rest. DoubleVerify has the only shipped agentic rails in the panel, with MCP and AdCP support, a Claude-connected agent and an autonomous activation agent scheduled for Q3 2026, and it is about to be fused with Nielsen's audience currency. Price compression is arriving from both directions at once, with enterprise verification bundling from above, cheap ubiquitous pre-bid data from below, scene and emotion segments given away free with Wurl's supply, and foundation models commoditising base classification. And the emotion differentiator has a fast-follower: IAS already runs named-emotion classification in 45 languages, while Seedtag's emotion proof rests on one company-commissioned EEG study.

The white space is the mirror image of chapter 5. Content-level in-app contextual, where every player in the panel scores between 0 and 3 and the best asset has not moved in five years. Audio and out-of-home, absent panel-wide beyond IAS's nascent podcast segments. A single cross-screen contextual and emotion currency spanning web, app and CTV, where Seedtag is closest and missing only in-app. And independently audited proof, where all ten players self-publish their performance numbers and GumGum has already demonstrated that accreditation converts into share.

Not everyone in the panel is a competitor. Illuma's paid Guardian and News UK licences answer the publisher-compensation exposure that threatens a 30,000-publisher crawl, though the acquisition case needs diligence on whether those licences are exclusive or transferable, because a paid API licence reads like something any competitor could buy directly. Anoki is a buy-or-partner candidate for live-TV scene science. KERV is a shoppable-format partner. Peer39 is a rails frenemy and a permanent price anchor.

Chapter 10

Why contextual vendors die

Contextual data licensing looks more like a feature than a business. In every case examined here the survivors attached context to media execution, and the pure signal-sellers sold, shrank, or were quietly retired. This is a pattern across a small number of cases, offered as a hypothesis with its evidence rather than as a demonstrated causal law.

Moments.AI is the most useful case in this chapter because it is not a death. It is a demotion, and demotion is what the common outcome actually looks like.

The facts, from filings. Verve's parent acquired Beemray Oy on 17 May 2021 in an asset deal for less than €1M, with the parties agreeing not to disclose the exact figure. For scale, Verve paid €25.6M for Captify in September 2025, so the contextual asset cost roughly 4% of what the company later paid for search-intent data. Exactly one patent traces to Beemray, and it covers battery-efficient geofencing rather than any semantic or language technology. The acquisition's "IP rights to patents" resolves to location technology.

The product is still alive and still sellable. Verve's FY2025 Annual and Sustainability Report, published April 2026, names it as an award-winning contextual targeting solution for web, CTV and mobile applications, and places it in the DSP segment. Verve's live developer documentation, timestamped 28 January 2026, lists MomentsAI as a selectable audience name in deal targeting, priced as a segment CPM data fee.

What changed is its standing. The phrase "innovative products such as ATOM and Moments.AI" appeared in the Q3 2025 and Q1 2026 interim reports and was deleted from the Q2 2026 interim report on 27 August 2026, which rewrote the company description to "mobile advertising intelligence technology company" with no product names at all. It is absent from Verve's current homepage and its DSP product pages. There has been no trade coverage since a February 2024 case study, and Verve's flagship March 2026 launch of conversational intent signals does not mention it. It has no named brand customers in any public source, and its single case study has an unnamed advertiser and two competitors identified only as Alternative Vendor 1 and 2.

The methodological point survives, and it is stronger in this form than it would have been as a death notice. Trade press records bankruptcies and does not record demotion. A capability can remain purchasable inside a platform's own interface while ceasing to exist as a business, and nothing announces the transition. Anyone studying this market has to read investor filings and developer documentation, because that is the only place the change is visible.

The five lessons

Data licensing behaves like a feature. Segments piped into DSP marketplaces earn pennies per thousand impressions with no workflow lock-in, and the demand side can rebuild or reroute the capability whenever it wants. IRIS.TV was the best data-only asset in CTV and could only exit into a DSP for an undisclosed and likely modest sum on around $30M raised.

Corporate shelter is not survival. The two cleanest deaths here, Grapeshot inside Oracle and Moments.AI inside Verve, were not product failures. They were parent-company strategy shifts. Oracle recalculated its GDPR liability and its relevance threshold. Verve pivoted to on-device targeting and its performance DSP.

Capital structure kills before product does. Private equity debt at MediaMath and Big Village, and subscale public listings at Kubient and Mirriad, are recurring proximate causes. It is worth separating proximate causes, meaning debt, fraud and cash-out, from the ultimate cause, which is the absence of a defensible monetisation layer.

The real-time thesis was right about value and wrong about capturability. Same-day and live contextual is now consensus direction. But language models collapsed the cost of building good-enough classification, so the capability commoditises faster than any point vendor can price it. Being right about the future is not the same as being able to charge for it.

The next deaths will also be invisible. The remaining independents face a three-way fork: attach to media or curation, sell to a platform or verification consolidator, or wind down quietly. Watching for the third requires monitoring domains and filings rather than press releases.

Chapter 11

What is actually missing

Four things this field does not have, in the order someone should build them.

0. First, define proof

This paper is called Proof and Story, and it has so far used proof loosely. Accreditation, ontology validation and campaign lift get cited as if they were the same currency. They are not, and a buyer asking a vendor for proof should be clear which of four different things is being asked for.

Classification validity. Does the system label content correctly? Precision, recall, calibration, and performance by language. This is what an MRC-style content-level accreditation examines, and it is the only layer where the field has any independent audit at all.

Decision consistency. Does the same input produce the same output over time? Stability, drift, and reproducibility across model updates. Nothing in this study publishes evidence on this layer, which matters more than it sounds: a classifier that silently re-labels the same page after a model change breaks every campaign built on the old labels.

Delivery fidelity. Did the impression actually run in the context that was bought? This is the layer where the January 2026 inter-vendor disagreement study and the news over-blocking evidence in chapter 7 both bite, and it is separate from whether the classifier was right in the lab.

Causal effectiveness. Did it produce an incremental outcome against a fair baseline — not against no targeting, but against plain contextual and against identity-based targeting? Chapter 3's effect sizes almost all fail this test, because they compare against weak baselines and are funded by the party being measured.

1. Audited meaning

Every vendor in this study self-publishes its performance numbers, and the ones with the best evidence are the ones measuring whether a moment was seen rather than what it meant. GumGum has already proved that accreditation converts into commercial share, and its accreditation covers safety and suitability rather than emotion. Nobody has submitted an emotion or attention engine to independent audit.

The argument against doing it is that publishing methodology helps competitors. That argument was stronger when derivation was expensive. Now that a foundation model can approximate a vendor's base classification, the methodology is worth less than the trust that auditing it would buy. It requires no new science, which is not the same as requiring nothing: independent accreditation is slow, costly and operationally demanding, and it constrains how quickly a model can then be changed. The point is that it is the only differentiator here that depends on no invention.

2. Content-level in-app

Chapter 5 showed the gap and chapter 9 showed why it persists: the only content-level in-app stack in the panel has not moved since 2021. In-app is where reach and meaning are furthest apart, which makes it the largest addressable capability gap rather than the largest greenfield. The hard part is architectural. Without a page to crawl, understanding has to happen either on-device, where it cannot easily emit a portable signal, or through publisher-side instrumentation, which requires the app developer to cooperate.

3. A signal object that carries its provenance

Chapter 6's figure is the specification. Six parts: intent, meaning, provenance, policy, execution, evaluation. AdCP standardises four of them and the Tech Lab's runtime work standardises two, and the join is unspecified. The part that gets dropped most often in practice is provenance, which is also the part that decides whether this machinery produces a context economy or a faster surveillance economy.

Verve is the clearest illustration of the cost of not having this. It owns a deep web contextual engine and the only production on-device in-app inference engine in the study, sitting inside the same company on the same balance sheet. They emit incompatible signal formats. The result is that a company with both halves of a web-plus-app contextual product cannot sell one portable object across surfaces it already reaches, and it retired the web half rather than joining them.

4. The story layer

This is the one with no competition. The pieces exist in separate industries: narrative arc detection in academia and in StoryFit's studio tooling, emotion-scored creative at DAIVID and Reticle, emotion-scored content at Wurl and Anoki, and archetype detection at Quilt.AI. Nobody has joined them, and the join is the product.

Two honest cautions belong with that. The first is that archetype layers built on psychographic personas drift straight back into audience profiling, which is the thing contextual was supposed to move away from. Built on content, they do not. The second is that this is a media-planning construct throughout. It classifies what content expresses and which brand identity it suits. It does not claim to know what any individual viewer feels, and the moment it starts claiming that, chapter 3's legal boundary stops protecting it.

Chapter 12

Method, and what it got wrong

How this was built, what the verification pass caught, and the full ledger of corrections. Published because an analysis that hides its errors is asking to be trusted rather than checked.

How it was built

The research ran as parallel streams rather than as a reading list. Nine thematic agents covered meaning derivation, emotion, narrative and archetype, agentic architecture, the contextual economy, omnichannel expansion, the vendor landscape, a dated timeline of the last twelve months, and the machine-learning foundations underneath. Four sweeps profiled vendors by segment: web, CTV, in-app and infrastructure. Ten agents built company profiles for the competitive panel.

Every one of those streams was then handed to a separate agent whose instructions were to refute it. That agent researched the claims independently and was told to be strict about numbers, dates, funding, ownership and superlatives, and to mark a claim partially confirmed when it was directionally right but wrong in a specific. The competitive analysis got an additional pass from an agent playing a hostile review panel, attacking the logic rather than the facts.

Two things about this process are worth stating because they affect how much weight the paper deserves. The Seedtag profile was built twice by independent agents so the two could be reconciled. And the verification pass corrected the brief this paper was commissioned from, not just its findings, which is the Moments.AI correction in chapter 10.

What it caught

The corrections fall into recognisable classes. Wrong attributions, such as the media-context meta-analysis credited to the wrong authors. Numbers that were real in the literature but traced to a different source than the one cited, as with the GoEmotions figures. Vendor claims contradicted by the vendor's own other pages, as with Peer39's throughput. Dates off by days or weeks, such as the Vistar close and the AAMP naming. Counts that were simply incomplete, such as Seedtag's acquisitions, where a fifth deal had been missed.

Two corrections changed arguments rather than facts. The first is Moments.AI. The second is the vendor count: the widely repeated figure of about 3,000 contextual vendors has no countable source behind it, and does not appear in the No Fluff Advisory corpus either, so chapter 8 replaces it with a two-tier estimate and its method.

Reproduce it yourself

Every derived statistic in this paper comes from the scored data below. It is published so a reader can recompute the numbers rather than take them on trust, which is the standard the paper asks of vendors.

Sampling and scoring, stated plainly

The 51 vendors are not a census. They are the nameable core of companies whose primary product is contextual intelligence, plus the verification, curation and channel-specialist vendors that compete for the same budget, assembled from the vendor-segment sweeps described above and cross-checked against published landscape sources. Chapter 8 gives the two-tier estimate this sample sits inside. A vendor absent from the matrix is absent because the sweeps did not surface it, not because it was judged too small.

Scores run 0 to 5 and are evidence-bounded rather than absolute. A 5 means the public record documents the capability in operational detail; a 3 means it is real but thinly evidenced or narrow in scope; a 1 means adjacent or incidental. 0 means the record shows the capability is absent. n/e means the record gives no basis to score either way. Those two are deliberately not merged, because in emotion and narrative the difference between proven absence and absent evidence is most of the finding.

Where evidence conflicts, the ladder is: independently verified, then third-party measured, then vendor-claimed, then vendor-claimed with internal contradictions. Chapter 9 documents the two occasions where applying that ladder unevenly was caught and corrected.

What to distrust in this paper

Five things, named so they are not discovered later.

Two claims were flagged as time-critical by the completeness review and have since been checked against primary sources.

EU AI Act. The European Commission confirms that the transparency obligations, which cover disclosure for applicable emotion-recognition systems, apply from August 2026, and that the Article 5 prohibitions have applied since February 2025. The AI Omnibus moved certain high-risk obligations to 2 December 2027 and product-integrated high-risk systems to 2 August 2028. Chapter 7 rests on the transparency date and the Article 5 line, and both hold.

Nielsen and DoubleVerify. Now confirmed from Nielsen's own newsroom and an SEC filing rather than secondary reporting: announced 6 August 2026, approximately $2.15B enterprise value, all cash at $13.60 per share, a 30% premium to the 60-day volume-weighted average. The transaction is pending, expected to close by the first quarter of 2027 subject to shareholder approval and regulatory clearance. It is described as pending throughout this paper; any reading of the two companies as already combined is premature.

The remaining three are structural.

Scores are judgments on an evidence ladder, not measurements. A 4 and a 3 differ by the quality of available evidence as much as by capability, which is deliberate given the audit gap this paper documents, and it means a vendor can be underrated for being private rather than for being weak.

Vendor-published figures are marked as such throughout and are still vendor-published. Where a company's own properties contradict each other, this paper uses the architectural argument instead of the number, and says so.

The competitive analysis was written from Seedtag's vantage point. That shapes which questions got asked, even with the red-team pass. A reader at a different company should re-run chapter 9's logic from their own position before using its conclusions.