Read this first
This is the short, plain form of a long research paper. The full paper is called Paying Attention: The State of Attention Measurement in 2026. It has the sources, the tables and the math. This page has the same findings in fewer and shorter words. The counts come from the same data file as the paper. The other numbers are copied from it. Each part links to the place in the paper that backs it up.
The research runs to 1 October 2026. It looks mostly at the US.
What an attention score is
An ad can load on a page and still go unseen. So the ad trade first learned to count ads that could be seen. That is called viewability. It asks: was the ad on the screen, and for how long? That "how long" is time in view.
Attention asks the next thing. Did a person look at the ad, or hear it? Firms that sell attention scores say that is what counts. Buyers now use those scores to plan, to bid and, now and then, to set the terms of a deal. Chapter 1 sets out what gets measured.
The paper keeps four steps apart:
- The chance to see or hear the ad.
- Attention: did a person look or listen?
- Memory: did they recall the ad or the brand?
- What they did: did they search, visit or buy?
A score at step 2 does not prove step 3 or 4. That gap is the heart of this paper.
What we found
1. Almost every score you buy is a guess made by a model
Few scores come from watching the people who saw your ad. Most come from a model. The model learns from a small group of people who were watched, with eye tracking or a webcam. Then it guesses attention for the ads you buy. It works from signals like ad size, place on the page and time on screen.
So a campaign score is a model's guess, or an index of what the device did. It is not a count of eyes. One firm, TVision, puts sensors in homes that signed up, and counts who looks at the TV. Even its market numbers are scaled up from that panel. Each firm has its own unit, too, so you can't compare one firm's score with another's. Chapter 1 has the details.
2. There is a rule book now, but only one score has passed the check
In November 2025 the trade's two main bodies, the IAB and the MRC, put out final rules for attention scores. A draft came out in May 2025. The MRC is the body that checks how ad numbers are made. It calls that check accreditation.
By late September 2026, one attention score had passed: DoubleVerify's. It first passed in December 2022, before the rules were final. It was renewed in April 2026. It passed on computers, the mobile web and phone apps. It has no pass for TV screens, and the MRC's lists showed no review of that under way. The MRC's letter does not list the new attention rules. It lists other rules. So, by the paper's rule, that test came out "no" at that date. Whether the new rules shaped the check is not public. A second firm, IAS, left the MRC's review in January 2026. A third, Adelaide, is at an early step, and its full check has not begun. A fourth, XPLN, is part way through its full check, for the mobile web and phone apps. Chapter 2 walks through the record.
In the paper's view, passing the check shows how the number is made and told. It does not show that the score predicts memory, brand or sales.
3. No one has shown attention beats time in view
Here is the key test. Take the same ads. Predict memory or sales from time in view. Then add attention and predict again. Do it on data the model never saw. Does adding attention help?
This research looked for such a test with its design and results made public. In the sources it checked, it found none. The closest studies test on the same data they learn from, and they can't be compared with each other. A trade group had firms score the same ads. Two firms that both use eye tracking hardly agreed.
No fair test was found of buying on attention, with the cost counted, against buying the usual way. So the answer is "not shown", not "shown to fail". Chapter 3 goes through the studies.
4. Most of the proof comes from the firms that sell the scores
The firms do publish studies. But none the paper checked counts as independent. Most come from the firm itself, or from a partner, with the method kept private. Two cut against the pitch. In a review that Lumen paid PwC for, attention was not clearly better than viewability on memory. In one IAS case, high and low attention showed no real gap in sales for plain web ads. No study the paper checked says what the extra cost of buying on attention was. Chapter 4 has the list.
5. The big firms changed hands, or agreed to
The market has moved fast. Oracle shut its Moat tools on 30 September 2024. A buyout firm, Novacap, took IAS private on 23 December 2025, at $10.30 a share. Viant, which runs a tool for buying ads, bought TVision on 1 May 2026, for about $37.5 million. On 6 August 2026, DoubleVerify agreed to be bought by the parent of Nielsen, at $13.60 a share. That deal was not done by the cutoff.
Some ask a hard question. Do the big checking firms hurt the small ones when they add attention to their other tools? No source the paper checked links the two. So that stays open. Chapter 5 has the deals.
6. Attention is traded, but no outside check was found
Attention firms and sellers say you can bid on their scores in the main ad buying tools today. Of the tools, only Google's own help pages were found to say so. A few sellers go further and promise a level of attention, or set the price by it. Ten such offers were found. Only two say what the level is. Only a few say what you get if they miss it, and the answer is more ads for free. Where an offer says who checks, it is the firm that sells the score. No outside check was found.
Two offers price ads by attention itself. SeenThis, from June 2026, sells by the attentive second, with Lumen as the scorer. It says it sells around the world. An older offer, from 2024, does not say how it measures a "seen" ad. Neither shows sales in the US.
No offer says how much is spent on it, out of what total. So the paper finds attention is traded, but no one says how much. A true trading unit would pass four tests. It would pass the MRC check where it is sold. It would set a price or a promise, with the level and the fix in writing. Someone else would check it. And how much is sold on it would be made public. No score found passes all four. Chapter 6 lays out the offers.
7. Buyers mostly use it to plan
Each of the big ad agency groups has a deal with an attention firm. Most use the scores to plan where to buy. Some have used them to steer bids. 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.
In the IAB's survey of US ad buyers, 36% said they would focus more on attention in the next year, in late 2022. A year later, 47% did. A year after that, it was 47% again. That is a plan, not money spent. Chapter 7 covers the buyers.
8. It is measured most on screens with a web page or video
Scores are live on web pages, online video and TV screens. On TV, one firm counts real people and the rest make a model. On social apps, scores come through deals with Snap and TikTok. In the sources checked, none is in Meta's or YouTube's partner lists. One firm says it measures Facebook and Instagram through their reporting tools. Sound, games, street screens and store sites are thin, or seen only in one-off studies. Chapter 8 has the map.
9. The law cares about faces and feelings
Two things decide the legal risk. Can the data tell who a person is? And is it used to guess how they feel? In the EU, guessing feelings from a face or a body for ads is not banned, but it counts as high risk. The duties for high-risk tools start in December 2027. Since August 2026, people must be told when it is used on them. Some US states have strict rules on face data. The paper's searches found no privacy lawsuit or regulator action against an attention firm. Not every firm was searched by name. Chapter 9 sets out the rules.
10. AI agents can hold an attention score, but no live use was found
More and more ads may be bought by software agents. The paper checked the main agent rule sets. AdCP, in its release of 23 September 2026, has a place to hold an attention score. No real firm's score was shown in a live deal. Only a test with a fake attention firm passed. In late September 2026, the AdCP list of 34 agents had none that measures. Chapter 10 checks each step.
How the products were scored
The full paper scores 20 products on up to 12 areas. There are two groups. One group measures attention on the ads you buy. The other tests an ad before it runs. Each area is scored from 0 to 5, only from what is on public record. A missing fact is marked "no evidence", not 0. The scores roll up into the Focus Quotient.
The Focus Quotient shows what firms have put on record. It does not show which product works best. Chapter 11 has the scores and how they move.
What a buyer should ask
Before you pay for attention, ask for these things:
- The MRC's own record for the score, in the place you buy.
- What is seen and what is guessed by a model.
- A test on fresh data that beats time in view, for the result you want.
- The level promised, the fix if missed, and who checks.
- What it costs you on top, all in.
- The model version the deal settles on, and a fresh check after each update.
- Which signals the score uses, whether any can tell who a person is, and the consent behind them.
Chapter 12 turns this into a checklist.
How this was done
AI research agents did the work from one brief. A second agent checked each part and tried to prove it wrong. In all, 985 sources were opened and logged, and the paper holds 584 claims. Four outside reviews by another AI model and two passes of our own tested the work. Each fix is logged. This is AI-assisted research with a public method. It has not been audited or peer reviewed. No vendor saw it before it came out.
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, one of the agent rule sets in chapter 10. 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. Chapter 13 has the method and the full list of fixes.