Read this first
This is the short, plain version of a long research paper. The full paper is called Proof and Story: The State of Contextual Advertising in 2026. It has the sources, the scores and the math. This page has the same findings in fewer and shorter words. The numbers are copied from the paper. Each part links to the place in the paper that backs it up.
The research runs to August 2026.
What contextual ads are
Many online ads are aimed at people. A system tracks what you do and guesses what you might buy. Contextual ads work the other way. They are aimed at the page, the show or the song. An ad for hiking boots goes next to a story about a trail, because of what the story is about, not because of who you are.
To do that, a system has to read the page or watch the show and work out what it means. That is the skill this paper looks at. Chapter 1 tells the story of how it grew.
How the reading got better
The first tools just matched words. They put airline ads next to plane crash news. They blocked the word "shot" in stories about vaccines. Then came shared lists of topics, so a buyer and a seller could use the same words. Then tools that read for meaning. Then tools that can also see pictures and hear sound. Now there are AI agents that decide what to do with what they read.
The way the reading is done has changed four times. But to be bought, each page must still get a code from a shared topic list. That part has not changed since 2015. Most of the mix-ups in this market today come from that gap. Chapter 1 has the five steps.
What we found
1. Reading content got cheap, so the fight moved to proof
Big AI models made it cheap to read the web. Brian O'Kelley, who runs Scope3, a firm that sells this kind of tool, gave one estimate. He said it cost about $45 million in 2023 to sort the web the way one brand sees it. By March 2025 it cost about $270,000. For 2026 he put it under $100,000.
When the hard part gets cheap, the edge moves to what is still hard. Right now that is two things. Proof that the reading is right. And a way to get the ads bought and shown. Chapter 2 goes through it.
There is a catch. No shared, public test exists that scores how well these tools sort ad content. So a buyer can't line up two firms and see which one reads better. The paper sees that gap as something the firms prefer, not an accident. Chapter 2 has the detail.
2. Ads that aim at feelings sold faster than they were proven
Between 2024 and 2026, aiming ads at the mood of a scene became a real product. Some tools score the scene just before a TV ad break for joy, fear, trust and so on.
The proof is thin. The best published tests of fine-grained emotion reading get about half right. Almost all of the proof is paid for by the firms that sell it. Five firms in the study score 4 or 5 out of 5 on emotion. We found none of them with its emotion model checked by an outside auditor. Only two firms, GumGum and Zefr, hold the MRC audit seal for how they read content. Both seals cover safety, not feelings. So on this evidence, you can't buy an audited emotion engine today. Chapter 3 has the evidence.
3. We found no one who sells ads matched to the story
Stories have shapes. A show builds, turns and ends. Brands have types too, like the hero or the rebel. Research on this has been around since 2020. But of the 51 firms studied, 27 showed no way to use it. Another 18 gave too little to score. Only six showed anything close, and just one scored high.
This research found no product you can buy that aims ads at a point in a story, or at a brand type. In this study, that is the biggest open space. It looks like a gap in how the market is set up, not a race between firms. It may also be empty because it is hard to do. Many TV ad breaks are set before anyone reads the story. Chapter 4 shows the empty column.
Where it reaches, and where it doesn't
Contextual moved past web pages in 2025 and 2026. Streaming TV now reads the scene, not just the show. Audio reads the words in each episode. Screens in public places know where and when they are.
Phone apps are the big gap. They reach lots of people, but there is no page to read, so the tools know very little about what is on screen. No firm covers web, apps, TV, audio, outdoor screens, games and store ads all at once. Chapter 5 covers each channel.
AI agents and the rules for them
AI agents that buy ads went from idea to real tools between October 2025 and August 2026. Two groups are writing the rules. One set of rules, called AdCP, covers how agents plan and run a buy, with checks and logs. The other, from the IAB Tech Lab, covers how the signal itself is packed and sent. Each group is building half of one thing, and they have not agreed how the halves join. Chapter 6 explains the gap.
The money underneath
Google said in 2025 that it would keep third-party cookies. Then it shut down its plan to replace them. But that did not make it easy again to aim ads at people on the open web. Safari and Firefox still block those cookies, and many people say no to them.
In one big study of ads bought by machine, most of the tracked money now goes through private deals. Their share went from 64.5% to over 92%. About three quarters of what flows through ad auctions is now pre-packed deals. The profit moved to the side that sells the ad space. Numbers for how big the contextual market is are close to useless. Rules on privacy, not cookie news, are what keep pushing buyers to context. Chapter 7 follows the money.
How many firms, and who leads
You often hear that 3,000 firms sell contextual ads. No count backs that up. The paper's answer: about 500 to 1,500 firms say they have a contextual feature. About 50 to 150 have it as their main product. About 30 to 60 are the real core. This study scored 51 of them on 11 areas. Chapter 8 has the full table.
The scores roll up into one index, the Contextual Quotient. It groups the areas into three parts: how deep the reading goes, how far it reaches, and where the firm stands, meaning its size, its proof of results and how ready it is for AI agents. A score counts only what a firm shows in public. Read the bands, not the order. When the index is built a bit differently, the top group barely moves. But firms in the middle can jump by up to 19 places. Chapter 8 explains the index.
The paper then looks closely at ten firms in five groups. The two that do the whole job, from reading the page to showing the ad: Seedtag and GumGum. The two big checkers now moving into aiming ads: DoubleVerify and IAS. The TV scene experts: KERV, Anoki and Wurl. A data rail found inside many buying tools: Peer39. And two that come at the same budgets from the side: Ogury and Illuma. Chapter 9 has the ten profiles.
Why firms in this field fail
Across the cases the paper looked at, firms that only sold the data did not last. The ones that lived tied the reading to the buying and showing of ads. A big parent firm was no shield. Money trouble killed firms before their product did. And the deaths are getting harder to see, because a tool inside a big platform can be shut down with no news at all. These come from a small number of cases. The paper offers them as an idea to test, not a proven rule. Chapter 10 has the cases.
What someone should build
Four things, in this order:
- An outside audit of how a firm reads content. No new science needed. It is the cheapest way to stand out. But a real audit is still slow and costly.
- Real reading of what is on screen in apps, where the reach is biggest and the reading is weakest.
- Rules that let a signal carry where it came from and what it may be used for. That needs the two rule groups to work together.
- The story layer. It is empty and the tools exist. But no one has shown it works yet, so prove it before selling it.
How this was made
AI agents did the research. The author drew the conclusions. The method is public. Nine research streams, four sweeps of firms by type and ten firm profiles fed it. A separate AI agent checked each stream and was told to try to prove it wrong. Another played a hostile review panel for the chapter on the ten firms. 168 findings survived. 117 claims were wrong or partly wrong and were fixed, and the paper lists every one. One pass set every score. No second person scored the firms to check. Chapter 12 has the method and the list of fixes.
The author, Evgeny Popov, works at Samba TV. Samba TV is one of the 51 firms scored, and it scores 5 out of 5 on how deep its reading goes. He used to run Moments.AI at Verve Group, which chapter 10 discusses. Verve is also one of the 51 firms scored, and it comes top of the index. None of the firms scored paid for this research, gave facts for it or saw it before it came out. Chapter 9 is written from one firm's point of view: Seedtag's. So read it as a case study, not a neutral view of the market.
Read the full paper
The full paper has the sources, the scores and each step of the math. Read Proof and Story in full.