ChatGPT ads are matched to the intent of a live conversation rather than bought against a keyword, so there is no term to research and no keyword auction to read.
There is still an auction, and there are still bids. What is missing is the thing every paid search workflow is built on: a phrase you can look up, size and compete for. That absence changes the first question you ask about this channel. It is no longer which keywords to buy. It is which of the answers you already win can now be bought out from under you.
OpenAI announced its intention to test ads on 16 January 2026 and began the United States pilot on 9 February, for logged in adults on the Free and Go tiers. The rollout since then has been quick and well documented.
| Date | What happened |
|---|---|
| 16 January 2026 | Test announced for Free and Go users in the United States |
| 9 February 2026 | US pilot begins |
| 26 March 2026 | Early update: no measured impact on trust metrics, low dismissal rates |
| 5 May 2026 | Self serve Ads Manager beta, CPC bidding, Measurement Pixel and Conversions API |
| 11 August 2026 | Ads launch in the UK, Mexico, Brazil, Japan and South Korea |
| 18 August 2026 | Expansion announced across 31 European countries |
| 31 August 2026 | OpenAI states availability in over 40 countries |
The May change is the one that reset the entry price. The pilot ran on CPM and, according to Digiday's reporting at the time, a 50,000 dollar minimum spend. Self serve replaced that with CPC and oCPC bidding and published daily minimum budgets: 25 dollars in the United States, 15 pounds in the United Kingdom, 25 Canadian dollars, 2,500 yen in Japan. Those are daily minimums rather than realistic campaign budgets, but the barrier went from enterprise to expense claim.
Ads may appear for users on the Free and Go plans. Plus, Pro, Business, Enterprise and Edu accounts have no ads, and OpenAI also excludes accounts identified as belonging to users under 18.
Almost every marketing team evaluating this channel is on Plus, Business or Enterprise. That means the people making the decision structurally cannot see the ads in their own account, which is one practical reason tracking what ChatGPT actually answers has to happen outside your own login, and any audit that starts with "let me check in ChatGPT" produces a clean screen and a false conclusion.
There is a second wrinkle specific to Europe. Personalised ads were initially unavailable in the European Economic Area and Switzerland. With personalisation off, ads can still be selected using the current chat thread, but not other threads, ad history or topic profiles. In practice European delivery leans even harder on the immediate conversation, which sharpens the point of the next section.
Ads are selected through a relevance weighted second price auction. Bids are one factor, weighted against relevance and other delivery signals rather than deciding the outcome alone. So far, familiar.
What breaks the familiar workflow is what advertisers can point that bid at. Matching is contextual and intent based, using the current conversation and its intent, the ad's landing page, title and copy, and advertiser supplied context hints. Those hints may describe relevant conversations, topics, customer needs, use cases and, yes, keywords. OpenAI is explicit that they are not exact match keywords, not audience targeting rules, and no guarantee of delivery for any particular word, topic or conversation.
Targeting beyond context exists but is coarse. Geographic targeting arrived during 2026, along with custom audiences built from email addresses, phone numbers, hashed identifiers or GAIDs. Those carry a floor: at least 25,000 matched users for inclusion targeting or bid adjustments, with 100,000 recommended. Below 25,000 an audience can only be used for exclusions.
The practical consequence is that no keyword research tool can tell you whether your category carries ads, who is buying them, or what they say. A conversation is not a keyword and never resolves to one. The only way to find out is to run the questions your customers actually ask and read what comes back, which is what AI ads tracking records: the advertiser, the prompt that triggered the placement, the landing URL behind it, and the date it first appeared.
This is where published figures diverge sharply, and the gap is worth showing rather than smoothing.
Seer Interactive's updated analysis of 28 July 2026 covered roughly two weeks of tracking and reported ads in 77% of tracked responses, visibility across 96 of 97 client AI tracking workspaces, and 3,697 unique advertisers.
Our own tracking, over 60 days and 30,467 distinct prompts, found a sponsored placement on 1,382 prompts, or 4.5%.
Both numbers are real and neither is the population rate. Seer's sample is client workspaces, which are built almost entirely from commercially loaded prompts. Ours spans a broader mix that includes a large share of informational questions where no advertiser has any reason to bid. The honest reading is that ad density tracks commercial intent almost perfectly, and that a category level number tells you nothing until you know which prompts went into it.
Which is the actual lesson: measure your prompts, not the market.
Between 25 June and 26 August 2026 we recorded 20,556 sponsored placements inside ChatGPT answers across 1,382 prompts, from 2,736 distinct advertisers. Over a comparable period Google AI Overviews produced 1,397 placements across 615 prompts from 702 domains.
| Measure | ChatGPT | AI Overviews |
|---|---|---|
| Placements recorded | 20,556 | 1,397 |
| Prompts with at least one placement | 1,382 | 615 |
| Distinct advertisers | 2,736 | 702 |
| Placements per answer, average | 1.42 | n. a. |
| Placements per answer, maximum | 4 | n. a. |
An answer that carries advertising usually carries one placement, occasionally two, at most four. This is not a results page with ten paid slots above the fold. It is a single block below the answer, clearly labelled sponsored, and OpenAI states that it does not influence the answer above it.
One block also means the competition for it is sharper than the format suggests.
| Figure | Value |
|---|---|
| Distinct advertisers per affected prompt, average | 9.4 |
| Distinct advertisers on the most contested prompt | 31 |
| Share of all placements held by the top 10 advertisers | 16.9% |
| Share held by the top 50 | 37.4% |
Nine advertisers circling one prompt is a competitive channel. But the concentration figures say the field is still open: the ten largest advertisers account for 16.9% of everything we recorded, and the top 50 for 37.4%. Compare that with paid search, where a handful of brands and aggregators own most commercial categories outright. Nobody has claimed this yet.
An AI answer names sources in its body, then sometimes shows a sponsored block underneath. Two contests, two mechanisms, decided independently.
Seer's data puts numbers on how rarely one brand holds both. In its sample the advertiser appeared in the organic ChatGPT response 5.4% of the time and was cited as a source 3.3% of the time. On branded prompts, only 7.3% of responses featured the brand's own ad, up from 2.2% in its earlier report.
Read that from the defensive side. You can be the source an answer is built on and still lose the click to whoever bought the space below it. Earning the citation does not protect the click, and buying the placement does not earn the citation. And because there is no keyword, there is no auction insight report that will warn you. The signal exists only inside the answers themselves.
OpenAI supplies a browser Measurement Pixel and a server side Conversions API covering page views, leads, registrations, subscriptions, trials, checkouts, orders and custom events. Attributed conversions can take 24 to 48 hours to appear.
The attribution windows deserve precision, because they are widely misquoted. The view through window is fixed at one day after an eligible impression and cannot be configured. Click through attribution uses a configurable click window, and OpenAI's own documentation does not publish a default length. The widely repeated "seven day click" figure traces to a third party guide from March 2026, not to OpenAI. Where both a qualifying click and impression exist, the click takes credit, and view through conversions are reported separately as VTA and do not feed bidding, billing or optimisation.
Even read generously, this counts clicks. It cannot count the person who reads an answer, decides inside it, and arrives four days later by typing your brand into Google. In our attribution sample of 44,110 self reported responses, AI search accounted for 14.4% of reported discovery and 16.3% of reported deal value. In one comparable subset of 472 responses, 35 people named an AI assistant as where they decided, while analytics recorded AI as the source for 2.
That is not an argument against the pixel. It is an argument for reading its numbers as a floor, and for adding a self reported source question at the point of purchase when the decision is worth money. AI search attribution closes that loop by matching post purchase answers to orders, which is how a sale that analytics filed under direct gets traced back to the answer that produced it.
Our figures come from Finseo tracking data, aggregated across accounts, with no customer, project or private domain identifiable in any number. Sponsored placements were recorded between 25 June and 26 August 2026 for ChatGPT and between 24 June and 31 August 2026 for AI Overviews. The 4.5% coverage figure compares distinct prompts carrying a placement against 30,467 distinct prompts tracked in the preceding 60 days.
Prompt sets are chosen by the customers who run them, so the corpus skews toward the categories those customers sell in. It is a large sample, not a random one. That is also why our 4.5% and Seer's 77% are both defensible: they describe different prompt mixes over different windows, which is stated above rather than reconciled away.
Every external figure is attributed with its source and date. Where a widely quoted number comes from a third party rather than OpenAI, notably the seven day click window and the 50,000 dollar former minimum, that is flagged in the text.
Why does ChatGPT have ads now? OpenAI announced the test in January 2026 and began the US pilot in February. Ads fund the free tiers, which is also why Plus, Pro, Business, Enterprise and Edu accounts remain ad free.
Where do ChatGPT ads show up? Below the end of the response, in a separate block labelled sponsored. The ad unit can hold one or more products, and OpenAI states that ads do not influence the answer above them.
Who sees ads in ChatGPT? Logged in adults on the Free and Go plans. Accounts identified as belonging to users under 18 are excluded, as are all paid tiers above Go.
How does ChatGPT ad targeting work? Through a relevance weighted second price auction, matched contextually to the conversation using the ad's landing page, title, copy and advertiser supplied context hints. Those hints are explicitly not exact match keywords and guarantee delivery for no particular term.
How much do ChatGPT ads cost? There is no published rate card. Daily minimum budgets start at 25 dollars in the United States and 15 pounds in the United Kingdom. Q1Media reported an average CPC of roughly 5 dollars with click through rates around 0.5% to 2.5% across the accounts it observed, which is an agency observation rather than an audited benchmark.
Can I stop a competitor advertising on my prompts? No. You cannot prevent anyone buying a placement under an answer that cites you. What you can do is know when it happens and decide whether to answer it with budget or with a better page.
Ours
External
Disclosure: the tracking figures in this article come from Finseo, an AI visibility platform. External figures are attributed to their original source.
