ChatGPT began showing ads on 9 February 2026, as a US-only test. The self-serve Ads Manager opened on 5 May. It now covers more than fifty countries, including the UK.<\/p>\n\n\n\n
Most writing about the channel is either a launch recap or a pitch. This is the mechanical version: where the ad sits, who sees it, how it gets picked, what it costs, and what you can see afterwards. Everything below comes from OpenAI’s own documentation, which is worth reading directly, because it has changed several times already.<\/p>\n\n\n\n
Ads appear for users on the Free and Go plans. Plus, Pro, Business, Enterprise and Edu accounts don’t get them at all. Accounts identified as belonging to under-18s are excluded, using stated age plus behavioural prediction. Temporary Chats show no ads. During the test, ads don’t appear in the Atlas browser either.<\/p>\n\n\n\n
There’s also an Ads-Free option on the Free plan: users can drop ads in exchange for lower message limits and reduced feature access.<\/p>\n\n\n\n
Add those up and your addressable audience inside ChatGPT isn’t “ChatGPT users”. It’s the ad-eligible subset, in specific markets, on specific surfaces. That distinction matters later, because when you go looking for your own ads, you have to be looking from an account that could have seen one.<\/p>\n\n\n\n
Below the end of a response. Labelled sponsored, visually separated from the answer itself. The unit carries an advertiser name, favicon, headline, description, an image asset and the landing page.<\/p>\n\n\n\n
There is no results page. No position one through four, no above-the-fold argument, no ten organic listings underneath. One answer, and a sponsored unit below it. A user may see one or more units under a response, and a single unit can carry several items, or more than one advertiser.<\/p>\n\n\n\n
Not by keyword match. The selection system weighs the context and intent of the current conversation, the ad’s landing page, its title and copy, advertiser-supplied context hints, and, where the user has ad personalisation switched on, signals from their broader ChatGPT history.<\/p>\n\n\n\n
Context hints live at ad group level. You describe, in plain language, the conversations where your product is relevant. They guide matching. They are explicitly not exact-match keywords and carry no delivery guarantee.<\/p>\n\n\n\n
Eligible ads then go through a relevance-weighted, second-price auction.<\/p>\n\n\n\n
The practical consequence is that your landing page is targeting infrastructure, not just a destination. Two advertisers bidding the same amount with similar copy can see different delivery because one page describes the product in the language customers actually use when they ask about it.<\/p>\n\n\n\n
Two objectives. Reach optimises for impressions and bills per thousand. Clicks optimises for clicks and bills per valid click. You set a maximum bid at ad group level: CPM for Reach, CPC for Clicks. OpenAI suggests starting somewhere between $3 and $5 for CPC campaigns, and Ads Manager gives bid-strength guidance on whether your number is likely to limit delivery.<\/p>\n\n\n\n
No minimum spend, which is why the channel is testable on a small budget.<\/p>\n\n\n\n
Ads Manager Beta reports impressions, clicks, spend, click-through rate, average CPC, average CPM and conversions. You can configure conversion measurement, and static tracking parameters such as UTMs persist through the click into your own analytics.<\/p>\n\n\n\n
That’s a complete account of your money. It’s a complete account of nothing else.<\/p>\n\n\n\n
On Google, an impression is a known thing: your ad, on a results page, in a known position, against a known query, next to nine other listings you can go and look at.<\/p>\n\n\n\n
In ChatGPT an impression is one person, one conversation, one answer. The number is honest \u2014 it just doesn’t carry the context that made impressions useful on older channels. It doesn’t tell you which question produced it, what the answer said, or whether a competitor appeared in the same footer.<\/p>\n\n\n\n
Your totals answer “did it run”. They don’t answer “on what”.<\/p>\n\n\n\n
Ads aren’t eligible near sensitive or regulated topics, including personal health, mental health and politics. Some advertisers in regulated verticals can run at all if they clear stricter eligibility, but placement near sensitive user content stays excluded regardless.<\/p>\n\n\n\n
This matters more for measurement than most people expect. If you check a question and no ad appears, that’s not automatically a lost auction. It might be a question the surface doesn’t carry ads on at all. Those two outcomes look identical in a report and mean completely different things. One is a competitive problem, the other is a category boundary. You can’t separate them from inside Ads Manager. You can only separate them by observing the conversation repeatedly and seeing whether anyone at all ever shows up.<\/p>\n\n\n\n
Self-serve is live in over fifty countries, the UK and US among them. The catch is in the fine print: the legal entity being billed has to be based in a listed country. Advertisers in some markets, including the newer European ones, currently go through OpenAI’s ads team or an agency or technology partner instead.<\/p>\n\n\n\n
Worth rechecking the country list before you plan anything. It has moved repeatedly since May.<\/p>\n\n\n\n
So: you can buy, set bids, pace budget, and read your totals back.<\/p>\n\n\n\n
You cannot see which of your customers’ questions your budget reached. You cannot see which questions carry ads at all. You cannot see who appeared in the conversations where you didn’t. There’s no public ad library for ChatGPT the way there is for Meta or Google. Ads Manager reports your account, and only your account.<\/p>\n\n\n\n
The only route to those answers is from the user side: ask the same questions repeatedly, on ad-eligible accounts, in each market you care about, and record every outcome, including the checks that come back empty. Those are findings, not gaps.<\/p>\n\n\n\n
That’s what Metistic does. On one headphones question in one market, an ad was present in 62% of checks (n=92). The rate is the finding; the n is what makes it usable.<\/p>\n\n\n\n