ChatGPT Ads: How They Work and Whether to Buy In

Advertising inside ChatGPT stopped being a thought experiment this year. OpenAI began testing ads with US users in February 2026, extended them to the UK, Mexico, Brazil, Japan and South Korea in May, and on August 19 announced the jump to 31 European countries - its largest expansion so far, with self-serve access following. For scale: ChatGPT reported 900 million weekly active users in February 2026, of which 50 million were paid subscribers. Ads show only to the free and Go tiers, which is to say almost everyone.

I think this is the most interesting new ad surface since TikTok, and I also think most teams should not be first through the door. Both halves of that sentence need the details, so here they are.

What the ads actually look like

The format is deliberately conservative. According to OpenAI’s own documentation, an ad appears as a clearly labelled unit below a ChatGPT response, when the conversation is a relevant match. Paid tiers - Plus, Pro, Business, Enterprise, Edu - see no ads at all. Accounts identified as under 18 see none. Political advertising is banned outright, and ads are kept away from sensitive conversations such as personal health.

Two structural promises matter more than the format. Advertisers do not receive users’ conversations, and advertising does not influence the answers ChatGPT generates. The ad lives next to the answer, and the answer is computed without it. Whether that wall holds over years of revenue pressure is a fair question, but as launched, you cannot buy your way into the response itself - and anyone who tells you otherwise is describing a product that does not exist.

That distinction cuts both ways for marketers, and I keep it front of mind when the channel comes up in planning. Buying an ad under an answer is now possible. Being the brand the answer recommends still is not for sale, which keeps earning citations a separate, unpaid game.

Targeting, buying and measurement

The mechanics, per Search Engine Land’s coverage of the European launch:

• Matching is conversational, not keyword-based. Ads surface when what someone is discussing matches an advertiser’s targeting. There is no query report to mine, which will feel like flying blind to anyone raised on search campaigns.
• Bidding runs on CPM and CPC, with conversion optimisation added as the product matured through 2026.
• Targeting options include geo-targeting and custom audiences.
• Measurement is OpenAI’s pixel, a Conversions API, and third-party integrations.
• Buying started through OpenAI’s Ads Solutions team and agency and technology partners, with the self-serve Ads Manager rolling out to advertisers through the summer - OpenAI says tens of thousands of marketers have already run ads.

My read-through: this is a mid-funnel awareness surface wearing performance-channel clothes. The conversion tooling exists, but the intent signal is a conversation, not a search, and a person mid-conversation with an assistant is not a person scanning a results page for somewhere to click.

Whether you should buy in

My honest framework, since I can’t run your budget for you:

Test now if you sell to consumers in a category people research conversationally - travel, software, finance, health-adjacent products that clear the sensitivity rules - and you have the creative discipline to make one card under an answer feel like a useful suggestion rather than an interruption. Early surfaces historically underprice attention before the auction thickens, and this one comes with a billion-scale audience attached.

Wait if your measurement culture cannot tolerate ambiguity. Attribution here is young, the channel’s incrementality is unproven, and the conversation-matching means you cannot fully predict adjacency. If your CFO wants every channel defended to the last click, this one will lose that argument for a while yet - the same problem zero-click attribution has, wearing a paid badge. Run the numbers with the same discipline you would use to calculate ROI on any AI spend: count the setup and learning time, credit only what you can baseline.

Either way, write for the surface. A response-adjacent card competes with an answer the user just got for free, which makes flat feature copy even weaker than usual. The craft overlaps heavily with what already works in AI-drafted ad copy: one concrete benefit, platform limits respected, compliance fenced in from the start.

What I’m watching from here

Three things will decide whether this becomes a line item or a footnote. Whether the answer-integrity wall survives contact with revenue targets. Whether OpenAI ships anything resembling a query-level report, because contextual matching without visibility is a trust ask that platforms historically abuse. And whether users on free tiers simply acclimatise, the way they did on every prior platform, or whether ads push more of them into paid tiers and shrink the inventory.

Until those resolve, my position is the boring one: small test budgets for good-fit categories, no migration of core spend, and a hard look at the post-click data before scaling anything. New channels reward early learning, and they punish early confidence.

More paid media guides live in the paid media topic - the same spend discipline applied to the automation you already run is in my Smart Bidding guide and the Performance Max guardrails - and for how a paid test fits the broader stack, start with the complete guide to AI marketing.