ChatGPT Ads Just Launched: The Format, the Tiers, and What It Means

It shipped. On February 9, OpenAI began testing ads inside ChatGPT — US only, logged-in adults only — and published a short post titled “Testing ads in ChatGPT” to confirm it. After years of “will they, won’t they,” the largest consumer AI product now carries advertising. That part was inevitable. A product with hundreds of millions of weekly users and a model bill that doesn’t sleep was always going to find a revenue line beyond subscriptions.

The interesting part isn’t that ads arrived. It’s the shape they took.

The format choice is the story

For months the speculation pointed one way. Back in December, The Information reported on internal explorations: ads woven directly into answers, post-click sponsored links, “generative ads” assembled on the fly, personalization mined from ChatGPT’s memory of you. That’s the version everyone braced for — the assistant quietly steering you toward whoever paid, with your own conversation history used as targeting fuel. The dystopian read wrote itself.

What actually launched is none of that.

The ad is a sponsored card in a clearly labeled section below the answer. Not inside the response. Not a phrase nudged toward a brand. A separate, visually distinct block that sits underneath ChatGPT’s reply, marked “sponsored,” carrying an advertiser name, favicon, headline, a short description, an image, and a link. The model writes its answer; then, in a separate lane, an ad may appear. And the line OpenAI drew hardest is this: ads cannot influence the assistant’s answer. The content of the response is decided before any advertiser enters the picture.

That is the conservative format. After all the leaks suggesting the opposite, OpenAI shipped the trust-preserving version. The restraint is the news.

What actually shipped

Strip away the framing and here’s the mechanism, point by point.

• Where ads run. Free and Go tiers only. Plus, Pro, Business, Enterprise, and Education stay ad-free. If you pay past the entry plan, you don’t see ads. That keeps the most engaged and highest-intent users — and the ones companies put on the product — out of the ad pool entirely.
• The unit. One labeled sponsored card below the answer: advertiser name, favicon, headline, short description, image, link. Always labeled, always visually separated, never merged into the text the model produces.
• The wall between ad and answer. Advertisers cannot shape what ChatGPT says. The recommendation and the ad are separate systems by design, and OpenAI is making that separation the headline feature rather than a footnote.
• What advertisers get back. Aggregated metrics only. No chat logs, no personal data handed to the buyer. The advertiser learns that the campaign performed; it does not learn who you are or what you typed.
• What users get. You can dismiss an ad, view and delete your ad history and interest data, and toggle personalization off. No ads to anyone predicted to be under 18. No ads in sensitive chats — mental health, politics, and the like are walled off.

On pricing, the launch model is CPM — cost per view, charged on impressions. Early test pricing was reportedly steep: on the order of a ~$60 CPM, with reported minimum buys around $200K for the first holding-company testers. Treat those numbers as reported, not gospel, but the direction is clear — this is being sold as premium, scarce inventory to large buyers first, not a self-serve free-for-all.

And the buyers are exactly who you’d expect for a managed launch. Adobe is in early, putting Acrobat Studio and Firefly through the format. The four big holdcos — WPP, Omnicom, Publicis, and Dentsu — are all named partners, with Omnicom reportedly placing 30-plus clients. This is a controlled rollout through the agencies that already move most of the world’s ad money, not an open beta.

Why the conservative version is the right one

Anyone who watched display advertising grow up knows how this normally goes. A new surface appears, the early format is restrained, then the pressure to monetize grinds the restraint down. Labels shrink. Ads creep closer to content. “Sponsored” becomes a five-point gray word nobody reads. The surface gets more profitable and less trustworthy at the same rate, until users route around it.

ChatGPT’s answer is a fundamentally different kind of surface, and OpenAI clearly understands the risk. The whole value of an answer engine is that you trust the answer. The moment a user suspects the model is recommending a product because someone paid — not because it’s the right answer — the trust that makes the product worth using starts to drain. You can’t claw that back with a disclosure line. So the only durable design is the one that makes manipulation structurally impossible: decide the answer first, show the ad separately, label it, and never let the advertiser touch the response. That’s not generosity. It’s the only version that survives contact with users.

It’s worth holding this next to the other AI-search ad model in market. We’ve written before about what Perplexity’s ads mean — sponsored follow-up questions and related-media units sitting alongside answers. Different surface, same instinct: keep the paid unit adjacent to the answer, not inside it. The two largest players in answer-engine advertising have independently landed on roughly the same rule. When competitors converge on restraint, that’s the category telling you what the format actually is.

Where Elo fits

We build Elo, an adserver for AI chat. Our entire thesis has been that the only conversational ad worth building is one that’s labeled, context-relevant, and incapable of bending the assistant’s answer — a native message in the thread that reads the live conversation, matches it to a real advertiser offer, runs brand-safety filtering, and gets delivered server-side. So watching the biggest product in the category ship the labeled, answer-safe, below-the-answer version is not a threat to that thesis. It’s confirmation of it.

This is also the difference between intent and audience. The leaked version leaned on who you are — your memory, your history, your profile — to decide what to sell you. The shipped version leans on the moment: a sponsored card relevant to the conversation in front of you, with the targeting data kept out of the advertiser’s hands. Matching to live context instead of mining a profile is the more honest mechanic, and it tends to perform better anyway, because intent in the moment beats a guess about identity.

A couple of things the launch deliberately leaves on the table. CPM is a fine starting point — it’s how real-time bidding priced display inventory for years — but it pays for attention, not outcomes, and managed CPM buys at a $200K floor are not how the long tail of advertisers eventually shows up. And a card below the answer is closer to a native ad than a banner, but it still sits beside the conversation rather than living inside the thread as a message. Those are evolution points, not flaws. You start conservative and earn your way toward more.

For now, the takeaway is simple. The format that shipped is the right one, and it sets the bar for everyone building in this space. Labeled. Below the answer. Untouchable by the advertiser. After months of leaks pointing at the worst version, OpenAI shipped the disciplined one — and in doing so, defined what a conversational ad is allowed to be.

The era of ads in AI chat didn’t start with a land grab. It started with a fence around the answer. Good.