ChatGPT ads, read by someone who buys Meta ads for a living

Every media buyer I know has had the same message this month, usually from a client: should we be on ChatGPT? It is a fair question, and it deserves a better answer than yes.

I buy Meta ads for a living. What follows is ChatGPT ads read from that chair: what the product actually is, according to OpenAI rather than the fifty guides written about it, what carries over from Meta, and the part nobody seems to be writing about, which is whether the thing Fadar exists to catch even happens there.

What ChatGPT ads actually are

Strip away the launch coverage and the product is narrow. A sponsored unit, labelled as sponsored, sits visually separate from ChatGPT's answer. OpenAI's position, repeated in every announcement it has made on the subject, is that ads do not influence the answer itself.

Who sees them is equally specific. Logged-in adults on the Free and Go plans. Paying subscribers on Plus, Pro, Business, Enterprise and Education see none, and free users can opt out of ads in exchange for fewer daily messages. Ads are not shown to users OpenAI knows or predicts to be under 18, and they are not eligible to appear near sensitive or regulated topics like health, mental health or politics.

Seven months from closed test to self-serve

The timeline matters because it explains why so much of what you read about this is already wrong.

• 9 February 2026: testing begins in the US, on the Free and Go plans only.
• 26 March: pilots announced for Canada, Australia and New Zealand.
• 5 May: a beta self-serve Ads Manager opens, and cost-per-click bidding arrives alongside CPM.
• 11 August: live in the UK, Mexico, Brazil, Japan and South Korea.
• Late August: 31 European markets, with self-serve access through Ads Manager from 31 August.
• 16 September: Sponsored Agents enter a US test, AI creative tools land in Ads Manager, and HubSpot and Shopify integrations launch.

OpenAI names Germany, France, Spain, Italy, Sweden, Norway, Denmark, the Netherlands and Austria among the 31 European markets without printing the full list in its announcement, so check Ads Manager for your own country before building a plan around it. And anything you read that is more than a few weeks old is probably describing a different product. That will eventually include this post.

How you actually buy it

The mechanics will look familiar, which is part of the trap.

• Ads Manager at ads.openai.com, or through agency partners and technology partners such as Criteo, Pacvue and StackAdapt.
• You set budgets, bids and pacing. OpenAI's system makes every delivery decision.
• CPM for reach, CPC for clicks, and since the summer, conversion optimisation.
• Geo-targeting and custom audiences.
• Measurement through the OpenAI Pixel, a Conversions API, and third-party integrations.

Pricing is where I would slow down. OpenAI does not publish a rate card, and the third-party estimates in circulation disagree with each other by a wide margin. That is not a criticism, it is what a young auction looks like. Treat any benchmark CPM you read as a rumour until your own account produces one.

What carries over from Meta

More than you would expect on the plumbing. Less than you would hope on the thinking.

• Measurement. A pixel plus a server-side Conversions API is exactly the setup you already run on Meta. If your Meta tracking is clean, you can be measuring by the end of the afternoon.
• Conversion optimisation. Handing the system a goal and letting it find the people is the same bargain Advantage+ asks you to make.
• Testing discipline. Several distinct ideas beat one idea in several colours, on every platform that has ever existed.

What does not carry over is almost everything about attention

On Meta you are interrupting someone who came to scroll. The whole craft is the first second: the hook, the thumb-stop, the frame that earns the next frame. In ChatGPT nobody is scrolling. The person has typed a question, often a long one, explaining what they want and what constrains them. Your ad appears below an answer they asked for. You are not interrupting anything. You are the footnote to a conversation that is already about your category.

Targeting works differently for the same reason. OpenAI matches ads to the topic of the conversation and, where ad personalisation applies, to past chats and past interactions with ads. There is no interest graph to build an audience from, and OpenAI says it keeps guardrails against narrow ad targeting. The lever you pull hardest on Meta, who sees the ad, is largely replaced by a different one: what they were asking about when they saw it.

And the copy may not stay yours. Since September, advertisers can opt into AI-powered text customisation, which adapts existing headlines and descriptions to the context of each conversation and translates them into the user's preferred language. It is optional. It also means the thing you wrote and the thing a given person saw can differ, which matters a great deal for the next section.

On Meta you are the interruption. On ChatGPT you are the footnote. Footnotes get read by people who already care.

Does creative fatigue even happen there

This is the question I actually wanted to answer, since spotting fatigue is the whole of what Fadar does. The honest answer is that nobody knows yet, including me, and anyone who says otherwise has at most seven months of data from a platform that changed its buying model twice in that time.

What I can do is reason from the mechanism. On Meta, fatigue comes from repetition: a fixed creative shown to a finite audience until that audience stops responding. Two things about ChatGPT ads push against it.

• Delivery follows the conversation, not the person. Someone sees your ad when they ask about your category, which is occasional by nature, rather than every time they open the app.
• The creative can change per impression. With text customisation switched on, the headline adapts to each conversation, which is a partial answer to repetition built into the format itself.

Two things push the other way. Categories with a small pool of people asking about them will see the same people come back, and a narrow audience is exactly where fatigue arrives first on Meta. And text customisation adapts headlines and descriptions, not the visual, so the image half of the ad can still wear out on its own schedule.

My working guess, labelled as a guess: fatigue in ChatGPT ads will be slower and patchier than on Meta, concentrated in narrow categories and in ads running with customisation switched off. The first thing I would watch is click-through rate against the ad's own trailing baseline, exactly as I would on Meta. The signal travels between platforms. What changes is how long it takes to move.

Who should wait

The section most guides leave out.

• If nobody asks ChatGPT about what you sell, there is no conversation for your ad to appear in. Impulse and entertainment products that live on the feed belong on the feed.
• Health, mental health and political advertisers are kept away from those topics regardless.
• If your Meta tracking is a mess, fix that first. A new channel does not repair broken measurement. It copies it.
• If you need proof before spending, wait a quarter. The platform is changing monthly, and early benchmarks will be obsolete by the time you trust them.

The ones I would test now are anyone selling something that gets researched and compared before it gets bought. Software, travel, furniture, B2B services, anything where the buyer already opens a chat window to make up their mind.

Where Fadar fits, which is nowhere yet

Worth saying plainly. Fadar monitors Meta ads. It does not read ChatGPT ad accounts, and I am not going to pretend otherwise in a post about which claims are real. Meta only is a choice: doing one platform properly beats doing four badly.

If ChatGPT ads turn out to fatigue the way Meta ads do, it will show up in click-through rate against baseline first, and that is worth building for. If they turn out not to, it will be one of the more interesting findings in performance marketing this decade. Either way, I would rather wait for the data than write the feature description in advance.