If you still think of ads in ChatGPT as a rumour from a leaked Android build, your mental model is about nine months out of date. The strings that Tibor Blaho found in the beta back in late 2025 shipped. OpenAI started testing ads with logged-in adult users on the Free and Go tiers in the US in February, opened a self-serve ads manager in May, and has spent this month pushing the product into new markets.
Here is the current state of it, and the one number in the data that matters more than any of the rollout news.
What actually exists now
The format is small. One sponsored card below the conversation: a title of 3 to 50 characters, up to 100 characters of body copy, an image, a favicon and a target URL. No video, no carousel, no interactive units. It is closer to a link than to a banner.
Who sees it. Free and Go users in supported markets. Plus, Pro, Business, Enterprise and Education tiers get nothing, and neither do users OpenAI predicts to be under 18.
Where you can buy it. Self-serve is live in the US, UK, Canada, Australia, New Zealand, Japan, Korea, Brazil and Mexico. Thirty-one European markets were announced this month, opening through agency and technology partners first. The $50,000 minimum spend that gated the pilot is gone.
What the platform can do. Product feed campaigns for retail catalogues, a conversions objective billed per click, custom audiences from email or phone lists that need at least 25,000 matched users to run, geo and platform targeting, and a first-party pixel with hashed advanced matching on by default. It is a normal ad platform. That is the point of it.
What it costs. The default max bid is $60 CPM with $3 to $5 suggested for CPC, and pilots have reported effective CPMs well below the default as inventory scaled. OpenAI is explicit that it has no cross-advertiser benchmarks yet, so treat every number floating around as one account's experience, not a market rate.
The number that matters
SE Ranking looked at around 97,000 ad impressions and found a click-through rate of 1.30%, with smaller tests landing under that. Interesting, but not the important figure.
This one is: in those placements, the advertiser appeared as a cited source inside the answer 3.63% of the time. In Google's AI Mode, the comparable figure was 11.53%.
Buying the ad slot does not buy you into the answer. That is not an accident or a beta limitation, it is the design. OpenAI's stated principle is answer independence: ads run on separate systems from the model, advertisers cannot shape or rank responses, and by default ChatGPT cannot even see the ad that was shown next to it.
Rough arithmetic
What the slot buys, and what it does not
Move your monthly budget. The numbers run on published third-party benchmarks, not OpenAI figures, so read them as an order of magnitude rather than a forecast.
The card costs $5,000 a month and stops the day you stop paying. The sentence above it costs nothing and takes a year.
So the page splits cleanly in two. The card at the bottom is for sale. The paragraph above it, the one that names three tools and explains which fits your case, is not.
What this changes for AI visibility
Paid and organic are now separate budgets with separate mechanics. On Google you could at least tell yourself that the same keyword research fed both sides. Here they barely touch. A campaign gets you a card. Getting named in the body still comes from what the model has read about you.
Nothing about the citation game changed, which is the news. Documentation that answers the question directly, reviews with specifics in them, forum threads where real people compare you to alternatives, comparison pages that admit what you are bad at. Models synthesise consensus, and consensus is still assembled from the open web over months.
Attribution gets murkier while spend gets easier. A pixel and a conversions objective make the paid side measurable and therefore defensible in a meeting. The organic mention that actually decided the purchase shows up as direct traffic, if it shows up at all. The channel that is easiest to measure will eat the budget of the channel that works, unless someone in the room says so out loud.
Cheap inventory is a real window, and a short one. Early ad products are underpriced because demand has not arrived yet. If your niche is in a live market, testing now costs less than testing next year. Just do not confuse renting the slot with earning the recommendation.
The honest summary
Ads in ChatGPT are not the end of AI visibility. They are proof that the surface is worth money, which was the argument for working on it in the first place.
The card under the answer has a price. The sentence inside the answer does not, and there is no team at OpenAI you can call to change it. That sentence gets written by everything about you that already exists in public, which is slow, unglamorous, and the only part of this that a competitor cannot outbid you for.