The story that ran everywhere last week was that OpenAI stopped accepting advertising from rival AI image and audio tools, and that Adobe found out the hard way. True, and incomplete. The part almost nobody connected: in the same window, ChatGPT also stopped surfacing outbound links and citations for action oriented image queries, the searches where a reader is trying to find a photo editor or an image generator. Paid placement and organic referral closed in the same categories at the same time, immediately after OpenAI shipped its own products into those categories. That is not an advertising policy. That is a distribution channel being sealed from both ends by the company that competes inside it.
What Was Actually Restricted
The change was communicated privately to advertising partners rather than announced, which is itself worth noting for a company running a platform that now reaches more than a billion monthly users. Campaigns promoting standalone image generation and voice generation products will no longer be approved to run in ChatGPT. Adobe was among those caught out, having joined the original pilot to promote Acrobat Studio and its Firefly image generator.
The boundaries of the restriction are where this gets specific:
- Image generation ads: blocked.
- Voice and audio generation ads: blocked.
- Video generation ads: still permitted.
- Outbound citations on image tool queries: also pulled.
Video is the tell. OpenAI does not yet have a shipping consumer video product sitting in that exact slot, and video advertising is still welcome. The categories it did close are the two it had just entered, having launched ChatGPT Images 2.5 and its ChatGPT Live voice product. PYMNTS framed the result plainly, noting that OpenAI is now simultaneously the platform advertisers depend on and the rival they are forced to compete against. The line between those two roles was drawn by the party that benefits from where it sits.
The Money Explains the Timing
OpenAI’s advertising business is not a side experiment any more. Ads launched inside ChatGPT in February 2026 and reached a $1 billion annualized run rate in under 200 days, against a 2026 target in the $2.5 billion range and total company revenue tracking past $40 billion annualized. The pilot went from roughly $100 million annualized in under two months to 600 plus advertisers, then expanded into the UK, 31 European markets and self serve access across India and the Middle East.
We wrote in April that the interesting thing about OpenAI selling cost per click ads inside ChatGPT was not the revenue but the structural position it created. This is that position being used. A business growing that fast has every commercial incentive to sell inventory to everyone, and OpenAI just turned away a category of well funded buyers anyway. Companies do not leave advertising revenue on the table for tidiness. They do it when the strategic value of the exclusion exceeds the cheque, which means OpenAI has priced the competitive threat in image and voice generation higher than the ad dollars Adobe and its peers were willing to spend.
This Is Self Preferencing, and Europe Already Built the Hook
Here is where we come down. Blocking competitors from buying ads on your own property is aggressive but defensible on its own. Plenty of platforms refuse categories. Pulling organic outbound citations for the same product category in the same period is a different act entirely, because those links are not inventory OpenAI is choosing not to sell. They are the answer to a user’s question. A person asking ChatGPT for a good image editor is receiving a worse, narrower answer so that a first party product faces less competition, and they have no way of knowing the category was quietly narrowed.
That is textbook self preferencing, and it is the conduct that produced a decade of antitrust litigation against Google Search and Amazon’s marketplace. The usual defence, that an AI assistant is a product rather than a gateway, has already been weakened. We covered the designation when it landed: Europe regulated ChatGPT as a search engine rather than as an AI model. Once a regulator treats the interface as a search surface, degrading organic results in a category where you sell a competing product stops being a product decision and becomes regulated conduct with an existing body of law behind it.
Our view is that this warrants a formal look now rather than after the ad business is worth ten times what it is today, and that the burden should sit with OpenAI to publish what it changed. The company is entitled to run its ad platform how it likes. It is not entitled to quietly thin the organic answer set in the categories where it sells, tell nobody, and have that treated as a routine policy update. Adobe learned about a material change to its distribution from a private note. Users learned nothing at all.
What Comes Next
Watch three things. Whether video advertising stays open once OpenAI ships a consumer video product, which would confirm the rule is written around the product line rather than around any principle. Whether European regulators treat the citation change as a search behaviour under the designation already in force. And whether Adobe and the other affected buyers say anything publicly, which they have commercial reasons not to do while they still need the channel.
The broader point for anyone building on top of a model provider: distribution you do not own can be withdrawn in a private email, and you will find out when your campaigns stop clearing review.