OpenAI’s advertising system inside ChatGPT is not a simple keyword auction. A report published on 15 October, built on two test ad campaigns bought inside the platform, shows that sponsored placements can be matched to accounts using uploaded email addresses, phone numbers or Android advertising IDs — and that outside Europe the system can also lean on previous conversations and stored memory. Every detail matches OpenAI’s own help-centre documentation, which the company has published as the ads business scales.
- Context first: the default signal is the subject and intent of the prompt you are typing right now, plus the advertiser’s landing page, copy and targeting hints.
- Optional personalisation: switching on “Personalize ads” adds the current thread, your past interactions with ads, and your chat history and memory.
- Custom audiences: advertisers can upload customer lists and need at least 25,000 matched users to run an inclusion campaign; smaller lists can only be used for exclusions.
- Europe differs: in the EEA and Switzerland, OpenAI says personalisation and custom audiences are not available, so only conversation context applies.
What actually decides which ad you see
Per OpenAI’s help pages, the base layer of ad selection combines three things: the topic and intent of the ongoing chat, the material supplied by the advertiser, and the user’s general location and language. That advertiser material includes the destination page, the ad text and a set of targeting “hints” describing the sort of conversation in which the brand wants to appear. Critically, OpenAI states that these hints guide matching but are not exact-match rules — advertisers describe a scenario, and the system decides who fits it.
Turning on the personalisation toggle widens the funnel considerably. Three additional sources come into play: the live conversation thread, prior ad interactions, and the user’s chat history and memory. Personalisation can be switched off in the ad settings, but OpenAI is explicit that ads will still be served on the basis of the current chat. Opting out reduces targeting; it does not remove the ads.
Custom audiences and the 25,000-user floor
The most sensitive component is the custom audience tool. Advertisers can upload lists containing email addresses, phone numbers or Android advertising identifiers belonging to their own customers, either to reach those people inside ChatGPT or to exclude them. OpenAI’s developer documentation sets firm limits: inclusion requires at least 25,000 matched users, with 100,000 recommended; uploaded files are deleted after processing, “usually no later than 24 hours”; and advertisers never see which individuals matched or pick out specific people from a list. Only first-party data the advertiser has the right to use is permitted — lists bought from data brokers are banned.
An OpenAI spokesperson told reporters that appearing on an advertiser’s list does not, by itself, trigger an ad. What the company has not clarified is whether raw, unhashed contact data is converted into an irreversible code before processing, or whether opting out of personalisation blocks an ad aimed at your email address or phone number. The documentation is silent on both points.
Zach Edwards, co-founder of DecryptAds, framed the strategy bluntly: OpenAI is “boiling the frog” and counting on users not noticing what it means for their privacy.
The European carve-out — and what stays off limits
OpenAI launched ads in the United States on 9 February 2026, extended them to Canada, Australia and New Zealand in April, then to markets including the UK, Mexico, Brazil and Japan, and announced on 18 August that 31 European markets — Spain among them — would follow the next week. Inside the EEA and Switzerland, however, the company says personalisation is not available “initially”, and custom audiences cannot be used in campaigns aimed at those regions. The most striking part of the investigation, the contact-list matching, therefore does not currently apply to a Spanish user.
Several boundaries apply everywhere. OpenAI insists it does not share conversations, chat history, memory, names, emails or precise location with advertisers, who receive only aggregate data such as impressions and clicks; ads are also said not to influence answers. Advertising appears solely on Free and Go accounts, while Plus, Pro, Business, Enterprise and Edu remain ad-free, and the free tier can trade ads for fewer daily messages and no image generation. No ads are shown to accounts identified as belonging to under-18s, none appear alongside conversations about health, mental health or politics, and temporary chats carry no ads. Data used for advertising can be reviewed and deleted, with full removal taking up to 30 days.
Scale explains the urgency. OpenAI says ChatGPT reaches 1.2 billion people a week and is preparing visual formats alongside the images the model generates. The real story is not contextual advertising, which behaves much like a search engine’s sponsored results, but the marriage of memory, historical chats and client lists. An assistant that users confide in is a far more intimate surface than a results page. The European carve-out proves ChatGPT can be monetised without crossing those data streams — the open question is how long OpenAI will wait before arguing that the rest of the world should be treated the same way.