How does an advertiser reach you inside ChatGPT? A newsroom decided to find out by becoming one. 404 Media ran its own advertising campaigns in the chatbot and reported this week that advertisers can target people in two ways at once: by what they are talking about, and by who they are.
The findings matter because of scale and habit. 404 Media puts ChatGPT at roughly one billion monthly users, many of whom treat a conversation with a chatbot as a private exchange. The ad system treats that same conversation as a place where a sponsored message can be matched to a person.
Route one: target the conversation
Advertisers can describe the kinds of conversations where their product might fit. OpenAI calls these descriptions context hints. In OpenAI’s documentation, as described in reporting on the tests, the hints guide how ads are matched rather than working as exact-match keywords, and they do not guarantee that an ad will appear in any particular conversation.
The effect is still direct. If a conversation matches what an advertiser described, an ad can be selected for that session because of what the user is typing about at that moment. That is a different surface from a search page or a social feed: the input is a prompt a user may have written expecting a one-to-one exchange.
Route two: target the person
The second route will sound familiar to anyone who has bought ads on a large platform. According to 404 Media’s tests, advertisers can upload lists of email addresses, phone numbers or Android advertising IDs — in readable form or hashed — and OpenAI matches those identifiers against ChatGPT accounts. A shop that already has a customer’s email address from an earlier purchase can, in other words, use that address to help define an audience inside ChatGPT.
404 Media itself noted that this kind of customer-list targeting is standard practice across the advertising industry. The newer element is the combination: an identity advertisers already hold, matched inside a conversational product, alongside targeting based on the conversation itself. Gizmodo, summarizing the tests, noted that OpenAI did not respond to its request for additional detail by the time it published.
The guardrails, as stated
OpenAI’s documentation sets limits. An uploaded audience can be used for active targeting only once it reaches at least 25,000 matched users; smaller lists can be used to exclude people instead. Advertisers cannot pick a specific individual, and OpenAI says they are not shown which accounts matched their list.
OpenAI also states that uploaded files are deleted after processing, usually within 24 hours in the reporting, that conversations are not shared with advertisers or marketing partners, and that ads do not influence ChatGPT’s answers. One question remains open: 404 Media reported that it is unclear whether OpenAI hashes readable email addresses and phone numbers that advertisers upload. That is an unanswered question in the reporting, not evidence that uploaded data was exposed.
Turning personalization off does not turn this off
Users do have controls. OpenAI says people can manage ad personalization, dismiss an ad, send feedback and see an explanation of why a particular ad appeared. Ads in the current rollout are shown to users on the Free and Go plans, while the paid tiers listed in reporting do not receive them.
But the off switch is partial. If a user disables ad personalization, OpenAI’s own wording, quoted in reporting on the tests, is that ads may still be based on the current chat. Other conversations and history are not used in that case, according to OpenAI’s help documentation — yet the conversation happening right now remains a targeting signal for anyone eligible to see ads.
That gap between expectation and mechanics is the core of the story. The formal controls exist, and OpenAI’s stated policies are clear. What 404 Media’s experiment showed is that most users have no reason to know an advertiser can arrive through their inbox identity and their prompt at the same time. Until that becomes common knowledge, the practical advice from the reporting is simple: treat a ChatGPT conversation on a free plan less like a private notebook and more like a page that can carry advertising matched to what is on it.
