The Maze: ChatGPT advertising combines conversation topics with familiar customer-list controls. An October 15 experiment by technology newsroom 404 Media shows how advertisers suggest suitable conversations for their ads. Its investigation also documents audience matching using personal identifiers, with an important limit: advertisers cannot choose or identify individual matched users.
The Details:
A topic suggestion, not a keyword order. 404 Media ran two ads promoting its reporting about human review of some ChatGPT conversations. It supplied “context hints” about privacy and technology. These tell OpenAI which conversations might suit an advertisement; they do not require an exact keyword match. The experiment makes a normally invisible buying process tangible. It demonstrates contextual advertising, where the subject of a conversation helps choose an ad. It does not demonstrate that a particular person received an ad because their email address was uploaded.
The customer list is a separate lever. Advertisers can supply email addresses, phone numbers, obscured versions called hashes, or Android advertising identifiers. OpenAI matches these against users to form an audience. A campaign can include a group, exclude it, or change its bid—the maximum amount it is willing to pay—when a group member is eligible. Inclusion and bid changes require at least 25,000 matched users; smaller groups can be used for exclusions. This is a count after matching, not simply rows in an uploaded file. Duplicates, invalid entries and identifiers without matches do not count. The practical size of a merchant’s contact database therefore overstates its usable advertising audience.
The privacy boundary has several parts. OpenAI told 404 Media that advertisers cannot see which individuals matched or select people within an audience. Merely appearing on a list does not cause an ad to be shown. The company says uploaded files are deleted after processing, usually within 24 hours. It did not directly answer the reporter’s question about whether raw identifiers are hashed on receipt. Separately, switching off ad personalization still permits ads based on the current chat. That limits the use of past activity; it does not remove conversation context from advertising. OpenAI says advertisers receive aggregate performance information, without access to users’ chats.
MarketMaze Take: The useful dividing line is between targeting a group and exposing a person. OpenAI’s minimum audience size and hidden individual matches are meaningful restrictions. They also leave a commercial system in which customer identifiers and conversational intent can influence advertising. Both facts belong in the same privacy discussion.
Customer data changes the buying decision. For a retailer, excluding existing customers can serve a different purpose from paying more to reach valuable ones. Those choices remain relevant to marketers outside current ChatGPT ad markets: a new interface can still reward an existing customer database. Availability needs its own check. OpenAI’s guide warns against custom audiences in the European Economic Area and Switzerland while ad personalization is unavailable there.
Two ads cannot establish a sales advantage. The test explains buying controls, not conversion lift, reliable delivery to every listed customer or individual surveillance. OpenAI says advertising helps fund broader access to ChatGPT, including its Free and lower-cost Go plans. The commercial question is whether useful relevance earns enough trust to support that model. This experiment does not settle it.
Sources: 404 Media, OpenAI Help Center, OpenAI
