SAN FRANCISCO, October 17 — OpenAI now lets advertisers upload their own customer email and phone lists and match them to ChatGPT users, a 404 Media investigation revealed this week, formalizing a targeting mechanic that looks a lot like the one Meta spent a decade building. Inclusion campaigns require at least 25,000 matched identifiers, according to OpenAI documentation cited by Gadget Review, which puts the floor well above anything a small list can clear.

Four identifier types are supported: raw email addresses, phone numbers, SHA-256-hashed versions of both, and Google Advertising IDs. OpenAI states that advertiser-uploaded files are deleted after processing and that advertisers cannot identify which individuals on their list matched, according to Gadget Review. Advertisers also supply what OpenAI calls “context hints,” descriptions of the conversations, topics, or keywords where a sponsored card would be relevant. The current conversation remains a targeting signal even when a user disables ad personalization.

The commercial context is doing most of the work here. OpenAI’s ad business hit a $1 billion annualized run rate on August 31, per The Information via TokenPost, with an advertiser base in the tens of thousands and availability in more than 60 countries by October 15, up from 40-plus in August. The 140-person ad group has reorganized seven times since the U.S. test began February 9. Meta is the largest single source of hires. The self-serve Ads Manager beta rolled out May 5; CPC bidding was added after.

Ads reach only Free and Go users. Plus, Pro, Business, Enterprise, and Edu subscribers are excluded, which quietly concentrates ad inventory on the users least likely to be paying for anything else.

The 25,000-match floor is the operational detail worth sitting with. It rhymes with earlier reporting on AI-search conversion rates and the fragility in ChatGPT referral measurement exposed last week: the channel converts, but the instrumentation and the entry ramps are still tuned for advertisers who already look like Meta’s.

Sources