OpenAI Is Now Following You Around the Rest of the Web
A researcher pulled apart OpenAI's ad measurement stack and found the thing everybody eventually finds. ChatGPT mints a signed JWT containing your account identifier, drops it into a cookie called __obi scoped to .openai.com with a one year expiry, and that cookie then gets sent back to OpenAI from ordinary websites you visit. The writeup is at https://www.buchodi.com/chatgpt-now-knows-what-you-do-on-other-websites-via-ad-collector/ and it hit 455 points on Hacker News.
The delivery mechanism is the boring one. Advertisers who buy placement in ChatGPT install a pixel, same as they have installed Meta and Google pixels for fifteen years. Loading that pixel sends __obi to a collector at bzr.openai.com along with the page path, location data, and whatever email addresses and phone numbers the pixel can scrape off the page. The researcher observed the identifier leaving twelve commercial sites including Chewy, Wayfair and Coursera, and noted that most of the identity data is scraped rather than deliberately handed over by the advertiser.
The detail that should bother you is the classification. Under OpenAI's cookie policy __obi is an analytics cookie, not a marketing cookie. So the consent banner where you accepted analytics and rejected marketing, the one you thought was the whole point of those banners, does not stop this. Safari on desktop and every browser on iOS break the mechanism through tracking prevention, which means the users protected are the ones who happened to pick the right vendor.
Why this belongs on an agents site and not just a privacy site: the entire pitch for assistants doing your shopping and booking and research is that they sit between you and the web. That position is worth enormous money precisely because it sees everything. Ad attribution is the first product that monetizes the position, and it arrived before agentic commerce did.
The advertising business model did not get bolted onto the assistant. The assistant is a better version of the advertising business model.
← Back to all articles
The delivery mechanism is the boring one. Advertisers who buy placement in ChatGPT install a pixel, same as they have installed Meta and Google pixels for fifteen years. Loading that pixel sends __obi to a collector at bzr.openai.com along with the page path, location data, and whatever email addresses and phone numbers the pixel can scrape off the page. The researcher observed the identifier leaving twelve commercial sites including Chewy, Wayfair and Coursera, and noted that most of the identity data is scraped rather than deliberately handed over by the advertiser.
The detail that should bother you is the classification. Under OpenAI's cookie policy __obi is an analytics cookie, not a marketing cookie. So the consent banner where you accepted analytics and rejected marketing, the one you thought was the whole point of those banners, does not stop this. Safari on desktop and every browser on iOS break the mechanism through tracking prevention, which means the users protected are the ones who happened to pick the right vendor.
Why this belongs on an agents site and not just a privacy site: the entire pitch for assistants doing your shopping and booking and research is that they sit between you and the web. That position is worth enormous money precisely because it sees everything. Ad attribution is the first product that monetizes the position, and it arrived before agentic commerce did.
The advertising business model did not get bolted onto the assistant. The assistant is a better version of the advertising business model.
Comments