ChatGPT now knows what you do on other websites via ad collector

OpenAI's ad collector, accessible at bzr.openai.com, sets a cookie named __obi, which is tied to a user's ChatGPT account. This cookie is shared with other websites visited by the user, allowing OpenAI to connect their online activities to their ChatGPT account. Every company that buys ads on ChatGPT installs OpenAI's code on its site, sending the __obi cookie along with data about the page being browsed. This includes user searches, articles read, and purchase behaviors.

To verify this mechanism, the author reproduced the process on their phone, validating it with two independent capture methods and analyzing over 936 distinct advertiser pixels across 1,029 hostnames. The process begins when ChatGPT creates a 16-byte identifier, which is then sent as a JWT (JSON Web Token) to OpenAI's ad collector. This identifier becomes a cookie on OpenAI's domain, with the cookie configuration allowing cross-site requests and a one-year expiration period.

The __obi cookie is configured with SameSite=None, meaning it can be sent on cross-site requests without being restricted to the originating website. OpenAI blocks most sensitive data in the cookies, but still collects email, phone, name, and location information. They hash the email and phone data before transmission, but send location data in the clear. Three request classes are made from an advertiser's page to OpenAI's hosts, with the largest source of email harvesting being the tag-manager bus.

OpenAI has a denylist that excludes certain sensitive fields, but some of the observed paths contained sensitive information, such as medical conditions, debt-solutions funnels, and litigation intake forms. Despite the policy, OpenAI allows automatic matching on 638 out of 881 pixels, including credit and lending advertisers. The __obi cookie is classified as an analytics cookie, even though users who grant analytics consent but refuse marketing consent still receive it.