OpenAI ad collector links cross-site browsing to ChatGPT accounts via __obi cookie

OpenAI has deployed an advertising tracking system that connects user activity on external websites to their ChatGPT accounts. The core mechanism relies on a cookie named __obi, which is scoped to the .openai.com domain but configured with SameSite=None and Secure attributes. This specific configuration allows the browser to attach the cookie to cross-site requests, distinguishing it from other OpenAI cookies that are blocked by standard SameSite=Lax policies.

The tracking flow begins when a user visits ChatGPT, where the client generates a random identifier and exchanges it for a signed JWT via a backend API. This token binds the user's account subject to the tracking ID and is used to set the __obi cookie on the OpenAI domain. The cookie has a one-year expiration, ensuring the identifier persists across sessions. Once set, the identifier is ready to be transmitted whenever the user interacts with sites that have installed OpenAI's advertising SDK.

On third-party websites, the installation of OpenAI's ad code triggers the transmission of the __obi cookie back to OpenAI's servers. The SDK collects data from the page, including form fields, rendered text, and tag-manager events. In observed traffic, scraped identity data outnumbered advertiser-supplied data, with email and phone numbers being hashed before transmission while location data like postal codes were sent in the clear. The mere loading of the SDK script can disclose the identifier before any explicit user action occurs.

The system's reach is significant, with the identifier observed being sent from numerous commercial websites to OpenAI's endpoints. The data collected includes specific browsing behaviors such as product searches and article reads, which are then linked to the user's ChatGPT profile. This creates a detailed profile of user interests and behaviors that extends beyond the ChatGPT interface itself. The implementation mirrors traditional ad-tech patterns but integrates them directly with a major AI service's user base.

For engineers, the key takeaway is the reliance on browser cookie behavior for cross-site tracking. The use of SameSite=None is critical to the system's function, as it overrides the default protection against cross-site request forgery and tracking. The system also includes a denylist for sensitive fields like passwords and SSNs, but the scope of data collection remains broad. Understanding this mechanism is important for developers building privacy-focused tools or auditing third-party scripts on their own sites.