ChatGPT Ad Tracking Mechanism Exposed: Cross-Website Data Relayed to OpenAI

Overview

Recent investigations have revealed that OpenAI established a cross-site tracking mechanism through its internal advertising platform, "Bazaar." Although this feature is nominally categorized under "analytics," its actual operation allows OpenAI to link users' browsing behavior on third-party advertisers' websites with their ChatGPT account identities. This discovery has sparked new controversy regarding user privacy boundaries and the transparency of data collection.

Technical Mechanism Analysis

The core of the tracking lies in a Cookie named __obi. When a user logs into ChatGPT, the client generates a random identifier and binds it to the user's account via an RS256-signed JWT (JSON Web Token). This Token has a validity period of only 60 seconds, after which it is written as a Cookie under the .openai.com domain with a validity of up to one year.

1. Identifier Generation: The ChatGPT backend generates a JWT containing the account subject identifier (sub) and the object identifier (obi).
2. Cookie Setting: The client sends the Token to bzr.openai.com, and the browser saves the __obi Cookie with SameSite=None and Secure attributes.
3. Data Return: Companies purchasing ChatGPT ads embed OpenAI's code snippet on their websites. When users visit these sites, the browser automatically sends requests to OpenAI carrying the __obi Cookie.

Crucially, even if the advertising SDK itself has a code path without attached credentials, the initial request to load the SDK script still leaks this Cookie. This enables OpenAI to continuously collect cross-site data without active user interaction.

Scope of Data Collection

The data collected by this mechanism extends far beyond simple page click statistics. The SDK not only captures identifiers (in) intentionally provided by advertisers but also obtains more sensitive personal information by scanning form fields, rendered text, and tag manager buses (such as Adobe Data Layer and Google Tag Manager).

• Mixed Sources of Identity Information: Among the observed traffic, the number of identity data events obtained through scraping (685) was significantly higher than those proactively provided by advertisers (255).
• Leakage of Sensitive Data: Email addresses, phone numbers, and names are hashed using SHA-256 before transmission, but information such as country, region, and city postal codes is transmitted in plaintext. Postal codes are one of the most frequently collected form fields.
• Details of Browsing History: Although URL query parameters are removed, path information is retained. Observed data shows that paths containing medical conditions, debt solution funnels, and legal case intake forms were all uploaded to OpenAI servers.

Privacy Policy and Compliance

OpenAI's Cookie policy classifies __obi as an "Analytics Cookie," claiming it is used to help the company understand service performance and usage. However, the technical implementation demonstrates the ability to identify users across websites, which clearly exceeds the scope of traditional website analytics tools.

Notably, OpenAI separates analytics and marketing into two distinct consent options. The investigation found that all decoded synchronous Tokens carried the consent_decision: analytics_allowed flag. This means that even if users reject marketing uses, as long as analytics collection is permitted, their cross-site tracking mechanism remains active. Currently, OpenAI has established a blocklist for specific sensitive fields (such as passwords, Social Security numbers, and medical diagnoses), but this does not alter the core architecture of large-scale identity association.