OpenAI Tracks User Behavior on External Websites via Cookies Linked to ChatGPT Accounts

OpenAI's ad collection system sets an identifier cookie (__obi) linked to ChatGPT accounts, tracking user behavior on external websites they visit. This mechanism follows a structure similar to existing ad tracking technologies from Meta or Google, and the fact that such cross-site tracking occurs within ChatGPT, an AI conversational product, is noteworthy.

Tracking Mechanism and Data Collection

When a user accesses chatgpt.com, the backend issues a token in RS256 JWT format, which sets the __obi cookie on the .openai.com domain. This cookie has SameSite=None and Secure attributes and a validity period of one year. When advertisers install the OpenAI measurement pixel on their sites, data such as search products, articles being read, and purchase behavior are sent to OpenAI along with __obi upon page loading, linking it to ChatGPT account activity.

The collected data is classified into four sources: advertiser intentional transmission (in), form field scraping (fm), rendered page text (ht), and tag manager bus (js). In particular, email addresses and phone numbers are collected via the tag manager bus, while country, region, city, and postal code are transmitted in plaintext. URLs are reduced to origin and path for transmission, but paths containing sensitive information such as medical conditions or lawsuit filing forms are collected as is.

Privacy and Consent Policy Issues

OpenAI classifies the __obi cookie as 'Analytics cookies', but all observed sync tokens include consent_decision: analytics_allowed. This means that tracking occurs when analytics cookies are allowed, even if users reject marketing cookie consent. Additionally, anonymous identifiers are generated even in logged-out states and persist for at least 27 days, having the same stability as logged-in accounts.

The Automatic matching feature is enabled in 638 out of 881 pixels with known settings, controlled via OpenAI Ads Manager. While extremely sensitive information such as passwords, card numbers, SSNs, and medical history is excluded via a denylist, other behavioral data is collected extensively. Cases have been observed where a single __obi value is sent to OpenAI from more than 12 commercial websites, including Chewy and Wayfair.