OpenAI's ChatGPT Ad Tracking Cookie Raises Privacy Concerns

OpenAI's advertising measurement system has been found to employ a cross-site tracking cookie, identified as __obi, which can correlate user activities on external advertiser websites with their ChatGPT accounts. This mechanism, observed across numerous advertiser pixels, collects detailed browsing information, including page paths and potentially sensitive data like hashed emails, even for users who are not logged into ChatGPT.

The tracking process begins when a user visits ChatGPT, prompting the system to generate a random identifier and request a signed token from OpenAI's backend. This token contains an account-linked subject value, the obi identifier, and a consent status for analytics. Subsequently, ChatGPT sends this token to bzr.openai.com, an OpenAI advertising collector internally referred to as "Bazaar." In response, the collector sets the __obi cookie for the .openai.com domain. Crucially, this cookie is configured with SameSite=None and Secure attributes, enabling it to be included in cross-site requests, and has a maximum lifetime of one year.

This specific configuration distinguishes __obi from other OpenAI cookies, which are often restricted by domain settings or SameSite=Lax policies, preventing them from being sent in third-party requests. Advertisers who purchase ads on ChatGPT can integrate OpenAI's measurement pixel into their websites. When a visitor with the __obi cookie present on their browser accesses such a site, the browser automatically includes the cookie in requests to OpenAI's infrastructure. This allows OpenAI to gather data on the pages a user visits on these third-party sites, including product pages, article URLs, and conversion funnels.

While query strings are reportedly removed from observed URLs, the page paths themselves remain visible. These paths have been noted to include sensitive information related to medical conditions, debt solutions, and legal intake services. Furthermore, the advertising software is capable of collecting data directly from advertiser pages, either through data intentionally provided by advertisers or by scraping information from forms, rendered page content, and tag manager systems.

According to independent analysis, the data collected can include hashed email addresses, phone numbers, names, and location details such as country, region, city, and postal code, which may be transmitted in plain text. The tracking pixel also reportedly monitors data layer elements within systems like Adobe's data layer and Google Tag Manager. Notably, automatically scraped identity information appeared more frequently than advertiser-supplied data in observed traffic.

The tracking functionality is designed to operate even for anonymous users. In such cases, a device-linked subject identifier is generated, which has been observed to remain stable for at least 27 days. However, the effectiveness of this tracking method is limited in browsers with robust third-party cookie blocking, such as Safari and iOS browsers based on WebKit, due to their Intelligent Tracking Prevention features.

OpenAI's official cookie policy classifies __obi as an analytics cookie, not a marketing cookie. However, the observed sync tokens were issued even when users had explicitly denied marketing consent, provided they had granted analytics consent. OpenAI has acknowledged a privacy inquiry regarding this matter but has not yet provided a detailed response concerning the cookie's classification or consent behavior.