OpenAI’s advertising measurement system can send a one-year browser identifier created during a ChatGPT session back to OpenAI when a user visits unrelated websites carrying its ad pixel, according to a network-traffic analysis published September 20. The finding matters because the cookie is not merely local conversion bookkeeping: the observed setup bound its identifier to a signed-in ChatGPT account before the browser carried it across sites.
This is a carefully bounded finding, not a finished audit. Independent researcher Buchodi reproduced the flow on Chrome for Android and recorded the identifier arriving at OpenAI from 12 commercial sites. No reputable outlet has yet independently reproduced the capture, and the researcher did not observe what OpenAI’s servers did after accepting the requests. OpenAI acknowledged the disclosure but had not answered the researcher’s consent and classification questions at publication.
One token crosses three boundaries
The reported route begins on ChatGPT. Its web client asks an OpenAI backend for a short-lived, cryptographically signed token containing an account subject and a separate 22-character identifier. The browser then sends that token to OpenAI’s advertising collector, which sets a cookie called __obi on the wider openai.com domain for one year.
Two cookie attributes do the important work. Secure limits transmission to encrypted connections. SameSite=None permits the browser to attach the cookie when an OpenAI resource is requested from a different website. When a retailer loads OpenAI’s measurement script, the request still goes to an OpenAI domain, so the browser can include __obi even though the visible page belongs to the retailer. The retailer cannot read the cookie directly, but OpenAI receives it.
The researcher decoded 932 synchronization tokens: 736 carried an account subject and 196 were anonymous. Across a broader traffic set covering 1,029 hostnames, one observed __obi value appeared under as many as 10 advertiser pixel IDs. OpenAI’s cookie policy confirms that __obi lasts one year and classifies it as analytics on ChatGPT and OpenAI domains. The policy does not explain the account-binding sequence described in the capture.
The pixel is designed to find conversions
OpenAI’s conversion documentation says its pixel can connect website events to eligible ad clicks. Its automatic advanced matching feature detects supported customer details in forms and other page sources, hashes fields such as email addresses and phone numbers in the browser, and submits them with conversion events. OpenAI tells advertisers to provide clear notice and obtain consent where required.
Buchodi’s capture adds uncomfortable implementation detail. The observed pixel traffic included page paths and some location fields; paths in the sample exposed visits to a medical-condition page, debt services and litigation intake. The report says query strings were removed and a denylist excluded passwords, payment-card numbers, Social Security numbers, diagnoses and several other sensitive fields. Guardrails exist, but a path can reveal plenty without asking a form for permission.
The strongest counterargument is that this resembles ordinary ad-tech attribution. OpenAI needs to know whether an ad produced a sale, advertisers receive aggregated reporting rather than a readable ChatGPT identity, and Safari’s tracking protections block the demonstrated third-party-cookie route. The test also covered Android Chrome, not desktop browsers, and only about one in five observed ChatGPT sessions minted the synchronization token.
TINA’s view: call advertising advertising
TINA’s view: the technical mechanism is less surprising than the consent label. A persistent identifier that leaves ChatGPT through an advertising collector and returns from advertiser pages belongs behind marketing consent, even if OpenAI also uses it for analytics. Filing it under analytics asks users to understand an internal taxonomy when the browser behavior is the part that matters.
This judgment would change if OpenAI shows that __obi cannot be used for ad attribution or account resolution, demonstrates that rejecting marketing consent prevents the cross-site flow, or replaces the shared identifier with per-advertiser values that cannot connect activity across sites. It would strengthen if independent captures reproduce the account-bound route on desktop Chrome or show that it operates broadly without an eligible ChatGPT ad interaction.
The next signal is OpenAI’s response: what __obi is for, which consent switch controls it, how long event data is retained, and whether an account-bound value is ever joined server-side to browsing events. Until those answers arrive, the measured fact is narrower but still consequential: one ChatGPT-created identifier crossed from OpenAI’s own service into the commercial web and came back carrying context from other sites.
