ChatGPT's Ad Pixel Tracks What You Do on Other Sites, and Advertisers Can't See It

Any brand running ads inside ChatGPT has to install a small script on its own website: bzr.openai.com. It is presented as a standard conversion pixel, the same category of code Meta and Google have shipped for a decade. An independent researcher's writeup published September 20 shows it does something those pixels were never built to hide from the advertiser installing them. The ChatGPT ad pixel tracking mechanism sets a cookie scoped to OpenAI's own domain, not the advertiser's, and that cookie follows a visitor's behavior back to a specific ChatGPT account on OpenAI's side. The advertiser who put the code on their site cannot read it.

The writeup carried real weight on Hacker News, 639 points and 336 comments, and the researcher says the mechanism was reproduced directly and cross-checked against several months of traffic covering 936 distinct advertiser pixels. That is not a claim resting on OpenAI's own documentation. It is an outside party reverse-engineering what the code actually does once it is running on a live site.

How the ChatGPT ad pixel tracking mechanism works

The flow has three steps. First, ChatGPT itself generates a signed JWT token that binds a random identifier to the logged-in user's account, and sends it to bzr.openai.com/v1/obi/sync. Second, OpenAI's server responds by setting a cookie called __obi, configured with SameSite=none so it travels across domains, and set to expire a year out. Third, when that same person later visits any site running the advertiser's pixel, the __obi cookie rides along automatically, and whatever behavioral data the pixel is collecting, products viewed, pages read, form fields touched, gets tied back to that ChatGPT identity on arrival at OpenAI's servers.

The researcher's data shows the pixel is not passive. Scraped signals outnumber anything a visitor actually typed into a form, 685 scraped events against 255 intentional submissions in the sample reviewed. Postal code turned up as the single most-harvested field, showing up 100 times across 28 different sites. Email and phone numbers get hashed before they leave the browser. Geographic data does not; it travels unencrypted. Retailers named in the writeup as running the pixel include Chewy, Wayfair, and Eventbrite.

Why the advertiser installing the code can't see any of this

This is the detail that separates ChatGPT's ad pixel tracking setup from an ordinary Meta or Google pixel, where the advertiser at least owns the dashboard the data lands in. Because __obi is scoped to openai.com, not to the advertiser's own domain, standard same-origin cookie rules keep the advertiser's own code from ever reading it. A brand can watch its own analytics all day and never see the cookie that is quietly resolving its visitors to individual ChatGPT accounts. The data has a destination. It just is not the company that paid to put the tracking code on its own site.

That asymmetry is not new in advertising generally, it is closer to how Nielsen's household ratings meters have worked since the 1950s: Nielsen installs the meter, Nielsen reads what it reports, and the network buying the ratings only ever sees the aggregate number Nielsen chooses to hand over. The difference here is that Nielsen's arrangement was disclosed and negotiated. A pixel presented to marketers as a conversion tracker, without disclosure that a second, invisible cookie is riding along scoped to someone else's domain, is a different kind of arrangement.

The scope narrowed, and the questions went unanswered

OpenAI has run ads inside ChatGPT since earlier this year, and the collector script's own scope was reportedly wider before August 27, when it narrowed to the current set of fields. That narrowing suggests OpenAI was already aware the collection was broader than it needed to be. It does not resolve the core question: whether an advertiser knowingly consented to a cookie it cannot itself inspect being set on its own site's traffic.

The researcher filed a report with OpenAI on September 14, addressed to both the press and privacy contacts, asking two direct questions about how the __obi cookie is classified under OpenAI's own data handling policy. The response, per the writeup, was an acknowledgment with no answer to either question. Six days is enough time for a substantive reply if one exists. Its absence is itself a data point.

What this means if you are running or planning ChatGPT ads

For anyone buying or evaluating ChatGPT ad placements, this is the same due-diligence question marketers already had to learn to ask about Meta and Google pixels years ago, just showing up under a newer product with a chat identity wired into it. Three things are worth checking before installing OpenAI's pixel on a live site:

• Ask directly what data bzr.openai.com collects beyond stated conversion events, and get the answer in writing rather than relying on the SDK's own comments.
• Assume any cookie an ad platform sets scoped to its own domain is unreadable to you by design, not by accident, and factor that into what you are willing to disclose to your own users about third-party tracking on your site.
• If a customer or a compliance team asks what data ChatGPT's ad SDK sends back to OpenAI, "we don't know, we can't see it" is currently the honest answer, and that is worth flagging to legal before, not after, someone asks.

None of this means ChatGPT ads are a mechanism to avoid outright. It means the standard privacy checklist marketers apply to any ad pixel needs to run again here, with the added wrinkle that part of what the pixel does is, by construction, invisible to the party installing it.

Sourced from the researcher's writeup on the ChatGPT ad-collector mechanism, independently reproduced and cross-checked against observed pixel traffic.