ChatGPT now carries advertising, and advertising needs measurement: an advertiser wants to know whether someone who saw an ad later bought something. The standard tool for that is a small snippet of tracking code, a pixel, that the advertiser installs on its own website. A security researcher who goes by Buchodi has published a technical teardown of how OpenAI's version behaves, and the findings have set off a privacy debate. This article separates what the researcher observed from what he inferred, and what OpenAI has and has not said.
A note on sourcing: the findings come from a single independent researcher's analysis, published on 20 September, and have not been independently reproduced in what I could find. OpenAI has not confirmed or disputed them in substance.
What the researcher reports observing
According to the teardown, OpenAI's ad system sets a cookie named __obi. It is scoped to openai.com, lasts a year, and is configured so that it is sent back to OpenAI whenever a visitor loads a page running OpenAI's advertiser pixel. The researcher observed the identifier on twelve commercial websites, among them Chewy, Wayfair, ThriftBooks, Eventbrite, HelloFresh, Coursera and SeatGeek.
The cookie, he says, is linked to a ChatGPT account. In his sample, 736 of the tokens observed belonged to identified account users and 196 to anonymous ones, and the anonymous identifier was stable on a device for at least 27 days. That means the tracking is not limited to logged-in ChatGPT users.
Data sent alongside it, in his account, included the pages visited (with query strings removed) and identity details: email addresses, phone numbers and names, which were hashed before transmission, and location data such as country, region, city and postal code, which were sent in readable form. He also reports that the pixel picked up identity information from form fields and page text, and that the amount scraped from pages exceeded the amount advertisers supplied deliberately. Among the events he examined were pages touching on health conditions and debt help.
Hashed does not mean anonymous. A hashed email can still be matched against a database that holds the same email.
What he did not show
The teardown is careful about its limits, and they matter. The behaviour was confirmed on Chrome for Android. Safari and Chrome on iOS were not affected, because their tracking protections block this kind of third-party cookie. Desktop Chrome was not tested. Not every ChatGPT session produced a tracking token, roughly one in five did, and the mobile web client served ads without one. And the link between the cookie and a ChatGPT account is partly inferred from how the system is designed, not directly observed on OpenAI's servers. He adds that advertisers cannot see or use the identifier.
What OpenAI has said
The researcher sent his findings to OpenAI on 14 September. OpenAI's support team acknowledged the message and said the observations would be shared internally. According to the teardown, it did not answer two specific questions: how it classifies the cookie, and whether people who decline marketing consent are still tracked. Reports note that OpenAI's cookie policy lists the cookie among analytics cookies, which the researcher argues is a mismatch with what it does, since it appears to function as an advertising identifier.
OpenAI's published position is that logged-out visitors can use the Manage Cookies or Your Privacy Choices link in a site footer, and that it recognises the Global Privacy Control signal as a valid opt-out. Whether that opt-out stops the cookie in practice is the question that remains unanswered.
How advertising pixels normally work
To judge what is unusual here, it helps to know what is routine. When you click an ad and later buy something, the shop's confirmation page loads a small piece of code from the ad network. That code tells the network that a purchase happened, and a cookie previously set by the network lets it link the purchase to the ad you saw. This is how Meta's and Google's advertiser tags have long worked, and it is the reason browsers have spent years restricting third-party cookies.
The controversy is therefore not that a pixel exists. It is about scope and labelling: what else the pixel collects from the page, whether it is tied to a person's chat account, and whether the cookie's classification tells users the truth about its purpose. Those are the details on which the researcher and OpenAI have not yet been matched by a public answer.
The question is not whether ad tracking exists. It is whether people were told what this particular tracker does.
Why this matters beyond one company
Advertising pixels from Meta, Google and others already work this way, so the technique is not new. What is new is the combination: a company that holds people's private conversations with an AI assistant now also has a way to see what those people do on other sites. Even if the two data sets are kept apart, the possibility raises questions that regulators in the EU and elsewhere will care about, such as whether consent for one purpose covers the other. India's data protection law and its rules on consent apply here too, though this article does not claim any violation.
What readers can do
None of the following depends on the research being right, and none is a substitute for OpenAI's own answers. Use a browser with strong tracking protection. Safari and the iOS versions of Chrome were not affected in the researcher's tests. Review cookie choices on the sites you visit, and use the privacy links in ChatGPT's own footer and settings. Turn on the Global Privacy Control signal in browsers that support it, since OpenAI says it honours it. And avoid pasting sensitive personal or medical details into any chatbot you have not read the privacy terms for.
The most useful thing to watch now is OpenAI's reply. A clear answer on how the cookie is classified and whether it respects consent would settle much of the debate. Until then, treat the findings as credible but unconfirmed, and the privacy questions as real.
