ChatGPT Is Tracking You on Other Websites: What the __obi Cookie Does, and How to Block It in Five Minutes
What's in this guide
On September 20, a security researcher who writes under the name Buchodi published a teardown of ChatGPT's new advertising system, and it went straight to the top of Hacker News with more than 600 upvotes. The claim: ChatGPT plants a cookie in your browser called __obi, ties it to your account, and then collects that same cookie from ordinary shopping sites you visit later. OpenAI can, in principle, connect your pet supply, furniture and concert ticket browsing back to the account you use to ask about your health, your money and your job.
The big outlets have not written the plain English version yet, so here it is. I read the research, checked it against OpenAI's cookie policy and its developer documentation for advertisers, then went looking for the cookie in my own browser. Here is what it does, who it affects, and three ways to shut it down.
The short answer
If you use ChatGPT on the Free or Go plan, in Chrome or Edge, and you have not blocked third-party cookies, a one year cookie called __obi is probably in your browser right now. Any website that advertises on ChatGPT (the tracking firm Bloomberry counts more than 12,000 of them as of September 19) can send that cookie back to OpenAI the moment its page loads, along with the page you are on and, on some sites, a hashed copy of your email, phone number and name.
Paid subscribers on Plus, Pro, Business and Enterprise do not see ads, and Safari users are protected because Safari blocks third-party cookies by default. Everyone else should spend five minutes on the fixes below.
What the researcher actually found
The mechanism has three steps, and it is the same trick Facebook has used for a decade, bolted onto a chatbot.
Step one: ChatGPT creates an ID for you. When you use chatgpt.com, the site generates a random identifier and has OpenAI's servers sign it together with your account. That signed token is posted to an OpenAI ad server at bzr.openai.com, which answers by setting the __obi cookie. It lasts one year, covers the entire .openai.com domain, and carries a flag called SameSite=None. That flag is what allows the cookie to be sent from other people's websites, and it is the only OpenAI cookie set up that way.
Step two: advertisers install OpenAI's code on their sites. Anyone who buys ads in ChatGPT is told to add a script called the Measurement Pixel to their website so OpenAI can count which ad clicks turn into purchases. OpenAI's developer documentation describes this, including a feature called automatic advanced matching, where the pixel scans the page for email addresses and phone numbers, hashes them, and sends the hashes to OpenAI.
Step three: your browser hands the cookie over. When you land on an advertiser's page, the pixel script loads from OpenAI's servers and your browser attaches __obi automatically. The researcher watched this happen from 12 commercial sites on one phone, including Chewy, Wayfair, HelloFresh, Coursera, SeatGeek, Eventbrite and ThriftBooks.
Three details stand out. It works logged out: anonymous visitors get an ID that stayed stable for at least 27 days. The pixel scrapes more than it is handed: it pulled identity details out of web forms and page text more often than advertisers passed them in deliberately, and while email, phone and name are hashed, postal code and city go in the clear. And page paths survive: paths the researcher saw reaching OpenAI included a medical condition, a debt relief funnel and a legal intake form.
Then there is the label. In OpenAI's cookie policy, updated September 10, __obi is the only cookie in the "Analytics" section. Every token the researcher decoded carried a consent value of "analytics allowed," which means someone who accepted analytics cookies but refused marketing cookies still gets it.
Two honest limits. The research was done on Chrome for Android, not desktop. And the final step, OpenAI joining the off-site data to your account on its servers, was not observed directly. It follows from how the system is built, but nobody outside OpenAI has seen it. The researcher sent questions to OpenAI's press and privacy teams on September 14 and has an acknowledgement and no answers. Reported in detail, then, but not confirmed by OpenAI.
Who this affects (and who it does not)
Your exposure comes down to your plan and your browser.
Plan. Ads run only on Free and Go. OpenAI's help center says Plus, Pro, Business, Enterprise and Edu accounts get no ads, so there is no ad system to sync a cookie for. Ads also skip Temporary Chats and under 18 accounts. The ad test began in the US on February 9, 2026, and personalized ads are not yet switched on in the EU or Switzerland.
Browser. Safari on Mac, iPhone and iPad blocks third-party cookies outright, and every iOS browser runs on Safari's engine, so the cookie never travels. Firefox with Enhanced Tracking Protection on Strict blocks it too. Chrome and Edge on desktop and Android send third-party cookies by default, and those are the browsers most people use.
One wrinkle: only about one ChatGPT session in five produced a sync token in the researcher's captures, so not every Free account on Chrome has the cookie today. But it only has to happen once, and then it lasts a year.
What I found in my own browser
I use ChatGPT on a Free account in desktop Chrome, so I was the target audience. Checking takes a minute: open chatgpt.com, press F12 for developer tools (Option, Command and I on a Mac), click the Application tab, expand Cookies in the left column, click https://chatgpt.com and type obi in the filter box.
There it was. A cookie named __obi, domain .openai.com, expiry almost exactly a year out, value a short string of random looking characters.
Then I opened a new tab, went to one of the retail sites named in the research, and watched the Network tab with the filter set to openai. Two requests appeared as the page loaded: one fetching oaiq.min.js from bzrcdn.openai.com, one posting to bzr.openai.com. The first carried my __obi cookie. I had never bought anything from that site or clicked a ChatGPT ad, and OpenAI was told I was there.
For the record, "Personalize ads" was switched on in my ChatGPT settings at the time, because that is the default. Turning it off did not remove the cookie.
Three levels of fix: pick yours
As with the smart TV tracking in my guide to turning off ACR on any TV, there is no single off switch. There are three layers, and each closes a different door.
Level 1: block third-party cookies (two minutes, fixes the core problem)
If your browser refuses to send cookies to any domain other than the one you are visiting, __obi stays on chatgpt.com and never leaves.
In desktop Chrome: Settings, Privacy and security, Third-party cookies, then "Block third-party cookies." In Chrome for Android: Settings, Site settings, Third-party cookies, Block. Edge has the same option under Cookies and site permissions. In Firefox, set Enhanced Tracking Protection to Strict. Safari users can skip this step.
After switching it, I reloaded the retail site and repeated the Network check. The script still loaded, but the request no longer carried the cookie. A few sites (some bank logins, some embedded tools) misbehave with third-party cookies blocked, and Chrome offers per-site exceptions for those.
Level 2: tell ChatGPT no (five minutes, cuts off the source)
Level 1 stops the cookie travelling. Level 2 stops it being created.
First, cookie consent. On chatgpt.com, find "Your privacy choices" or the cookie preferences panel in the page footer, and decline analytics cookies as well as marketing cookies. This is the part most people get wrong, because the instinct is to refuse "marketing" and allow "analytics." The researcher's data says __obi is issued under analytics consent, so both need to be off.
Second, ad controls. Go to Settings, then Ad Controls (or open chatgpt.com/#settings/Advertising). Turn off "Personalize ads" and "Past chats and memory," then tap "Delete ads data." OpenAI says deleted ads data takes up to 30 days to leave its servers.
Third, marketing privacy. Under Settings, Data Controls, a separate "Marketing Privacy" setting (on by default for free accounts, as WIRED reported in May) lets OpenAI promote itself to you on other platforms using data like your email. Switch it off too.
Finally, delete the cookie that already exists: Chrome Settings, Privacy and security, Third-party cookies, "See all site data and permissions," search for openai, delete. With consent declined it should not come back.
Level 3: leave the ad system entirely (the full lockdown)
If you would rather not rely on settings OpenAI can rearrange, there are three exits.
Upgrade to Plus or Pro. No ads, no ad cookie. My comparison of Claude, ChatGPT and Gemini covers what the paid tiers actually get you.
Switch the Free plan to Ads-Free. OpenAI offers a free, no ads option with lower message limits and fewer tools (no image generation or deep research), under Settings, Ad Controls, "Change plan to go ad-free." You trade capacity for privacy.
Add a blocking rule. If you run uBlock Origin or another blocker that accepts custom filters, two lines shut the pixel down everywhere:
||bzr.openai.com^
||bzrcdn.openai.com^
That stops the script loading on advertiser sites at all: no cookie, no page path, no scraped form data. Chatgpt.com itself is unaffected.
What the "Personalize ads" toggle does not do
OpenAI's help center is careful with its wording. Turning off personalization means OpenAI will not use your other chats, memories or ad history to choose which ad you see. It says nothing about the cookie. In my testing __obi was still present with personalization off, and still sent to the retail site until I blocked third-party cookies.
The FAQ's other reassurance, "we do not share your conversations with advertisers," is almost certainly true and beside the point. The concern runs the other way: advertisers' websites sharing your visits with OpenAI. Nothing in the help center addresses that.
If you followed my guide to stopping ChatGPT, Claude and Gemini from training on your chats, you know the pattern: the friendly toggle and the switch you actually care about are often two different things.
Is this worse than what Meta and Google do?
Mechanically, no. The Meta Pixel and Google's conversion tags have worked this way for years, and several Hacker News commenters made that point.
I still think it is different, for two reasons. First, the content. What people type into ChatGPT is not what they post on Instagram. It is medical worries, salary negotiations and legal questions, typed to something that feels private. Linking that account to a browsing trail that includes a debt relief site is a new kind of profile, held by the same company you confide in. Second, the framing. Meta never called the Pixel "analytics." OpenAI's cookie policy does, and its consent screen lets you refuse marketing and still get the cross site cookie. That is the kind of thing European regulators notice, which may be part of why personalized ads are not switched on there yet.
The verdict
Block third-party cookies in your browser today. It takes two minutes, it stops __obi from ever leaving chatgpt.com, and it protects you from the Meta and Google versions of the same trick as a bonus.
Then spend five more minutes in ChatGPT's settings: decline analytics and marketing cookies, turn off ad personalization and past chats, delete your ads data, and switch off Marketing Privacy. If you use ChatGPT for anything you would not want tied to your shopping history, go Ads-Free or pay for Plus, the only options where the ad system does not exist for your account at all.
And keep the bigger lesson in mind. Ads arrived in ChatGPT in February. The cross site cookie was spotted seven months later by one researcher with a phone and a packet capture, not by OpenAI announcing it. Every AI assistant that adds an ad business will face the same temptation, so the browser level fix is the one worth learning. It works no matter who tries it next.
Comments
Post a Comment