ChatGPT now knows what you do on other websites via ad collector
TL;DR Highlight
OpenAI's ad pixel sets an __obi cookie that links activity on other sites to your ChatGPT account
Who Should Read
Web or backend developers who embed ad/conversion pixels on their sites or are evaluating ChatGPT ads
Core Mechanics
- __obi cookie is tied to your account
- ChatGPT generates 16 random bytes and puts them in a 60-second signed token (JWT) alongside the account.
- bzr.openai.com then sets it as a .openai.com cookie lasting one year.
- The cookie flows out via advertiser sites
- Advertisers install OpenAI code on their sites, which sends __obi to OpenAI along with page data.
- The author says this includes products searched, articles read and purchase behavior.
- Just loading the script leaks it
- The SDK has a no-credentials path, but it doesn't help.
- The browser attaches cookies to the request that loads the SDK script (oaiq.min.js).
- It scrapes identity from forms and pages
- In observed traffic, scraped identity events numbered 685 versus 255 supplied by advertisers.
- Email, phone and names are SHA-256 hashed, while country, region, city and postal code go in the clear.
Community Reactions
- Browser protection info was shared
- Commenters said Firefox, Brave and Safari block this while Chrome and Edge do not.
- Another said strict cookie partitioning stops the link between your account and ad views.
- Users reported DNS blocking
- One commenter's DNS server with Hagezi blocklists returns NXDOMAIN for bzr.openai.com.
- Others suggested uBlock Origin plus Privacy Badger.
- Many said the tech is old but the context is different
- Some noted it's the same standard adtech Facebook has used.
- Others found it worse because people share more private things with chatbots and pay for subscriptions.
- Spread of ChatGPT trackers was noted
- One comment cited Bloomberry data saying sites with ChatGPT ad trackers are on track to double from last month.
- Another linked an AdExchanger report saying this has existed since April.
How to Apply
- Check for __obi in the network tab
- Today, open DevTools on an advertiser site and filter requests to bzrcdn.openai.com and bzr.openai.com.
- Check whether the request headers carry the __obi cookie.
- Audit third-party pixels on your site
- Search your tag manager for the OpenAI pixel (oaiq) and check the automatic matching setting in Ads Manager.
- Turn off automatic identity collection if you don't need it.
- Exclude the pixel from sensitive paths
- Don't load the pixel on pages like medical, lending or legal intake whose URL path alone reveals sensitive info.
- The article says origin plus path is sent and such paths were collected.
- Update consent wording and cookie policy
- Classify the OpenAI pixel as marketing, not analytics, in your consent banner.
- Tell visitors their activity may be linkable to a ChatGPT account.
Try It
# Paste in the browser console (F12): a quick check for whether the page loaded the OpenAI pixel
performance.getEntriesByType('resource').map(r => r.name).filter(n => /openai\.com/.test(n)).forEach(n => console.log(n));
// If bzrcdn.openai.com/sdk/oaiq.min.js or bzr.openai.com/v1/sdk/events appears, the pixel loaded.
// Next: in the Network tab, check whether the Cookie header on those requests has __obi. (It is HttpOnly, so document.cookie won't show it)Terminology
Related Papers
The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge
Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.
90% of US working hours are spent on tasks not covered well by AI benchmarks
Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage
Postgres 19 delayed after AI reviewers find more bugs than humans can fix
LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted
Show HN: Gemini 4 Argon plays StarCraft
LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.
A Plausible Explanation Is Enough to Mislead Multimodal Judges
Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.
Neural networks are clean
Looking at neurons directly shows most have one main job, challenging superposition.