A post sat on the Hacker News front page yesterday: "ChatGPT now knows what you do on other websites via ad collector". 446 points, 246 comments, a number that qualifies as a phenomenon in a community where the median story rarely clears 50.
The short version: ChatGPT can now see which sites you visit, not just what you type into the prompt box.
What Happened
The reporting points to a specific mechanism: the ad attribution SDK, a toolkit advertisers use to track what happens after someone clicks an ad. This category is not ChatGPT's invention. Streaming platforms like Roku and FuboTV have run on it for years, all chasing the same question:
"I bought an ad. Did the user actually convert somewhere else?"
The classic flow: a user clicks an ad on site A, site A's SDK logs it, the user lands on site B, site B runs the same SDK, and the two ends get stitched together.
OpenAI has now plugged into that network. The SDK sits inside the ChatGPT web app or desktop client, and it can collect:
- Timestamps and device info for each ChatGPT visit
- Behavioral signals from other sites running the same SDK on the same browser or device
- A stitched-together user profile spanning multiple domains
A commenter on HN posted reverse-engineering notes claiming ChatGPT's web client loads scripts like rta-collect.js, tied to a domain belonging to the ad tech firm The Trade Desk. That detail remains unverified, and the thread treats it with appropriate skepticism.
The specific company is not the point. The point is that the LLM has moved from "you ask, it answers" to a data collection endpoint.
How the Roku Model Actually Works
An attribution SDK runs on three primitives:
- Device fingerprinting and cookies. The browser or client generates a persistent unique ID that survives across sites.
- Event reporting. Every user action (ad impression, add to cart, purchase) fires a ping to the attribution platform.
- Data stitching. The attribution platform joins events from every site running the same SDK into a single user journey.
Roku pioneered this playbook. When it started selling OTT ads, it noticed a gap: people watched the ads, but the connection between viewing and buying was broken. Roku built its own attribution SDK, got every advertiser site and every post-click landing site to install it, and stitched the data back together.
FuboTV runs a similar setup, just in a more niche corner of live sports streaming.
At the code level it looks roughly like this (simplified; real SDKs are messier):
// Attribution SDK reporting user behavior
window.rtaCollect = window.rtaCollect || [];
window.rtaCollect.push({
event: "page_view",
site: "example.com",
userId: "abc-123",
referrer: document.referrer,
// ...more fields
});
// Attribution platform match
fetch("https://rta.example.com/collect", {
method: "POST",
body: JSON.stringify({
event: "page_view",
deviceId: "device-fingerprint-xyz",
timestamp: Date.now()
})
});
If the ChatGPT web client loads the same script, and an e-commerce site you visited earlier runs the same SDK, the two sides line up.
This is a categorically different data collection from anything ChatGPT has done before:
- Chat context: you volunteer it.
- Training data usage: there's a toggle (Settings → Data Controls).
- Cross-site behavior: you never consented, and most users don't know it's happening.
Why This Is a Watershed for AI Assistants
Old ChatGPT was an information island. You said something, it answered, and it had no idea what you were doing on a Saturday night.
Now it knows. It knows you searched for flights on Expedia last night, browsed used cars on AutoTrader this morning, and have been lurking in a particular subreddit all week. That data flows in from the ad attribution network straight into OpenAI's reach.
The implications are obvious: the monetization path is welded shut.
OpenAI's current business model rests on two legs:
- Subscriptions (ChatGPT Plus, Pro, enterprise tiers)
- APIs (developers paying for inference)
Subscription growth is flattening. APIs are under pressure from open-weight models (DeepSeek, Qwen, Llama are all pushing in). Where does the next dollar come from? Ads are the obvious answer. But "what the user asked" is not enough. Advertisers want "what the user wants to buy."
Cross-site behavior data fills exactly that gap.
Someone will say: "OpenAI wouldn't do that. It goes against their values."
I don't buy it. Here's why:
- Wiring in an SDK is a few lines of code. Marginal cost is close to zero.
- Attribution data directly lifts ad CPM. Revenue is immediate.
- Even if internal voices object, the ad team's KPIs tend to win the argument.
This is not an "if" question. It's a "when" question.
What Can Users Actually Do
A few options that exist today:
1. Turn off OpenAI's data sharing.
ChatGPT Settings → Data Controls → toggle off "Improve model for everyone." But this affects training data, not cross-site tracking.
2. Disable browser ad identifiers.
- Chrome: Settings → Privacy and security → Ad privacy → Site-suggested ads
- Safari: Settings → Privacy → Hide IP address from trackers
- Firefox: Settings → Privacy & Security → Enhanced Tracking Protection (Strict is on by default)
3. Use Brave or Tor Browser.
Brave blocks third-party tracking scripts by default. Tor is more aggressive, but ChatGPT frequently throws errors.
4. Inspect ChatGPT's network requests.
Open DevTools (F12) → Network tab → filter for rta, collect, attribution. Strange-looking domains are almost certainly attribution SDK traffic.
5. Use the ChatGPT desktop app instead of the web client.
The desktop app runs outside the browser context, so cross-site stitching is harder. Whether OpenAI ships an SDK into the desktop client next is anyone's guess.
The honest answer: ordinary users will never know this is happening. People reading the HN thread already have technical backgrounds. The median ChatGPT user doesn't know what DevTools is.
Editorial Take
This deserves more alarm than the usual "AI company behaving badly" story, because it is not a data leak. It is a connection to an entire privacy-violating infrastructure that has been running for over a decade.
Over the last ten years, Google, Meta, and Amazon built the most granular user-tracking apparatus in history through ad attribution networks. GDPR and CCPA have tried to regulate the practice, but attribution SDKs have always lived in a gray zone, technically not "selling personal data," just "sharing anonymous IDs."
LLMs now have access to that infrastructure for the first time.
My prediction (with a 30% margin of error): regulatory action within a year. The European Data Protection Board will almost certainly send inquiries. The FTC may follow. The reasoning: LLM training data and user prompts have already been classified as "sensitive" (Italy's Garante banned ChatGPT in 2023 before reversing course, setting the precedent). Feeding cross-site behavioral data into an LLM reads as an escalation in regulators' eyes.
What I'm less sure about is whether OpenAI did this deliberately or whether an SDK shipped by default through a third-party integration. If the latter, PR can absorb the hit. If the former, the framing changes entirely. The original report and the HN thread don't pin this down, so I lean toward "intentional" without being able to confirm it.
One last thing: this is not an OpenAI-only problem. Anthropic's Claude, Google's Gemini, and Perplexity are all weighing or already experimenting with ad-based monetization. The day LLMs fully plug into the ad attribution network, the entire AI assistant industry becomes the next generation of super-trackers.