What ChatGPT actually knows about you

Two sources, one answer

The first source is training data — a snapshot of public text taken before the model shipped. It is frozen. Nothing you publish today changes it, and nothing you delete today removes it.

The second is a live web search the assistant runs while answering. ChatGPT search, Claude, and Perplexity all fetch pages at question time and cite them. That layer follows the public web within days or weeks, which is the layer you can actually move.

The same question can hit either one. An answer with links came from the live search. An answer with no links, delivered confidently, came from memory — and that is the answer most likely to be years out of date.

Why every assistant tells you something different

Each assistant reads the web through its own crawler and its own index. OpenAI fetches with OAI-SearchBot and ChatGPT-User, Anthropic with ClaudeBot, and each keeps a separate index with separate coverage.

So disagreement between two assistants is normal and tells you something useful: a claim only one of them makes is usually sitting on a single source that only that crawler reached. A claim all of them make is on a page that everyone indexed — which is the one worth fixing first.

Agreement is not proof. Two models repeating the same wrong line usually means they read the same wrong page, not that the line is true.

How to test it without fooling yourself

Asking your own ChatGPT account about yourself is the least reliable version of this test. Personalization and saved memory feed your own words back to you, and a chat that already knows your name answers differently from one that does not.

Run it clean instead: a logged-out or temporary chat, memory off, the exact same prompt in each assistant, three times each. Write down the claim, the assistant, and the URL it cited. A claim with no citation is a memory claim; note it as such.

Then repeat it monthly. One reading is an anecdote. The same prompt over time is the only thing that shows whether a correction actually landed.

What this site does with that

OpenProfiles runs the identifier you enter through OpenAI and Claude as separate researchers, plus Exa and Perplexity people search, and shows you each source's answer next to the URL it came from.

A source that was not called stays visible as not called. "We did not ask" and "we asked and found nothing" are different facts, and collapsing them is how audits start lying.

Does ChatGPT store a profile of me?

No. There is no per-person record to open or delete. Claims about you are reconstructed at answer time from training data and, for questions that trigger it, a live web search.

Why does ChatGPT know things about me that are not on my LinkedIn?

Because the crawlers read far more than the profiles you maintain — conference pages, GitHub, press mentions, aggregator sites that republished an old bio. The claim usually traces to one page you forgot about.

Why does it get me confused with someone else?

Name collision. Assistants match on a string and rarely have enough evidence to separate two people with the same name, which is why every claim should be judged against its cited source rather than the summary.

How to remove your information from AI answers

You cannot edit a model. You edit the page it read, then the index that cached it. Here is the order the layers actually come off in, and how long each one takes.

How to correct wrong information AI says about you

Most bad AI claims are not lies, they are stale facts. Correcting is a different job from removing, and it needs a better source rather than a deleted one.

A repeatable AI visibility check

A ten-minute audit you can run monthly on a person or a brand, and the four pieces of evidence worth writing down each time.