2026-07-22 · Jack Stovell
How does Claude know who a person is?
Ask an AI who you are. Go on. If you're not famous, one of three things happens: it's never heard of you, it hedges so hard the answer is useless, or, worst case, it confidently describes someone else with your name.
None of those are the model being stupid. They're the predictable result of how models learn about people. This guide was written with Claude, about Claude. Consider it a first-hand account.
Where a model's knowledge of you comes from
Two places. First: training. Models read a snapshot of the public web — Wikipedia, news, personal sites, GitHub, published work. If you exist there, a compressed version of you exists in the model. If you don't, you don't.
Second: live retrieval. When you ask ChatGPT or Claude about someone today, it usually searches the web and reads what comes back. This is where most 'who is X' answers actually come from now. And here's the part almost nobody has clocked: the page that ranks #1 for your name is probably useless to it.
Why your LinkedIn is invisible
LinkedIn sits behind a login wall. An AI search tool that hits your profile gets a sign-in page. The canonical professional profile of the last decade is unreadable by the exact systems people now ask instead of Google.
Same story for most social profiles, and for any site that only renders through JavaScript. Bots read plain HTML. If your page needs a browser to assemble itself, a crawler sees an empty shell.
What actually makes you legible
- Disambiguation first. The single most valuable thing you can give a model is proof of which same-named person you are: your site, your GitHub, your handles, all cross-linked. In schema terms, sameAs. Models fail by conflation far more often than by ignorance.
- A dense, dated summary. One paragraph: who you are, what you do, since when, where. Models digest that shape almost losslessly.
- Claims with dates and sources. 'CTO at Acme (2021–2024)' with a link beats three paragraphs of vibes. Uncited facts get hedged; cited facts get repeated.
- Freshness you can see. A visible last-updated date decides conflicts between old and new information.
- Ungated plain text. No login, no JS required, explicit welcome in robots.txt. Boring. Decisive.
What we built
This is exactly what a hilyt profile is: a stable URL that renders as real HTML for crawlers, schema.org Person markup with your sameAs graph, dated and sourced claims, and a plain-text llm.txt card any AI can read in one gulp. No wall. That's the whole trick.
Make yourself legible once, properly, and every AI that anyone asks about you reads from the same page. Yours.