Someone is going to ask an AI assistant about you this week: a prospective client checking before a call, a conference organizer, a journalist, a hiring manager. The answer they get is assembled from whatever the machine can find and trust. Most people have never looked at it. This is the test, it takes about fifteen minutes, and the result tells you exactly what to fix.
The test
- Open five assistants, signed out where you can. ChatGPT with web search on if your plan offers it, Claude, Gemini, Perplexity, and Google itself, looking for an AI Overview at the top of the results and, where available, AI Mode. A fresh session matters because your own history shapes what some of them say.
- Ask each the same three questions. "Who is [your full name]?" Then "What is [your full name] known for?" Then "Who is [your full name] at [your company]?" Use the exact name form you use everywhere else. Do not help the assistant with extra context; a stranger would not.
- Screenshot every answer, fifteen in all. Put them in a folder named with today's date. You will run this again in ninety days and the comparison is the whole point.
- Ask one more: "What are the sources for that?" An assistant that cites your own site, your Wikidata entry and a couple of interviews is reading your record. One that cites a directory you have never heard of has just told you which page to fix first.
How to score the answers
Each answer lands in one of five buckets.
Empty. "I don't have information about that person." The machine has nothing it trusts.
Wrong person. It describes a namesake: the dentist, the footballer, the person with your name in another country. This is the most damaging result because it is confident and incorrect.
Outdated. It has you, but at a job you left, with a book from years ago and nothing recent. Usually traced to one stale third party bio that outranks everything else.
Thin. It names you correctly but with a sentence or two and no specifics. It knows you exist, not what you do.
Correct. It names your role, your company, your work and recent appearances, and it cites sources, ideally your own site.
Note which bucket each of the fifteen answers falls in. The pattern across tools is what matters, not any single answer, because the assistants that search the live web will reflect what is online today, while the ones answering from training data reflect what was online months ago.
Why each failure happens
Empty answers mean there is no canonical page about you that the machine can find and trust. Your facts live in a LinkedIn it cannot read and a company bio it does not connect to you.
Wrong person answers mean the namesake has a stronger record than you do. They have a Wikipedia page, or a decade of consistent press, and you have a profile and a few mentions. The machine resolved the name to the better documented entity.
Outdated answers mean one old source is winning. A conference bio, a directory listing, an author page from a previous book, something that states your facts plainly and has been sitting there for years.
Thin answers mean you have a home page but it does not say much, or it says things the machine cannot verify anywhere else.
The fix for each
For empty and thin: build one page on your own domain that states every fact about you in plain sentences, with Person schema underneath it and a sameAs list of your profiles. That is the page every assistant will read first once it exists.
For wrong person: the same page, plus disambiguation. State the facts that separate you from the namesake: city, company, field, middle name if you use it. Make sure every profile uses the identical name form and the identical photo. Then add corroboration from third parties, interviews and articles, because the namesake is winning on volume and you need enough consistent sources to tip the resolution your way.
For outdated: find the stale source. Search your name plus the old job title and it will surface. Email the site and ask for the current bio with a link to your home page. Most will do it. Then make sure your home page and your profiles all state the current facts, so the old page is outvoted even if it never changes.
For correct: keep it that way. Add new appearances and work to the page as they happen, and run the test quarterly.
What to expect in ninety days
Assistants that search the web live, which includes Perplexity, Google's AI features and ChatGPT with search on, will reflect a new canonical page within weeks of it being indexed. Assistants answering from training data will catch up when their models are refreshed, which can take months. So the second run will usually show improvement on some tools and none on others. That is normal. The ones that lag are reading the web as it was; the ones that improve are reading the record you built.