How AI assistants decide what to say about a person.
When someone asks an assistant who you are, four steps happen in a few seconds. Knowing them tells you where an answer comes from and what you can do about it.
First, the model checks what it already holds.
Each assistant is a language model trained on a large body of text up to a cutoff date. If your name appeared often in that text, the model holds a rough picture of you. A profession, a place, a few associations. If it appeared rarely, the model holds nearly nothing. It may fill the gap with the nearest match, which is often a namesake. Most individual professionals are likely in the second group. It is a count of pages, and no judgment of you.
Second, it decides whether to search.
With web search on, the assistant decides whether the question needs fresh pages. A question with a name, a profession, and a city usually triggers a search. The assistant writes one or more search queries, often close to the question itself, and reads the top results. Which results come back depends on the search index it uses. ChatGPT, Gemini, Perplexity, and Claude can read different pages for the same question.
Third, it reads a handful of pages, fast.
The assistant does not read your whole site. It reads the opening of a few pages, favoring ones that state the answer early. A directory profile with your name, profession, and city in its first line can outweigh your own About page. That happens when your About page opens with a story. If two pages disagree, the assistant may pick one, blend them, or report the disagreement. It rarely says which page it trusted.
Fourth, it writes the answer and cites what it read.
The answer is composed from the pages it read and the picture it already held. The citations tell you which pages did the work. When your own site is not among them, the assistant described you from somebody else's pages. That is the finding behind the Sources score in a reading.
Why namesakes get mixed up.
Nothing in the process checks identity. Suppose a page about another person with your name states a profession and a city clearly, and your pages do not. The assistant has no way to know it has the wrong person. A disambiguation line on your own pages, naming your city and firm and noting the namesake, gives it that way.
Why the same question gets different answers.
The search can return different pages from one minute to the next. The model samples its words with some randomness. The assistant may search on one try and answer from memory on the next. A reading asks each question three times to show that spread rather than hide it.
What you can change, and what you cannot.
You can change the pages. One consistent bio sentence. An About page that states the answer in its first lines. Person schema that links your profiles. Current directory profiles. A few mentions in publications the assistants read. Those steps improve the odds that the next search reads your pages first.
You cannot change the model, the search index, or the randomness. No one outside those companies controls what an assistant says, and even they do not pick each answer. A reading is a dated record. The work is in the sources.