What an AI assistant says when someone asks about you

Ask a colleague to look someone up before a first meeting now and there is a reasonable chance they open a chat window instead of a search bar. "Who is [name], what do they do, anything I should know before the call" — typed into an assistant the way a search query used to be typed into Google. The assistant answers in a paragraph, with confidence, and the person reading it rarely checks the sources, if sources are even shown. This is already happening before interviews and before first calls with a supplier, and there is, at the moment, no established way to manage what gets said.

That is worth sitting with before offering any advice, because most of what gets written about this treats it as search-with-extra-steps: build a good page, get it linked, watch your reputation improve. Some of that carries over, and this piece explains why. But the honest starting position is that this is a genuinely new problem, badly served by existing playbooks, and it fails in a different way than search results do.

Where the answer actually comes from

An AI assistant answering "who is this person" is doing one of three things, and from the outside you generally cannot tell which.

The first is training data. The model was trained on a large snapshot of text from the internet up to some cutoff, and if your name appeared often enough in that snapshot — a company bio, a conference listing, a news mention, a GitHub profile — the model has some compressed, statistical trace of that in its weights. Ask it about you and it produces a plausible continuation based on that trace. There is no lookup involved; it is closer to the model recalling a pattern than reading a page.

The second is retrieval. A growing number of assistants, asked about a specific person, run a live web search first and summarise what came back, sometimes showing sources, sometimes not. This behaves much more like search, because the answer changes as the web changes and a page published last week can show up in it.

The third, and the one that actually explains most of what you see, is both at once: the model reasons from its training-data impression of your name while also folding in whatever a live search just returned, and it does not clearly separate the two in the answer it gives you. A sentence that reads as one coherent fact can be training data and retrieval spliced together without a seam. This is why the same question, asked of the same assistant a month apart, can come back different — not because you changed, but because the live half of the answer changed while the trained-in half did not, and you cannot see which words came from which source.

The practical upshot: you cannot fully separate what the model already believed about you from what it just read about you, and neither can the person asking. Both inputs are doing work in whatever answer they get.

The failure that matters most: being someone else

If you only remember one thing from this article, it should be this. The dominant failure mode of asking an assistant about a person is not that it says something false about you. It is that it confidently describes a different person who shares your name, and states it as if there were only one of you.

This happens for a mechanical reason. A search engine, imperfectly, keeps results as discrete pages you can tell apart — two different links, two different snippets, and you do the disambiguation yourself. A language model is not retrieving discrete pages when it answers; it is predicting the most statistically likely description of "a person named X" from everything it has absorbed about that string of characters, and if two people share the string, the model can blend attributes from both into one description that belongs to neither. The result reads as a single coherent biography, unhedged, because the model has no signal telling it to hedge — from its perspective there was one entity called your name, and it described that entity.

This is worse than the classic search-results problem of "someone with my name outranks me," which the wider question of what shows up when someone searches your name already deals with. Outranking is at least visible and honest: you see the other person's page and you know it isn't you. A conflated AI answer looks like one true paragraph about one person, and unless the reader already knows enough to spot the seam — a wrong job title, a city you've never lived in, an award that belongs to someone else — they walk away with a confidently delivered wrong impression and no reason to doubt it.

Test it yourself, and expect it to move

There is no substitute for doing this yourself, and it takes about five minutes across two or three assistants. Ask a plain version of the question a colleague might ask — your name, plus enough context to be specific if your name is common — and read the answer as a stranger would, not as the subject who already knows what's true and what isn't.

Two things will likely surprise you. First, how confidently wrong or half-wrong an answer can sound even when parts of it are accurate — the fluency of the sentence gives no signal about its reliability. Second, how unstable the results are: ask the same question of the same assistant a week later, or ask a second assistant the identical question, and you can get a noticeably different answer, sometimes describing entirely different work history or location. That instability is itself informative. It tells you the model has no settled, verified fact about you sitting in one place — it reconstructs an answer each time from whatever combination of training-data trace and live search happened to be available at that moment, closer to how a rumour shifts with each retelling than to a database record staying fixed.

Do this the way you would run a name search audit — periodically, not once, written down so you can compare — and treat it as diagnostic rather than something you're trying to "pass." The step-by-step search audit is the closest existing model for running this kind of check without it becoming a daily compulsion.

Why the same fix that helps search also helps here

The genuinely useful piece of continuity with ordinary search is this: consistency and outbound linking help an assistant the same way they help a search engine, and for a related reason. When an assistant retrieves live pages, it is looking, in effect, for signals that resolve which fragments of text on the web belong to the same entity — matching name, matching self-description, pages that link to and confirm each other. A single page stating plainly who you are, what you do, and where, linked out to and from your other profiles, gives both search engines and retrieval-based assistants exactly the disambiguating evidence they are built to use. It does not fix the training-data half of the problem — you cannot retroactively edit what a model absorbed before some past cutoff — but it gives the live-retrieval half something accurate to draw on instead of a blended guess. A page that functions as the anchor other mentions of your name point back to is the same asset described in the piece on using a personal site as the anchor for your name, and none of that work is wasted here even though the underlying mechanism is different.

What this cannot do is guarantee anything. A model's trained-in impression of your name, formed from whatever was in its training snapshot, sits underneath any live retrieval and can still surface regardless of how good your current page is. This is not a lever you get to pull. It is closer to sediment than to a setting.

There is no correction mechanism, and that changes what's worth doing

Search engines, for all their faults, at least have a request-a-review process, a way to report a specific harmful page, something you can point at and ask to be reconsidered. AI assistants, as of now, have nothing equivalent for "the biography you just gave a stranger about me is wrong." You cannot flag an individual answer and expect it corrected before the next person asks the same question, and you cannot appeal to a training-data snapshot already baked into a released model. Feedback mechanisms exist in most of these products in some generic form, but there is no visible path from "a user reported this" to "this specific error about this specific person is now fixed," and no timeline for the ones that might quietly get addressed in a future model version.

Practically, this means most people are simply not in the training data at all, in which case the model has nothing to hallucinate about them specifically and tends to say so or say very little — the safer position to be in, in this narrow sense. For the people who are in it, and especially for anyone who shares a name with someone more visible, the honest plan is not "get the model fixed." It is to know what it currently says, keep the accurate, consistent version of your information visible enough that retrieval finds it when it looks, and accept that the trained-in half is out of your hands.

The one case worth treating differently is when what's being said is not a stale fact or a blended biography but something that actually causes harm — a specific negative claim, a legal matter, or a serious accusation attributed to you that belongs to someone else entirely. That is not a search-optimisation problem and no amount of publishing consistent pages resolves it on any useful timeframe. At that point the right move is to stop trying to out-rank or out-signal the model and instead raise it directly, in writing, with the company that operates the specific assistant, describing the exact query and the exact answer. It is slow and unglamorous, with no guaranteed outcome, but it is the only route that addresses the actual claim rather than trying to drown it in better material — which works for a stale job title and does nothing for a false accusation.