What ChatGPT, Gemini, Perplexity and Claude say when somebody asks about you

Why this is suddenly worth asking

Somebody deciding whether to take your call used to type your name into a search box. A growing number now ask an assistant instead, and an assistant does not hand back ten results for them to judge. It hands back a paragraph.

That paragraph is assembled from whatever the assistant can find and is willing to rely on. You cannot see which sources it leaned on unless you go looking, and the person reading it certainly cannot.

So the useful question is not whether the answer is flattering. It is which sources produced it, because those are the only part of this anybody can do anything about.

What we measured, and how

On 24 September 2026 we put three questions about the same person to four assistants: who he is, whether he is credible, and what he is associated with. Three questions, four engines, twelve answers. The subject was our own founder, so there is no permission question and nobody else is named.

Three questions rather than one, because they fail differently. An assistant that confidently answers who somebody is will sometimes decline the credibility question entirely, and the one about associations is where another company with a similar name tends to arrive.

What the check records for each answer:

• Whether the person was named at all.
• Every source the answer leaned on, by host.
• Which of those hosts the subject actually controls, matched against addresses they gave us rather than against their name.
• Whether the engine answered, declined, or could not be reached, which are three different results and not two.

It costs about 34p a person for the full twelve answers. That is worth saying plainly, because the thing being sold elsewhere as an AI visibility subscription is in large part this measurement.

Twelve answers, four different pictures

All four named him, every time. That is the good half and it is the half most people expect.

The sources behind those answers were not remotely alike. One engine drew on twenty two distinct hosts. Another drew on exactly one, and that one was not a publication at all but an internal redirect belonging to the company that makes the assistant.

Same person, same question, same day. Twenty two sources against one.

Three of the four cited pages he owns, including the hosted page this business sells. Behind the rest were a public register, a journalist database, a social profile and a scatter of small titles. Nothing an assistant reads is a surprise once you can see it, which is the whole argument for looking. Where the answers come from goes through the sources properly.

The run that was not flattering

A company we checked at the end of September is the better illustration, and it is the reason this pillar is worth reading rather than skipping.

Three assistants answered. Between them they leaned on fifty one distinct sources, and the company's own website was not one of them. It was cited zero times, and that zero is real rather than a default: the check was run with the list of addresses the company controls, so there was something for it to match against.

Twenty of those citations belonged to seven other companies that share the first word of its name. One of the seven sells the same service it does. An assistant asked about this company could assemble a confident paragraph out of a competitor's material with nothing on the screen to flag it. When an assistant names the wrong company is the full case.

What can be changed, and what cannot

You cannot make an assistant say a particular sentence, and anybody selling you that is selling something they do not have. There is no control panel and there is no submission form.

What can be changed is what is available to be read. Every source we have seen cited falls into a small number of kinds, and most of them are things you can either edit directly or get corrected. The one that is yours outright is a page of your own, and on our own founder's run three of the four engines cited his. What you can actually change sets out which is which.

The work is also the same work that produces a knowledge panel, which is covered in what actually produces one. That is not a coincidence. Both are assembled from sources about you rather than from pages you wrote.

Two things we will not claim

We will not promise what any assistant will say about you next month. The models change, the engines change which sources they will rely on, and a reading is a reading on the day it was taken.

And we will not tell you an assistant ignored you on the strength of a check that was never given anything to match against. We made exactly that mistake in a draft in September, caught it before it went anywhere, and rewrote the method so the output is a list of what was cited rather than a claim about what was not.

If you want your own reading, the free audit includes one. How to check it yourself covers the free route first.