How AI engines choose the companies they name

Ask an AI engine which accounting firm to use in your city and you get three names. Not twenty links to sort through. Three names, inside a sentence, sometimes with two sources underneath.

Then comes the question everybody asks: why those three.

The engine does not search, it retrieves and then writes

A classic search engine hands you an index. An answer engine does something else: it fetches a few documents, reads them, and writes a text out of what it found there.

That retrieval step is the first filter, and it is the harshest one. The engine does not read fifty pages, it reads a handful. If your site is not in that handful, nothing downstream applies to you. You do not exist for that particular answer, however good your page is.

What gets into the handful depends on the engine. Perplexity and Google AI Mode run a real web search at question time. ChatGPT sometimes does, depending on the wording and the model. Either way, ranking well on the matching query helps. It is not sufficient, but it helps, and that is the only good news for anyone who has already done search work.

What is written about you elsewhere counts more than what is on your own site

This is the second mechanism, and the one that surprises clients most.

When an engine has to name three companies, it leans on pages that already name several. A comparison article, a sector directory, a trade publication, a forum thread, a "best X in Y" page. Those pages are a gift to a model: the answer is already pre-chewed into a list.

Your own site only talks about you. It asserts that you are good. Every home page on earth asserts the same thing, and the model knows it.

The practical consequence: a company can have an excellent site and never be named, while a competitor with a mediocre one comes up every time because it sits in three directories and two articles. I have seen that often enough to stop treating it as an exception.

It is also what makes the source list useful. In a Scan, the column of cited domains is often worth more than the score itself: it tells you where the game is actually being played.

The model also has a memory, and it is old

Third mechanism: part of the answer comes from no retrieved page at all. It comes from training.

If your brand was already well known when the model was trained, it can surface without a single source being consulted. That is why the big names appear everywhere, including where they are not relevant, and why a company founded last year starts with a handicap that has nothing to do with merit.

You cannot act on that memory. You can act on retrieval. In practice, that is where all the work ends up.

What it changes about the method

If those three mechanisms hold, a few consequences follow, and they are not the ones you hear at conferences.

Working on your own site stays necessary. A page that does not plainly state what trade you are in and where is a page a model can do nothing with. That is the baseline, not the goal.

Being present where the lists get made matters more than winning a position on one query. Serious sector directories, trade press, comparison pieces, associations, partners who mention you. Each of those pages is a place the engine can find you at the moment it composes its answer.

And you have to measure, because none of this is visible otherwise. On a search engine, being invisible is visible: a results page, a rank, a console that tells you so. In an AI answer there is nothing to look at. Three names, two sources, and the conversation moves on without you.

A measurement, not a score

Last point, and it matters if you want to avoid fooling yourself.

Ask the same question again five minutes later: the wording changes, sometimes the names change too. An answer engine is not deterministic. A single reading is therefore worth very little, and anyone selling you an "AI visibility score" that holds steady to the decimal is selling you an illusion.

What makes sense is repetition. The same set of questions, to the same engines, at a regular interval, and you watch the trend. Less spectacular than a single number. It is the only thing that holds.

One figure to close, measured on one of the agency's client accounts: traffic arriving from AI engines converted at 12%, against 4.2% for search traffic. One account, one sector, so take it for what it is. But it explains why the question is worth attention now rather than in two years.