How does ChatGPT decide which businesses to recommend?

Ask ChatGPT for a good roofer in York and it will name two or three. The question I hear next, on nearly every call where this comes up, is some version of: how did it choose? Who decided my competitor gets named and I do not?

The honest answer is that nobody decided. There is no list, no ranking, and no one at an AI company scoring the roofers of York County. There is a machine doing four things in a row, in a few seconds — and every one of the four is something you can influence. I will say ChatGPT throughout because that is the one people ask about, but the other assistants work the same way.

First, it decides whether to look anything up

An assistant can answer from what it absorbed during training, or it can search the web right now and read what it finds. For a local recommendation it almost always searches, because “good roofer near me” is exactly the kind of question where stale knowledge embarrasses it. I have written about those two mechanisms before — the short version is that the live search is the part that matters, and the part you can reach.

Then it writes its own searches

It does not pass your customer’s question along word for word. It turns “who should I call about a leaking roof in York, they need to be able to come out fast” into two or three ordinary search queries — roof repair York PA, emergency roofer York reviews — and runs them.

Two things follow from that. The words on your pages should be the words people actually use on the phone, because those are the words the machine searches with. And underneath every assistant there is still a plain search engine deciding what comes back — not always Google’s. Different assistants lean on different indexes, some on Bing’s, some on their own crawlers, and they do not announce which on any given day. If you have only ever checked how you look on Google, spend ten minutes seeing whether Bing knows you exist.

Then it reads a handful of pages

Not the whole web. Not even the whole first page of results. A handful — directory pages, review profiles, a local paper’s “readers’ choice” piece, your site, your competitors’ sites.

This is the gate most businesses fall at, and it is worth being blunt about: if ordinary search does not surface you, the assistant never meets you. Most of what gets sold as AI visibility is decided here, by the boring, established mechanics of search ranking.

Then it looks for names it can trust

From those pages it pulls out candidate names, each with whatever facts came attached — services, hours, reviews, the neighbourhoods they serve. Then it cross-checks. A business that appears in three places, with the same name and phone number and roughly the same story in each, reads as real. A business that exists only on its own website has nothing backing it up. A business whose hours disagree between two directories gets hedged or quietly dropped, because the machine’s job is to sound reliable, and contradictions are how it ends up wrong.

Reviews carry weight here for the same reason — not because the machine admires you, but because a steady stream of recent reviews is hard-to-fake evidence that a business is open and functioning.

Then it writes — and the writing has habits

The answer has to be a useful sentence, not a list of links. So the machine favours businesses it can say something about: “does emergency repairs, mentioned by the York paper, well reviewed.” A page that states its facts in plain text hands the machine its sentence ready-made. A page whose hours live inside a graphic, or whose services only appear after a script runs, gives it nothing to lift, and it moves on to a competitor who was easier to quote.

Why the answer changes between Tuesday and Thursday

Ask twice and you may get different names. Different searches get run, pages change, and the same model answers the same question differently on different days. That is not a fault you can fix — it is the reason nobody can sell you “position one in ChatGPT,” and the reason checking once tells you little. Ask the assistants what your customers would ask every month or two, from a fresh session, and write down what comes back. That is the whole measurement, as of August 2026.

What the four steps add up to

Each step is an instruction. Be findable in ordinary search, on more engines than one. Be corroborated somewhere that is not your own site. Keep your name, hours and phone number in agreement everywhere they appear. Keep the reviews coming. And state what you do in plain sentences a machine can lift. The trade has a name for courting these systems — GEO, generative engine optimisation — but the name matters less than the mechanism, and the working list is already written out.

The short version

ChatGPT does not consult a ranking, because there isn’t one. It searches, reads a few pages, keeps the names it can corroborate, and writes about the ones it can quote. Every step favours a business that is findable, consistent, reviewed, and written in plain text.

The free site check reads your site the way one of these machines does and tells you what it could and could not find. Or tell me your trade and your town and I will tell you which of the four steps you are losing at.