How do AI assistants choose which businesses to recommend?

For most questions about businesses, the assistant searches the web, reads a handful of pages, and names the businesses those pages name. So you are mostly recommended because other pages recommend you. Your own site matters as the place it confirms what you do — but the shortlist is largely assembled from what others have written.

Two ways an assistant knows things

Training. Everything the model absorbed before it was released. This is fixed until the next version and fades for anything recent or local. A business that opened last year may simply not be in it.

Live retrieval. For questions about current, local or commercial things, assistants increasingly search the web at the moment you ask, read the top results, and write an answer from them — often citing the pages they used. Perplexity does this for nearly everything. The others do it more and more.

For "recommend me a…" questions, retrieval dominates. That makes the answer traceable: you can see which pages it read.

What 142 citations show

We asked ChatGPT, Perplexity, Gemini and Claude three questions about tools that check AI visibility — a category we know well — and recorded every page they cited. Twelve answers produced 142 citations. They fell into two groups:

• Product pages for the tools themselves, especially ones with a page dedicated to exactly the thing being asked about.
• Guide articles — and strikingly, many were from small specialist UK sites, not large brands. The most-cited guides had the question in their URL, word for word: "check if AI recommends your business", "how to make ChatGPT recommend your business".

The citations were not going to whoever had the biggest name. They were going to whoever had written a page that answered the question directly.

Why the four assistants disagree

They search with different engines, read different pages, weigh them differently, and were trained on different data. The same question regularly produces four different shortlists. In our measurements a business named every time by one assistant can be named never by another — which is why checking one tells you about that one.

Why the same assistant disagrees with itself

Answers are generated, not looked up. Ask the same assistant the same question twice and the shortlist can change; we measured this at about 4.7% of answers. A pattern across questions is real. One answer is not.

What this means for a business

1. You are recommended because others recommend you. Directories, guides, reviews and lists in your category are where the shortlist comes from.
2. Your site confirms rather than introduces. It is where the assistant checks what you do and charge — so it has to be readable by a machine.
3. Specific questions favour specific pages. A page answering exactly what a customer asks competes for that question in a way a general homepage never does.

Questions people also ask

Does ChatGPT use Google to find businesses?

When ChatGPT searches the web it uses a search index and then reads the pages returned. The exact provider varies and changes; what matters for you is that it reads ordinary web pages.

Do AI assistants read reviews?

Yes, when they appear on pages they retrieve. Review sites are structured and comparable, which makes them useful to an assistant building a shortlist.

Can I see which pages an assistant used?

Often. Perplexity shows its sources by default, and ChatGPT and others show citations when they search. Those links are the pages shaping the answer about your category.

Is it the same for every country?

No. Assistants answer for the person asking, so location changes the pages they read and the businesses they name.