AI assistants choose businesses in two steps: they retrieve — running web searches and reading what comes back, your Google Business Profile, directories, review platforms, business websites — and then they synthesize, composing a short answer from the handful of sources they just read. A business gets named when it appears in those sources, the facts about it agree with each other, and its details are stated in forms machines parse reliably. It gets skipped when it’s absent, contradictory, or unreadable. What nobody outside the AI labs can tell you is the exact formula for who gets named first — and anyone claiming to have that formula is guessing.
This article walks through both steps in shop-owner language, because once you see the machinery, most “why did it recommend them?” mysteries turn into specific, fixable gaps.
Step one: retrieval
When a customer asks “who does ceramic tint near me,” the assistant doesn’t consult a stored directory of tint shops — it doesn’t have one. Instead it does a fast, automated version of what the customer would have done: it runs a few searches, opens a handful of results, and reads them.
What comes back is the assistant’s entire working knowledge for that answer. You don’t have to guess what those sources are — grounded assistants cite them, and the same kinds show up over and over: Google Business Profiles and the maps results built on them, review platforms, general and industry directories, local press and “best of” lists, and business websites. If you run the check yourself, your own citation trails will show you the exact reading list for your market.
The first filter is brutally simple: businesses in the retrieved sources can be in the answer; businesses that aren’t, can’t. Not as punishment — the assistant literally has nothing to say about a business it didn’t read about. Most absences from AI answers are retrieval problems, not quality-of-your-work problems.
Step two: synthesis
Now the assistant has, say, a profile, two directories, a review page, and a couple of websites open in front of it — and a customer waiting for a short answer. It composes one: names a business or two, says something about each, and stops.
Composing means choosing, and this is where source quality starts to matter as much as source presence:
- Fit to the question. The customer asked for weekend availability, or walk-ins, or a specific service. If a source states that fact about you, you can be the answer to that exact question. If no source anywhere states it, you can’t — the assistant won’t assume it.
- Something to say. An assistant recommending a business usually says why — well-reviewed, long-established, offers the thing asked about. Businesses whose sources are rich in specifics give it material. A bare listing with a name and a phone number gives it almost nothing to work with.
- Confidence. The assistant is, in effect, deciding how sure it is about each candidate. And that’s where consistency comes in.
Consistency: the quiet gatekeeper
Here’s the mechanic that surprises most business owners. An assistant reading multiple sources compares them — it has no other way to judge what’s true. When your sources agree, each one corroborates the others, and the assistant can state your facts plainly. When they conflict — two phone numbers, an old address still live on one directory, a duplicate profile, a slightly different business name in each place — the assistant can’t tell which version is right, or even whether it’s looking at one business or two.
A machine that isn’t sure which business you are can’t recommend you with confidence. In practice, low confidence tends to surface as hedged descriptions, wrong details repeated from whichever source it happened to trust — or as the safest option available to it: leaving you out and naming a competitor whose story checks out. Conflicting facts don’t just risk a wrong answer about you; they risk no answer about you.
This is why the unglamorous cleanup work — one exact name, one address, one phone number, everywhere; duplicates resolved; stale listings corrected at the source — moves AI answers more than any clever trick does.
Why structured data helps
Machines read prose worse than they read structure. Your hours stated in a styled paragraph, an image, or a PDF might be obvious to a human and invisible to a parser. Structured markup — schema.org — lets your website state its facts in a labeled, machine-readable form: this is the name, these are the hours, this is the phone, these are the services.
Structured data doesn’t make an assistant prefer you. What it does is remove ambiguity at the exact moment the assistant is deciding how confident it is in your facts — and hand it a clean, quotable version of the details it’s about to compose into an answer.
Citations: the part you can verify
Grounded assistants attach citations to their answers — the actual sources each answer was built from. Treat them as receipts. When an answer names a competitor, the citations show which sources earned it for them. When an answer gets your hours wrong, the trail usually leads to the one listing still holding the old hours. Every mystery about step one and step two above is partially inspectable through the citation trail, which is why we build our measurement around capturing it — and why our case write-ups trace findings source by source.
The part nobody can predict
Now the honest limit. Everything above describes the machinery; none of it is a formula. Assistants vary their answers — the same question asked twice can name different businesses. Which sources get retrieved shifts with wording, location, and day. How a model weighs one source against another isn’t published, differs across ChatGPT, Gemini, Perplexity, Claude, and Meta AI, and changes as the platforms update. Ranking inside an answer — who gets named first — is not fully predictable by anyone.
So the honest version of “how do I get recommended?” is not a trick, it’s a posture: be present in the sources assistants read, be consistent everywhere you appear, be specific about what you offer, be machine-readable — and then measure repeatedly, because single answers are anecdotes and nobody can guarantee a mention. What you can do is remove every reason an assistant currently has to skip you or get you wrong. In a local market where most businesses haven’t done that work, it’s a real edge — and an inspectable one.
If you want to see how the two steps play out for your business specifically, the free check shows you what three assistants answer today, which sources they cited, and the first gap we’d fix.