How AI decides which local business to recommend

The short answer: AI assistants recommend businesses that are unambiguous, corroborated, and citable. When someone asks ChatGPT, Gemini, or Google's AI Overviews "who should I call about a burst pipe?", the model doesn't rank ten options — it names one or two. Those one or two are the businesses the model can clearly identify, verify across multiple sources it trusts, and quote from directly. Everything in this post is about becoming that business.

Where the answer actually comes from

Two places, and the distinction matters. First, the model's trained knowledge — what it absorbed about businesses, brands, and reputations during training. Second, live retrieval — most assistants now ground local answers by running searches behind the scenes and reading the results before responding. That means your visibility in an AI answer is part long-term reputation and part right-now search performance. You can't fully control the first without the second: the pages that get retrieved are, overwhelmingly, the pages that already rank and the profiles that already dominate local results.

This is why "SEO is dead" takes miss the mechanics. The assistant is a new front door, but the house behind it is still built from crawlable pages, structured data, and local signals. What's changed is the shape of the outcome: a results page had room for everyone; an answer has room for almost no one.

The five signals that decide it

1. Entity clarity. The model has to know, without guessing, exactly who you are: one business name, spelled one way, tied to one address, one phone number, one website, one set of services. Schema markup (LocalBusiness or your specific type, plus Organization and Person for the people behind it) turns your site from prose the model interprets into facts the model can state. Inconsistency is the silent killer here — if your name appears three slightly different ways across the web, you're three weak entities instead of one strong one.

2. Corroboration. Models weight claims they can verify across independent sources: your Google Business Profile, the major directories for your industry, local press, supplier and association pages, and reviews. One self-published claim is an assertion; the same fact echoed across ten sources the model reads is knowledge. This is old-school citation building with a new payoff.

3. Citable pages. When retrieval happens, the model quotes pages that answer plainly. A location page that says, in clear prose near the top, who you are, what you do, where you do it, and why you're trusted gives the model a sentence it can lift. Pages that bury the answer under sliders and slogans don't get quoted — they get skipped for a competitor who wrote it straight.

4. Review evidence. Assistants lean hard on review volume, recency, and specificity when recommending local services, because reviews are the closest thing to ground truth they have. A steady velocity of detailed reviews — ones that mention the service, the city, the situation — feeds the exact language models look for when matching a business to a question.

5. Structure. Clean site architecture is how everything above gets found. Location silos that map to how people actually search, internal links that concentrate authority on the pages that matter, and a crawlable, fast site determine whether your best evidence is even in the room when the answer gets written.

What to do about it, in order

• Audit your entity. Search your own business the way a stranger would — and ask the assistants directly. Note every inconsistency in name, address, categories, and services across your site, GBP, and top citations. Fix them everywhere.
• Ship the schema. LocalBusiness (or your specific subtype), Organization, and Person markup, wired together and matching your visible content exactly.
• Rewrite your key pages to be quotable. Top of page, plain language: who, what, where, why trusted. Then the detail.
• Build review velocity, not review bursts. A repeatable ask that produces steady, specific reviews beats a one-time campaign every time.
• Corroborate. Get the same facts about your business onto the directories, associations, and local sources your industry actually uses.
• Measure it. You can't manage what you check once. I use a fixed-panel benchmarking method — the same locked queries, the same assistants, every month — to turn AI visibility from a vibe into a trend line.

Common questions

Can you pay to be recommended? Not the recommendation itself. Assistants are experimenting with sponsored placements around answers, but the organic recommendation is drawn from what the model knows and retrieves. That's earned, not bought — which is bad news for deep pockets and good news for well-run businesses.

How long does it take? Months, not days. Entity signals compound the way reputation does. Some retrieval-driven gains show up faster when you fix quotability on pages that already rank, but durable AI visibility is built.

Does this replace SEO? No — it sits on top of it. The assistants ground themselves in search. Strong traditional SEO is the foundation; AEO is the layer that turns rankings into recommendations.