How AI Assistants Decide Which Local Businesses to Recommend

A customer asks ChatGPT for “a reliable emergency plumber in Leeds” or Gemini for “a dog groomer near Ballard that takes anxious dogs”, gets three names, and calls one. You were not one of them, and nothing in your Google Business Profile dashboard tells you why. This guide explains how the assistants actually arrive at those names, which sources they lean on, what wastes your time, and the work that gets a local business into the answer. By the end you will know how to measure it, starting with the free AI Visibility Checker, and what to fix first.

Two mechanisms, not one

Every assistant answers a local question in one of two ways, and the fix differs depending on which one is in play.

Model memory. A language model is trained on a large snapshot of the web up to a cutoff date. Businesses that were written about often before that date, in directories, review sites, press and “best of” lists, leave a trace in the model, and it can name them without looking anything up. The trace is fuzzy: the model recalls what was said about you more than what is true, confuses similar names, and can invent an address or a phone number with complete confidence. A business that opened after the cutoff, or was rarely written about, does not exist here.

Live retrieval, or grounding. The assistant runs a search at answer time, reads the results and writes from them. Now the question is not “does the model remember you” but “are you in the pages it retrieved, and do those pages say what the user asked about”. This is where most of the winnable work sits, because it changes when the sources change, not when a model is retrained.

Which mechanism each product uses, at the time of writing (these products change often):

Assistant	How it answers local questions
ChatGPT	Memory when it does not search; when search is on or triggered, live web results and map listings, with links to sources
Gemini	Grounded in Google Search and, for places, Google Maps data, so Business Profile facts appear directly
Perplexity	Live web search on every answer, with numbered citations you can open
Google AI Overviews and AI Mode	Built on Google’s Search index and Business Profiles; local answers often carry a local pack

The practical consequence: for Google’s own features your Business Profile is the source. For the others it is one step removed, reached through the directories, review platforms and articles that copy your profile data, plus whatever your own website says in plain text.

Why answers vary from one run to the next

Ask the same question twice and you can get two different lists. Language models generate text with deliberate randomness, so the wording and the order change even when the sources do not. The retrieved results change too: the search behind the answer can return a different page set for the same words, the assistant may decide not to search at all, and location signals from the account or device shift which businesses count as “near”. Model versions update quietly.

This is why one screenshot proves nothing, in either direction. A business that appears once is not “recommended by ChatGPT”, and one that is missing once is not invisible. Judge visibility the way you would judge a rank grid: several phrasings, repeated over time, and a pattern.

What the assistants consistently lean on

Read enough cited answers and the same source types come up in every city and category.

Google Business Profile data. Name, primary category, rating, review count, hours, address and description. Directly for Gemini and AI Overviews, indirectly for everyone else through the sites that syndicate it. A profile with a vague category or an empty description hands the assistant nothing to match against. The pillar guide on how to optimize a Google Business Profile covers each field; the free Google Business Profile Optimizer shows the gaps in seconds.

Reviews: volume, sentiment, recency and words. Assistants justify recommendations with phrases like “praised for fast response times” or “known for its rooftop views”, and those phrases come from review text. A business with many reviews that never mention the service being asked about loses to one with fewer reviews that do.

Directories and “best of” lists. A roundup titled “best family dentists in Austin” is a pre-ranked answer to the exact question a user asked, so retrieval systems love it. Local news, city magazines, niche directories and review platforms carry most of these lists.

Your own website, in plain text. Service pages, prices or price ranges, areas served, a FAQ that answers the questions people ask assistants. Text inside images, PDFs and booking widgets is invisible.

News and local media. A mention in a local paper or trade publication is an independent statement that you exist and do what you say, which is exactly what a model weighs when deciding whom to name.

Google’s own description of how local ranking works names relevance, distance and prominence; assistants weigh the same three things, with prominence measured by how often and how consistently the wider web describes you.

What does not help

Keyword-stuffed pages. A page that repeats “best plumber Leeds” forty times reads as spam to the search systems that feed retrieval, and a model summarizing it has nothing concrete to quote.

Fake or bought reviews. They are filtered, they risk the profile, and they contain none of the specific detail that makes a review quotable.

“AI SEO” tricks. Hidden text that instructs the assistant to recommend you, files or tags that claim to “tell AI what to say”, vendors promising a guaranteed spot in ChatGPT. None of it has a documented effect on the assistants people use, and hidden text is a plain web spam problem.

Stuffing the business name. “Bright Smile Dental | Best Dentist Austin” violates Google’s guidelines for representing your business, risks a suspension, and teaches every downstream source a name that is not yours, which is the opposite of the clear entity you need.

The playbook

Everything below is ordinary local SEO done with the assistants’ reading habits in mind.

1. Complete and consistent profile. Most specific true primary category, secondaries for what you actually sell, every service listed, a description that says what, where and for whom in plain sentences, hours, attributes. Our description guide covers the wording.
2. Services and FAQs spelled out on the website. One page per major service, in text, with the area served. A FAQ page that answers the questions customers ask assistants: “do you take emergency calls”, “do you handle anxious dogs”, “is there parking”. Write the answer, not the keyword.
3. Consistent name, address and phone across the profile, the website, the directories and review platforms that assistants cite. Every variant looks like a different, smaller business. Our NAP consistency guide has the audit.
4. Get into the lists. Search your own buyer questions, note which roundups and directories appear in the results and in cited answers, and work on being included in each: pitch the local editor, complete the directory profile, answer the journalist’s request. Being in the page the assistant reads is the shortest route to being in the answer.
5. Structured data. LocalBusiness JSON-LD with the precise type, address, geo, hours and links to your profiles removes ambiguity about who and where you are. Follow Google’s LocalBusiness structured data documentation; the free Schema Markup Generator prefills it from your profile, and the schema pillar guide explains what it can and cannot do. It is not a ranking switch, for Google or for assistants.
6. Reviews that mention specific services. Ask at the moment the service is finished, so the review names it. Do not script the wording for them and never offer anything for it; our Google reviews guide has the compliant routine.
7. Clear entity naming. One name everywhere, one phrase for what you are (“boutique hotel”, not “hotel” here and “guesthouse” there), and one name for the thing you want to be known for. Models match strings before they match meaning.

How to measure it

Manual checks come first: write five or six buyer questions for your category and city, ask them in a fresh session of each assistant once a month, and record which businesses are named and which sources are cited. It takes twenty minutes and it is the ground truth.

The free AI Visibility Checker automates a larger sample. Pick your business from the Google suggestions; it detects your category and city, generates twelve buyer-intent questions (“best…”, “emergency…”, “near me”, price, booking) and runs them through four engine personas modeled on ChatGPT, Gemini, Perplexity and Google AI. You get the share of answers that named you, mentions per engine, the competitors named instead with evidence-based reasons, an action plan, and a chat to ask follow-up questions.

Its limits, honestly: the personas are one underlying model instructed to behave like each assistant, not a live crawl of each product; twelve questions is a sample, not the universe; and the result is a snapshot that a different city, category or day changes. Use it for direction and trend, re-run monthly, and confirm what matters in the real assistants.

Worked example: a Lisbon hotel that never appears for “rooftop”

A fourteen-room boutique hotel in Alfama, call it Casa do Miradouro, had a rooftop terrace with a bar and a river view, strong reviews, and never appeared when anyone asked an assistant for “a boutique hotel in Lisbon with a rooftop”. The owner ran the checker: two mentions in twelve answers, both for generic “boutique hotel Alfama” questions, none for rooftop, and the competitors named instead were the hotels that appear in every “best rooftop bars in Lisbon” article.

The diagnosis took an hour of reading cited sources. The hotel’s website called the space “the terrace”, and the terrace page was a photo gallery with two lines of text. The Business Profile description did not mention it at all. Booking platforms listed “terrace” as an amenity. Guests wrote about “the view” but rarely the word “rooftop”, because nobody at the hotel used it. And the hotel was in none of the roundups the assistants cited; it had never been pitched.

The plan: rename the space “the rooftop terrace and bar” everywhere, on the website, in the profile description, on every booking platform and in staff vocabulary; rewrite the terrace page as text, with opening hours, what is served and whether non-guests are welcome; add Hotel schema with the amenity listed; email the two Lisbon travel editors whose rooftop lists the assistants kept citing, with photos and the facts; and keep asking guests for a review at checkout, on the rooftop if that is where they were. Then re-run the checker and the manual questions monthly, watching the rooftop questions specifically. The roundup pitch is the slowest item and the one most likely to move the answer.

Common mistakes

• Checking once, from your own logged-in account, and drawing a conclusion. Personalization and randomness make a single check meaningless.
• Buying “AI ranking”. Ads inside assistants are labeled; nobody sells the organic recommendation itself.
• Services hidden in PDFs, images or booking widgets. Retrieval reads text.
• Inconsistent names for the thing you want to be known for. Terrace on the website, rooftop in the reviews, nothing in the profile.
• Reviews that praise “great service” and nothing specific. They count, but they cannot be quoted for a service question.
• Treating the checker score as a ranking. It is the share of a sample of answers that named you.
• Expecting change in a week. Retrieval-based answers move when the sources move, typically weeks; memory-based answers move only when a model is retrained.

Do this now

1. Run the AI Visibility Checker and note the score, the engines that named you and the competitors named instead.
2. Ask the real assistants your five most valuable buyer questions in a fresh session; record names and cited sources.
3. Fix the profile: precise primary category, all services, a plain-sentence description, hours and attributes; re-run the Google Business Profile Optimizer to confirm.
4. Write or rewrite one page per major service and a FAQ page that answers the recorded questions in text.
5. Pick one phrase for what you are and for your signature service; apply it to the website, profile, directories and staff scripts.
6. Audit name, address and phone across every cited source and fix the variants.
7. List the roundups and directories the assistants cited and start on inclusion, one pitch a week.
8. Add LocalBusiness JSON-LD to the homepage with the Schema Markup Generator and validate it.
9. Put the monthly re-check in the calendar, and read the pillar guide on AI search optimization for local businesses for the wider program.