Local AI visibility: how a local answer is built and what to fix first

What is local AI visibility?

It is AI visibility for a question with a place in it. Local SEO is the work of appearing in the map results and the local listings when someone searches for your kind of service in your area; local AI visibility is whether an assistant names you when someone asks it the same thing in a sentence. The two share raw material, your name, address, phone and category, but the output is different. A map pack ranks; an assistant writes a paragraph that names two or three businesses and stops. There is no page two.

With web retrieval on, the assistant is not consulting a stored ranking. When a question arrives it fetches pages that seem to answer it and writes from what they say, which is what retrieval-augmented generation means in practice. It reads those pages as a machine: no browser, usually no scripts run, no patience for a map it cannot read. What is in the HTML is what it knows about you at that moment.

So the question for a local business is not how to rank. It is whether, at the moment a page of yours is fetched, that page states plainly who you are, where you are, what you do and why anyone should believe it.

How is a local answer assembled?

First comes the place. The question names a town; the page has to name the same town, and one address, one phone number, the area served and the hours, written the same way each time they appear. Written that way, they describe one entity rather than three near-matches nobody can reconcile. If the page carries no address, an answer has no address of yours to give, and a question about a place finds nothing on the page that places you.

Then comes evidence. Services stated in words, so "furnace repair" is on the page rather than implied by a photo. Reviews quoted in the text, with the platform named. Mentions elsewhere that agree with the site about the name and the address. The table sets out each fact, where it is stated, and what usually breaks it.

What did the audit find on local business sites?

In our audit of 9,483 small business websites (40 trades, 90 metros in the US, Canada, Europe and the UAE, September 2026), the pinning-down step failed more often than anything else. 43.1% of the sites were inconsistent about who and where they are across their own three main pages. 36.9% published more than one phone number. 34.1% had no postal address a machine could parse anywhere.

The evidence step was worse. 64.7% showed no review signal a crawler can read, no rating markup and no review quoted in the text, and only 8.9% carried rating schema. Most of these businesses have reviews; they live on the platforms, and the site does not say so. 17.8% had no schema.org markup at all, and of the rest most of it described the website rather than the business. LocalBusiness appeared on roughly a quarter of sites.

We also asked. One buyer question from each business's trade and metro went to Claude with web retrieval on, once, for 9,240 of them, and 14.9% were named; 17.7% among the 5,928 answers that gave a differentiated list rather than a hedge. That is a Claude number from one question on one day, directional and not a rate for any assistant. Two things from those answers are worth holding on to. The businesses that were named were, on inspection, the ones with the consistent identity and the answer-shaped content described above. And of every name the assistant gave, 99.4% were actual businesses and 0.6% were directories. The assistant's list is not Google's first page.

The four things to fix, in order

1. One identity, repeated exactly. Pick one form of the name, one address written the way the post office writes it, and one phone number, and put them in text on every page, in the footer at least. List the towns you serve by name. Then make the about page and the contact page agree with the footer to the character.
2. Mark up the business, not the website. Add LocalBusiness schema, or the specific subtype for your trade, as JSON-LD, with the name, address, telephone, openingHoursSpecification, geo and areaServed filled in. Google's local business documentation lists the properties it reads, and its structured data introduction says to mark up only what a visitor can see on the page. The markup states the facts; the text has to carry them too.
3. Put the reviews on the page as text. Quote two or three real reviews with the reviewer's first name and the platform they came from. Keep the platform's rating out of your rating markup; Google's review snippet rules say not to aggregate ratings from other websites, and a business's reviews of itself on its own site earn no stars.
4. Write the five questions customers ask on the phone. Each as a heading, in the customer's words, with the answer in the next two or three sentences: do you come out on weekends, what does a first visit cost, how far do you travel. That is the content an assistant can lift, and 61.5% of the audited sites had none of it.

The order is the order of dependence. Markup that names a business the text describes three ways adds a fourth version. Reviews on a page the assistant cannot place are praise for nobody in particular. Answers under headings are the most valuable step and the last one, because they only count once the assistant knows whose answers they are.

How is this different from a Google Business Profile listing?

The profile matters, and it is not the same thing. Google's Business Profile guidelines require the listing to represent the business as it exists: its real name, its real address, the categories that fit, its hours. That listing is what Google's map results are built from, and it is where many of your reviews live. Keep it accurate; nothing on your site substitutes for it.

But an assistant answering "who can fix a furnace in Duluth tonight" is reading pages, and your site is the page you control. For Google's own AI features, the company's guide says AI Overviews and AI Mode are rooted in core Search ranking, and its page on those features says no special markup or file is required and a page has to be indexed and eligible for a snippet to be a supporting link. The other assistants document crawlers that fetch your pages by name; our crawler access check reads your robots.txt for all of them. None of these documents describes an assistant reading a business profile in place of a site.

Where the profile and the site disagree, an assistant sees the same defect the audit measured inside the site, from outside it: two addresses, two phone numbers, two names. So the order is the site first, then the profile brought into line with the site, then the directories. The profile has no room for a service described in a paragraph or a question answered in three sentences. The site does.

What to measure

Write the questions your buyers ask, one for each town you serve and each service you want to be named for, in their words. Put each question to ChatGPT, Claude, Gemini and Perplexity three times, because one answer varies from run to run, and record whether you were named, whether you were recommended, whether you were first, and who else was named. Take the same reading every month on the same questions. The procedure we use for a paying client is narrower: twelve questions, each put to one assistant, Claude, at the baseline, again once the fixes are live and every month with monitoring, with a record of whether the business was named, whether it was named first, and who else was. Yours can start at six questions and a spreadsheet.

Before any of that, find out what a crawler receives from your own site, because 7.2% of the audited sites served a challenge page instead of the site and never got as far as being read. The free visibility check does that fetch, asks Claude the buyer questions for your trade and town, and a person sends you the result within 48 hours. What it cannot do is promise a place in the answer. Nobody outside the assistants can, and the measurement is how you find out whether the fixes above did anything.