AI Visibility for Local Businesses: Which Signals Actually Reach the Assistants

The Statistics Everyone Quotes Come From People Selling the Fix

Three numbers anchor almost every article on this subject: a jump in the share of consumers using AI to find local businesses, a figure for how rarely ChatGPT recommends local businesses compared with Google's map pack, and a share of restaurants said to be invisible to AI. We traced each to its origin. All three were published by companies that sell local marketing or AI visibility software. Two reached the trade press through paid press-release distribution. The restaurant figure was computed from the vendor's own customer base, which makes its denominator unknowable.

The clearest tell is that two of them cannot both be true. One implies that assistants recommend somewhere around one percent of local businesses. The other implies that roughly one in six restaurants is visible. Those describe different worlds, and they are routinely cited in the same paragraph as if they support each other. Neither discloses its prompt set, sample geography, or how many times each query was run, which matters enormously: a survey of 45 GEO studies found substantial run-to-run variability and no technique with a stable cross-platform effect.

What is solid is thinner and more useful. Pew, using a probability panel of 5,119 US adults, finds 49% of US adults now use AI chatbots, with information-seeking the most common use. Pew has published nothing specific to local-business discovery, and neither has anyone else without something to sell. Google reports that the average AI Mode query is three times longer than a traditional search, with "where to" and "where should I" phrasings growing quickly. Directionally, local intent is moving. Precisely how much, nobody credible has measured.

The behavioural consequence is measured, though, and it is the number that should actually worry a local operator.

When an AI summary appears, people click a result on 8% of visits instead of 15%, and end the session outright on 26% instead of 16%. Being named in the answer is increasingly the whole transaction.

There Is No "AI Layer." There Are Four Different Pipes

Most local AI advice treats the assistants as one system reading one public record. They are not. Two of them have documented, named sources of local place data, and those sources are different from each other.

Google's Maps grounding reaches over 250 million businesses and places, exposes timely details such as a restaurant's current hours, and can answer subjective questions using context from Maps user reviews. ChatGPT took a different route: Yelp's own announcement describes Yelp reviews, ratings, photos and business details appearing in the chat, with table booking attached.

One detail in Google's documentation is worth dwelling on, because it explains a lot of the behaviour people find baffling. The application calling the model supplies latitude and longitude explicitly. The model is not quietly working out where the user is. Proximity, which is a hard ranking input in the map pack, becomes a choice made when the query is constructed. That is why the same question can return a business twenty miles away, and why a screenshot from your phone tells you very little about what a customer across town sees.

Which Local SEO Signals Actually Transfer

This is the question every local operator has, and the usual answer is a list of six things to do with no indication of which assistant any of them reaches. Graded strictly on what platforms document:

Two conclusions follow, and both cut against the standard advice. Your Google Business Profile is close to a Google-only asset for AI purposes. And your Yelp presence, which a lot of local operators have quietly neglected, is now a documented input to ChatGPT. Reviews are the one signal with a documented path into both.

Notice how much of the table reads "unproven." That is not hedging. Google states plainly that there are no special optimizations necessary to appear in AI Overviews or AI Mode. And Google's local ranking guidance, which still describes relevance, distance and popularity, does not mention its AI surfaces at all. Anyone selling you a proprietary local AI optimization method is selling something Google says does not exist.

Independents Start Behind, and There Is Real Evidence Why

A single-location business is competing against national chains inside a model that has read far more about the chains. This is not a hunch: peer-reviewed work on brand bias in language models finds they systematically favour global brands over local ones. A separate study found that LLM restaurant recommendations shift measurably with the dialect the question is asked in, though it tested selection from a supplied list rather than open discovery.

One clarification, because this figure gets misused. Our own 102-brand study, published openly on arXiv, found a three-tier stature ladder in which niche brands appeared in 11.4% of unbranded answers. That describes niche brands in a brand-tracking cohort. It is not a measurement of single-location local businesses, and we are not going to relabel it as one. The direction it points is sound; the number is not yours.

The practical read is that stature is accumulated third-party coverage, and local coverage is the one kind you can go build this quarter. You are not trying to outrank a chain nationally. You are trying to be the obvious answer in one city, where the citation set is small enough to move.

Absent Is Not the Only Failure. Present and Wrong Is Worse

Visibility tooling almost universally measures one thing: were you mentioned. For a local business there are three states that matter, and the middle one does the most damage. Absent means you lost a customer. Present and wrong means an assistant sent someone to a closed shop, an old address, or the wrong branch, and they blame you rather than the model.

The best-documented case is a government one. New York City's official business chatbot told business owners they could break the law, including that a restaurant could go cash-free in a city that banned it in 2020. Reporters and independent experts got inconsistent answers from the same system. If a city cannot keep its own assistant accurate about its own rules, the odds on your opening hours are not good.

Part of the problem is that the underlying record is contested at a scale most operators never see.

In one year Google removed 240 million policy-violating reviews and 12 million fake Business Profiles. That is the data layer AI answers about your category are built on. And the stakes are rising: Google Search will now call local businesses to check pricing and availability on a user's behalf. Wrong information stops being a reputational annoyance and starts being a phone that rings for the wrong reason. If an assistant is already saying something false about you, the correction playbook is its own piece of work.

Nearly Every Claim in This Category Rests on One Screenshot

Local AI answers vary by the location the app passed in, by whether the user is logged in, and between runs of the identical prompt. Yet the evidence base for this entire topic, including the trade press, is people asking once and screenshotting the result. We have written before about why one measurement is not a reading, and local is the worst case for it, because the location input is itself a variable.

It also means "we track city-level visibility" deserves a follow-up question, including when we say it. Here is what geo-targeting actually means per engine in our own pipeline:

Three engines take a real location parameter. Two are told the city in the prompt text and may do as they please with it. Google AI Overviews is country-level only on our side, so we do not claim city-level numbers there. Any tool that reports a single blended "local visibility score" across engines is averaging those different mechanisms into one figure, which is a large part of why two tools give you two answers.

The Playbook, Ordered by What Is Actually Documented

Ranked by evidence rather than by habit. The first three have documented paths into an assistant. The rest are reasonable and unproven, which is a fine reason to do them cheaply and a bad reason to buy a programme around them.

• 01Reviews, everywhere that feeds a pipe. The only signal documented to reach both major surfaces. Steady and recent beats a burst. Google's own local guidance says more reviews and positive ratings help, and Maps grounding can draw on those reviews for AI answers.
• 02Google Business Profile, kept accurate. Complete categories, services and above all current hours, because hours are explicitly exposed to Google's AI surfaces. Treat it as an accuracy obligation, not a ranking lever.
• 03Your Yelp page. Documented input to ChatGPT. If you have not looked at it in two years, that is now a gap in the assistant with the largest consumer reach.
• 04Local "best of" lists. Both surfaces retrieve from the open web, and ranked lists are the single highest-cited content format in our study. You cannot write your way onto them, so earning a place is the work.
• 05Location pages and consistent details. Worth doing, and honestly reported as unproven for AI answers. Only about 3% of citations point at a brand's own domain, so this is hygiene rather than leverage.
• 06Measure per location, not per brand. Being named as a business is not the same as the right branch being named. If you have several branches, the interesting failure is the model defaulting to the wrong one.

Check It Yourself This Week

You do not need a tool to get an honest first reading. You need a method, because the thing that makes local answers hard to trust is exactly the thing a casual check ignores.

One thing not to bother with: asking the assistant why it recommended a competitor. Models produce fluent explanations of their own output that are reconstructed after the fact, not a record of what happened. Score what it said, not what it says about what it said.

How Ranqo Helps

The protocol above is the job. Ranqo is that protocol run on a schedule, at whatever number of locations you have. You set the prompts and the location per prompt, and it runs them weekly across the engines, records whether you were named and where you placed, tracks which competitors appear beside you, and shows the sources the answers leaned on so you know which pipe to work.

What it does not do, stated plainly: it does not verify whether the hours or address an assistant quoted are correct. That accuracy layer is a real gap in this category, ours included, and it is the one we think matters most for local. Nor can it geo-target Google AI Overviews below country level. When we can do those honestly, we will say so.