Talk to MapsExperts

Information about how we process your data is available in our Privacy Policy.

Insights

AI & Search · Guide

How to measure AI visibility for local businesses: beyond rankings and mention counts

A mention count tells you that a business appeared somewhere. It does not tell you whether the right location was recommended, with correct facts, to someone nearby who then called or came in. The framework we use to measure AI visibility for local businesses, from one location to a thousand.

By
, Founder, MapsExperts
Published
Reading time
18 min
Tags
AI visibility · AI search · Data accuracy · Multi-location

In short

  • For a local business, AI visibility is a set of rates measured across repeated, place-anchored customer questions: not a position, and not a count of mentions.
  • Mention, recommendation and citation are different events. A business can be cited without being recommended, and recommended without its own website being cited.
  • Accuracy is part of visibility. The share of answers with a wrong status, address, phone number or opening hours matters more than average accuracy across all facts.
  • A composite AI Visibility Score works as the summary KPI, read alongside a core set of individual metrics: per market and surface, the recommendation rate for the decisions that matter, share of recommendations against named competitors and the critical-error rate, each with its sample size, plus AI-attributed outcomes with a note on what they miss (Google's AI clicks count as organic; calls and visits carry no referrer).

AI visibility is a rate, not a rank

Most AI visibility reporting still comes down to a mention count: how often a business is named in AI answers. It is easy to collect and hard to manage by. This article sets out what we use instead, the MapsExperts AI Visibility Measurement Framework: what to count, how to sample it, and where each method stops being reliable.

Rankings are positions in a list. Organic positions are reported by the platforms themselves, as an average in Search Console; Local Pack and Maps positions are not, so local SEO tracks them across a grid of searcher locations. AI answers have no such list. An assistant composes an answer for one person, names a few businesses, may or may not cite a source, and asked again can choose differently. No AI surface we know of reports to businesses how often they were recommended, so that has to be sampled. AI visibility therefore has to be measured as rates, and at its core are three, each of which can fail on its own:

The three fail for different reasons, and that is the point of separating them. A location that is never included has a relevance or coverage problem. One that is included but rarely recommended loses on fit or reputation for that particular decision. One that is recommended with the wrong opening hours has an operations problem: its data is out of date somewhere an assistant looks. Each needs a different response, and a single score, however useful for tracking, cannot show which one you have.

The surfaces behave differently, so measure them separately

The AI environments a local customer meets differ in where their place data comes from, how they locate the user, and whether they show sources.

What is documented about local answers, by surface
SurfaceWhat is documented about local answers
Google Search: classic results, Local Pack, Maps Local results are ranked mainly on relevance, distance and prominence; distance uses the searcher's location, or what Google knows about it if none is shared[1]
Google AI Overviews and AI Mode No requirements beyond being indexed and snippet-eligible; both may run several related searches ("query fan-out")[2]. Answers can include local business information; Google names the Business Profile as one way to appear[3]
Gemini and Ask Maps The Gemini API can ground answers in Google Maps place data, including hours and reviews (developer documentation)[4]. Ask Maps, conversational recommendations inside Google Maps, began rolling out in the US and India in March 2026; a rollout to 150+ countries and territories in English was announced in August[5]
ChatGPT May localize with an approximate IP-based location, or device location if the user enables it; may share location with unnamed providers of information such as nearby business listings[6]
Perplexity Numbered citations link answers to their sources[7]
Microsoft Copilot and Bing Reports how often a site is cited in Copilot, Bing's AI summaries and some partner integrations (citation counts; no click data documented), and classifies a sample of grounding queries, the AI's own search phrases rather than users' prompts, by intent, including Local[8]

Check each surface's availability in the market, language and account state you will test; Ask Maps' broad rollout, for example, is in English. Assistants not covered here, such as Siri with Apple Maps, draw on other place data and need their own checks.

Three consequences follow. First, an average across surfaces hides which surface fails. An assistant grounded in map data and one that leans on web pages can describe the same location quite differently, and only per-surface rates show where the gap is. Second, measure in the consumer interfaces customers use. APIs suit volume and trend work, but an API call runs outside the customer's app and account, so its answer is not necessarily what a customer sees; a companion article looks at the difference. Third, platform reports measure less than visibility: Search Console counts impressions of links to your site[9], Bing counts citations[8], and a recommendation drawn from a Business Profile may carry no link to your site at all.

Define the unit before counting anything

A run is one answer to one phrasing of a customer decision, asked in a new conversation, from one place, on one surface, at one time. Runs are grouped into location-decision cells (one location, one decision, one surface, one period), and every rate is reported per cell or group of cells. A run is valid if the surface answered the question as asked, from the intended place; refusals, clarifying questions and errors stay out of the denominator and are reported per surface as their own share.

Decisions

Group questions by the decision behind them: category ("dentist near me"), need ("emergency dentist open on Saturday"), service, suitability and fact checks ("is the Favoriten clinic open on 24 December?"). Discovery questions, which do not name the business, measure inclusion and recommendation, and the accuracy of whatever an answer states about the location; fact questions measure accuracy only. Customers do not all phrase a need the same way, so write three to five phrasings per decision, in the languages customers use: in Vienna, German and English.

Place

Naming the place in the prompt ("near Westbahnhof") is reproducible and should form the core of the panel. An implicit "near me" is more realistic but only as precise as the location the surface receives: device location where allowed, otherwise an approximate location inferred from the IP address[6], which can be some distance from where the customer actually is. Run implicit checks on a real phone, on site, with the location permission recorded, and report them separately; only surfaces that receive device location can show how visibility changes with distance.

Repetition and personalization

Ask an assistant the same question twice and the list of businesses can change: answers are generated, not looked up. Part of that is the model, part is retrieval, and part is the person asking: signed-in accounts, memory, history and location can all shape answers[6][5], and none of these effects is documented in enough detail to predict. A position in an AI answer is therefore not a stable fact, while the rate at which a business appears across many runs can be measured. Keep a clean baseline (logged out, or fresh accounts with memory off) as a fixed yardstick, add a smaller set from realistic accounts, and record the model shown, account state, location method and date of every run.

How many times should each question be run?

Enough to tell the rate from noise. Detecting a change from 30% to 45% (two-sided 5% level, 80% power) takes about 160 runs per period for one cell, and the worked example below shows how wide the interval still is after 30 runs. Such figures assume independent runs, so start each in a new conversation and spread runs across phrasings and days; where a few phrasings are repeated many times, resample whole questions and expect wider intervals.

The AI Visibility Measurement Framework

Brand-level measures (mention counts, share of voice, sentiment) are a reasonable start for a brand, and the more careful approaches already separate mentions from citations and recommendations and track factual errors and run-to-run variation. We agree with that direction. It is not enough for a network, which does not win or lose customers as a brand but location by location, decision by decision. Our framework applies the same discipline there. It has five layers, and each can move while the others stand still.

The MapsExperts AI Visibility Measurement Framework: five layers and their metrics
LayerMetricDefinition and calculationHow to collect itMain limitation
1. Inclusion: are we in the answer? Inclusion rateDiscovery runs naming the business (any location) ÷ valid discovery runsRepeated question panel per surfaceCounts passing and negative mentions too
1. Inclusion Correct-location rateRuns naming the location that serves the place asked about ÷ runs naming the businessPlace-anchored panel; agreed catchmentsNeeds a rule for which location should win; catchments overlap
2. Recommendation: are we chosen? Recommendation rateDiscovery runs presenting the location as suitable for the need ÷ valid discovery runsClassify each mention (recommended, neutral, with a caveat, unsuitable) against a rubric that states which classes countClassification needs a measured agreement check
2. Recommendation Top-recommendation rateRuns naming the location first or as explicitly preferred ÷ valid discovery runsSameOrder in prose is a weak signal; cards and text can disagree
2. Recommendation Share of recommendationsThe location's recommendations ÷ all recommendations to the defined competitive set, same runsLog every business recommendedDepends on the competitive set; report it alongside
3. Accuracy: are the facts right? Field accuracyCorrect facts ÷ facts stated; one fact = one field (address, phone, status, one weekday's hours, one special-hours date)Compare with the business's own record as of the run dateFlatters: many easy facts can hide one harmful error
3. Accuracy Critical-error rateAnswers with at least one wrong status, opening hours, address or phone ÷ answers naming the locationSameNeeds a dated source of truth, special hours included
3. Accuracy Omission rateAnswers leaving out a fact the question asked for ÷ answers to questions that asked for itSameBrevity is not always an error
4. Sources: what supports the answer? Source mixCited sources by type: own site or location page, map profile, directory or review site, media, forum, videoClassify every cited URLVisible citations are not all sources used; some surfaces show none
4. Sources Owned-source citation rateRuns citing the business's own domain ÷ valid runs on surfaces that display citationsSame; Search Console and Bing show related, not identical, metricsA citation can support a fact without supporting a recommendation
4. Sources Citation–recommendation gapOf runs naming the location on surfaces that display citations: cited and recommended, cited only, recommended only, neither ("cited" = own domain)Cross-tabulate layers 2 and 4Depends on each surface's citation design
5. Outcomes: does it lead to action? AI referral sessions and key eventsSessions, leads and bookings from identifiable AI referrersAnalytics channel rules; location landing pagesMisses in-app traffic, copied names, and Google's AI features (counted as organic)[10]
5. Outcomes Offline-action proxiesCalls, direction requests, bookings; branded search demandBusiness Profile performance; call tracking; Search ConsoleNot attributable to AI answers specifically
Across all layers StabilityConfidence interval per rate; agreement between repeated runsRepetition across days and phrasingsSmall panels give wide intervals; model changes break trends

Four rules make it usable. Publish every rate with its denominator. Report per surface and location group; a composite score can summarize the result, but never replaces the individual rates. Check classification: two people code a random sample of answers and report their agreement (for example with a chance-corrected statistic such as kappa), and any automated classifier is re-checked against human coding after model changes. And judge accuracy against the business's own record on the day of the run, which means special hours and closures need dates in the master data.

A composite score as the summary KPI

Individual metrics are where diagnosis happens: they show where visibility changes. Most teams also want one number to track. A composite score that combines several AI visibility metrics serves that purpose; ours is the MapsExperts AI Visibility Score, a composite index on a 0–100 scale. It shows whether overall performance is moving and is the natural headline for executive reporting. It works under three conditions: its components and their definitions stay fixed from one period to the next; it is reported to the people who act on it together with those components and their sample sizes; and every change in it is traced back to the components that moved. A score that rises because inclusion improved while accuracy slipped is not an improvement a customer would recognize, and only the components show that.

A companion study applies this approach, alongside UTM-attributed lead tracking, to ten small local businesses that published no YouTube Shorts in September 2026 and started in October. Across the two months, the average change in their composite MapsExperts AI Visibility Score was +17 points on a 0–100 scale: an observation rather than proof that the videos caused it, and a summary that, as this guide argues, needs its components to explain it.

A worked example

A dental group runs three clinics in Vienna: Mariahilf, Favoriten and Döbling. Six discovery questions are asked as if from each clinic's district, with the place named in the prompt, on one AI surface, five times each over two weeks: 30 runs per district.

Simulated example: six questions asked five times from each district, 30 runs per district
Asked fromGroup named (any clinic)Right clinic namedRight clinic recommendedAnswers naming the group with a critical error
Mariahilf 18 of 3015 of 1812 of 30 (40%)3 of 18
Favoriten 12 of 307 of 126 of 30 (20%)2 of 12
Döbling 6 of 302 of 61 of 30 (3%)3 of 6
All districts 36 of 90 (40%)24 of 36 (67%)19 of 90 (21%)8 of 36 (22%)

Simulated values. The last column is a group-level critical-error rate: answers with an error ÷ answers naming any clinic of the group.

Recommendation rate, Mariahilf: 12 ÷ 30 = 40%, with a 95% confidence interval of about 25% to 58% if the runs are treated as independent; because they are six questions asked five times each, resampling whole questions would give a wider one. Even on that optimistic basis, a later reading of 45%, or even 60%, from 30 runs would not show a real change: two samples of 30 differ significantly only when they are about 25 points apart, and are reliably told apart (80% power) only at about 35.

Share of recommendations, Mariahilf: across the same 30 answers, 104 recommendations went to the competitive set; the clinic received 12 (11.5%), the most-recommended competitor 21 (20.2%).

Accuracy: the 36 answers naming the group stated 118 facts, 106 correct, a field accuracy of 89.8%; yet 8 of the 36 answers (22%) contained a critical error, most often Saturday hours that had changed.

A mention dashboard would report that the group appears in 40% of answers, mixing relevant and irrelevant questions, right and wrong branches, current and outdated facts. Separated, the runs say more. From Favoriten, five of the twelve answers naming the group missed the right clinic. Döbling is close to invisible where it operates; with 30 runs per district, only its gap to Mariahilf is clearly beyond noise, and the other differences are signals to retest with a larger panel. About one answer in five that names the group gets a critical fact wrong. Each finding raises its own question: whether Favoriten's information distinguishes it from its sister clinics, why Döbling is rarely named, and where the outdated Saturday hours come from.

Sources: what citations can and cannot tell you

No. A citation shows that a page was used as a source for some part of an answer. A recommendation means the business was presented as a suitable choice. The two can diverge, for the reasons below, so measure them separately and look at where they disagree.

A location page can be cited for an address in an answer that recommends a competitor; a business can be recommended on the strength of map data and reviews without its own website being cited. Citation counts are often treated as the measure of AI visibility. For a local business that equation does not hold, because the facts that decide a visit (status, hours, address) often come from place data that carries no link at all.

That is why the framework cross-tabulates the two. Cited but rarely recommended suggests a trusted source of facts that loses on relevance or reputation. Recommended but never cited suggests an answer resting on platform data the business only partly controls, which is where accuracy risk sits. For a network, the gap is also a governance signal: if the brand's own location pages are never cited, the owned location data layer is not doing its job, as our article on location pages for AI search explains.

Classify cited sources by type, and expect the mix to differ by surface: answers grounded in map data lean on place records, while others may draw on websites, directories, review sites and video. Platform citation metrics need the same care: Bing's citation share describes a site's share of the citations for a grounding query[8], not how often a business is chosen.

Outcomes: what analytics can and cannot attribute

Can Search Console or Google Analytics show traffic from AI Overviews and AI Mode?

Partly. Search Console's generative AI performance report shows impressions of your pages in AI Overviews and AI Mode by page, country, date and device, but not queries or clicks[9]. Clicks from those features are counted as ordinary Google organic traffic in Search Console[2] and in Google Analytics, whose AI Assistant channel covers assistants such as ChatGPT, Gemini and Copilot but not Google's AI features[10].

Annotate the AI Assistant channel's May 2026 introduction in every trend: before it, these visits fell under the other default channel rules, typically Referral, and Google does not say whether earlier data was reclassified[10], so check your own property. Where AI referrals are still small, a few sessions can swing the rate, so compare periods of several months.

The larger gap for local businesses is offline. Many AI-assisted decisions end with a name typed into Maps, a call or a visit, none of which carries a referrer. Business Profile calls, direction requests and bookings can now be reported in Google Analytics[10], but they show that an action happened, not that an AI answer caused it. Three steps narrow the blind spot: make each location's own page the website link on its profiles and listings and the link target in the Store Locator, so that when an assistant links to the business, the link can lead to a location-specific page; track calls, bookings and route requests as key events by location; and add an AI option to "How did you hear about us?" at booking, often the most direct evidence a local business has.

From one location to a thousand

At every size, changes are interpretable only against a stable panel: a fixed core of questions, places and surfaces plus a smaller rotating set, both versioned, because prompt sets drift toward questions the business already wins. Annotate each period with changes on the business's side (hours, pages, reviews) and the platforms' (launches such as Ask Maps, model changes). When a model update lowers every business's inclusion, share of recommendations should hold up better than absolute rates, because both sides of the ratio come from the same answers; but if an update changes how many businesses an answer names, every share shifts, so report the average number named per answer alongside it. Screenshots document a specific error; they do not measure visibility.

What changes with size is how much can be sampled, what has to be automated and who owns the result.

Panel design, cadence and ownership by network size
NetworkPanel designCadenceOwnership
1 location 10–15 decisions, three to five phrasings each, two or three surfaces; facts checked by hand against the business's own dated record, and the Business Profile against the same recordMonthly, and after changes to hours, services or addressOwner or agency
10 locations One shared decision set, each location tested for its own catchment; correct-location rate becomes a headline metric where catchments overlapMonthlyMarketing lead, with one source of truth for hours
100 locations Accuracy monitored at the source: place records on Google, Apple, Bing and main directories checked against the master record; discovery questions on a rotating sample by market and format, around a fixed coreSource accuracy weekly to monthly; recommendation monthly or quarterly per marketCentral team, regional review
1,000+ locations Source records checked continuously around the change calendar (holidays, closures, relocations); official APIs for volume and trends, reported separately and calibrated against a manual, stratified consumer-interface sample; exceptions routed to location operationsSource accuracy continuous; share of recommendations quarterly by marketLocation operations and analytics

These are our recommended starting points, not standards or measured benchmarks. A one-location panel of a few dozen questions finds large gaps and persistent errors, but one location's monthly sample cannot show changes of a few points, so pool several months before reading a trend. Before automating queries of a consumer AI interface, check that service's terms of use; if you buy monitoring, ask how answers are collected and how location is set. At scale, many small comparisons produce false alarms: flag a location only when a change persists across two periods, and read a small location's rate against its market and format average rather than on its own. When rolling samples up to a market or the network, weight each stratum by its number of locations, and compare rotating samples only on their shared core. At every size, time accuracy checks to the change calendar: check special hours in the week before a public holiday.

For executives, a composite AI Visibility Score is a sensible headline, as long as it arrives with the numbers in the summary at the top, per market and surface and each with its sample size and trend. The score says whether things are moving; the numbers beneath it say why. Behind them, the location-level framework replaces an impression of visibility with evidence to act on: which location, which decision, which surface, which fact.

Our Location Intelligence work also starts from a defined set of customer decisions rather than a prompt list, and traces recurring gaps to the most plausible controllable signals: location data, profiles, pages and reviews.

References

  1. Google Business Profile Help. Tips to improve your local ranking on Google (opens in a new tab). Accessed 9 November 2026. Platform documentation.
  2. Google Search Central. AI features and your website (opens in a new tab). Last updated 10 December 2025. Platform documentation.
  3. Google Search Central. Optimizing your website for generative AI features on Google Search (opens in a new tab). Last updated 10 July 2026. Platform documentation.
  4. Google AI Studio. Grounding with Google Maps (Gemini API documentation) (opens in a new tab). Accessed 9 November 2026. Platform documentation.
  5. Google, The Keyword. How we're reimagining Maps with Gemini; Ask Maps gets more helpful with food ordering and more (opens in a new tab). 12 March 2026 and 6 August 2026. Platform announcements.
  6. Perplexity Help Center. How does Perplexity work? (opens in a new tab). Accessed 9 November 2026. Platform documentation.
  7. Bing Webmaster Blog and Bing Search Blog. Introducing AI Performance in Bing Webmaster Tools (Public Preview); New AI Visibility Insights in Bing Webmaster Tools: Intents, Topics, Citation Share, Compare (opens in a new tab). 10 February 2026 and 16 June 2026. Platform documentation.
  8. Search Console Help. Generative AI performance report (Search) (opens in a new tab). Announced 3 June 2026; available to all sites since 31 August 2026. Platform documentation.
  9. Google Analytics Help. What's new in Google Analytics (entries of 13 May and 8 June 2026); [GA4] Default channel group (opens in a new tab). 13 May 2026 and 8 June 2026; default channel group accessed 9 November 2026. Platform documentation.

How this article was sourced. Platform functionality was checked against current Google, Microsoft and Perplexity documentation on 9 November 2026 and OpenAI's on 6 November 2026. The framework, definitions and recommendations are MapsExperts' own. The clinic example and its figures are simulated. The panel sizes and cadences in the scaling table are our recommendations, not standards. Confidence intervals are score intervals at 95%; the sample-size estimate is a standard two-proportion calculation (two-sided 5% level, 80% power), assuming independent runs. For networks, location-level flags can additionally be controlled with a false-discovery-rate correction, and small-location rates estimated with partial pooling toward market and format. Sources last checked . Research standards

Related insights

Next step

Measure AI visibility per location, not per brand.

MapsExperts connects AI discovery with location data, profiles, local pages, competitors and reviews, starting from the customer questions that matter for each location. We do not guarantee inclusion, ranking, citation or recommendation in a specific AI answer.

Location Intelligence