AI Visibility Tracking for Local Businesses: A Field-Test Method

A local business can appear in a ChatGPT answer on Monday and be absent from a similar Google AI Mode response on Tuesday. That single result is useful for investigation, but it is not a visibility metric.

AI visibility tracking gives local businesses a repeatable way to measure brand mentions, recommendations, citations, and competitor gaps across AI answer surfaces. The concrete payoff is better reporting: we can identify which service-and-location questions produce durable evidence, which findings need more repeat runs, and which apparent wins should never be presented as a client KPI.

This is a narrowly scoped measurement method, not a claim that every AI product uses the same index or returns a stable local ranking. We separate the products being observed, preserve the full answer and context, and calculate share of answer in a way that does not erase businesses outside a predefined competitor list.

What we mean by AI visibility tracking

For a local business, AI visibility tracking means recording whether an AI-generated response names, recommends, or cites the business when a prospective customer asks a realistic question. A prompt such as “EV charger installer in Burnaby” is more useful than a broad vanity query because it contains a service and a location.

The first rule is to distinguish AI answer surfaces from underlying systems. In this article, we monitor five answer surfaces:

• ChatGPT
• Google Gemini
• Google AI Mode
• Google AI Overviews
• Perplexity

ChatGPT, Gemini, and Perplexity are separate products. Google AI Mode and Google AI Overviews are separate Google Search experiences, not independent search engines. They may be compared as distinct answer surfaces because a buyer can encounter different responses in each, but we should not imply they are five equivalent indexes or five interchangeable ranking systems.

Google says AI Overviews and AI Mode use its core Search systems, while also applying its AI experience to the query and response. That makes separate observation worthwhile: a conventional Google ranking, an AI Overview, and an AI Mode answer can each present information differently. For a deeper Google-specific comparison, see AI Overviews tracking vs AI Mode tracking for SEO.

The field-test design: what to save on every run

We do not treat a screenshot as a dataset. A usable local AI visibility tracker needs an answer-level record that another person can inspect later.

For every prompt run, save these fields:

1. Timestamp: Record date, time zone, and time of day. A run at 9:00 a.m. Pacific time is not equivalent to an undated screenshot.
2. Answer surface: Label ChatGPT, Gemini, Perplexity, Google AI Mode, or Google AI Overviews precisely.
3. Exact prompt: Preserve punctuation, service wording, city, and modifiers such as “best,” “licensed,” or “near me.”
4. Location and settings: Note the location context, signed-in state, browser or app context where relevant, and whether search or web access was used.
5. Full response: Save the entire text, not just the sentence containing the brand.
6. Business appearances: Extract every business named, including untracked businesses that appear unexpectedly.
7. Recommendation and citation labels: Record whether each business was recommended and whether a visible citation or source link appeared.

This is deliberately more detailed than a monthly scorecard. The extra evidence matters when a competitor appears in a list one week and disappears the next, or when a response gives buying advice without naming any provider.

The published research brief for this topic emphasizes repeated prompt sampling and engine-level variance. We use that as a methodological warning, not as a universal numerical benchmark. No single local-business test can establish how frequently every city, trade, or AI product will change.

Define mention, recommendation, and citation before scoring

A business name in an answer does not always mean the same thing. We use three separate events so a local brand is not credited with visibility it did not actually earn.

Mention

A mention occurs when the answer identifies a business by name. For example, “Company A and Company B serve Vancouver” gives each business one mention. A mention can be neutral, negative, or incidental.

Recommendation

A recommendation occurs when the answer presents the business as an option that fits the user’s request. Terms such as “a good choice,” “consider,” “top option,” or an ordered shortlist may qualify, but we document the wording rather than relying on an automated assumption.

For example, “For EV charger installation, consider Company A because it lists residential charger work” is a recommendation. “Company A is mentioned in a local directory” is not necessarily one.

Citation

A citation is a visible source reference, link, footnote, or source card associated with the answer. It may point to the business’s website, a Google Business Profile, a directory, a review site, or an editorial roundup.

A citation is evidence that a source was surfaced; it is not proof that the cited business was endorsed. We verify the cited URL or domain and record its source type. Our guide to ChatGPT local search citations and top-list pages covers why third-party list pages deserve separate inspection.

Why the same prompt can produce different local businesses

“Best electrician in Burnaby” appears simple, but it combines a subjective qualifier, a trade, and a geographic qualifier. Different answer surfaces can interpret any of those elements differently.

Google’s documentation describes AI features as connected to Google Search’s established ranking and quality systems. OpenAI documents that ChatGPT search can search the web and show cited sources. Perplexity’s product documentation describes web-grounded answers and citations. Those are useful product-level facts, but they do not establish that the products retrieve the same pages, use the same local signals, or format recommendations the same way.

Three practical sources of disagreement are common:

Retrieval and source selection

One response may rely on an official service page, while another highlights a local directory, review platform, trade association listing, or editorial “best of” page. Source selection can change which companies are available to be named.

Intent interpretation

“Emergency electrician near me” can favor availability and proximity. “Who installs Level 2 EV chargers in Vancouver?” can favor service specificity, licensing language, installation examples, and local proof. The business set should change when the buyer need changes.

Answer format

A response can offer three businesses, cite sources without recommending a provider, or explain selection criteria without naming anyone. A no-business answer is an answer-format outcome, not automatic evidence that every local company lacks visibility.

The operational lesson is simple: do not combine all five surfaces into one undifferentiated leaderboard before looking at the underlying answers.

Build prompt clusters around real buyer intent

We recommend prompt clusters because one generic “best [service]” query cannot represent a local market. For a Vancouver-area electrical contractor, a 16-prompt exploratory set could include four clusters:

• Discovery: “best electrician in Burnaby,” “top-rated electrician in Vancouver”
• Urgent need: “emergency electrician in Burnaby,” “electrician for power outage repair in Vancouver”
• Specific service: “EV charger installer in Vancouver,” “electrical panel upgrade Burnaby”
• Trust and fit: “licensed electrician for older-home rewiring,” “residential electrician for condo EV charger installation”

The number 16 is an example, not a research-backed universal requirement. A plumber with three service areas may need a different set from a multi-location retailer or B2B consultancy.

Prompt clusters expose useful competitor gaps. Suppose Brand A is recommended in 7 of 10 saved EV-charger answers but only 1 of 10 emergency-repair answers. That is not one vague “AI visibility” problem. It is a potential emergency-service gap that merits checking for service-page clarity, service-area coverage, reviews, and the sources appearing beside competitors.

This also prevents us from optimizing for questions customers do not ask. Before expanding a set, map every prompt to a service, location, sales conversation, or competitor decision.

Calculate share of answer without hiding untracked businesses

Share of answer measures how often a brand occupies recommendation placements relative to competitors. The denominator must be explicit, because answers may name businesses that were not in the original monitoring list.

Our default is open-set share of answer:

Open-set share of answer =
recommendation appearances for one brand ÷
all recommendation appearances for every business named in the saved answers

If 10 answers contain 24 total recommendation appearances, and your business appears 6 times, its open-set share of answer is 25%. That remains true even if 5 of the 24 appearances belong to businesses we did not initially plan to track.

We also report monitored-set share of answer, but label it clearly:

Monitored-set share of answer =
recommendation appearances for one monitored brand ÷
recommendation appearances for the defined monitored brands only

This metric is useful for a fixed client-versus-competitor comparison. It is not a market-wide share measure. If untracked businesses account for 8 of 24 placements, monitored-set share can make a tracked brand look stronger than it is in the observed answer set.

For both calculations, segment results by answer surface, prompt cluster, and date range. We also use a first-position metric where answer ordering is clear:

First-position share =
answers where the brand is the first recommended business ÷
answers containing at least one recommendation

A report should show raw counts alongside percentages. “6 of 24 appearances” is easier to audit than “25%” alone.

A proposed starting framework for repeat runs

We cannot claim that one business, one trade, or one city proves an ideal sample size for everyone. The following is our proposed starting framework for local AI visibility tracking, intended to reveal obvious patterns rather than produce statistical certainty.

A 28-day baseline produces four weeks of observations and reduces the temptation to report a single favorable run. It is still only a starting point. A seasonal business, sparse prompt cluster, or highly variable answer surface may require a longer period before we draw a conclusion.

After the baseline, add prompts only when the evidence points to a decision: a competitor repeatedly appears for panel upgrades, a suburb produces materially different results, or sales teams hear a recurring buyer question. Do not add hundreds of weakly related prompts simply to create a larger dashboard.

What not to infer from a one-off result

One manual test remains valuable. It can reveal unexpected competitors, source domains, awkward business descriptions, or prompt wording worth tracking. It cannot reliably establish a rate, a durable ranking, or a platform-wide claim.

A single result does not tell us whether the outcome is stable across:

• another date or time of day;
• another local prompt with the same commercial intent;
• a second answer surface;
• changed source availability; or
• businesses outside the initial competitor set.

This matters especially when reporting to clients. “Our brand appeared in ChatGPT on October 11, 2026, for this saved prompt” is an accurate observation. “We rank number one in AI” is not, unless it specifies the product, query set, location context, time period, scoring rule, and comparison group.

For a broader evaluation checklist, see how to test whether AI search visibility tools are worth using. The central question is whether a tool preserves prompt-level evidence and repeatability, not whether it can produce an attractive aggregate score.

Turn a persistent competitor gap into an investigation

Measurement should lead to a restrained next step. If a competitor appears repeatedly in one prompt cluster, inspect the full answers and cited sources before changing your website.

We would investigate five questions:

1. Is the competitor actually recommended, or merely named in a long list?
2. Which source types recur: the competitor’s site, a directory, reviews, local news, or an editorial page?
3. Does our relevant page clearly explain the service, city, qualifications, proof, and customer fit?
4. Are citations pointing to accurate, accessible pages, or to stale third-party information?
5. Does the gap persist across enough comparable runs to justify action?

The answer may be a better service page, clearer service-area details, accurate structured business information, stronger reviews, or inclusion in a legitimate local resource. It may also be no change if the evidence is thin or unstable.

Where our product fits—and where it does not

Our AI Visibility Tracker is a local-first desktop workflow for saving prompt-level results across ChatGPT, Claude, Gemini, Perplexity, and Grok using the customer’s own API key. It is designed to make the evidence above inspectable: prompt, run date, mentions, citations, and named competitors rather than a black-box score alone.

That product description is separate from the measurement evidence in this report. A tracker can organize repeated observations; it cannot prove that an AI surface has a single fixed local ranking or guarantee that changing a page will change a future answer. To understand the web inputs behind AI answers, read how LLMs index the web and what AI search actually uses.

FAQ

How do you track whether a local business appears in AI answers?

Create fixed buyer-intent prompts that include a service and location, then run them repeatedly in ChatGPT, Gemini, Google AI Mode, Google AI Overviews, and Perplexity. Save the full answer, timestamp, settings, every named business, visible citations, and recommendation wording. Report mention and recommendation rates over a defined period, not from one screenshot.

Why do ChatGPT, Gemini, Perplexity, AI Mode, and AI Overviews recommend different businesses for the same prompt?

They are different answer surfaces with different product designs, retrieval behavior, source selection, and response formats. AI Mode and AI Overviews are both Google Search experiences, but they are still distinct surfaces to observe. A service-specific prompt can also produce a different shortlist than a broad “best” prompt on the same product.

What is the difference between an AI mention, a recommendation, and a citation?

A mention means the business name appears. A recommendation means the response frames it as a suitable choice for the buyer’s request. A citation is a visible source reference or link. A business can receive a citation without an endorsement, or be recommended without a visible citation, so we track all three separately.

How many prompts and repeat runs are needed to measure AI visibility reliably?

There is no universal threshold supported by one local-business study. Our proposed starting framework is 12–20 service-and-location prompts run daily, or on consistent weekdays, for an initial 28 days. Extend the period when results are sparse, seasonal, or inconsistent, and add prompts only when they map to a real business decision.

How do you calculate share of answer against local competitors?

Use open-set share of answer: divide your recommendation appearances by every business recommendation appearance in the saved answers, including businesses that were not originally tracked. You can also report monitored-set share for a fixed competitor list, but label it clearly. Always include raw counts and segment the result by prompt cluster and answer surface.