Why can't I just track rankings?

Because there is no ranking. An assistant names three businesses in a paragraph — there is no position four, no results page, and no stable ordering. The equivalent metric is whether you get named at all, measured across many questions and repeated runs.

How often should I measure?

Monthly is right for most local businesses. AI answers drift as sources change, and monthly measurement catches a slide before it costs you a season. Weekly is noise; quarterly is too slow to react.

Why measure the same question more than once?

Because assistants sample rather than return a fixed result. Asking once tells you almost nothing. Asking the same question several times and counting how often you appear turns a coin flip into a rate.

What's a good mention rate?

It depends on market size and competition, so judge against your own baseline and your named competitors rather than an absolute number. Going from 2 of 25 to 12 of 25 over a quarter is meaningful progress regardless of what any benchmark claims.

Guides/Measurement & buyingupdated 2026-08-04 · 5 min

AI Visibility KPIs for a Local Service Business

The short answer

The core AI visibility metric is mention rate — how often assistants name your business across a fixed set of customer questions. Track it monthly alongside share of voice against named competitors, per-assistant breakdown, accuracy of what is said about you, and self-reported AI-sourced leads.

The first thing an SEO does is set up rank tracking. That instinct fails here, because there is no rank.

An assistant names three businesses in a paragraph. There is no position four. There is no results page. Two people asking the same question can get different answers.

So the metrics have to be different. Here are the five worth keeping.

1. Mention rate — the core number

Definition: across a fixed set of customer questions, run across the major assistants, what percentage name your business?

How to collect it: define twenty to twenty-five questions your customers actually ask — a mix of generic local, problem-led, branded, and comparative. Run each across ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews. Count mentions. Divide.

Why repeated runs matter: assistants sample from a distribution. One run of one question is a coin flip. The same question five times gives you a rate, and rates are what you can track.

What to do with it: record it monthly. The trend is the whole point. An absolute number without a baseline tells you nothing; a move from 3 of 25 to 14 of 25 over a quarter tells you the work is working.

2. Share of voice against named competitors

Definition: across the same question set, which businesses get named, and how often, relative to you?

This is the most actionable number in the set, and it is more useful than your own mention rate in isolation.

Keep a running tally. If the same four competitors appear in every answer and you never do, you have both a target and a research subject: go look at what they have that you do not. Usually it is directory breadth, review specificity, or a local roundup listing.

A useful sub-metric: how many distinct businesses get named across your question set. A market where the same three names appear every time is consolidated and hard to break into. A market where twelve different businesses appear is fragmented and winnable.

3. Per-assistant breakdown

Definition: your mention rate on each assistant separately.

This is diagnostic rather than a scorecard. The assistants weight different things, so the pattern of your failures tells you where to work:

Weak on	Likely cause	Where to work
Gemini and AI Overviews	Incomplete Google Business Profile	Google profile depth, reviews, proximity signals
Perplexity	Missing from the directories and roundups it cites	Directory presence, local "best of" lists
ChatGPT	Thin entity, few third-party mentions	Directory breadth, community mentions, press
Claude	Weak verifiable credentials	Licensing, certifications, consistent facts
All five	Crawler blocked or JavaScript-only site	Technical fixes first

Scoring lower on Claude than on the others is normal — it names fewer businesses overall. Scoring zero everywhere usually means a technical problem, not a marketing one.

4. Accuracy of what is said about you

Definition: when an assistant describes your business, what does it get wrong?

Run the branded questions — "is [business] any good," "tell me about [business]," "is [business] licensed" — and note every error. Wrong phone number, old address, services you no longer offer, wrong hours, confusion with a similarly named business.

Every error exists somewhere public. This metric is a direct list of listings to fix, and it is the fastest-closing gap in the set.

Track it as an error count. Getting it to zero is achievable in a few weeks.

5. Self-reported AI-sourced leads

Definition: how many callers say they found you through an AI assistant.

How to collect it: a "how did you hear about us" dropdown on every form with "ChatGPT / AI assistant" as an explicit option, plus whoever answers the phone asking and logging it.

This is the only metric that captures the zero-click calls, which are the majority of the actual business impact. Analytics will never see them.

It is also the number that connects the leading indicator to revenue, which is what makes the rest of the exercise defensible.

What to skip

AI referral traffic as a headline metric. Track it — sessions from chatgpt.com, perplexity.ai, claude.ai, gemini.google.com, copilot.microsoft.com — but understand it undercounts badly. Google AI Overview clicks are indistinguishable from ordinary organic, and the highest-value path produces a phone call with no session at all.

Sentiment scoring. Interesting, rarely actionable at this size.

"AI visibility score" from a tool with an undisclosed formula. If you cannot see the questions asked and the raw answers, you cannot act on the number. Insist on seeing the verbatim responses.

A simple monthly review

Thirty minutes, once a month:

Run the question set. Record mention rate.
Record share of voice — who else got named, how often.
Note the per-assistant breakdown and which is weakest.
Run branded questions, list every factual error.
Pull self-reported AI leads from your CRM.
Pick one thing to fix based on the weakest number.

Then do that one thing. The point of measurement is the decision it produces, and one fix a month compounds faster than a spreadsheet nobody acts on.

What good progress looks like

Rough shape, for a business starting near zero and doing the work:

Month 1: technical fixes land. Mention rate may not move yet.
Month 2: Google Business Profile and directory work starts showing on Gemini and AI Overviews first.
Month 3: Perplexity moves as directory and roundup presence registers.
Months 4–6: ChatGPT and Claude follow as the entity thickens and third-party mentions accumulate.
Ongoing: community mentions and review specificity compound slowly and durably.

The assistants that move first are the ones tied to Google data. The ones that move last are the ones that depend on entity breadth. That ordering is consistent enough to plan around.

What to do this week

Write down twenty-five customer questions in the four categories.
Get a baseline mention rate and share of voice.
Add "ChatGPT / AI assistant" to your lead-source options.
Tell whoever answers the phone to ask and log it.
Put a monthly calendar reminder on the review.

The fast way to get the baseline

Doing this by hand across five assistants is a hundred and twenty-five prompts and a spreadsheet. Our free scan runs all twenty-five questions across all five assistants at once, grades each separately, and lists every competitor named in your place.

About fifteen seconds, free, no account. Run it monthly and the trend takes care of itself.

Common questions

Why can't I just track rankings?: Because there is no ranking. An assistant names three businesses in a paragraph — there is no position four, no results page, and no stable ordering. The equivalent metric is whether you get named at all, measured across many questions and repeated runs.
How often should I measure?: Monthly is right for most local businesses. AI answers drift as sources change, and monthly measurement catches a slide before it costs you a season. Weekly is noise; quarterly is too slow to react.
Why measure the same question more than once?: Because assistants sample rather than return a fixed result. Asking once tells you almost nothing. Asking the same question several times and counting how often you appear turns a coin flip into a rate.
What's a good mention rate?: It depends on market size and competition, so judge against your own baseline and your named competitors rather than an absolute number. Going from 2 of 25 to 12 of 25 over a quarter is meaningful progress regardless of what any benchmark claims.

Stop guessing

See exactly what the five AIs say about your business.

Twenty-five real customer questions across Claude, ChatGPT, Gemini, Perplexity, and Google AI Overviews. Free, about fifteen seconds, no account.

Run the free scan