Skip to content
Mentioned First

Methodology

How we measure AI visibility, in plain English.

AI answers are easy to screenshot and hard to measure. These are the rules every scan, report and export follows, including the limits of what anyone can see, and the things we never do.

Metric definitions, version 1

The headline number

AI visibility score

= 100 ×answers that recommend youcomplete discovery answers

Worked example

Recommended in 12 of 40 complete answers

30= 100 × 12 ÷ 40

  • Errors and incomplete answers are left out of both numbers
  • The count behind every rate is always shown

Engines switch on one at a time. Each engine is measured only once its integration has been verified, and every report lists the engines it asked. Numbers marked as samples on this site come from fictional businesses. The rules below apply to every live answer.

Measurement methods

AI answers reach people in different ways, so each engine is measured the most faithful way available, and every result is labeled with its method. An API test is never presented as what a shopper saw in an app.

  • App observed

    The closest to what people see inside the app.

    Captured from the consumer app through a data partner, logged out, with the market written into the question.

    Used for

    • ChatGPT
    • Gemini
  • Search observed

    What a real search page showed for your market.

    Captured from a live search results page for the chosen location. If no AI answer appears, that is recorded as its own result.

    Used for

    • Google AI Overviews
    • Google AI Mode
    • Microsoft Copilot
  • API benchmark

    A consistent test of the model with live search on.

    Asked through the official API with live web search turned on. Close to the app, but not identical to what every consumer sees.

    Used for

    • Perplexity
    • Claude
    • Grok

Question design

What you ask decides what you measure. Questions are written for each industry and intent, always name the market, and come in three families that are kept apart.

  • Discovery

    Feeds your visibility score

    Unbiased questions about your category and market. They never include your business name, website, phone number, address, reviews or a list of competitors.

    Example“Which roofing companies in Plano, TX help homeowners with insurance claims after hail damage?”

  • Brand accuracy

    Separate report

    Questions that name you on purpose, to check facts like hours and service areas. They run on their own and never count toward your score.

    Example“Is Summit Ridge Roofing open on Saturdays?”

  • Informational

    Tracked separately

    Topic questions that may cite a useful source without recommending anyone. Useful for spotting citation chances, kept apart from recommendations.

    Example“How do I file a homeowners insurance claim after hail damage in Texas?”

Intent families

A good set covers the different reasons people ask. We only use the ones that fit an industry: there are no “urgent” questions for restaurants or gyms.

General
Who's best overall
Service
A specific service you offer
Urgent need
Something that can't wait, only where that's realistic
Decision factor
What people filter on, like reviews or upfront pricing
Comparison
Who to compare before choosing
Nearby
Nearby neighborhoods and service areas
Customer concern
The worry that decides who gets the call

  • Written for the market

    Every discovery question names the city or region being measured, so the right market is checked even when an engine can't be told where the customer is.

  • Frozen for tracking

    Tracked questions don't silently change. If your services, market, wording or an engine's method changes, we start a new baseline or clearly mark the change.

How answers are read

A result is only as good as the reading. Nothing is decided until the complete answer has been read.

  • Complete answers only

    We wait for the full answer and read it start to finish. A mention in the last line counts. Refusals, cut-off answers and failed searches are set aside, never scored as “not visible”.

  • Careful identity matching

    Your name is matched together with context like your website, city or phone number. Similar names and businesses in other cities aren't counted as you. When it's unclear, the answer is held for review instead of guessed.

  • Three different facts

    “I can't vouch for them” names a business without recommending it. Your website being used as a source doesn't mean you were recommended. Recommended, mentioned and cited are recorded separately.

  • Never your own competitor

    Old names, aliases and alternate spellings are resolved to one identity, so you never show up as a rival to yourself.

The eight answer statuses

Every answer ends up with exactly one status. Each one has an icon and a label, so the meaning never depends on color.

The eight answer statuses and whether each one counts toward scores
StatusCounts in scores?
RecommendedNamed as a recommended option in a complete answer.Yes. Adds to your score
MentionedNamed in the answer, but not as a recommendation.Yes. Counted, not as a recommendation
CitedYour website was used as a source, but you weren't named.Yes. Also feeds your citation rate
Not visibleA complete answer was read in full and you weren't in it.Yes
No AI answerThe search page showed no AI answer for this question and location.Yes, as an observed absence
Needs reviewA name was found but identity is uncertain. Excluded until reviewed.No. Held until reviewed
IncompleteRefused, cut off or missing search. Not counted as "not visible".No. Shown in coverage
Scan errorThe engine failed. Excluded from scores, shown in coverage.No. Shown in coverage

Metric definitions v1

Every surface, from the dashboard to the PDF to a CSV export, reads the same numbers from one set of definitions. Each rate is shown with the counts behind it, and when there's nothing to measure it says “Unavailable”, never 0%.

Metric definitions, version 1
Metric and definition
AI visibility scoreYour headline number, kept simple on purpose.100 × recommendation rateNo hidden weights, sentiment bonuses or difficulty adjustments.
Recommendation rateHow often AI recommends you.Answers that recommend you ÷ scorable discovery answers
Mention rateHow often you're named at all, recommended or not.Answers that name you ÷ scorable discovery answers
Citation rateHow often your website is used as a source.Answers citing your site ÷ scorable answers that listed sourcesAnswers with no sources aren't counted as zero citations.
Mentioned firstHow often you're the first business recommended.Answers where you're recommended before any other business ÷ scorable discovery answers
Share of voiceYour slice of all the recommendations in the same answers.Your recommendation appearances ÷ every business's recommendation appearancesEach business counts once per answer. It describes sampled answers, not market share or demand.
Average rankWhere you sit in numbered lists.Mean position in numbered lists where you were recommendedUses numbered lists only. Unordered answers never get a rank, and the count is shown.
CoverageHow much of the plan we could measure.Scorable answers ÷ planned answers, overall and per engine
Publication policyWhen a combined score is shown.Overall coverage ≥ 90% and every engine ≥ 80%Below that, the combined score is withheld and each engine's results are shown on their own. It's a product rule, not a statistical guarantee.

Scorable discovery answers are complete answers, plus search pages that showed no AI answer at all. Errors, incomplete answers and answers held for review are left out and shown in coverage. When these definitions change, they get a new version number, and past reports keep the version they were made with.

Variability and tracking

Ask the same question twice and AI can give two different answers. That's why one run is a snapshot, and tracking is the measurement.

  • One run is a snapshot

    A free scan shows where you stand today. It's a starting point, not a verdict.

  • Trends compare like with like

    Each run uses the same questions, engines and market. Changes are calculated on answers that were scorable in both runs.

  • Your first scan is a baseline

    No improvement or decline is shown until there's a comparable second scan.

  • Small moves stay small

    We show how many answers sit behind each number, and we don't call a change significant when the sample can't support it. Retries never count as extra samples.

Percentage points, not percentages

Rates move in percentage points. A change from 20% to 25% is shown as +5 points, because calling it +25% would make a small shift sound bigger than it is.

Worked example

Your mention rate goes from 20% to 25% between two weekly scans.

  • We say: +5 points
  • Not: +25%

Limitations

No tool can see every answer every person gets. Here's what we can't see, so you can weigh the numbers properly.

  • API answers aren't the app

    API benchmarks use the model with live web search on. They're close to the consumer app, but not identical to what every person sees.

  • Answers are personal

    Signed-in history, memory and preferences change what people see. We ask without personal history wherever the method allows.

  • Location matters

    We write the market into every question and use location settings where an engine supports them. An API call doesn't know a customer's real location.

  • Answers change over time

    Models, search results and sources update constantly. Each result describes the moment it was measured.

  • Sometimes there's no AI answer

    Search pages don't always show an AI answer. When they don't, that's recorded as its own result, not as an error.

  • A sample, not a census

    We measure a fixed set of questions. Customers ask many more, in many more ways.

Data handling

Measurement is read-only. We look; we don't touch.

  • We never change your listings or website

    Scans only read public information and AI answers. You don't need to connect your Google Business Profile, and nothing is ever edited on your behalf.

  • Full answers are kept as evidence

    So every number can be traced back to exactly what the AI said, when, and how it was measured.

  • Required attribution stays visible

    Data and AI provider attribution stays on every report, including agency-branded ones. White labeling never removes it.

  • Reports are private unless you share them

    Sharing is your choice, and shared links can be turned off. We don't ask for sensitive details like patient records or case files.

More detail in our privacy policy, which is still a draft pending legal review.

What we never do

Short list, firm rules.

  • We never Put your business name in a discovery question
  • We never Count an error, timeout or refusal as “not visible”
  • We never Show 0% when there's no data to measure
  • We never Present an API benchmark as what shoppers saw in an app
  • We never Swap in a different engine when one is unavailable
  • We never Change your listings, profiles or website
  • We never Guarantee that AI will recommend you
  • We never Invent a weakness to sell you something
  • We never Rewrite past reports when our methods change

See every rule in the sample report.

The sample report shows every rule on this page at work: full answers, statuses, coverage and the counts behind each number.

Open the sample report Run a free scan

The sample report uses fictional businesses.