Noma by Nurdd (NomaHQ)Noma
Sign up

Free standard · PDF

How to measure AI visibility with numbers you can defend

A measurement standard for AI visibility: a metric dictionary with explicit denominators, prompt-set design, the sample-size arithmetic for reruns, engine settings to control, GA4 referral tracking, and a report specification with worksheets.

  • Metric dictionary with numerator and denominator for every KPI
  • Sample-size table: runs needed for a ±5 or ±10 point margin
  • Prompt-set register with versioning rules
  • GA4 channel group for AI assistant referrals
  • Report checklist and monthly scorecard template

Get your free copy

Instant download, plus a copy by email. No spam: one email with your link, and a reply from a person if you write back.

The short answer

Measure AI visibility by running a fixed, versioned set of unbranded buyer prompts on each engine several times, then reporting mention rate, share of voice, citation share and description accuracy, each with its numerator and denominator stated. Give every figure a margin of error, record engine settings and dates, and call a change only when it exceeds that margin.

For SEO leads, analysts, marketing ops and insights teams at brands and large companies

What you will learn

  1. Chapter 1

    Why AI visibility numbers disagree

    • AI answers are generated fresh each time, so any visibility figure is an estimate from a sample, not a fixed fact.
    • Most disagreements between tools come from different denominators, prompt sets, competitor sets or engine settings, not from errors.
    • Store every response with its prompt, engine, settings, timestamp, text and citations, so any figure can be recomputed.
    • Freeze definitions for a reporting period and log changes, or the trend line will measure your edits instead of the market.

  2. Chapter 2

    The metric dictionary

    • Every metric needs a written numerator and denominator, and the same matching rules applied to you and to competitors.
    • Share of voice depends on the competitor set, so publish the set and restate history when it changes.
    • Conditional metrics such as average position when named can improve while visibility falls; report them next to mention rate.
    • Own-domain citation rate has at least three common denominators; name the one you use.
    • Model-based sentiment and accuracy scores need calibration against human scoring before they go in a report.

  3. Chapter 3

    Designing a prompt set for measurement

    • Keep a fixed measurement set separate from exploratory research prompts, and promote prompts only at version changes.
    • Build prompts from buyer evidence and stratify them by stage, topic, persona, market and intent.
    • Branded prompts measure reputation and accuracy, not visibility; never mix them into the visibility headline.
    • Version the set, keep a never-changing core panel, and run old and new sets side by side when you change it.

  4. Chapter 4

    Variance, sample size and calling a change

    • The 95% margin on a rate is 1.96 × √(p(1 − p) ÷ n); at a 50% rate, 100 independent responses give about ± 9.8 points and 400 about ± 4.9.
    • Halving the margin needs four times the sample, so report the headline to about ± 5 points and accept wider margins on sub-groups.
    • Reruns of the same prompt are correlated; compute margins across prompts, not across raw responses.
    • Use at least three runs per prompt to separate held, contested and absent prompts.
    • Call a change only when it exceeds the margin on the difference, and confirm it in a second period.

  5. Chapter 5

    Engines and settings: control what you can, record the rest

    • List the variables that change an answer, fix the ones you can, and record the rest on every response.
    • Location, memory and account state affect answers in ChatGPT, Gemini, Perplexity and Claude according to their own documentation; use clean, fixed conditions.
    • Apps and APIs are different instruments; choose one per engine and say which in every report.
    • Record whether each answer used a web search, because a drift in search share can look like a change in visibility.
    • Report engines separately; any combined figure needs stated weights.

  6. Chapter 6

    Connecting visibility to traffic and pipeline

    • Visibility, referral traffic and pipeline influence need three different instruments; report them separately.
    • A GA4 custom channel group with a source regex, placed above Referral, isolates AI assistant visits; check the regex against your own referral list.
    • Traffic from Google's AI Overviews and AI Mode is counted inside ordinary search traffic and cannot be split out with referrer rules.
    • Self-reported attribution captures influence that never clicks; ask without leading, and report it with its denominator.
    • Never convert a mention rate into revenue; show the evidence side by side and say "consistent with" rather than "caused".

  7. Chapter 7

    The reporting standard

    • Every report states its prompt set and version, runs, dates, engines, settings, search state, definitions, competitor set and margins.
    • Use the same monthly structure every time, with a method appendix that makes the first page checkable.
    • A change to the method must be run side by side or restated, never shown as a change in results.
    • Cut one dataset for different readers rather than producing different numbers for each.

6 worksheets inside

The chapters explain the method. The worksheets are what you fill in to run it: print them, or copy the tables into a document of your own.

  • W1Metric dictionary card
  • W2Prompt-set register
  • W3Sample-size table and margin calculator
  • W4Report checklist
  • W5Monthly scorecard
  • W6Change log
Get the standard

Questions

How do you measure AI visibility?
Run a fixed set of unbranded buyer prompts on each AI engine several times, store every answer, and count how often your brand is named, how it is described and which sources are cited. Report each metric with its numerator, denominator, sample size, settings and margin of error.
What are the main AI visibility KPIs?
Mention rate, recommendation rate, share of voice, first-named rate, citation share, own-domain citation rate, description accuracy and sentiment. Report the presence and share metrics on unbranded prompts, and the accuracy and sentiment metrics on branded prompts.
How is share of voice in ChatGPT calculated?
Count the responses that name your brand and divide by the total of responses naming each brand in a declared competitor set. Because the denominator depends on which competitors are included, always publish the competitor set, and restate history if it changes.
How many prompts and runs do I need?
It depends on the precision you want. At a rate near 50%, about 100 independent responses give a margin of roughly ± 10 points and about 385 give ± 5 points. Reruns of the same prompt are correlated, so prefer more prompts, and use at least three runs per prompt to tell stable prompts from contested ones.
Why do different AI visibility tools give different numbers for the same brand?
They usually use different denominators, prompt sets, competitor sets, engine surfaces or settings. Two figures are only comparable when all of those match, so ask each tool for its method before comparing.
How can I track traffic from ChatGPT and other assistants in GA4?
Create a custom channel group with an "AI assistants" channel defined by a source regex covering the assistants' domains, and place it above Referral, because GA4 assigns traffic to the first matching channel. Check the regex against your own referral sources, and treat the result as a floor.
Can I see traffic from Google AI Overviews separately?
Not with referrer rules. Google's documentation says traffic from its AI features is included in overall search traffic in Search Console under the Web search type, and in GA4 it arrives as Google organic search.
When is a change in AI visibility real?
When the difference between two periods is larger than the margin of error on the difference, and it holds in a second period. Check model releases and changes to your own method before attributing it to your work.
Is The AI Visibility Measurement Standard free?
Yes. Give a first name and a work email and the download starts at once. We also email you the link, and one person from the Noma team follows up a day later; reply and say so if you would rather not hear from us.
Why do you ask for a work email?
The playbooks are written for teams at companies, and a work address tells us which company is reading, so a follow-up can be about your brand rather than generic. Personal and temporary inboxes are not accepted.
What format is it?
A PDF of 34 A4 pages, laid out to print, with worksheets you can fill in by hand or copy into a document of your own.

More playbooks

Updated 20 October 2026 · Written by the Noma team · All resources