How to measure your brand's AI visibility: a repeatable workflow

You can measure AI visibility with a spreadsheet and an afternoon. This workflow works at that scale and scales up unchanged.

#1. Choose the scope

Pick one business area: a product line, audience or market. Visibility is a property of a question set, not a company.

#2. Build the prompt panel

Write 20–50 questions across the buying journey — problem, shortlist, comparison, constraint, purchase — using language from sales calls, support and search data. Keep branded and non-branded questions separate. See how to build a prompt panel.

#3. Decide the collection conditions

Fix and record:

• Engines you will check.
• Country and language.
• Number of runs per prompt per period (more than one).
• Date range.

#4. Collect and store answers as given

For each run, save the prompt, engine, date, the full answer text and the list of cited URLs. Don't summarise or edit. Mark failed runs as failed.

#5. Code each answer

For every valid answer, record:

• Brand mentioned? and position in any list.
• Owned domain cited?
• Recommended? (explicit recommendation for the prompt's use case)
• Competitors mentioned.
• Accuracy: is what the answer says about you right?

Two people coding a sample independently is the cheapest quality check.

#6. Compute the metrics

Show a range across repeated runs. See share of voice for the sample-size math.

#7. Diagnose

Pair the metrics with:

• Access: can the relevant crawlers reach your pages? (crawler guide)
• Sources: which domains are cited, and who's missing? (citation tracking)
• Demand: what do people search for around these questions? (Search Console)
• Outcomes: identifiable AI referrals and what they do (GA4). (GA4 guide)

#8. Report

A trustworthy report states:

• The panel, engines, location, language and dates.
• Valid, failed and excluded runs.
• Metrics with ranges.
• Three example answers (one good, one bad, one surprising).
• What you will change, and when you'll re-measure.

#9. Change one thing, then repeat

Make a bounded change, record the date, and re-run the same panel under the same conditions. Describe results as evidence, not proof.

#Reporting template

Scope: <business area>
Period: <dates>
Panel: <n> prompts (<stages>)
Engines: <list> · Locale: <country/language>
Runs: <valid>/<failed>
Mention rate: <x>% (range <a–b>%)
Owned-citation rate: <x>%
Recommendation rate: <x>%
Top cited domains: <list>
Gaps: <domains citing rivals not us>
Accuracy issues: <list>
Next change: <what, when>

#Sources

• Aggarwal et al., GEO: Generative Engine Optimization.
• Google Search Central, Optimizing for generative AI features in Search.