How to Measure AI Visibility Without Buying a Meaningless Score

One score hides several different outcomes

A store can be mentioned without a link, cited without a recommendation, or recommended with an incorrect price. Those outcomes should not all become “visibility = 72.”

Use four separate measurements:

Measurement	Question
mention	was the brand or product named?
citation	did the answer link to the canonical store domain?
accuracy	were decisive facts correct?
competitive position	which alternatives appeared for the same prompt?

Add a fifth status for provider failure. A timeout or API error is not a non-mention.

Build a repeatable prompt set

Create 15 to 25 prompts across:

• category discovery
• product use case
• compatibility
• comparison
• purchase constraints
• brand verification

Keep wording stable. Record language, market, platform, visible mode, and date. Personalized chat history can change results, so use a consistent test setup.

Store the evidence

Your spreadsheet needs:

test_id
prompt
platform
run_time
market
completed_status
brand_mentioned
product_mentioned
store_domain_cited
competitors
fact_errors
source_urls
notes

This is a simple observation log, not an index of everything the platform shows every user.

Report what changed

Good monthly reporting:

Across 20 fixed prompts, 18 completed.
The store was mentioned in 7 completed answers and cited in 4.
Two answers used stale availability.
Three cited marketplace pages instead of the canonical store.

Bad reporting:

Your GEO score increased from 61 to 74.

The second statement has no meaning unless the scoring method, denominator, failure handling, and evidence are visible.

Connect observations to store work

A cited marketplace page suggests a canonical evidence or authority issue. Wrong stock suggests page, Offer, or product-feed drift. A competitor that consistently answers a missing attribute suggests a content gap.

Do not change the store after every fluctuating answer. Make a change only when the evidence points to a merchant-controlled input.

The technical monitoring framework is in Build an AI visibility spreadsheet and Avoid conclusions from one AI answer.