How to measure AI visibility without fooling yourself

Rank is the wrong unit

In search, a page has a position for a query at a given time. AI answers do not work that way. The same question produces different lists on different runs. SparkToro's research led its authors to conclude that tools giving a "ranking position in AI" are misleading. Parse found the top recommendation changed in 43.6% of 161,023 consecutive answers to the same prompt.

The useful unit is a rate: across N runs of a question, how often were you mentioned?

Five measures worth tracking

1. Mention rate. Share of runs in which the company is named, per question and overall.
2. Description accuracy. Whether the assistant describes your category, product and customers correctly. Score it simply: correct, partly correct, wrong.
3. Citation sources. Which URLs are cited when you are mentioned, and when competitors are. This shows where to be present.
4. Founder mention. Whether the founder is named, especially on direct questions about the company.
5. Share of shortlist. For recommendation questions, how often you appear among the named options compared with specific competitors.

How many runs

For a monthly check, 3 to 5 runs per question per assistant gives a rough picture. For a decision you will spend money on, use more. Keep the method the same each time: same questions, same assistants, same mode, clean sessions.

Separate signal from noise

Small movements are usually noise. A move from 2 mentions out of 10 to 3 out of 10 is not evidence of anything. Look for changes that persist across two or more measurement periods and appear across several questions.

Watch the inputs too

AI visibility is an outcome. It helps to track the inputs you control alongside it:

• Number of pages that directly answer a question in your set
• Consistency of company descriptions across your site and profiles
• New third-party mentions: reviews, interviews, partner pages
• Crawl access for OAI-SearchBot, PerplexityBot and Googlebot

If inputs improve and visibility does not, wait a cycle before changing course. If neither moves, the work is not getting done.

Connect it to buyers

AI visibility only matters if it reaches buyers. Ask new prospects how they found you and what they checked, and record the answers exactly. In TrustRadius's 2026 survey, 94% of buyers said they fact-check AI output. Being mentioned is the start; what the buyer finds when they check matters too.

Questions to ask a vendor

• How many runs per question, per assistant?
• Do you show the raw answers?
• How do you handle personalisation and location?
• What change do you treat as significant?
• Which assistants and modes are covered?

If a report cannot answer these, treat its numbers as illustrations.