Measuring AI visibility honestly

Watch · 3:03

• 0:00Why it is hard to measure
• 0:52Six rules
• 1:48Connect visibility to outcomes
• 2:08How AI Rank Lab helps

In shortMeasuring AI visibility means running a fixed set of buyer questions on the engines you care about, more than once, and comparing like with like over time. AI answers vary between runs, so a single run is a sample, not a score. Keep the question set, engines and competitor list stable, report each engine separately, and treat anything an AI estimated, rather than measured, as an opinion.

Why AI visibility is hard to measure

A search ranking is fairly stable from one day to the next. An AI answer is generated fresh each time: the engine may run different searches, retrieve different pages and phrase its recommendation differently. The same question, asked twice in an hour, can name different brands.

AI Rank Lab's own data shows how large that effect is. We looked at 991 successful runs of 45 prompt-and-engine pairs that were each run at least five times between February and October 2026:

• In 20 of the 45 pairs, whether the tracked brand was mentioned changed at least once. Only 19 pairs mentioned the brand every time, and 6 never did.
• Between two consecutive runs of the same pair, the cited domains overlapped by a median of 0.41 (Jaccard similarity), so consecutive answers typically shared well under half of their sources.
• Over all its runs, a typical pair cited about 2.7 times as many distinct domains as a single run did.

This is a small, anonymised sample from accounts on our platform, not a study of the whole web. But the direction is clear: one run tells you what happened once.

Six rules for honest measurement

1. Fix your question set. Choose 10 to 30 questions real buyers ask: category questions ("best X for Y"), comparisons ("X vs Y"), problems ("how do I fix Z") and a few branded ones. Keep the set stable. Report branded and unbranded questions separately, because branded ones inflate every number.
2. Pick engines on purpose. Run the engines your buyers use and report each one separately as well as the total.
3. Repeat before you conclude. Run each question more than once and compare averages across runs, not single results. The smaller the sample, the bigger a change must be before you believe it.
4. Compare like with like. Same questions, same engines, same competitor list, similar time windows. If anything changed, say so next to the number.
5. Separate measured from estimated. A mention counted in a stored answer is a measurement. A score an AI model assigns to your brand is an assessment. Don't mix them in one chart.
6. Say what wasn't measured. If an engine wasn't run this month, the result is "not measured", not zero.

Common mistakes

Connect visibility to outcomes

Mentions and citations are leading indicators. Check them against what happens on your site: AI-referred visitors in your analytics (lesson 4) and, for Google, Search Console's Search Generative AI performance reports, which since 2026 show how often your URLs appeared in AI Overviews and AI Mode (Google Search Central (opens in a new tab)). Those reports show impressions, not which prompts or competitors were involved, so they complement prompt tracking rather than replace it.

How AI Rank Lab helps

In Brand visibility, prompts run when you run them, and every run is stored, so you can repeat a question and compare. Failed engine calls are saved but not charged. Date filters cover whole days in UTC. Progress reports compare two points in time like for like and label anything that wasn't measured instead of showing zero. The only automatic re-check is the weekly citation tracker for prompts you opt into from Action plan.