Measurement · 9 min read
How to measure AI visibility (a practical method)
AI visibility is measurable, but not with rank trackers. You need a repeatable prompt set, a sampling schedule, and three scores you can defend in a meeting.
What exactly is AI visibility?
AI visibility is how often, how favourably, and on what evidence an AI assistant names your brand when someone asks a question in your category. It replaces the old question 'where do we rank?' with three sharper ones: are we in the answer, how are we framed, and which sources put us there.
It is not a single platform metric. Each engine builds answers differently, so visibility is always measured per engine, per market and per question type.
Step 1: build a prompt set that mirrors real buying questions
Start with 40–80 prompts grouped by intent: category discovery ('best tools for X'), comparison ('X vs Y'), evaluation ('is X good for enterprise teams'), and branded ('what does X do'). Write them the way a buyer speaks, not the way a keyword tool phrases things.
Freeze the set. The moment prompts change, month-over-month comparison stops meaning anything. Add new prompts as a separate cohort instead of editing the baseline.
Step 2: sample answers on a schedule
Generated answers vary between runs, so a single response is an anecdote. Run each prompt several times per cycle, on the same weekday and cadence, and store the full answer text plus every cited URL.
Keep the raw answers. Scores get recalculated as your definitions improve; the underlying answers are the asset.
Step 3: score share of voice, sentiment and citation coverage
Share of voice: the share of sampled answers in a prompt group that name your brand at all. Sentiment: how the answer characterises you — recommended, mentioned in passing, or framed with a caveat. Citation coverage: the share of answers that link to a page you own or influence.
Report all three together. High share of voice with weak citation coverage means the engine is repeating hearsay about you, which is fragile. Strong citations with poor sentiment means the evidence is there but reads badly.
Step 4: add first-party evidence
Server and CDN logs show which AI crawlers fetched which pages and what response they got. Search Console shows what the wider search surface already associates with you. Together they explain whether an absence is a content problem or an access problem.
This is also where most quick wins hide: blocked crawlers, redirect chains and thin pages that no engine can safely quote.
Step 5: report movement honestly
Always publish the estimate with its sample size and a plausible range. A move from 14% to 17% on twelve observations is noise; the same move on three hundred is a result.
Set review cycles long enough to outlast model updates — monthly for reporting, weekly only for diagnostics.
Frequently asked questions
Can I use rank tracking tools to measure AI visibility?
Not on their own. Rank tools measure link positions; AI visibility depends on whether a generated answer names you and what it cites, which has to be sampled from the answers themselves.
How many prompts are enough?
Enough to cover each intent group with several prompts and repeated runs — commonly 40–80 prompts with multiple samples each. Coverage of intents matters more than raw prompt count.
How often should I measure?
Monthly for trend reporting, with lighter weekly checks on your highest-value commercial prompts to catch sudden changes.
Why does my score change when nothing on my site changed?
Answer generation is probabilistic and engines update frequently. That is exactly why estimates should carry a range and why a frozen prompt set matters.
Sources
- OClever sample measurement methodology, 2026
- Google Search Central: Creating helpful, reliable, people-first content
- Google Search Console documentation