How to Measure AI Search Visibility
By ReleVantage. Last reviewed August 2026.
Short answer
Measure AI search visibility by defining a fixed prompt set, running it against each platform on a schedule, and scoring every answer on eight separate dimensions: visibility, mention position, sentiment, accuracy, citations, source diversity, qualified traffic, and business outcomes. A single visibility score hides all of the useful detail, so keep the dimensions apart.
Step one: build a prompt set you will not keep changing
The prompt set is the measurement instrument. If it changes every month, nothing is comparable. Build it from real buyer language and freeze the core set, adding new prompts as a separate tracked group.
- Category prompts: what the buyer asks before knowing any brand
- Comparison prompts: your brand against named alternatives
- Factual prompts: hours, coverage, pricing model, integrations, capacity, credentials
- Constraint prompts: location, budget, company size, dietary need, urgency
- Branded prompts: what the model says when asked about you directly
Step two: score eight dimensions separately
Visibility
The share of tracked prompts where your brand appears in the answer at all, reported per platform.
Mention position
Where in the answer you appear. First named carries different weight than a closing list, so record order rather than presence alone.
Sentiment
How the model characterises you. Being named as the expensive option is visibility with a problem attached.
Accuracy
Whether the stated facts are correct. Track errors individually, because each one traces back to a source you can usually correct.
Citations
Which URLs the answer attributes, split between your own pages and third-party pages that discuss you.
Source diversity
How many independent domains support your presence. Dependence on one source is fragile, since a single change can remove you.
Qualified traffic
Sessions arriving from assistant referrers, judged on engagement and conversion rather than volume.
Business outcomes
Enquiries, bookings, demos, and pipeline associated with those sessions, reported with their attribution limits stated.
Step three: keep the run conditions consistent
- Run the same prompts, in the same wording, on the same schedule
- Record the full answer text, not a summary, so scoring can be re checked
- Note platform, date, and any location context used in the run
- Run each prompt more than once, because generated answers vary between runs
- Score with a written rubric so different reviewers reach the same result
- Keep every historical run: the trend is the finding, not any single answer
Step four: connect answers to site behaviour
Assistant referral data is incomplete by design. Some platforms send an identifiable referrer, some send none, and some answers create demand that arrives later as a branded search or a direct visit.
Use first party tracking to capture landing page, referrer, and campaign parameters at the session level, pass that context into your enquiry forms, and add a self reported source field. Treat the result as directional evidence rather than a closed attribution model, and say so in reporting.
Step five: report change, not a single number
A useful report shows the baseline, the current run, what changed, which specific work preceded the change, and what is scheduled next. Where a movement cannot be explained, say that plainly.
Correlation is not proof. Platforms update independently of anything you ship, so avoid claiming credit that the data does not support.
Frequently asked questions
How often should prompts be re run?
Weekly is a reasonable default for most categories. Volatile or highly competitive categories justify a shorter interval for a smaller priority group.
Why do answers differ between two runs of the same prompt?
Generated answers are probabilistic and retrieval can change between runs. Running each prompt multiple times and reporting the pattern is more reliable than reading a single response.
Can I attribute revenue directly to AI search?
Only partially. First party session data and self reported source fields give directional evidence, and any report that claims complete attribution is overstating what the data supports.
Keep reading
Want this applied to your brand?
We start with a recorded baseline of how assistants describe you today, then work through the fixes in priority order. Results vary and placements are not guaranteed.