To measure AI share of voice, define a keyword set that reflects your category, list the competitors you compete with for those answers, query each AI engine for each keyword, and count how often each domain is cited — then express your citations as a percentage of the total. The non-negotiable step is labeling your signals: data you can read directly from an engine (e.g. Perplexity's source links) is 'measured', while presence inferred from a search proxy is 'estimated'. Keep the methodology constant over time so trends are comparable, and treat the number as a directional competitive signal rather than a precise census.
How do you choose keywords and competitors?
Your keyword set defines what the metric means, so choose it deliberately. Use the language buyers actually use when they ask an AI engine — question-shaped, intent-rich phrases — rather than only head terms. A focused set of high-intent keywords tells you more than a large set of vanity terms.
Pick competitors you genuinely compete with for those answers, not just the biggest names in your space. Share of voice is comparative, so the result is only as useful as the field you measure against. Three to five real rivals usually give a clearer read than a long, noisy list.
- Use buyer-language, intent-rich keywords, not just head terms.
- Keep the set focused — high-intent keywords beat vanity volume.
- Choose real, comparable competitors for the field.
- Hold the set stable so you can compare over time.
Why must you separate measured from estimated signals?
Because they are different kinds of evidence and blending them invents false precision. When an engine like Perplexity returns the source links behind an answer, you can read citations fairly directly — that's measured. When an engine doesn't expose its citations at scale, tools approximate visibility using a proxy such as organic search presence — that's estimated.
A report that merges both into a single decimal-point percentage implies a certainty it doesn't have. The honest approach is to show the breakdown and label the method, so a reader knows which part of the picture is observed and which is inferred. This is a credibility issue as much as a technical one.
Our scan follows this rule: it reports share of voice and labels the methodology behind it, rather than presenting an estimate as if it were ground truth.
How do you turn raw results into a share-of-voice number?
For each keyword, record which domains were cited or surfaced across the engines you query. Sum the citations per domain across the whole keyword set, then divide each domain's total by the grand total to get its share. Your share is your slice of that pie.
Layer on the two outputs that make it actionable: gap keywords (where rivals are cited and you aren't) and domination keywords (where you're cited and they aren't). The headline percentage tells you where you stand; these lists tell you what to do next.
- 1Record cited domains per keyword across each engine.
- 2Sum citations per domain across the full keyword set.
- 3Divide each domain's total by the grand total for its share.
- 4Extract gap keywords (rivals cited, you not) and wins (you cited, rivals not).
- 5Re-run periodically with the same method to read the trend.
Frequently asked questions
How many keywords should I measure?+
Enough to represent your category but focused on intent — a tight set of high-intent, buyer-language keywords is more meaningful than a large set of vanity terms. Whatever you choose, keep it stable over time so comparisons hold.
What does 'measured vs estimated' mean?+
Measured signals are read directly from an engine that exposes its citations (e.g. Perplexity's source links). Estimated signals approximate visibility using a proxy like organic search presence when an engine doesn't expose citations at scale. Credible reports label which is which instead of blending them.
How often should I re-measure?+
Often enough to see a trend rather than noise — many teams check monthly. The key is holding the methodology (keywords, competitors, engines, measured-vs-estimated mix) constant so changes reflect real movement, not changes in how you measured.