When buyers ask ChatGPT, Gemini, Perplexity, Claude, or Google AI Overviews for category advice, only a few brands make the cut. AI search share of voice is your slice of those answers relative to named competitors, across a fixed prompt set. Classic SEO share of voice tracks ranking positions. This one tracks presence inside generated answers, where results shift between runs and a single screenshot can lie to you.
Teams that mash every appearance into one percentage lose the signal. A brand can be named often, linked rarely, and recommended almost never. Split the work into mention share, citation share, and recommendation share, then measure each with a repeatable prompt-sampling method and clear uncertainty guardrails.
Why this metric is not classic SEO SOV
Search rank trackers assume a mostly stable results page. Answer engines do not. The same buyer prompt can return different brands on consecutive runs because retrieval timing, model settings, and source freshness all shift. That volatility is why a one-off ChatGPT check is closer to anecdote than measurement.
Two more differences matter for reporting. First, engines disagree: a brand that leads on Perplexity can lag on Google AI Overviews for the same prompt cluster. Second, the unit of measurement is an answer event, not a SERP slot. You count what the model said in a recorded response, then compare your count to the competitive set you defined up front.
If you already run competitor AI visibility analysis, treat these percentages as the competitive roll-up of answer-level observations, not as a replacement for page-level SEO reporting.
Mention share, citation share, and recommendation share
Think of three different jobs, not three labels for the same vanity score. Naming rate, source-backed rate, and preference rate answer different questions, so keep the math separate.
Mention share is a competitive naming rate that counts how often your brand name appears in answers for the tracked prompts, divided by total brand mentions across your competitive set (or by total answers, if you prefer an inclusion-style denominator). Use one denominator and stick to it. Similarweb's framing helps: visibility asks whether you appear at all; mention share asks how much of the competitive conversation you own once brands are named.
Citation share is a source-attribution rate that counts how often your owned domains are linked or explicitly referenced as sources, divided by total brand-attributable citations in the same answer set. A model can mention you without citing you. This stricter authority signal usually moves slower than raw mentions.
Recommendation share is a preference rate that counts how often the answer positions your brand as a primary pick, shortlist leader, or "best for" choice, divided by answers that make any clear recommendation. Passing comparisons do not count. Soft lists where five brands appear with equal weight should be scored carefully, or excluded, so the metric stays tied to preference rather than presence.
| Metric | What you count | What it diagnoses |
|---|---|---|
| Mention share | Brand-name appearances vs competitive set | Awareness inside answers |
| Citation share | Owned-domain links or explicit source credits | Source trust and retrievability |
| Recommendation share | Primary or shortlist "best for" selections | Purchase-influencing preference |
A brand that wins mention share but loses citation share is being remembered, not trusted as evidence.
For citation-focused workstreams, pair this split with your LLM brand citation tracking so content and PR fixes map to the metric that actually moved.
How to measure AI search share of voice with prompt sampling
Reliable scores start with a frozen measurement design. Change the prompts, engines, country, or competitor list mid-track and you measure the setup, not the brand.
1. Lock a buyer prompt set. Use 20 to 50 prompts drawn from category questions ("best X for Y"), problem questions, and comparison questions. Weight prompts by commercial value (for example 1 to 3) so a low-intent definition query cannot dominate the score. Document the set before the first run.
2. Freeze engines and locale. Run the same prompts across the engines your buyers use (commonly ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews) in one country. Per-engine scores come first; blended scores are optional and should be labeled as aggregates.
3. Sample repeatedly. Because answers vary, treat each prompt-engine pair as a distribution. Practical teams re-run the panel on a fixed cadence (weekly is common) and keep every raw answer event. For high-stakes baselines, collect multiple samples per prompt before you publish a headline percentage.
4. Score with written rules. For every answer, record: brands mentioned, domains cited, recommendation status, position of first brand mention, and sentiment if you need it later. Normalize product names to parent brands so "Acme" and "Acme CRM" do not inflate the denominator.
5. Compute three percentages, not one. Mention share, citation share, and recommendation share each get their own numerator and denominator. Optional weighting multiplies each event by prompt weight (and, if you choose, by position weight such as 1, 0.5, 0.33 for first, second, and third place).
Getspotted's public methodology walks through a weighted citation example that ends at 28.6% for one engine-country slice after three sample queries. Use that style of worksheet even if you automate collection later: the audit trail matters more than the dashboard chrome. Full step detail lives in Getspotted's measurement guide for AI answer visibility.
Uncertainty guardrails that keep the number honest
Report ranges and trends, not theatrical precision. A single weekly point can swing because engines refresh sources on their own schedule. A four-week moving average, plus a simple variance note (how often the same brand appears across repeats of one prompt), separates signal from noise.
Guardrails worth writing into the SOP:
- Never change the prompt set, competitor list, engine list, or country without starting a new series labeled as a break in continuity.
- Publish per-engine scores beside any blended category percentage.
- Separate brand-versus-competitor denominators from total-citation denominators that include publishers and review sites.
- Flag "ghost" or unverifiable mentions and weight them lower than named, source-backed appearances.
- Require a sustained move across more than one engine before you declare a GEO win or loss.
Similarweb's distinction between brand visibility and mention share is a clean check against overclaiming: presence is binary; competitive share is relative. See Similarweb's plain-language definition of brand share in AI answers when you need language that leadership already recognizes.
Feed the cleaned series into AI visibility reporting and analytics so editorial and technical work stay tied to the same scorecard.
What a healthy reading looks like
No universal good percentage exists. Against four tracked rivals, an even split is about 20% mention share; beating that average means you out-appear the typical rival on that panel. Fragmented categories make smaller shares look stronger. The actionable target is a rising delta on the metrics that match your goal: citation share for authority programs, recommendation share for shortlist influence.
Use the three-metric split to choose the fix. Weak mention share points to entity clarity, third-party corroboration, and topical coverage. Weak citation share points to source quality, structured evidence, and pages engines can retrieve cleanly. Weak recommendation share points to comparison content, proof points, and category framing that answers "best for" prompts without fluff.
FAQ
How do you calculate share of voice inside AI answers?
Divide your brand's scored answer events by the competitive set's scored events across a frozen prompt panel, then multiply by 100. Run the panel per engine and country, weight prompts by commercial value, and keep mention, citation, and recommendation counts as separate percentages rather than one blended vanity score.
What is the difference between mention share and citation share?
Mention share counts brand-name appearances. Citation share counts linked or explicitly credited owned sources. You can win mentions while losing citations when models know your name but do not treat your pages as evidence.
How many prompts do you need for a stable reading?
Most measurement playbooks land between 20 and 50 buyer-intent prompts. Sets under about 10 swing too hard between runs. Lock the set before baseline week one so later deltas stay comparable.
How often should you remeasure?
Weekly sampling with a multi-week moving average is a practical default in 2026. Daily checks help during launches, but treat single-day spikes as provisional until they hold across engines.