AI Share of Voice: How to Measure and Benchmark It

The metric, stated precisely

Most teams conflate two different things under the same label. Separate them and the diagnosis gets much easier.

Mention rate answers a binary question: do you show up at all? Share of answer answers a competitive one: of all the brands the assistant named, how much of the airtime was yours? A B2B tool can have a 90% mention rate and still lose, because it is named ninth in a list of ten and the first three get the buyer's attention.

Both are calculated per assistant and per prompt cluster. A cluster is a group of buying-intent questions that belong to the same decision — pricing comparison, alternatives to a named competitor, best-tool-for-X, integration questions. Averaging across clusters produces a number that moves for reasons you cannot trace.

• Mention rate (%) = runs in which your brand appears ÷ total prompt runs, for one assistant and one cluster, × 100.
• Share of AI answer (%) = your brand mentions ÷ all brand mentions across the same answers, × 100. This is the true share metric because it is zero-sum against a defined competitor set.
• Weighted share (optional) = the same ratio with each mention weighted by position, so being first in a list counts more than being a footnote. Use a simple weight (first = 1.0, mid = 0.6, last = 0.3) and document it.
• Recommendation rate = runs where you are named as a recommendation, not just referenced as an example. Worth splitting out for high-intent prompts.

The three inputs that decide whether your number means anything

An AI visibility metric is only as stable as its instrumentation. Three decisions do most of the damage when they are made loosely.

First, the prompt set. Fix it in writing — the exact phrasings, the market, the language, the persona framing — and version it. Every time you add prompts, your trend line breaks, so add them as a new cluster rather than editing the old one. Aim for enough prompts to cover the decision, not enough to look thorough: twenty to forty well-chosen buying questions per market beat several hundred generic ones.

Second, the run count. Assistant outputs are not deterministic; the same prompt asked twice can name different brands, especially where live retrieval is involved. That means a single run is an anecdote. Run each prompt multiple times per assistant per period — five is a workable floor, more if your category has a long tail of competitors — and treat small gaps on small samples as noise rather than movement.

Third, the mention rule. Decide in advance whether a bare URL counts, whether a citation in a source list counts as a mention, how you handle misspellings and parent-brand names, and whether a negative mention counts. Write the rule down; inconsistency here creates most of the disputes about whether visibility improved.

Why per-assistant benchmarks diverge — and why that is correct

A brand at 55% share of answer in Perplexity and 12% in ChatGPT for identical prompts is not a measurement error. The systems differ at the layer that decides which brands are eligible to be named.

Retrieval versus parametric knowledge is the biggest split. Assistants that lean on live web search surface whatever the current index and their ranking layer favour; answers that come mostly from model weights reflect how prominently a brand appeared in training data. OpenAI describes ChatGPT search as combining the model with information from the web and third-party providers (OpenAI), Anthropic ships web search as a tool Claude decides when to invoke (Anthropic), Perplexity is built around retrieval and citation by default (Perplexity), and Google documents its AI features as a surface layered on Search with its own linking behaviour (Google Search Central). Different eligibility, different brand lists.

Source preference is the second driver. One assistant may lean on editorial roundups and review platforms, another on documentation, forums and primary sites. If your category's third-party coverage is concentrated in a handful of listicles, the assistants that trust those pages will over-represent whoever ranks well on them.

Freshness and grounding rules are the third. A newly launched competitor can appear immediately in retrieval-heavy assistants and be invisible for months in weight-heavy ones. Region, language and whether a tool decides it needs to search at all add further legitimate variance.

The practical consequence: never report a blended cross-assistant average as your headline. Report a matrix. The spread between assistants is itself the insight, because it tells you whether your problem is source coverage, ranking on other people's pages, or plain absence from the corpus. That per-assistant breakdown is the core of what we build at Growli, and it is the reason a single global score is the least useful number in the report.

A benchmark table template you can copy

Build one row per assistant per prompt cluster per period. Columns, in order:

Read it two ways. Across a row, you see whether a cluster is healthy on one assistant and broken on another. Down a column, you see whether a competitor is winning everywhere (a corpus and authority problem) or on one surface only (a source problem you can target).

Add a second sheet holding the competitor set — five to eight named rivals, fixed for the period — plus the exact prompt list and the mention rule. Without those, next quarter's number is not comparable to this one.

• Period — week or month, with the run dates.
• Assistant — ChatGPT, Gemini, Claude, Perplexity, Copilot, Grok. One row each; no blending.
• Prompt cluster — e.g. best-tool-for-X, alternatives-to-competitor, pricing.
• Runs — total prompt runs behind the row. Anything under five per prompt should be flagged as low confidence.
• Your mentions — runs containing your brand under the documented rule.
• Mention rate % — your mentions ÷ runs.
• Total brand mentions — all mentions of you and the tracked competitor set in the same answers.
• Share of answer % — your mentions ÷ total brand mentions.
• Average position — mean rank when named in a list.
• Top competitor + their share % — who is actually taking the slot.
• Top cited domains — the three sources the assistant leaned on most; this is your action list.
• Delta vs previous period — in percentage points, not percent change, so small bases do not exaggerate movement.

Turning SoV gaps into actions

A share-of-voice report that ends in a percentage is a dashboard. The value is in the mapping from gap type to intervention, and there are only a handful of gap types worth naming.

Absent on every assistant for a cluster means you are not in the source material the answers are built from. The work is coverage: get named, described and compared on the pages assistants already cite for that cluster — review platforms, category roundups, community threads, industry publications — and make sure your own pages state plainly what you do, for whom and at what price.

Present in retrieval-heavy assistants, absent in weight-heavy ones points to recency. You exist on the live web but not in the deeper corpus. Expect a lag, and in the meantime concentrate on the surfaces where retrieval decides eligibility.

Named but ranked last is a comparative-evidence problem. Assistants order lists using the framing they find in sources. If competitors have explicit comparison pages, pricing tables and use-case statements and you do not, you get slotted as the also-ran. Publish the specifics — who you are not for is as useful as who you are for.

Named with wrong attributes is the most damaging and the easiest to miss if you only count mentions. Log the claims made about you, not just the fact of the mention: wrong pricing model, a missing integration, an outdated positioning. Correct it at the source that is being cited.

One competitor dominant on one assistant usually traces to a single page. Find it in the cited-domains column, then decide whether the play is inclusion in that page, a stronger alternative on a comparable domain, or a factual correction request.

Prioritise by commercial weight, not by gap size. A ten-point gap on high-intent alternatives-to prompts is worth more than a forty-point gap on definitional questions. If you want the wider framing of how mention rate, sentiment and citation sources fit together, see our primer on AI visibility.

Cadence, confidence and what counts as real movement

Monthly measurement suits most categories; weekly is justified during a launch, a rebrand or an active correction campaign. Whatever you pick, run the full prompt set in a tight window so model updates do not straddle your sample.

Set a movement threshold before you report. With five runs per prompt, a shift of a few percentage points on a single cluster is well inside sampling variance — call movement only when it holds for two consecutive periods or appears across multiple clusters on the same assistant. Annotate every report with the events that could explain a jump: a model update, a new competitor launch, a large third-party article going live, a change to your own site.

One last discipline: keep the raw answers, not just the tallies. Six months from now the tallies will tell you that share fell and the stored answers will tell you why — which is the difference between an AI visibility metric you can defend and a number that starts an argument.

FAQ