AI Share of Voice: Definition, Formula and Worked Example

AI share of voice is the target brand's share of all eligible brand-presence units observed across a defined competitor set, prompt cohort, engine scope and time window. The basic formula is: target brand presence units divided by total presence units for every included brand, multiplied by 100. A useful result always publishes the underlying counts and scope.

AI share of voice is a sampled competitive metric, not an estimate of every AI conversation or market share. It also differs from answer-level brand coverage: coverage asks in how many valid answers the brand appeared, while share of voice asks what portion of all observed competitor presence belonged to that brand. Both can be useful, but their denominators cannot be mixed.

Definition and required scope

In AI-search measurement, a presence unit is an eligible appearance of one approved brand entity in one valid answer. The clean default counts a brand at most once per answer, even if its name is repeated. This prevents verbose repetition from inflating competitive presence. If the analysis uses a different unit, it must be named and justified.

Before calculating, freeze five inputs:

• Prompt cohort: the exact decision-relevant questions and their version.
• Competitor set: the brands eligible for the denominator, including the target brand.
• Engine scope: each assistant or answer surface included and how it will be reported.
• Market and language: conditions that determine which observations belong together.
• Observation window: start, end and collection cadence.

A label such as “our AI SOV is 32%” is incomplete without these boundaries. The result describes the chosen sample, not all possible prompts or users.

The unweighted AI share-of-voice formula

Let each valid answer contribute one presence unit for each included brand it mentions under the entity rules. Then:

AI share of voice for Brand A = Brand A presence units ÷ presence units for all included brands × 100.

The denominator is not the number of answers. Because one answer can include several eligible brands, total presence units may exceed valid answers. It is also not search volume, estimated prompt volume, citations, clicks or impressions. Those are different units.

Use the AI share-of-voice calculator to apply the formula after the counting rules and inputs have been approved. The calculator cannot repair a biased cohort or an incomplete competitor set.

Worked example

Suppose a fictional software category is measured with 20 valid generated answers. The frozen competitor set contains Brands A, B, C and D. Each brand can contribute at most one presence unit per answer. After entity QA, the observed units are:

• Brand A: 12 presence units
• Brand B: 9 presence units
• Brand C: 6 presence units
• Brand D: 3 presence units

The denominator is 12 + 9 + 6 + 3 = 30 eligible presence units. Brand A's AI share of voice is 12 ÷ 30 × 100 = 40%. Its answer-level coverage is 12 ÷ 20 × 100 = 60%. Both results are correct because they answer different questions.

Brand A can therefore have 60% answer coverage and 40% competitive share in the same sample. Reporting only one number can hide the distinction. All figures in this example are invented for instruction and are not a Bee LLM benchmark.

How to build the dataset

1. Choose real questions. Include problem, category, comparison and implementation decisions instead of superficial keyword variants.
2. Freeze the cohort. Give the prompt list a version and retain a fixed view when the portfolio expands.
3. Declare conditions. Record engine, surface, language, market, time and collection state.
4. Resolve entities. Create approved aliases and exclusions for homonyms, subsidiaries and product names.
5. Count once per answer. Apply the default brand-presence rule consistently unless another unit is explicitly chosen.
6. Audit a sample. Have a second reviewer check ambiguous entities and missing runs.

The collection guidance in tracking brand mentions across AI search explains how to retain answer evidence. Share of voice should be calculated only from valid, reviewable observations.

Missing runs, answers with no included brand and the denominator

A failed or blocked collection is unavailable telemetry. It contributes no presence units and must be reported separately; it is not evidence that every brand was absent. An otherwise valid answer that genuinely mentions none of the included brands contributes zero units, but remains in the denominator for answer-coverage metrics.

For mention-based share of voice, a valid answer with no included brand adds nothing to the SOV denominator. This means SOV alone can look unchanged while overall category presence collapses. Always show valid answer count and target brand coverage beside competitive share.

Mentions, recommendations and citations need separate SOV variants

The basic metric counts eligible brand presence. A team may also define recommendation share of voice, using only brands actually recommended for the prompt's use case, or owned-citation share of voice, using exposed approved domains. These are distinct measures and must carry descriptive names.

Do not call a source URL a brand mention unless the answer text or entity rule supports it. Do not treat every neutral reference as a recommendation. Keep the raw fields so the metric can be recalculated when a coding rule changes.

When weighting is appropriate

An unweighted fixed cohort is easiest to audit: every valid prompt execution has the same potential contribution. Weighting may be appropriate when the business has credible evidence that some topics, markets or journey stages are more important. The weights should come from an external decision model, not be tuned to improve the result.

Calculate weighted presence units within stable strata, disclose every weight and retain the unweighted result. Do not apply conventional keyword search volume to conversational prompts without a validated mapping. A weighted metric that no reviewer can reproduce is not decision-ready.

Segment before aggregating

Calculate SOV separately by engine, market, language, topic and journey stage where sample size allows. Engines can differ in answer structure and source display. A blended figure requires explicit, stable weights; otherwise a shift in the collection mix can move the total even when no engine-level result changes.

The wider framework for AI search visibility metrics and KPIs places SOV beside accuracy, citation exposure, referrals and qualified outcomes. No single metric should carry the full performance narrative.

Interpret changes safely

AI share of voice is relative. Brand A's share can rise because its presence increases, because competitors decline, or because the competitor set changes. Show brand-level counts and preserve the frozen set so the driver is visible. When a competitor is added or removed, start a new version or restate history.

A temporal increase is an observation, not proof that a specific content edit caused it. Engine variability, source changes, competitor actions and collection conditions may coincide. Use annotations, repeated observations and cautious language.

Common calculation errors

• Dividing target mentions by valid answers and calling the result share of voice; that is coverage.
• Counting repeated mentions in one long answer as multiple presence units without disclosure.
• Changing competitors, prompts or engines while showing one continuous trend.
• Treating failed collections as brand absences.
• Mixing mentions, recommendations, citations and referral sessions in one numerator.
• Publishing a percentage without counts, scope or observation window.

Minimum reporting template

Publish the target SOV numerator and total denominator, valid answers, unavailable runs, target answer coverage, competitor set version, prompt cohort version, engines, market, language and observation window. Add segment tables and methodology changes. Link highlighted findings to saved response evidence.

That compact contract makes the metric comparable and challengeable. Use the calculator for arithmetic, but treat prompt design, entity QA and denominator governance as the real analytical work.