AI Share of Voice: The Metric That Replaces Market Share in AI Search

What is AI share of voice?

AI share of voice is the portion of AI-generated answers in a defined category that name your brand. The metric appears under several names — share of answer, AI citation share, AI visibility share — but they all measure the same thing: when an AI assistant responds to a buyer prompt in your category, what fraction of the time does your brand get mentioned?

The definition is precise. Given a locked prompt library (say, 141 buyer prompts for a specific vertical), a locked panel of assistants (typically ChatGPT, Claude, Gemini, Perplexity), and a locked sampling protocol (usually two samples per prompt per model), every response produces zero or more brand mentions. AI share of voice for a brand is that brand's mentions divided by the total mentions across all brands in that corpus.

The metric is a top-of-funnel signal, not a bottom-of-funnel one. It measures visibility inside the AI recommendation surface — the place where a growing share of buyers now form their first shortlist. It does not measure clicks, conversions, or revenue. Treating it as a proxy for those things is the most common misuse. Treating it as a proxy for share of the buyer's first impression is closer to correct.

For a broader tour of AI-specific visibility metrics and how they connect, see our pillar on <a href="/guides/ai-search-rankings">AI search rankings</a>.

How is AI share of voice different from market share?

Traditional market share measures a company's slice of category revenue, units sold, or customer count. It is a lagging economic measure. AI share of voice measures a company's slice of AI-generated attention. It is a leading discovery measure. The two often diverge.

A firm can hold significant market share and near-zero AI share of voice. This happens when the firm ranks well in Google and paid channels but produces little of the scenario-specific content AI assistants retrieve. The revenue is real; the AI visibility is not. Buyers who start their search inside an assistant will not encounter the firm.

A firm can also hold meaningful AI share of voice and modest market share. This happens when a smaller firm publishes exactly the kind of scenario answers assistants canonicalize on. The revenue lags because the firm is small, but the discovery surface is disproportionately in its favor. Over time, that discovery advantage compounds — or it does not, depending on how well the firm converts AI-driven inquiries.

Traditional share of voice, from PR and media planning, measures a brand's share of category media impressions. AI share of voice is closer in spirit but different in mechanics: the "media" is generated per query, the corpus is a defined prompt set rather than a media buy, and the audience is the assistant's user base rather than a demographic panel.

The practical implication: AI share of voice belongs alongside market share on a dashboard, not instead of it. They measure different points in the buyer's journey and diverging numbers are diagnostically useful. A firm with strong market share and weak AI share of voice has a specific, addressable problem — its content is not the content AI assistants prefer.

How to calculate share of voice in AI search

The calculation is straightforward. The methodology around it is where credibility lives.

The formula: AI share of voice for a brand = (mentions of that brand across the response corpus) / (total brand mentions across the same corpus). If your firm is named 47 times across 1,128 responses and total brand mentions across the corpus are 481, your AI share of voice is 47 / 481 = 9.8%.

Five methodology decisions determine whether that number means anything.

First, the prompt library. Prompts have to reflect real buyer intent, cover the buyer journey (situation, decision, validation, follow-up prompts), and stay fixed across measurement periods. A rotating prompt library makes drift unmeasurable. Second, the model panel. At minimum three or four assistants, because per-model variance is severe — the same firm can be cited five times more by one model than another. A single-model share number is a slice, not the metric. Third, the sample count. At least two samples per prompt per model, because a single draw from an assistant's distribution is noisy. Fourth, the deduplication rule. If the same brand is named twice in one response, does that count as one mention or two? Pick a rule and hold it. Fifth, the confidence interval. Wilson score intervals on small samples separate real gaps from apparent gaps.

For scale reference: a full vertical snapshot in the Viclaro Atlas runs 141 prompts across four assistants with two samples each, yielding up to 1,128 responses per firm per snapshot at roughly $14 in retail model API cost. That is the level of coverage needed for share numbers to be stable across months. Less than that and the metric moves too much between snapshots to trust.

What does a "good" AI share of voice look like?

There is no universal number. A good share depends entirely on the shape of the category. Three archetypes cover most cases.

Monopoly categories. One firm captures more than half of AI mentions and all four assistants converge on it. In the Atlas, RMA of New York holds 60.1% AI share of voice in NYC IVF across 629 citations — when a buyer asks any of the four assistants about IVF in New York, RMA is named more often than every other clinic combined. In categories like this, share of voice is not a competitive metric. It is a canonicalization result. The right question is not "how do we get to 30%" but "is there a niche scenario the leader has not canonicalized where we can lead?"

Leader-with-shortlist categories. The top firm holds around 15-25%, a defined shortlist of three to five firms shares another 30-40%, and the rest is a long tail. Aronson Mayefsky & Sloan sits at 21.0% in NYC divorce, with the #2 and #3 firms at 17.8% and 14.7%. Every top-three firm is cited by all four assistants. In categories like this, share is a real competitive metric. Peers are reachable, and moving from 5% to 10% to being on the shortlist is a tractable content-driven trajectory.

Fragmented categories. No firm has canonicalized the market. The top firm holds under 10%. In NYC personal injury, the top firm sits at 8.0% share and only 2 of 4 assistants cite it consistently. Firms in third, fourth, and fifth position sit within a Wilson interval of each other. In categories like this, a share of 5-8% is a serious position and 10%+ is a category-leading position. Ambitions calibrated against a "monopoly" benchmark would misfire badly.

The three anchor numbers — 60%, 21%, 8% — are all real Atlas share-of-voice figures measured under identical methodology. The gap between them is not measurement error. It is category structure. Any AI share of voice number needs to be read against the shape of its category, not against a universal benchmark.

How to increase your AI share of voice

Increasing AI share of voice is fundamentally different work from increasing traditional SEO rankings. The mechanics are different, so the moves are different.

Publish scenario-specific content, not practice-area content. AI assistants cite pages that match the specific language of buyer prompts. A page titled "Family Law Attorney NYC" loses to a page titled "What happens to a small business in a New York divorce when only one spouse is on the corporate paperwork," because the second matches the exact prompt shape a buyer types into an assistant. Scenario pages are the atomic unit of AI visibility content.

Mark up content with structured data. FAQPage JSON-LD, Article schema, Organization markup — these do not make a page rank inside AI, but they give the retrieval layers explicit signals about content type, authorship, and topical scope. Their absence removes signal.

Measure per-model, not blended. If your share of voice is 3% on Gemini and 15% on Claude, the two problems have different fixes. A blended 9% number tells you nothing about where to invest. Every credible share-of-voice measurement should report per-model breakdowns, and every improvement plan should target specific model-and-prompt combinations, not a blended average.

Run the loop. Measure baseline share, diagnose losing prompts by bucket and by model, ship a bounded content change, re-measure fourteen days later, and compare against confidence intervals. A one-time audit is a snapshot. A running loop is what moves share.

See what current AI share of voice looks like across eight NYC verticals on the <a href="/leaderboards">Viclaro Atlas leaderboards</a> — real firms, real shares, updated regularly. Or run a bounded share-of-voice measurement on your own firm with our <a href="/scan">free AI visibility scan</a>.