AI share of voice: definition and limits

99Visibility Research · Updated October 21, 2026 · 4 min read

Short answer

AI share of voice is the percentage of brand mentions or citations your brand earns across a fixed set of prompts, measured against named competitors. It is a relative number. It changes with the prompts you choose, the engines you test and the size of the brands you compare. Read it per engine and per prompt set, against like-sized competitors, and never as a ranking.

Key takeaways

AI share of voice is your brand's mentions or citations divided by the total for every tracked brand across one prompt set.
The number depends on the prompts you chose, and engines rewrite those prompts before they search.
Big brands dominate AI citations, so compare yourself with like-sized competitors or the metric mostly measures brand size.
Report share of voice per engine, because engines count sources differently and run different models.
Treat it as a trend inside one fixed method, not as a rank, a traffic figure or a score you can compare across tools.

Definition

AI share of voice (also called LLM share of voice) borrows a classic marketing idea: what fraction of the conversation is about you. In AI search, the conversation is the set of answers an engine gives to a list of prompts.

The basic formula:

Share of voice = your brand's mentions ÷ mentions of all tracked brands, across the same prompts and engines.

Two counting units get mixed up. A mention is your brand named in the answer text. A citation is a link to your page in the answer's sources. They measure different things, and a brand can be named without being linked. When the unit is citations, the metric is closer to citation share.

How it works

Every share of voice figure comes from four choices someone made:

The prompt set. Which buyer questions get asked.
The competitor list. Which brands count in the denominator.
The engines. ChatGPT, Perplexity, Gemini, Google AI Overviews and others.
The counting rule. Mentions, citations, or first position only.

Each choice moves the result. The prompts you type are also not the queries that retrieve pages. OpenAI's help page on ChatGPT search says it typically rewrites a question into one or more targeted queries before searching. As of October 2026, engines also decide per question whether to search at all, so some answers come from model memory.

Engines are not interchangeable either. Google says AI Overviews and AI Mode may use different models and techniques. Keep a separate figure for each engine. For the full method, see measuring AI share of voice step by step.

Why it matters for AI visibility

Share of voice answers a fair question: when buyers ask AI about your category, who gets named? Inside one fixed method, a change over time is a useful signal.

Brand size is the trap. A study of AI search citations (arXiv:2507.05301) found the top 20 news sources accounted for 67.3% of all citations for OpenAI models. Large, well-known brands win most answers by default. If you benchmark a ten-person startup against a category leader, the gap mostly reflects size.

Any prompt set is also a sample. In Otterly's 2026 experiment, Microsoft's own report credited one site with 42,817 citations, while Otterly's prompt-based tracker logged 3,934. The windows may differ, but the lesson holds: your prompts see part of the picture.

The next move: fix your prompt set, pick three to five competitors your size, and compare per engine. A spreadsheet method for free AI share of voice works for this.

Common misconceptions

"It's my AI ranking." No position data exists inside AI answers to check a rank against. Share of voice is a ratio, not a position.
"Higher share means more traffic." It counts appearances in answers you generated. It says nothing about clicks.
"Scores from two tools are comparable." Different prompts, engines and counting rules produce different numbers for the same brand.
"A falling share means I'm losing." A competitor gaining can shrink your share while your own mentions hold steady. See what a falling citation share means.
"Someone can sell me a higher share." Nobody controls what models say. Read whether share of voice packages are worth buying before paying for one.

Frequently asked questions

Should I count mentions or citations?

Track both, separately. Mentions show whether the engine names you in its answer text. Citations show whether it links to your pages as sources. A brand can be recommended by name while every link points to review sites or directories. Mixing the two into one figure hides which problem you have. If engines name you but cite other sites, being named on those sites is associated with appearing in answers.

How many prompts do I need?

No published study sets a minimum. What matters more is consistency: use the same prompts, engines and competitors every time you measure, so changes reflect the engines rather than your method. Choose questions real buyers ask at each stage, not only questions that include your brand name. Write them down and date the set. Changing even a few prompts makes the new figure a different measurement.

Why does my share differ so much between engines?

Engines draw on different indexes, models and retrieval rules. Copilot grounds answers in Bing, Google's AI features rely on Google's index, and ChatGPT searches at its own discretion as of October 2026. Engines also cite differently: Anthropic cites text spans while Perplexity cites whole sources. A blended cross-engine number averages things that are not counted the same way, so keep one figure per engine.

Can I improve my AI share of voice directly?

You can fix what blocks you, not what models choose. Check that AI search crawlers can reach your pages, that key facts appear in the HTML before JavaScript runs, and that no snippet controls hide your content. Beyond access, no tactic has evidence of reliably raising mentions. Treat share of voice as a reading of the outcome, then work on the access issues it might reveal.

Sources

This page is AI-assisted. It was written from a verified evidence base and checked against our editorial policy. Engine behaviour changes often; facts are as of October 21, 2026.