Calculate Share of Voice in LLM Responses: Formula, Method, and Metrics

By maxaeo.ai ｜ Published 2026-09-24 ｜ Updated 2026-09-24

To calculate share of voice in LLM responses, divide your brand’s tracked mentions by the total tracked brand mentions in the same AI answers, then multiply by 100. The result shows how much of the visible competitive conversation belongs to your brand across a defined prompt set, engine list, competitor group, and time period.

What is share of voice in LLM responses?

Share of voice, or SOV, is the percentage of brand presence your company receives compared with the total presence of tracked brands in AI-generated answers. Unlike a simple mention rate, it is a competitive metric: it tells you whether your visibility is growing relative to alternatives.

A practical baseline formula is:

LLM SOV (%) =
Your brand mentions ÷ Total tracked brand mentions × 100

For example, if your brand appears 42 times and all tracked brands appear 140 times across the same answer set:

42 ÷ 140 × 100 = 30% SOV

Industry definitions differ on whether to count mentions, answers, citations, or weighted prominence. That is why every report should state its counting rule before showing the percentage. (blimpp.com)

What should be included in the SOV denominator?

The denominator is the most important part of the calculation. Include only brand observations collected from the same scope:

• The same prompt set
• The same AI engines
• The same competitor list
• The same market and language
• The same observation window
• The same counting method

Suppose a SaaS company tracks 50 buyer prompts across ChatGPT, Perplexity, Gemini, and Claude. If the monitored brands produce 500 total mentions and your brand receives 75, your SOV is 15%.

Do not mix that result with a second report that uses different prompts or adds ten more competitors. The percentage may change even when your underlying visibility does not.

A useful reporting convention is to publish both:

1. Tracked-set SOV: your share among named competitors.
2. Open-market SOV: your share among every recognizable brand appearing in the answers.

Tracked-set SOV is better for direct competitive decisions. Open-market SOV is better for discovering unexpected alternatives that AI engines introduce into the category.

How to calculate LLM share of voice step by step

1. Define the commercial prompt universe

Build prompts around real buyer intent rather than only product names. Include categories such as:

• “Best project management tools for remote teams”
• “Alternatives to [competitor]”
• “Affordable CRM for a growing SaaS company”
• “Which analytics platform is easiest to implement?”
• “Tools with strong integrations for [use case]”

A strong prompt set normally combines category discovery, comparison, alternatives, use-case, and evaluation questions. The goal is to measure the market conversations where a buyer could realistically encounter your brand.

2. Select engines and repeat the same prompts

Run the same prompt set on each selected engine. Report engines separately before combining them because ChatGPT, Perplexity, Gemini, Claude, and other systems can use different retrieval sources, answer structures, and recommendation patterns.

For every response, store:

• Prompt text
• Engine name
• Date and time
• Full answer
• Mentioned brands
• Citation URLs or domains
• Mention position
• Sentiment or recommendation context

Keeping the raw answer is essential. It allows a team to verify whether a brand was merely listed, actively recommended, criticized, or cited as evidence.

3. Choose a counting unit

There are three common counting units.

Mention-level SOV is usually the clearest starting point. If one answer names a brand three times, it contributes three mentions. Answer-level measurement prevents verbose answers from dominating the result, but it may understate prominence.

The important rule is consistency. Never compare a mention-level score with an answer-level score and call the difference a performance change.

4. Apply the formula

Assume a 30-day sample contains:

• 100 prompts
• 4 AI engines
• 400 total answers
• 620 total brand mentions
• 93 mentions for your brand

Your raw SOV is:

93 ÷ 620 × 100 = 15%

If the competitor totals are:

This table reveals more than your score alone. It shows whether the problem is low overall visibility, one dominant competitor, or a fragmented category.

Should citations and recommendation position be weighted?

A raw SOV score should remain unweighted so that it is easy to reproduce. However, a second prominence score can explain the quality of your visibility.

One practical framework is the Visibility–Prominence–Evidence model:

• Visibility: Was the brand mentioned?
• Prominence: Where did it appear and was it recommended?
• Evidence: Was it supported by a citation or trusted source?

For example, assign internal diagnostic values:

Visibility = 1 if mentioned
Prominence = 1 to 3 based on position or recommendation strength
Evidence = 1 if supported by a relevant citation

Do not silently merge these values into official SOV. Instead, show:

Raw SOV: 15%
Recommended-answer rate: 8%
Citation-supported mention rate: 11%
Average recommendation position: 2.4

This separation is an important measurement improvement. A brand can have high mention SOV but weak buyer influence if it appears late, without evidence, or only in negative comparisons.

For broader measurement guidance, an AI search visibility dashboard framework can help organize these metrics into recurring reports.

How large should the sample be?

There is no universal sample size that guarantees a reliable SOV benchmark. A small sample can identify obvious gaps, but it is sensitive to prompt selection and answer variation.

For a practical SaaS baseline:

• Use at least 25–50 commercially relevant prompts.
• Run them across at least 3 engines when possible.
• Compare the same prompts for at least four weekly cycles.
• Keep a separate view for each engine.
• Label results as directional until the trend is stable.

The main source of error is usually not the arithmetic. It is sampling bias. If all prompts are written around your own positioning, your SOV may look stronger than it is in neutral buyer language.

How can teams improve their SOV after measuring it?

First, identify the prompts where competitors appear and your brand does not. Then classify the gap:

1. Entity gap: The AI system does not recognize your product category or use case.
2. Proof gap: Your brand is mentioned but lacks supporting evidence.
3. Source gap: Competitors are backed by review sites, comparisons, documentation, or community discussions that the model retrieves.
4. Positioning gap: Your product is described inaccurately or associated with the wrong buyer need.
5. Recommendation gap: You appear in lists but are not selected as a strong option.

Track these gaps alongside citations and sentiment. MaxAEO supports daily monitoring across eight AI engines, including ChatGPT, Perplexity, Gemini, Claude, Copilot, Grok, Google AI Mode, and Google AI Overviews. It also compares brand and competitor mention rates, citation sources, sentiment, and recommendation position.

You can begin with a free AI visibility diagnosis or review an AI citation optimization framework to connect SOV findings with specific content actions.

What is the difference between mention rate and SOV?

Mention rate measures how often your brand appears across answers. SOV measures how much of the competitive mention volume belongs to your brand.

For example:

Mention rate = 80 appearances ÷ 400 answers = 20%
SOV = 80 mentions ÷ 500 total mentions = 16%

The brand appears in one-fifth of answers, but owns only 16% of all brand mentions because several answers contain multiple competing brands.

Use mention rate to measure reach. Use SOV to measure competitive share. Use citation rate, recommendation rate, sentiment, and position to understand the quality of that share.

Frequently asked questions

Is LLM share of voice the same as SEO share of voice?

No. SEO SOV is generally based on rankings, search visibility, impressions, or estimated clicks. LLM SOV measures brand presence inside generated answers. A company can perform well in traditional search and still receive little AI recommendation visibility.

Should repeated mentions in one answer count multiple times?

They can, but the rule must be declared. Mention-level counting captures textual prominence; answer-level counting gives every response equal weight. For executive reporting, publish both when repeated mentions are common.

Should citations count as extra mentions?

Not in raw SOV. A citation is evidence of source support, not necessarily another brand mention. Keep citation rate as a separate metric so the score remains interpretable.

What is a good LLM SOV score?

There is no universal good score. Compare your result with the same competitors, prompts, engines, and time period. A lower score in a highly fragmented category may be strategically healthier than a higher score based on a narrow prompt set.

How often should SOV be monitored?

Daily collection is useful because AI answers can change, but weekly or monthly analysis is usually easier to interpret. Keep the raw daily data and use fixed reporting windows for trend comparisons.