AI share of voice: definition, formula, and tracking guide

AI share of voice measures your brand presence relative to selected competitors across a defined portfolio of AI answers. A common formula is your qualified mentions divided by all qualified mentions among the comparison brands. The result is meaningful only when the prompts, models, markets, competitors, and treatment of multiple mentions are documented. Keep that denominator stable so each reporting period remains genuinely comparable.

AI share of voice formula

If your brand receives 24 qualified mentions and the comparison set receives 96 mentions in total, your mention share of voice is 25 percent for that test. You can also calculate recommendation share, first-position share, or citation share. Do not combine these silently: appearing in a long list is a different outcome from being recommended first for a high-intent requirement.

Choose competitors and prompts fairly

• Include direct alternatives that serve the same audience, use case, and geography.
• Add emerging or substitute solutions when customers genuinely evaluate them.
• Keep a stable core set for trend comparisons and document additions or removals.
• Use unbranded category, use-case, comparison, and eligibility prompts based on real buyer language.
• Separate prompts where not every competitor qualifies instead of rewarding irrelevant appearances.

Count visibility at the right level

Decide whether each brand can earn one mention per answer or additional credit for prominence and recommendation order. A binary rule is easy to audit. A weighted rule captures nuance but introduces judgment, so publish the weights. Also decide how aliases, parent companies, product names, and misspellings are resolved. Otherwise, entity matching errors can create false movement.

Segment before interpreting the average

1. Calculate results separately by model because answer and retrieval behavior differs.
2. Split by market and language so a strong home-country presence does not hide international gaps.
3. Compare discovery, comparison, and verification intent instead of merging the whole funnel.
4. Break out product lines and customer segments when eligibility differs.
5. Use rolling periods or repeated runs to reduce the impact of ordinary response variation.

The AI visibility scan provides a starting observation, while recurring tracking shows whether patterns persist. When presenting a chart, include the number of prompts and runs behind it. A share calculated from ten generic questions should not be compared with one based on hundreds of qualified market-specific observations.

More mentions are not always better. A company may be named because the model repeats a known controversy, states an outdated limitation, or recommends it to an ineligible buyer. Track sentiment, factual accuracy, citations, and fit beside share of voice. Review the full answer whenever a material movement occurs before deciding that marketing performance improved or declined. Create an exclusion rule for navigational, duplicate, refused, or otherwise ineligible responses and report how many observations were excluded. That keeps denominator changes from masquerading as competitive gains.

Inspect the sources and claims associated with brands that appear more often. Look for legitimate differences: clearer documentation, stronger category education, verifiable credentials, useful research, or consistent public data. Use the findings to improve your own evidence. Do not copy claims, generate comparison spam, or try to suppress accurate competitor information. A transparent ROI calculator can help your team prioritize effort, but it should use your own assumptions rather than promised gains.

• Validate prompt coverage, entity matching, and competitor eligibility before reading the score.
• Identify the largest changes by model, intent, market, and requirement.
• Open representative answers and verify citations, fit, sentiment, and facts.
• Assign evidence-based actions to content, product marketing, technical, or communications owners.
• Annotate major launches and source changes, then review whether movement persists in later periods.
• When monitoring expands to another model, show its contribution separately until enough comparable history exists to decide how it belongs in the combined view. Recalculate both weighted and unweighted results during the transition, disclose the method, and avoid presenting a formula change as organic competitive growth.