Share of voice is only useful when the denominator, eligible brands, prompt families, and answer roles are published before collection. Mention-counting without a fixed denominator is not measurement, it's marketing.
What you should leave with
- Recommendation share, mention share, and citation share are three different metrics — never blend them into one score.
- The formula is simple; the denominator is where most SOV numbers are quietly gamed.
- Report ChatGPT, Google AI, Gemini, and Perplexity as separate rows on a common prompt spine, not an average.
- Only trust a change over time if the prompt set, competitor set, and classification rules stayed stable or the change is annotated.
What is the formula?
The simplest recommendation-share formula is brand qualifying recommendations divided by all qualifying tracked-brand recommendations in the same prompt set. Coverage and citation share use parallel formulas with their own denominators.
Three formulas cover the core cases. Brand recommendation coverage equals prompts where the brand is recommended divided by eligible prompts. AI recommendation share equals brand recommendations divided by all tracked-brand recommendations. Citation share equals brand-owned citations divided by the tracked citation set.
Position weighting, such as first-mentioned or top-three, is optional and can add nuance. It must be frozen before collection starts, though, or teams quietly adjust weights until the score matches the story they wanted to tell.
| Metric | Formula |
|---|---|
| Recommendation coverage | Prompts brand recommended / eligible prompts |
| AI recommendation share | Brand recommendations / all tracked-brand recommendations |
| Citation share | Brand-owned citations / tracked citation set |
What belongs in the denominator?
The denominator should include the brands, prompts, platforms, regions, and answer roles defined before collection, not whatever makes the score look better afterward. Locking this in advance is the difference between a metric and a marketing claim.
A published denominator checklist should list the competitor set with inclusion criteria, the prompt family and count, the platforms and regions covered, and the answer role being scored (recommendation, mention, or citation). Any change to these mid-project needs a dated note, not a silent edit.
The common gaming moves are prompt stuffing with easy-win prompts, cherry-picking weak competitors, and quietly dropping categories where the brand lost. A well-built prompt set, covered in our buyer-intent prompt set guide, closes most of these gaps before they start.
- Competitor set with inclusion criteria
- Frozen prompt family and count
- Platforms and regions covered
- Answer role being scored
- Dated change log for any mid-project edits
Method boundary: Prompt stuffing, competitor cherry-picking, and dropping losing categories after the fact all invalidate the denominator and the resulting score.
How should ChatGPT share of voice be reported?
ChatGPT share of voice should be reported as one platform-specific view, not treated as the whole AI search market. Buyers use multiple assistants, and each has different retrieval behavior worth tracking on its own.
ChatGPT, Google AI Overviews, Gemini, and Perplexity each pull from different sources and phrase answers differently, even against identical prompts. Google's own guidance confirms that crawlable, eligible content is what feeds its AI surfaces, which is a distinct pipeline from ChatGPT's retrieval behavior.
The practical structure is a common prompt spine, separate platform rows. Run the same prompt set everywhere, then report each platform's recommendation and citation share individually before ever averaging them, since an early average hides which platform is actually worth the fix effort.
Evidence used in this section
How do you interpret changes over time?
Treat share-of-voice changes as meaningful only when the prompt set, competitor set, platform access, and classification rules stayed stable or the change is clearly annotated. Otherwise the movement is likely a method artifact, not real progress or decline.
A baseline change happens when you deliberately expand the prompt set or add a competitor; that's a new baseline, not a trend point, and it should be logged as such. Real movement happens when the same prompts, brands, and rules produce a different answer distribution over repeated runs.
A jump from 12% to 30% recommendation share after adding five easy-win prompts is a method artifact. A steady climb across the same 60 prompts over eight weeks, confirmed by repeat runs, is real movement worth reporting to a client with confidence.
What should agencies show clients?
Agencies should show prompt-level evidence before the summary chart, because clients need to see which buyer questions competitors own and why. A scorecard without the underlying prompts invites doubt, not confidence.
A defensible client report leads with prompt rows and competitor reason themes before the scorecard, so the client sees the raw evidence behind the number. Add a source map showing which domains get cited and a flagged list of wrong facts the assistants are repeating about the brand.
Close with a fix queue that ranks the prompts and sources worth acting on first, ordered by buyer intent and gap size. This structure turns share of voice from a vanity chart into a work plan the client can execute against.
- Scorecard
- Prompt-level rows
- Competitor reason themes
- Source map
- Wrong facts flagged
- Fix queue
Questions that change the decision
Frequently asked questions
Do citations count as share of voice?
Citations count as their own measure, citation share, not as recommendation or mention share. Report brand-owned citations divided by the tracked citation set separately, since a citation without a recommendation isn't a buyer win.
Can multiple brands count in one answer?
Yes. AI answers frequently name several brands in a comparison, so each qualifying brand can count toward its own mention or recommendation share for that prompt. Decide this rule before collection, not after.
Should AI share of voice be weighted by platform?
Only if the weights reflect real audience distribution and are frozen before collection. Otherwise, report each platform as its own row on a common prompt spine and let stakeholders weigh them themselves.
What is a good AI share of voice?
There is no universal good number; it depends on category size, competitor count, and prompt set breadth. A published denominator and repeated measurement matter more than chasing a specific percentage.
Is AI SOV the same as SEO visibility?
No. SEO visibility tracks ranking positions in traditional search results; AI SOV tracks recommendation, mention, and citation behavior inside generated answers. They correlate loosely but require separate measurement setups.
Primary sources and research
Platform documentation supports factual statements. Where we describe an audit method or prioritization rule, that is AnswerMentions' operating judgment and is labeled as such.
- [1]Google Search Central: AI features and your websiteGoogle explains how AI features can surface links and how crawlable, eligible web content remains part of the Search foundation.
- [2]arXiv: repeated measurement of AI search resultsThe paper supports repeated observations and careful treatment of answer variability instead of relying on one run.
- [3]Aggarwal et al.: Generative Engine OptimizationThe GEO research provides precedent for measuring visibility in generative engines, while not acting as a universal ranking recipe.
- [4]NIST AI Risk Management FrameworkNIST frames AI measurement as a governed, monitored process with transparency, validity, and risk awareness.