AI Share of Voice: Formula, Denominators, and Competitor Tracking

AI share of voice sounds like one clean percentage. It is not.

One dashboard may count answers that mention a brand. Another may count how often the brand is recommended. A third may count citations to the brand's domains. If they all label the result "AI share of voice," three different outcomes can appear to be the same metric.

Short answer: Define AI share of voice as a family of ratios, not a universal score. For a defensible competitor comparison, freeze the Questions, competitor cohort, answer-engine surfaces, market, language, device or account conditions, time window, execution policy, and coding rules. Then report answer coverage, competitive mention share, recommendation share, and citation share separately, with their exact numerators and denominators.

This guide gives you formulas, coding rules, a worked example, a CSV schema, and a measurement contract you can reproduce.

AI Share Of Voice Is Not One Official Standard

In traditional media, share of voice often compares one advertiser's spend or exposure with the category total. Search teams have also used the term for impression share, ranking presence, or estimated traffic. AI answers create a different measurement unit: a generated response can name several brands, recommend none of them, cite sources without naming their owners, or fail to complete.

There is no single official cross-platform definition that resolves those choices for you.

Therefore a report should never stop at:

"Our AI share of voice is 34%."

It should say:

"Our competitive mention share was 34% among three fixed brands, using binary brand presence in 112 completed eligible answers from 20 unbranded Questions across two named surfaces and four matched Runs in US English."

The second statement exposes the instrument. Another analyst can review it, rerun it, and understand what the percentage does and does not mean.

The broader AEO metrics guide separates mentions, citations, source quality, and framing. Share of voice adds a competitor-relative denominator to some of those signals; it does not erase the distinctions.

Start With The Experimental Unit

Use this raw observation:

Question × answer-engine surface × Run

For example:

Q-014 × Perplexity web answer × Run 2026-07-31-01

One scheduled unit should end in one recorded outcome. A retry used to recover a network error is not a bonus observation. Keep the retry log, but retain only the result selected by the predeclared retry policy as the unit's final outcome.

The unit should preserve:

• Exact Question and Question class.
• Channel and product surface, not only a vendor name.
• Requested and returned model when exposed.
• Search or retrieval mode when exposed.
• Market, language, device, and relevant account state.
• Start and completion time.
• Attempt status and metric eligibility.
• Raw answer and displayed citations.
• Brand, recommendation, and citation coding.
• Coding version and reviewer.

Do not treat ChatGPT search, an API model without search, Google AI Overview, Google AI Mode, and Gemini as interchangeable simply because they all use generative AI. If a surface changes, record a contract break or start a new baseline.

The Four Metrics You Should Keep Separate

Each metric below answers a different business question. A brand can improve one while declining on another.

1. Answer Coverage, Or Brand Mention Rate

Answer coverage asks:

In how many completed eligible answers did the target brand appear at least once?

Formula:

answer coverage = completed eligible answers mentioning target brand ÷ completed eligible answers

Code the target brand as binary for each answer:

• 1 if the governed brand name or an approved alias appears.
• 0 if it does not.

Repeating the brand name ten times inside one answer still contributes 1, not 10. This prevents verbose answers and repeated product names from receiving accidental extra weight.

Answer coverage is not competitor share. Its denominator includes completed eligible answers where no cohort brand appears. That makes it a useful discovery metric: it shows whether the brand appears at all for the selected opportunity set.

Keep Question classes separate. A Question such as "Is Acme a good analytics tool?" primes an Acme mention. Do not mix it into the headline result for unbranded discovery Questions such as "Which analytics tools support enterprise data residency?" The AI search query set guide explains how to govern unbranded, branded, comparison, problem, and buying-stage Questions.

2. Competitive Mention Share

Competitive mention share asks:

Of all brand appearances earned by a fixed cohort, what share belonged to the target brand?

For target brand A and fixed cohort C:

competitive mention share(A) = appearances(A) ÷ sum of appearances for every brand in C

An appearance is binary per brand per completed eligible answer.

Suppose one answer says:

"Acme and Beacon are strong options. Acme integrates with..."

Code:

• Acme: 1
• Beacon: 1
• Other governed cohort brands: 0

Acme does not get two appearances because its name occurs twice. Because the answer names two brands, it contributes two cohort appearances to the denominator. That is intentional: share is allocated across brand opportunities, not forced into one winner per answer.

This also means:

• An answer can contribute to more than one brand.
• Cohort shares sum to 100% when every governed appearance is assigned once.
• An answer with no cohort brand remains in answer-coverage denominators but adds nothing to the competitive-share denominator.
• If the cohort has zero appearances in the window, report competitive mention share as N/A, not 0%.

Freeze the competitor cohort before collecting data. Adding a category leader after seeing the result changes every brand's denominator, even if no answer changed. Include direct competitors, important substitutes, and category leaders only when you can explain why each belongs. Govern canonical names, old names, parent brands, product names, aliases, and ambiguous terms.

For the operational setup, use the competitor AI search tracking workflow and competitor benchmarking feature.

3. Recommendation Share

Recommendation share asks a narrower, higher-intent question:

Of all affirmative recommendations given to the fixed cohort, what share belonged to the target brand?

Formula:

recommendation share(A) = answers recommending A ÷ sum of answers recommending each brand in C

Again, code at most one recommendation per brand per answer. A response can recommend more than one brand when it presents a shortlist or assigns different brands to different needs.

Write the coding rule before review. A workable strict rule is:

• Recommended: the answer affirmatively proposes the brand for the defined need or places it on a recommended shortlist.
• Listed, not recommended: the brand appears as an example or exhaustive alternative without endorsement.
• Compared: the answer contrasts brands but does not affirmatively recommend this one.
• Caveated: the brand is mentioned mainly as a poor fit, warning, or exception.
• Absent: no governed name or alias appears.

"Acme offers an API" is a factual mention, not automatically a recommendation. "Consider Acme if data residency is your priority" is a conditional recommendation. Keep the raw sentence so a reviewer can audit the judgment.

Report recommendation coverage beside recommendation share:

target recommendation coverage = completed eligible answers recommending target ÷ completed eligible answers

A 50% recommendation share can sound dominant, but it may mean one target recommendation out of only two cohort recommendations in 100 answers. Coverage exposes the scale.

If no cohort brand receives a recommendation, report recommendation share as N/A and show zero recommendation coverage. There is no recommendation space to allocate.

4. Citation Share

Citation share asks:

How much of the governed citation space belongs to the target brand's source set?

You must choose what constitutes one citation unit. Common options are:

• Displayed citation occurrences.
• Unique cited URLs per answer.
• Unique cited domains per answer.
• Answers containing at least one governed-domain citation.

Do not switch units between Runs. Unique cited domains per answer is often easier to review than raw link occurrences because one page repeated several times does not inflate the count.

With unique governed domains per answer:

custom citation share(A) = unique per-answer citations to A's governed domains ÷ unique per-answer citations to all cohort domains

This is your custom cohort metric. It is not necessarily the same as a first-party platform metric.

If the governed cohort receives no eligible citation, report the share as N/A and retain the zero-citation coverage result.

Microsoft's June 2026 Bing Webmaster Tools Citation Share announcement supplies an important official boundary. Bing defines its Citation Share as the percentage of citations attributed to a site out of all citations shown across all sites for the same grounding query. Microsoft explicitly describes it as an observational metric, not a ranking system or competitive scoreboard; it does not expose competitor domains, represent traffic share, or assign content quality.

Therefore label the source:

• Bing first-party Citation Share, when reporting Bing's metric.
• Tracked-cohort citation share, when calculating your own governed-domain ratio.

Do not combine them as if their data universe and denominator were identical.

A citation can also exist without a brand mention, and a brand can be recommended without an owned citation. The cited-but-not-mentioned funnel explains why source selection, answer absorption, entity mention, recommendation, and click are separate stages.

Define Completed Eligible Answers Before Looking At Results

The denominator is where many share-of-voice reports become unreliable.

Track at least these counts:

A completed answer that exposes a citation surface but displays no citations is usually citation-eligible with zero citations. A surface that never exposes inspectable citations is citation-ineligible. Those are different observations.

Do not:

• Count a timeout as "brand absent."
• Drop failed answers without showing the completion rate.
• Retry until the brand appears.
• Treat every retry as an independent answer.
• Exclude no-brand answers from answer coverage.
• Include only answers that cited somebody when reporting mention coverage.

If refusals are a meaningful behavior for the Question class, show a refusal rate. For routine operating errors, classify them as failed after the fixed retry policy. Whichever rule you choose, apply it before comparing brands and keep it stable.

Freeze The AI Share-Of-Voice Measurement Contract

Use this contract before the first Run:

The public AI visibility measurement methodology provides the wider Task, Run, report, citation, and score framework.

Version the contract. If you change Questions, cohort membership, aliases, surfaces, locale, eligibility, or aggregation, start a new baseline or draw an explicit break in the trend.

Choose Pooled Or Equal-Channel Weighting Deliberately

Suppose one surface completes 96 answers while another completes 42 because of availability problems. A pooled ratio lets the first surface dominate:

pooled share = target appearances across all surfaces ÷ cohort appearances across all surfaces

That can be appropriate when each completed answer is intended to have equal weight. It is not appropriate if the business promised equal representation for each channel.

For equal-surface weighting:

1. Calculate the share separately for every eligible surface.
2. Average the surface-level shares with equal weight.
3. Show each surface's completed and failed counts beside the summary.

Do not quietly move between pooled and equal-surface reporting. Keep the channel detail even when leadership wants one summary number. The Google-versus-Bing AI visibility report guide shows why first-party products measure different layers and should not be collapsed too early.

Worked Example: The Same Answers, Four Different Results

The following numbers are illustrative, not AEO Table customer data.

A team freezes:

• 10 unbranded buyer Questions.
• 2 named answer-engine surfaces.
• 3 matched Runs.
• US English.
• Desktop where the surface exposes device state.
• Target brand Acme and competitors Beacon and Cinder.

That creates 60 attempted units. Four fail under the predeclared retry policy, leaving 56 completed eligible answers.

Per-Run Competitive Mentions

Acme's aggregate answer coverage is:

18 ÷ 56 = 32.1%

Acme's aggregate competitive mention share is:

18 ÷ (18 + 25 + 12) = 32.7%

Those values are close here, but their meanings differ. In this illustrative coding, 11 completed answers contain no cohort brand, while some answers contain multiple brands. Answer coverage uses all 56 completed eligible answers. Competitive mention share uses 55 governed brand appearances.

Now suppose the coding shows:

Acme has:

• 32.1% answer coverage.
• 32.7% competitive mention share.
• 36.4% recommendation share.
• 35.0% tracked-cohort citation share.

None of these is "the true SOV." Together they say Acme receives about one-third of governed cohort presence, but its recommendation share is slightly higher than its raw mention share in this observed window.

Do not infer traffic, revenue, market share, or cause from the table. Review the answers:

• Which Questions produced Acme's recommendations?
• Were the recommendations favorable and accurate?
• Which sources supported them?
• Did one surface create most of the movement?
• Were competitor names already present in the Question?
• Did any aliases create false matches?

The evidence decides the action; the ratio tells you where to inspect.

Repeat Runs And Report The Observed Range

One Run is a timestamped sample, not a stable market ranking.

AI answers can change with retrieval, source freshness, model behavior, timing, surface design, and ordinary generated-answer variation. The 2026 preprint Don't Measure Once: Measuring Visibility in AI Search argues for treating AI visibility as a distribution across repeated observations. Another 2026 preprint, Quantifying Uncertainty in AI Visibility, observed substantial citation variation in repeated samples across three answer systems and three consumer-product topics.

These are preprints with defined study scopes, not universal sample-size rules. They do not prove that four Runs, ten Runs, or any other fixed number is sufficient for every category.

Use a practical progression:

1. First matched Run: establish the pipeline and inspect coding.
2. Repeated baseline Runs: learn the ordinary range for the frozen Task.
3. Confirmation Runs: check whether a surprising movement persists.
4. Matched pre/post windows: evaluate a content or positioning intervention under a predeclared plan.

Show:

• Every per-Run value.
• The observed minimum and maximum, or another declared descriptive interval.
• Attempted, completed, failed, refused, and ineligible counts.
• Per-Question and per-surface breakdowns.
• The raw answers responsible for large changes.

Be cautious with a textbook binomial confidence interval. Repeated answers to the same Question and answers from the same surface can share sources and behavior, so rows may not be independent. High-stakes statistical claims need an analysis that respects repeated Questions, channels, and time.

The AI search volatility guide explains how to distinguish a snapshot, repeated matched Runs, and a durable movement. If the goal is to test a content change, use the AEO content experiment protocol.

What First-Party Reports Can And Cannot Confirm

First-party webmaster reports can validate important parts of the source layer, but they do not automatically calculate your competitor SOV.

Bing Webmaster Tools

Bing's AI Performance reporting can show publisher citation activity across supported Microsoft AI experiences. Its 2026 preview adds Intents, Topics, Citation Share, and period comparison. As Microsoft states, its Citation Share is site-relative citation presence for a grounding query, not a revealed competitor scoreboard.

Use it to corroborate:

• Whether Bing observed citations to your site.
• Which grounding-query themes and topics are associated with citation activity.
• Whether first-party citation presence changed over comparable periods.

Do not translate it into:

• Brand recommendation share.
• Cross-engine market share.
• Traffic share.
• A quality or ranking score.
• Named competitor performance.

Google Search Console

Google's June 2026 Search Generative AI performance report announcement says the pilot reports impressions for AI Overviews and AI Mode in Search, plus a separate Discover view. Google's published dimensions include pages, countries, devices for Search, and dates.

The official Search Console help documentation defines an impression as a link to the site shown in a supported generative AI feature and notes that the report is rolling out to a subset of site owners.

That first-party impression evidence is useful. The published report does not give you the competitor cohort, raw generated answer, brand-mention coding, recommendation coding, or Question-level SOV denominator required by the formulas in this guide.

Use first-party and controlled evidence side by side:

No single row replaces the others.

A CSV Schema You Can Audit

Keep one row per scheduled unit. Array fields can use JSON in CSV or normalized child tables in a database.

Add a separate execution log for retry number, latency, provider errors, and request IDs. Do not let multiple retries become multiple SOV rows.

From this table, calculate each metric with code that is versioned beside the codebook. Save the aggregate export and the eligible row IDs behind every published number.

Use Decision Thresholds, Not Decorative Targets

There is no universal "good AI share of voice." A three-brand niche and a twenty-brand category create different denominators. A 10% movement can be ordinary variation for one Task and operationally important for another.

Predeclare a decision rule with four gates:

1. Data-quality gate: completion, failure, eligibility, and cohort definitions stayed inside declared limits.
2. Magnitude gate: the movement exceeds the smallest change worth acting on.
3. Persistence gate: matched Runs place the movement beyond the Task's ordinary observed range often enough for the decision's stakes.
4. Evidence gate: raw answers show the expected Questions, framing, and sources rather than a coding error or unrelated event.

An illustrative editorial rule could be:

"Open a high-priority content investigation when unbranded competitive mention share falls at least five percentage points below the established baseline range in three of four matched Runs, completion remains above the declared floor, and answer review identifies the same missing proof or competitor source pattern."

That is an example operating heuristic, not an industry benchmark or statistical guarantee.

Choose the response that matches the evidence:

• Low coverage, normal competitive share: the entire cohort rarely appears; review the Question set and category maturity.
• Stable coverage, falling competitive share: competitors gained governed appearances; inspect their framing and sources.
• Rising mention share, flat recommendation share: awareness improved, but buyer-fit or proof may still be weak.
• Rising citation share, flat mentions: owned content is being selected without carrying the brand into the answer.
• Rising recommendations, worsening accuracy: stop celebrating the share and fix the claim risk.
• Movement on one surface only: investigate that surface before changing the whole content strategy.

Turn recurring evidence gaps into comparison pages, current documentation, primary research, clearer product facts, third-party proof, or corrections. Do not copy the competitor page; answer the buyer's unresolved question with better evidence.

Common AI Share-Of-Voice Mistakes

Counting Words Instead Of Answers

A long answer can repeat one brand many times. Use binary presence per brand per answer unless a different occurrence-based method is explicitly justified.

Changing The Cohort Mid-Chart

Adding or removing a competitor changes the denominator. Version the cohort and break the trend line.

Mixing Branded And Unbranded Questions

Named-brand Questions prime mentions. Report each Question class separately.

Hiding No-Brand Answers

No-brand answers belong in coverage. Their absence from the competitive-share denominator should be visible through a separate no-cohort-brand rate.

Treating Failures As Absence

A timeout says nothing about brand visibility. Show failed and completed counts.

Pooling Channels Without A Weighting Rule

The surface with more completed units can dominate a pooled result. Publish channel results and declare the summary weighting.

Calling Citation Share Recommendation Share

A source can be cited while its brand is absent or criticized. Code citation, mention, recommendation, and accuracy separately.

Claiming Causation From A Before/After Chart

A movement after a page edit can coincide with model changes, recrawling, news, competitor activity, or ordinary variance. Use a predeclared AEO content experiment when the decision requires stronger evidence.

Comparing Vendor Scores Without Their Contracts

Tools can differ in Questions, models, search modes, locales, retry rules, aliases, competitors, eligibility, and aggregation. Compare raw definitions before comparing headline percentages.

The Bottom Line

AI share of voice becomes useful when the ratio is less important than the contract behind it.

Start with a stable unit: Question × surface × Run. Freeze the competitors, Questions, markets, languages, execution policy, coding rules, and time window. Preserve every completed, failed, refused, and metric-ineligible outcome. Then report:

• Answer coverage.
• Competitive mention share.
• Recommendation coverage and share.
• Citation coverage and share.
• Per-surface values and repeated-Run ranges.
• Raw answer and source evidence.

Call the result an observed share for that governed measurement scope, not a universal AI market share.

Create an AEO Table account to organize stable Tasks, repeatable Runs, competitor evidence, citations, and reviewable reports before the next share-of-voice decision.