AI Share of Voice: How to Measure Your Brand Visibility Across ChatGPT, Perplexity, and Claude in 2026
AI Share of Voice is brand mentions divided by all tracked-brand mentions in a declared prompt, engine, and sample frame. Use this evidence-bounded protocol to measure it without confusing mentions, citations, composite scores, or revenue.
On this page, AI Share of Voice means mention share: your brand-answer mention units divided by all tracked brand-answer mention units in the same declared prompt, engine, market, and sample frame. Count a tracked brand at most once in each eligible completed answer after applying the declared alias rule. AI Share of Voice is not recommendation share, Share of Citation, visibility rate, citation rate, position, an AI Visibility Score, or revenue attribution.
That distinction is the measurement contract. If a dashboard changes the event class, numerator, denominator, prompt set, engines, competitive set, geography, language, model mode, eligibility rule, or sampling window without saying so, the trend is not comparable.
Keep the event classes and denominators separate
| Metric | Numerator | Denominator | What it measures |
|---|---|---|---|
| AI Share of Voice / mention share | Eligible answers in which the brand appears, counted once per answer | All tracked brand-answer mention units in the declared comparison set | Competitive share of brand presence inside one frame |
| Recommendation share | Eligible answers classified as recommending or shortlisting the brand | All tracked-brand recommendation units under the same classification rule | Competitive share of recommendations, not passing mentions |
| Share of Citation | Citation appearances attributed to the brand under a declared ownership rule | All citation appearances attributed to tracked brands under that rule | Competitive share of attributed evidence |
| Visibility rate | Eligible completed answers in which the brand appears | All eligible completed answers, including answers with no tracked brand | How often the brand appears at all |
| Citation rate | Eligible completed answers citing the brand at least once | All eligible completed answers | How often the brand is cited in the sample |
| Position or prominence | Raw observed order, first-place count, or declared weighted position points | Present mentions, or all brands' weighted scores when reporting a share | Where the brand appears when order is measurable |
| AI Visibility Score | Components defined by the score owner | Weighting and scale defined by the score owner | A separately defined composite, not a synonym for any ratio above |
Use the formulas exactly:
AI Share of Voice / mention share = brand-answer mention units / all tracked brand-answer mention units in the same declared prompt, engine, market, and sample frame
Recommendation share = brand recommendation units / all tracked-brand recommendation units under the same classification rule
Share of Citation = brand-attributed citation appearances / all attributed tracked-brand citation appearances in the same declared frame
Visibility rate = eligible completed answers mentioning the brand / all eligible completed answers
Citation rate = eligible completed answers citing the brand / all eligible completed answers
A brand can be mentioned without being recommended. A recommendation can appear without a citation. A cited page can be attributed to a brand that is not named in the answer text. Position can improve while visibility falls. A composite score can combine these signals, but it must publish its own inputs, weights, scale, and version under the label AI Visibility Score.
Do not relabel visibility rate as Share of Voice. Visibility rate uses eligible answers as its denominator; AI Share of Voice uses tracked brand-answer mention units. Do not relabel citation rate as Share of Citation. Citation rate uses eligible answers; Share of Citation uses attributed citation appearances. Do not count a passing mention as a recommendation unless the classification rule says why.
Declare the measurement frame before collecting answers
Create a versioned frame specification before the first run. Record:
- the exact prompts and a stable prompt ID for each;
- the decision stage or intent tag assigned to each prompt;
- every engine, model or product surface, search/browse mode, and account state used;
- language, country or market, and any location or personalization settings;
- the tracked brands, aliases, products, and exclusion rules;
- the collection dates, times, and repeat-sampling schedule;
- what counts as a mention event;
- what counts as a citation and how a citation is attributed to a tracked brand;
- how duplicate links, repeated brand names, ambiguous entities, and missing citations are handled.
There is no evidence-backed universal prompt count, competitor count, cadence, or alert threshold for every company. Choose a frame large enough for the decision you need to make, document it, and hold its definition stable for period-over-period comparison. When the frame changes, publish a new series or show the old and new calculations side by side.
Build prompts from buyer decisions, not a fixed template
A useful prompt panel covers the decisions the measurement is supposed to inform. Start from first-party inputs such as sales questions, support logs, site search, customer interviews, request-for-proposal language, and known comparison tasks. Add discovery prompts only when they represent a real use case.
Tag prompts before collection. A practical operator-designed taxonomy might include problem definition, category education, comparison, vendor evaluation, implementation, risk, and troubleshooting. That taxonomy is a reporting choice, not an industry law. Keep the original prompt text and the tag history so a later edit does not silently rewrite the baseline.
Avoid padding the panel with easy prompts to raise the score. Report the result for each prompt and each prompt group so readers can see whether the aggregate is being driven by one narrow slice of the frame.
Repeat samples because generated answers are stochastic
One response is one observation, not a stable measurement. Generated answers can change across repeated runs, model versions, retrieval states, account contexts, and collection times.
Predeclare a repeat-sampling schedule for each prompt-engine pair. Store the raw answer, citations, timestamp, model or product label, search mode, and parser version for every run. Report the number of observations next to every percentage. If the design supports uncertainty estimation, report an interval or observed range rather than presenting a point estimate as exact.
Do not average away the structure of the sample. Always publish:
- Per-prompt results so prompt-specific gains and losses remain visible.
- Per-engine results because each engine is a separate measured stratum.
- Repeat-run counts and dispersion so output variance is not mistaken for a strategic change.
- A blended result only as an explicitly weighted roll-up, with the engine and prompt weights shown.
If one engine was sampled more often, a raw pooled percentage will overweight it. Calculate each declared stratum first, then apply the published weighting rule.
Count mentions and citations with auditable rules
A defensible mention rule is answer-level and entity-aware: count one mention event for a tracked brand in a sampled answer after resolving approved aliases, rather than rewarding an answer for repeating the same name. If you choose occurrence-level counting instead, label it and use it consistently. This is an operator-designed anti-duplication rule, not a universal standard.
A defensible citation rule records each source link, resolves the cited property to an owner or brand where possible, and preserves an unattributed state when ownership is unclear. Decide in advance whether repeated links in one answer count once or multiple times. Never force unknown citations into a tracked brand merely to make the denominator close.
Keep the atomic event table. At minimum, each row should contain frame version, prompt ID, engine, run ID, timestamp, named brands, cited URLs, attributed brands, mention context, citation context, and reviewer or parser status. Recalculate the metrics from that table rather than editing dashboard totals by hand.
What the retained sources can and cannot support
The sources below have bounded roles. None supplies a portable AI Share of Voice benchmark, a provider ranking rule, a universal prompt count, or a revenue forecast.
| Exact source | Source role and measured population | Unit, frame, and denominator | Evidence status | What it does not establish |
|---|---|---|---|---|
| From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms | Primary research preprint analyzing the public geo-citation-lab dataset: 602 controlled prompts across ChatGPT, Google AI Overview/Gemini, and Perplexity; 21,143 valid search-layer citations; 23,745 citation-level feature records; 18,151 fetched pages; and 72 extracted features. | Citation-level records and fetched pages, not tracked-brand mentions. The abstract does not state the prompt-topic mix, geography, collection dates, production account state, or a brand-competitive denominator. | Observational and analytical dataset study, not a randomized brand intervention. | It supports separating citation selection from answer-level absorption. It does not define AI Share of Voice, reveal a production provider mechanism, set a source-count rule, prove that a page feature causes citation, provide a current market benchmark, or connect citations to revenue. |
| Trakkr: AI Share of Voice | Canonical vendor methodology page separating mention share, recommendation share, citation share, visibility rate, and position. Its reviewed engine-coverage aggregate reports 825,343 prompts, 797,644 valid comparisons, and 6,439,133 model responses from August 12, 2025 through March 11, 2026 across eight tracked model families, with 43.3% average model agreement and 4.0% perfect agreement. | The engine benchmark counts the share of valid analyzed prompt responses returning at least one tracked-brand recommendation. The page does not publish raw prompts, raw responses, geography, brand-level rows, or exact model versions. | Vendor-published methodology and aggregate. It is a bounded vendor panel, not a portable brand benchmark or independent market census. | It supports keeping event classes and denominators separate and demonstrates a disclosed panel shape. It does not set a universal brand target, establish model market share, rank engines for every use case, expose production recommendation mechanisms, or prove business impact. |
| Meltwater: AI Search Visibility Report - May 2026 | Vendor-published GenAI Lens analysis of more than 8 million citations across eight major LLMs, comparing source composition between April and May 2026. | Citation appearances and monthly source totals inside Meltwater's panel. The article does not publish the raw citation-event rows, full prompt or query universe, geography, account state, customer composition, or a tracked-brand mention denominator. | Point-in-time source-composition panel published by the measurement vendor, not an independent experiment. | It can describe source composition and model-to-model variation in that vendor panel. It does not establish a universal source hierarchy, provider mechanism, causal content effect, portable AI Share of Voice benchmark, or revenue outcome. |
| Meltwater AI Visibility API overview | Vendor product documentation describing measurement across account-selected prompts and assistants such as ChatGPT, Gemini, Perplexity, and Microsoft Copilot, with results available per model. | The measured population is the customer account's selected prompts and configured providers. The overview does not publish one universal query set, geography, time window, competitive denominator, or benchmark population. | Product documentation, not an independent experiment or portable market panel. | It establishes what the vendor says its product can organize. It does not prove why an engine mentions or cites a brand, that one source type is preferred by mechanism, or that a product metric predicts pipeline or revenue. |
| Google Search Central: AI features and your website | First-party product documentation stating that links and traffic from Google AI features are included in Search Console's Web performance reporting. | The population is the verified Search property under the operator's selected date, query, page, country, and device filters. It has no cross-engine prompt panel or tracked-brand mention denominator. | Product documentation for Google Search, not a cross-provider visibility study. | It can support a Google traffic observation. It does not measure ChatGPT, Claude, or Perplexity visibility; expose a recommendation mechanism; identify which mention caused a visit; or establish conversion, pipeline, or revenue causality. |
The research preprint separates a citation being selected from its contribution to an answer. That is useful for audit design: keep mention, citation, and any separately coded answer-support signal as different fields. Trakkr's current methodology shows why recommendation share, visibility rate, citation share, and position need their own denominators; its aggregate remains a bounded vendor panel because the raw prompts, responses, geography, brand rows, and exact model versions are unpublished. Meltwater's May report is a point-in-time description of source composition across its own eight-model panel. The Meltwater API and Google documents describe product reporting surfaces. None of these sources should be merged into a claim about how all engines work.
Reporting template for a measurement cycle
Publish the frame before the headline percentage:
| Field | Example reporting shape |
|---|---|
| Frame version | 2026-Q3-v1 |
| Collection window | Start and end timestamps |
| Engines and modes | Exact product labels and search/browse state |
| Market | Language, country, location, account context |
| Prompt panel | Count, stable IDs, groups, additions, and removals |
| Competitive set | Tracked brands and alias rules |
| Sampling | Runs per prompt-engine pair and missing-run count |
| Counting | Mention, citation, attribution, and deduplication rules |
| AI Share of Voice / mention share | Mention units, tracked-brand mention denominator, percentage, by prompt and by engine |
| Recommendation share | Recommendation units, tracked-brand recommendation denominator, percentage, and classification rule |
| Share of Citation | Attributed citation appearances, attributed tracked-brand citation denominator, percentage, and ownership rule |
| Visibility rate | Brand-present answers, all eligible completed answers, percentage, by prompt and by engine |
| Citation rate | Brand-citing answers, all eligible completed answers, percentage, by prompt and by engine |
| Position or prominence | Raw order, first-place count, or weighted score with its denominator and missing-position rule |
| AI Visibility Score | Components, weights, scale, and version, if used |
| Uncertainty | Repeat-run range, interval, or other declared dispersion measure |
For a period comparison, show both the common-panel result and the full current-panel result. The common panel answers, "What changed under the same frame?" The full panel answers, "What does the current market-facing panel show?" Mixing those questions creates false trend precision.
Revenue and pipeline require a separate first-party join
Visibility metrics stop at observed answers and their cited sources. Revenue or pipeline requires a separate first-party attribution join.
Build that join from systems you control: tagged landing URLs where available, referrer and session data, self-reported discovery fields, lead and opportunity records, campaign or content identifiers, and closed-revenue data. Preserve the timestamp and join rule so an analyst can distinguish direct referral, assisted exposure, self-reported influence, and modeled influence.
Report the chain as separate proof layers:
sampled answer → mention → citation → referral or self-reported exposure → lead → opportunity → revenue
A result at one layer does not prove the next. A higher AI Share of Voice does not by itself establish more citations. More citations do not by themselves establish referral traffic. Referral traffic does not by itself establish pipeline or revenue. Public studies and vendor panels cannot substitute for this first-party join.
How to use the results without inventing a benchmark
Use your stable frame to diagnose changes, not to declare a universal score good or bad.
- A prompt-level decline may indicate a changed answer pattern, a competitor event, or ordinary sampling variation; inspect the raw runs before assigning a cause.
- An engine-level gap is an observation in that engine and frame, not proof of a permanent provider preference.
- High AI Share of Voice with low recommendation share means the brand appears in more category answers than it is shortlisted; it does not identify why.
- High AI Share of Voice with low Share of Citation means the brand receives a larger share of tracked mentions than tracked citations; it does not identify why.
- High visibility rate with low AI Share of Voice can occur when the brand appears often but competitors accumulate more total mention units across the same frame.
- High citation rate with low Share of Citation can occur when competitors accumulate more attributed citations per answer or across the frame.
- A composite AI Visibility Score can summarize a program only if its versioned components remain available underneath it.
Choose interventions after reviewing the cited sources and answer contexts in the affected prompt-engine strata. Measurement identifies where to investigate. It does not reveal a deterministic optimization mechanism.
Frequently Asked Questions
What is a good AI Share of Voice percentage?
There is no portable percentage that is good across every category. The denominator changes with the tracked brands, prompts, engines, sampling schedule, market, and counting rules. Use a stable internal baseline and a declared competitive frame rather than a vendor benchmark band.
How many prompts should an AI Share of Voice panel contain?
There is no evidence-backed universal count. Use enough prompts to represent the buyer decisions in scope, publish the list and groups, and preserve a common panel for trend analysis. A smaller audited panel is more defensible than a larger undocumented one.
How often should AI Share of Voice be measured?
Set cadence from the decision and the expected rate of change, then keep it stable. A launch or incident may justify a denser temporary sample; an executive trend may use a longer window. Do not call a movement meaningful until you have inspected repeat-run dispersion and frame changes.
Why does AI Share of Voice differ by engine?
The measured outputs can differ because the products, models, retrieval modes, indexes, interfaces, locations, and run times differ. Report each engine separately. Observed differences do not disclose a provider's source preference, weighting formula, or causal mechanism.
Can owned content or earned media improve AI Share of Voice?
Either may coincide with changes in a declared panel, but this page does not claim a universal earned-versus-owned share or deterministic effect. Test a specific intervention against a stable frame, preserve the before and after raw answers, and treat the result as evidence for that experiment rather than an industry law.
What is the difference between AI Share of Voice and Share of Citation?
AI Share of Voice is brand-answer mention units divided by all tracked brand-answer mention units in the same frame. Recommendation share uses recommendation units; visibility rate uses all eligible answers; Share of Citation uses attributed citation appearances; citation rate uses all eligible answers; and position uses a declared order or weighting rule. Keep every label and denominator distinct.