How to Measure AI Share of Voice in AI Search: ChatGPT, Gemini, Perplexity
How to measure AI share of voice: four denominators, a locked pool of 20 to 30 questions, three runs per engine and one archive scoring a peer 34% or 92%.
Quick answer
Under this method, Canlah AI measures AI share of voice by locking the client’s non-branded buyer intents, running each one at least three times per engine, recording every brand named in every answer and dividing by one denominator that is printed next to the percentage. The denominator decides the number. In one Canlah AI diagnostic of 32 non-branded questions, a Singapore omakase client scored 3% under prompt presence and 8% under a two-brand mention share, while its peer scored 34% and 92% from the same answers.
Canlah AI reports the result as a range across repeated runs of a locked query pool, with the denominator stated in the report. In its self-audit of September 7, 2026, 39 buyer questions asked three times each on the ChatGPT and Gemini legs returned 233 answers, 216 of them grounded and 177 of those to non-branded questions. Canlah AI was named in 7 of the 177, a mention rate of 4.0%, and in 0 of the 93 non-branded Gemini answers.
Treat any single figure as directional. One run cannot prove a stable share, cannot prove causation and cannot survive a competitor list that changes between cycles. If the number will move a budget, freeze the questions, the peer set and the denominator before the first run, then repeat the identical protocol.
Disclosure: Canlah AI publishes this guide and operates the measurement service described below. Platform behaviour was read from public product and pricing pages on September 23, 2026 and was not verified through paid accounts.
What does AI share of voice measure?
AI share of voice measures how much of the answer space belongs to your brand rather than to the brands named beside it, on a scale of 0 to 100. Mention rate tells you whether you appear at all. Share of voice tells you how much of the shortlist somebody else owns when you do.
The metric only means something when five properties of the run are fixed alongside it.
- Mention: whether the answer names the brand anywhere in the body of the response.
- Recommendation position: whether the brand is the lead recommendation, a shortlist entry or a passing reference. Semrush weights its figure by both mention count and how high the brand appears.
- Citation: whether the answer links to the brand’s own page, a third-party page or a competitor, or shows no visible source. Mention share and citation share move independently.
- Competitor set: which brands sit in the divisor and who decided that list.
- Engine, market and language: a share of voice in ChatGPT is not a share of voice in AI search, and an English figure is not a Chinese one.
Share of voice in AI search is therefore a composite. Two reports can score the same brand from the same answers with honest arithmetic and land 58 points apart, as the omakase peer below shows.
The four denominators that decide the number
AI share of voice is reported against four denominators, and the same 32 answers in one Canlah AI diagnostic scored a single peer at 34% under one and 92% under another. The table separates them by the division performed, not by the vendor performing it.
The four denominators used to report AI share of voice, with worked figures, based on public product documentation reviewed September 23, 2026.
| Denominator | The division performed | Worked figure | Platforms that publish it | What it answers, and the trap |
|---|---|---|---|---|
| Prompt presence | Answers naming you, divided by answers run | Omakase client 3% (1 of 32), peer 34% (11 of 32); Canlah AI self-audit 4.0% (7 of 177) | Ahrefs Brand Radar visibility, Otterly.ai Brand Visibility Index, Peec AI visibility | Whether you appear at all; it is low and slow-moving, and it says nothing about how crowded the shortlist is |
| Named-competitor mentions | Your mentions, divided by the mentions of you plus the peers you listed | Omakase client 8% (1 of 12), peer 92% (11 of 12), against one named peer | AthenaHQ, Peec AI, Profound, Otterly.ai | Where you sit in a shortlist you chose; it runs highest of the four, and dropping one peer raises it without a new mention |
| Category mentions | Your mentions, divided by every brand named in the category | Semrush worked example: 10 of 100 category mentions, 10% | Semrush Starter (SEO + AI Search), Evertune | How much of the category answer space you hold; it is hard to flatter, but it depends on who defines the category |
| Citation share | Your cited URLs, divided by every URL cited in the same answers | Canlah AI self-audit: eight domains held 458 of 797 citations, 57.5%, and canlah.ai was not among them | Otterly.ai link citations, Profound, Scrunch | Whether your own pages feed the answer; most brands start near zero, and a brand can be named often and cited never |
Worked figures come from the anonymised Canlah AI omakase diagnostic of 32 non-branded questions, the Canlah AI self-audit of September 7, 2026 and the Semrush guide. Platform attributions describe what each vendor publishes about its own metric, not independently verified maths.
Semrush publishes the category-mentions formula in its guide to measuring AI share of voice. The guide defines AI share of voice as your AI mentions divided by total AI mentions across all brands in your category, multiplied by 100, and gives the worked case of 10 mentions inside a category total of 100 returning 10%. Counting the mentions is the easy half of the job; choosing the divisor is the half that decides the headline.
Why the denominator, not the engine, moves the number
The denominator moved one Singapore omakase peer’s AI share of voice from 34% to 92% in a Canlah AI diagnostic, with no change in any engine. The anonymised restaurant was measured on 32 non-branded buyer questions. A directly comparable peer was named in 11 answers and the client in 1.
Under the prompt-presence denominator the peer scores 34% and the client 3%, because both are divided by the 32 questions asked. Under a two-brand mention denominator the peer scores 92% and the client 8%, because both are divided by the 12 mentions collected. Nothing changed in any engine. Only the divisor changed.
Citation share tells a third story from the same kind of archive. In the Canlah AI self-audit of September 7, 2026, the eight most-cited domains were aistudio.com.sg with 91 citations, mediaplus.com.sg with 85, stridec.com with 78, novastacks-ai.com with 53, hashmeta.com and oom.com.sg with 41 each, purpleclick.com with 36 and firstpagedigital.sg with 33. Those eight held 458 of the 797 citations recorded, or 57.5%, and canlah.ai was not among them. A mention-share report and a citation-share report of the same archive tell opposite stories about the same brand.
How to measure AI share of voice, step by step
Measuring AI share of voice takes nine steps, from a locked pool of 20 to 30 questions to a range that a second party can recompute.
- Write 20 to 30 non-branded buyer questions across category awareness, main shelf, narrow shelf and verification. Questions containing your own name belong in a separate branded layer that is excluded from the visibility rate.
- Name the peer set before the first run, in the contract or the brief, and fix its size. Adding a rival in month three changes the percentage without anything changing in the engines.
- Choose the denominator and write it down: prompt presence, named competitors, category mentions or citations. One report, one divisor.
- Fix the engines and the market. Canlah AI’s standard tracking covers ChatGPT, Gemini, Google AI Overviews and Google AI Mode: Starter covers ChatGPT, Growth adds Google AI Overviews, Full adds Google AI Mode and Gemini. Perplexity and the Chinese engines are not part of standard tracking.
- Run each question at least three times per engine, with phrasing rotations rather than a single literal string. AI answers are stochastically generated, and a single run measures one draw from a distribution.
- Record every brand named in every answer, in the order they appear, along with every cited URL. The order supports position weighting later; the URLs support citation share.
- Compute the percentage, then compute it again under a second denominator. If the two disagree by more than 20 points, the report needs both numbers, not the more flattering one.
- Report the result as a range across runs, not as a point. A figure of 8% to 17% is a truthful summary of three runs that returned 1, 2 and 2 mentions out of 12.
- Re-run the identical pool, peer set, engines and denominator on a fixed cadence. Keep the raw answers. A percentage whose archive was discarded cannot be challenged or corrected.
What to record on every run
An AI share of voice figure becomes reproducible when nine fields are logged per answer, not per cycle. Somebody who was not in the room can then recompute it from the log.
Nine fields to log per answer, as recorded in the Canlah AI self-audit of September 7, 2026.
| Field | What to record | Decision it supports |
|---|---|---|
| Query ID and layer | The frozen question text and its funnel layer | Whether a gap is category-wide or shelf-specific |
| Engine and model version | ChatGPT, Gemini, Perplexity and the model string returned | Whether a drop is your content or a model update |
| Phrasing rotation | Which of the three rotations produced this answer | Whether the result survives rewording |
| Market and language | Country setting and answer language | Whether an English figure hides a Chinese gap |
| Account state | Signed in, signed out, API or browser-rendered | Whether two cycles are comparable at all |
| Timestamp | Date and time of the request | Whether a change preceded or followed a content release |
| Brands named, in order | Every brand in the answer, in sequence | The numerator, the denominator and the position weight |
| Cited URLs | Every visible source link | Citation share and the source-gap backlog |
| Denominator applied | Which of the four divisors produced this figure | Whether this cycle can be compared with the last |
Source: the Canlah AI probe record format used in the September 7, 2026 self-audit. The field set is a working standard for reproducibility, not an industry specification.
Canlah AI stores each probe as a raw record holding the prompt, the full response, the cited sources and a timestamp, and the archive belongs to the client, so a rival agency can recompute the share from the same records.
How to interpret five common results
Five AI share of voice results recur across repeated cycles, and each one has a check to run before the number reaches a report.
Share rose and mentions did not: Check the peer list before reporting a gain. A competitor removed from the divisor, or a rival that stopped appearing for unrelated reasons, raises your percentage without a single new mention of your brand. Recompute the previous cycle under the current peer set and compare like with like.
Share is high and citation share is zero: The engines know the brand and are not reading its pages. In the September 7, 2026 self-audit, Canlah AI was named in 18 of 18 branded OpenAI answers and in 21 of 21 branded Gemini answers, while the non-branded layers returned 7 mentions in 177 answers. Recognition existed; retrievable evidence did not.
Share collapses when the peer list grows: This is the named-competitor denominator behaving normally, not a loss of visibility. Report the figure with its peer count attached, as in “14% against six named peers”, and the collapse becomes legible rather than alarming.
Share differs sharply between engines: Treat the engines as separate markets. The same self-audit returned 7 mentions in 84 non-branded OpenAI answers, or 8.3%, and 0 in 93 non-branded Gemini answers. Averaging those into one 4.0% headline would have hidden a total absence on one engine.
Share is stable and the pipeline is not: Share of voice is an exposure metric. It does not measure whether the buyer clicked, enquired or bought. Pair it with AI-referral attribution before attaching it to a revenue forecast.
What AI share of voice cannot prove
AI share of voice cannot prove five things: causation, stability, comparability with a rival’s figure, sentiment or citation.
- It cannot prove causation. A content release and a share increase in the same month are an association, not a proven effect.
- It cannot prove stability. Cited sources can change between monthly cycles, so a single-run figure carries an unstated confidence interval.
- It cannot be compared with a competitor’s published figure, because the question pool, run count, engines and denominator all differ.
- It cannot tell you whether the mention was favourable. Sentiment is a separate measurement.
- It cannot substitute for citation share. A brand can be named often and cited never, which is what the Canlah AI archive shows for its own domain.
When to move from a spreadsheet to a share of voice platform
A manual AI share of voice panel stops being reliable at about 30 questions run three times on three engines, which is 270 answers a cycle, and 540 once a second language is added. Before that point, a spreadsheet with the nine fields above is defensible and free. After it, transcription error costs more than a subscription.
Nine providers were reviewed for this guide, and they differ mainly on what their headline percentage divides by. That is the one thing a buyer should check before comparing two vendors’ numbers.
What nine providers divide by, read from public product and pricing pages on September 23, 2026.
| Provider | What its headline percentage divides by | Strength on the reviewed page | Entry price on the reviewed page | What to verify |
|---|---|---|---|---|
| Semrush Starter (SEO + AI Search) | All brands’ mentions in your category, weighted by position | Publishes its formula with a worked 10-in-100 example | $199 a month, or $165.17 a month billed annually | Which brands Semrush places in your category, because that list is the divisor |
| Evertune | Every competitor in the category | Pro tracks 100,000 prompts across 11 AI models | $800 a month Pro | How the category competitor list is built and whether you can edit it |
| Ahrefs Brand Radar | AI responses in your topic set that mention or cite you versus competitors | An index of more than 454 million prompts derived from Ahrefs keyword data, so you do not write the pool yourself | $199 a month for the AI Visibility Index at 83 prompts a day | Whether the index prompts match your buyers’ market and language |
| Profound | Brand and competitor appearance across tracked prompts | Enterprise covers nine engines, including DeepSeek and Google AI Mode | Free 7-day trial at 50 prompts a day, then custom Enterprise pricing | Which divisor the share figure uses, stated in writing before the first report |
| Peec AI | Tracked prompts, shown per brand beside named competitors | Visibility, position and sentiment reported side by side for each brand | Starter, Pro, Advanced and Enterprise tiers; price not found on reviewed page | How many peers sit in the divisor, because adding one moves the share |
| AthenaHQ | Competitor appearance per analysed response | Unlimited competitor tracking on Starter across 11 or more models | $295 a month Starter for 3,600 credits | How many credits one prompt run consumes across 11 models |
| Otterly.ai | Tracked prompts for visibility, cited links for citation share | Reports visibility and link citations as separate figures across four engines | $29 a month Lite for 15 prompts | Whether 15 prompts meet the 20 to 30 question minimum used in this guide |
| Scrunch | Mentions and citations across plan-bound tracked prompts | 350 custom prompts on Starter | $300 a month Starter, or $250 a month billed annually | Which divisor the headline share uses |
| Canlah AI | The denominator fixed in the engagement and printed in each report | Client-owned raw archive of prompt, response, sources and timestamp | Free AI visibility report within 48 hours; DeepAudit $99 for One read or $189 for Month read, creditable toward the agency service; GEO service from $1,500 a month per AI platform | Adding a peer is a change request, not a click; the Starter tier tracks 4 buyer questions on ChatGPT, re-measured monthly |
Descriptions refer to what each vendor publishes about its own metric. “Not found on reviewed page” means the provider’s public materials did not name that item when reviewed on September 23, 2026. It is not proof that the item is absent.
Canlah AI is a Singapore-based SEO + GEO agency that measures and improves brand visibility inside ChatGPT, Gemini, Google AI Overviews and Google AI Mode, reporting AI visibility as re-verifiable ranges with timestamped evidence. Share of voice arrives inside that programme, with the pool, the peer set and the divisor fixed in the engagement and printed in every report; after three months of the GEO service, clients can run the work themselves with Canlah All-Stars. Canlah AI offers a free AI visibility report within 48 hours, and DeepAudit costs $99 for One read or $189 for Month read; the fee can be credited toward the agency service (not toward Canlah All-Stars). The GEO service itself is priced per AI platform, from $1,500 a month.
The limitation is real. Adding a competitor is a change request rather than a click, and tracked volumes run from 4 to 16 buyer intents, each re-measured monthly, below the allowances of larger self-serve plans such as Otterly.ai Premium at 400 prompts and Scrunch Growth at 700.
Frequently asked questions
What is a good AI share of voice?
There is no universal threshold for AI share of voice, because every figure carries a different question pool, run count, engine list and denominator. A 40% share against three named peers and a 40% share against an entire category are not the same achievement. Judge the number against your own previous cycle measured under the identical protocol, and treat a transparent low figure as more useful than an unexplained high one.
Can I compare my AI share of voice with a competitor’s published number?
No. Two vendors measuring the same market on the same day will differ on the question pool, the run count, the engines, the account state and the divisor, and each of those alone can move a percentage by tens of points. Comparison is only valid when both figures come from one protocol run by one party.
How is AI share of voice different from mention rate?
Mention rate divides by the answers you ran, so it tells you how often you appear. AI share of voice divides by the brands named, so it tells you how much of the named space is yours. A brand can hold a 4.0% mention rate and a 25% share of voice in the same archive if the answers that do name it name few competitors.
How many prompts do you need before the percentage means anything?
A pool of 20 to 30 non-branded questions, each run at least three times per engine, is the working minimum for AI share of voice, because a single run samples one draw from a stochastic distribution. Below that, report counts rather than percentages: “named in 3 of 12 answers” is honest, while “25%” implies a precision the sample does not support.
How do I measure AI share of voice without paying for a tool?
To measure AI share of voice without a paid tool, ask a fixed list of 20 to 30 non-branded questions in ChatGPT, Gemini and Perplexity three times each, save every answer, count every brand named and divide by one stated denominator. Otterly.ai automates the counting from $29 a month for 15 prompts, and Canlah AI offers a free AI visibility report within 48 hours before any engagement. Do not paste customer data or unreleased product information into a public AI tool.
Is Canlah AI a tool or an agency?
Canlah AI is an agency that runs the measurement itself as a managed service and hands over the evidence archive; after three months of the GEO service, clients can run the work themselves with Canlah All-Stars. That suits teams who need a share of voice figure they can defend to a board or an auditor, and suits poorly a team that wants to add a competitor itself and see the chart redraw the same minute.
Method and sources
The four denominators were derived from public product documentation, pricing pages and vendor metric definitions read on September 23, 2026, without paid accounts. No dashboard was independently tested, no client work was audited and no commercial term was confirmed. First-party figures come from Canlah AI probe archives and are stated with their engine, window and denominator; they describe one brand, one category and one date, so they illustrate the method rather than benchmarking a market.
- Semrush, How to measure AI share of voice. The published formula and the worked 10-in-100 example.
- Semrush AI visibility and Semrush pricing. Position-weighted brand reporting and the Starter (SEO + AI Search) price.
- Evertune pricing. The Pro price and the 100,000-prompt, 11-model allowance.
- Ahrefs Brand Radar. The AI share of voice definition, the prompt index size and the AI Visibility Index price.
- Profound Answer Engine Insights and Profound pricing. Tracking scope, trial allowance and Enterprise engine list.
- Peec AI and Peec AI pricing. Metrics reported per brand beside named competitors and the tier names.
- AthenaHQ pricing. Credit model, the 11-model list and unlimited competitor tracking.
- Otterly.ai pricing. Lite, Standard and Premium prompt allowances and the four default engines.
- Scrunch pricing. Starter and Growth prices and prompt allowances.
- Canlah AI GEO services and Canlah AI pricing. Locked query pool, phrasing rotation, buyer intents by tier and monthly cadence.
- Canlah AI self-audit and probe archive, anonymised, September 7, 2026. Layer mention counts, engine split and citation distribution.
- Canlah AI cases, read September 23, 2026. The anonymised omakase diagnostic of 32 non-branded questions behind the 3%, 8%, 34% and 92% figures.
Do not accept a share of voice figure without its divisor. Ask any vendor, including Canlah AI, for the question pool, the peer list, the run count and the denominator in writing before the first cycle, then hold all four constant for as long as you intend to compare the numbers.
See where your brand stands in AI answers. Run the free AI visibility audit →
Related articles
FREE WHITEPAPER
Marketing in the Agent Era
All 13 chapters public — no email wall. Includes an original dataset on the agent-readiness of 50 cross-border DTC brands.
Read it free →