Run the same buyer question through ChatGPT a dozen times this week and you will get vendor lists that disagree with each other. SparkToro and Gumshoe's research found AI engines are highly inconsistent when recommending brands, and Search Engine Land's coverage of the same study reported that AI recommendation lists rarely repeat exactly. Any share-of-voice number built from one answer per question is noise wearing a percentage sign. The honest metric is an appearance rate, measured across repeated runs.
The definition fits in two lines. Appearance rate: answer runs naming your brand, divided by total runs of that question. AI share of voice: your total appearances divided by all tracked brands' appearances in the same runs, per engine, over a stated window. The stakes justify the discipline: 51% of software buyers now start research with an AI chatbot more often than with Google, G2's research found AI chatbots and review sites are now the two biggest influences on B2B shortlists, and buyers evaluate roughly five vendors with most of the list set before they ever contact sales.
Why is one answer per question not a measurement?
Because the engine is probabilistic, a single answer is a sample, and appearance rates are the only stable ground. A 2026 variance-components study found run-to-run noise large enough to swamp real brand differences in small samples, and work on quantifying uncertainty in AI visibility puts wide confidence intervals around small-sample rates. Our own floor is several weeks of answers per question, across phrasings and engines; neither study sets a number.
This is where AI share of voice parts company with the search version. Traditional share of voice is a position in a ranked list many brands occupy at once; an AI answer names a short list of vendors and everyone else is absent. Presence is close to binary, which makes the appearance rate, and who appears alongside you, the unit of competition.
Keep the rungs separate from the start: a mention is your name in the answer, a citation is your page used as a source, and a recommendation is the engine advising the buyer to pick you. Profound's citation research found more than 4 in 10 AI citations point to sources brands do not control, so your citation count and mention count can tell opposite stories.
How do you calculate AI share of voice?
Count appearances across repeated runs, then divide twice. Per question, appearance rate equals runs naming your brand over total runs. Across the panel, share of voice equals your total appearances over all tracked brands' appearances in the same runs. Log citations and recommendations as separate columns.
A worked example: five buyer questions, each run twelve times on one engine in a month, sixty answer runs. You appear in 21 runs, Competitor A in 48, Competitor B in 33, Competitor C in 12: 114 tracked appearances in total.
| Brand | Appearances in 60 runs | Appearance rate | Share of voice |
|---|---|---|---|
| Competitor A | 48 | 80% | 42% |
| Competitor B | 33 | 55% | 29% |
| You | 21 | 35% | 18% |
| Competitor C | 12 | 20% | 11% |
The single number says 18% and third place. The per-question split says more: those 21 appearances were 11 of 12 on one question, 8 of 12 on another, and 2 of 36 across the remaining three. That split is a work plan; 18% is only a scoreboard.
Two rules keep the number honest: state the window and run count with every report, and never compare it against one built on a different panel or counting rule.
Which questions and competitors belong in your measurement panel?
The questions a real buyer would ask before choosing, and the competitors the engines actually surface: both are empirical, and neither comes from your positioning deck. Semrush's AI Visibility Index analyzed 126 million AI search prompts, so any panel is a deliberate sample of an enormous space. Write five to ten questions in buyer language, skipping the brand-name lookups you already win.
Build the competitor set from the answers rather than from instinct. Run the panel once and log every vendor the engines name, including brands your sellers never meet: a third of buyers said an AI chatbot led them to choose a vendor they hadn't considered, and omitting those brands inflates every share number in the report.
Size the set for stability. Two or three brands make the metric jumpy, because one competitor launch swings every ratio; fifteen dilute it with vendors no buyer weighs.
What does a single universal score hide?
It hides the splits that tell you where you are losing, the entire point of measuring. Engines retrieve differently: Seer found 87% of SearchGPT citations matched Bing's top results, so a brand strong in Bing's index can lead on one engine and trail on another, and the same uncertainty research reports platform medians that differ by engine. Average across engines and the platform where you are losing vanishes into the one where you are winning.
The same erasure happens across rungs. A score that blends mentions, citations, and recommendations with fixed weights asserts that a citation is worth some exact multiple of a mention, a number nobody has measured. The IAB splits AI-era brand measurement into four groups, Presence, Prominence, Portrayal, and Persuasion, because one number cannot carry them all.
A dashboard selling one universal score is selling comfort: a blended number cannot fall on the question that funds your quarter while rising overall. The honest report is small and split: appearance rate per question, per engine, per rung, with run counts attached. If a tool cannot show you the runs behind its number, treat the number as marketing.
How do you read movement without fooling yourself?
Test any change against run-to-run noise first, because an appearance rate that moves within its variance band has not moved. At ten answers per question, one flipped answer moves a rate by ten points. Before escalating, re-run the panel and see whether the movement survives.
Read absolute and relative together: your appearance rate can climb while your share of voice stays flat, which means the category got more visible and you only kept pace.
Falling share despite good work usually means a competitor shipped the evidence the engines now quote, or an engine changed retrieval: a 2026 analysis of brand dynamics in LLM recommendation systems found even a well-known brand's edge disappeared when a competitor had a rating advantage of less than 0.1 stars. A sudden jump deserves suspicion before it earns a board slide.
What actually moves AI share of voice?
Nothing is proven to, but the strongest case is for evidence that engines can retrieve and quote, on your pages and in the sources they cite; phrasing tricks have even less behind them. The GEO study (Aggarwal et al., KDD 2024) found adding statistics, quotations, and citations lifted citation visibility by roughly 30 to 40% in its benchmark. The same literature carries the corrective: C-SEO Bench (NeurIPS 2025) tested ten conversational-SEO rewrite methods, the evidence additions among them, and found most of them largely ineffective and often harmful to citation rank, with gains shrinking as more sites adopted them. Add evidence buyers can check; do not rewrite prose for machines.
Some language actively hurts: scarcity and exclusivity framing measurably reduces how often an LLM recommends a product. Comparison and alternatives questions are the highest-value targets: they force the engine to name several vendors, so your absence is visible in every run. A comparison page with checkable claims about you and your rivals gives the engine something quotable.
Spend where the citations are: more than 4 in 10 AI citations point to sources brands do not control, so much of the answer can be decided off your site. Match the work to the rung you are losing: retrievability if you are absent, on-page evidence if you are mentioned without being cited, earned sources if a rival owns the answers.
How do you measure AI share of voice in ChatGPT, Perplexity, and Google AI Mode without a tool?
A spreadsheet, a question list you do not edit mid-cycle, and one row per answer run. Each row holds date, engine, question, run number, separate mentioned, cited, and recommended columns, and the other brands named. A row per question collapses the runs, so nothing can be recounted later.
Run the same wording in ChatGPT, Perplexity, and Google AI Mode, capturing each answer's source list, not just the answer text. Google states that AI Overviews and AI Mode use a query fan-out technique, and ChatGPT search issues one or more targeted queries, so the searches behind the answer are rarely just the words you typed.
Start each run in a fresh chat, signed out where the engine allows it: earlier turns in the same conversation steer later answers. Save every cycle as a dated sheet, or you can never show the number moved.
Manual work breaks when the question list or engine count outgrows the hours. Tooling then buys hours rather than certainty, since the same number of answers carries the same confidence interval either way.
How does Trovance measure your share of voice against competitors?
Trovance runs the protocol in this article as a standing system instead of a monthly spreadsheet exercise. You define tracked questions in buyer language, and it sweeps them across AI engines every week, preserving every answer run as a snapshot with its full context: which brands were mentioned, which were cited, which were recommended, and which sources carried the answer. Appearance rates and share of voice are computed from those preserved runs, per question and per engine, with run counts visible.
Because the raw runs are kept, the competitive picture stays inspectable. Share of answer shows where each competitor appears and where you are absent, and any number can be traced back to the snapshots behind it. When a rival holds a question, the citations in those runs show whether they won it on their own pages or on a third-party source, which is the difference between a content task and an earned-media task.
Measurement then feeds work. The Brain keeps the claims you are entitled to make with the proof behind each one, and recommended actions arrive as a ranked queue of content worth making, backed by answer evidence for the question you are losing. Drafts are written in your voice from your author voice guidance and research sources, a person reviews and approves everything before it publishes, and the next analysis cycle reruns the same questions so you can verify whether the asset moved the rate.
What Trovance will not do is report a universal score or promise a share-of-voice target. The engines are probabilistic, competitors keep publishing, and a single number blended across engines and rungs would hide exactly the failures this article describes. What it offers instead is a preserved, comparable record: the same questions, run after run, cycle after cycle, so the movement you act on actually happened.
What should you do this week?
Build the smallest honest panel and run a few weekly cycles of the spreadsheet workflow above: five buyer questions, two or three phrasings each, two engines, mentions, citations, and recommendations logged in separate columns. Compute appearance rates per question and share of voice across the set you actually observed.
Then resist the summary. Do not average engines or blend rungs, and put the run count on every number you circulate. Where the split shows a question you never appear on, that is next month's work; where it shows variance, run another cycle instead of calling a strategy meeting.
Appearance rates move as engines re-crawl and re-retrieve, and anyone promising a share-of-voice target on a schedule is selling against the evidence above. If you would rather define the questions once and let the runs accumulate, start a free Trovance trial and measure your share of voice from preserved answer runs instead of a one-off spreadsheet.
Measure without the theater
How to measure AI search visibility without one score: the framework this metric belongs inside
There is no such thing as an AI visibility score: why a universal number fails on contact with data
The prompt panel is measuring the wrong question: how to choose buyer questions worth tracking
The 10-minute audit of what ChatGPT tells buyers about you: the fastest first look before you build a panel
Close the gaps you find
Why ChatGPT recommends your competitor: the five failure modes behind a losing share number
Where AI citations come from across industries: which sources AI citations point to, by industry and engine
How to build a competitor comparison page that survives scrutiny: the asset comparison questions keep asking for
A visibility gap is not a content brief: deciding what the measurement tells you to do
