Share of Voice in AI Search

A fraction borrowed from advertising

Share of voice in AI search is the attempt to express a brand's presence in generative answers as a proportion: of all the times an AI assistant could have named or cited a brand in your category, what proportion went to you rather than to a competitor. A share is a ratio, and a ratio means nothing until both of its terms are defined, which is where this metric runs into trouble immediately.

The term arrives from advertising, where it has a precise and auditable meaning — a company's advertising spend as a proportion of total category spend, or the relative portion of ad inventory available to one advertiser in a defined market over a stated period. It works there because both terms of the fraction are countable and the market boundary is set by the buyer rather than guessed at.

Transplanted onto AI answers, the numerator becomes one of at least three things — brand mentions in the answer prose, citations of your domain as a source, or an undisclosed blend of both — and the denominator becomes one of at least three others: all mentions or citations across every brand, all prompts in the tool's prompt set, or all answers in which any brand of that kind appeared.

That is the whole problem in one sentence. The advertising metric works because the denominator is a real, finite, externally defined market. The AI version has no such denominator, because there is no fixed inventory of answers. A generative surface can produce an unbounded number of answers to an unbounded number of prompts, and nobody has enumerated the population. Every commercial AI share-of-voice figure is therefore a share of a sample the vendor chose, not a share of a market.

The one version whose denominator is real

There is exactly one exception as of 29 August 2026, and it deserves the space. On 16 June 2026 Microsoft shipped Citation Share in Bing Webmaster Tools, defined as "the percentage of citations attributed to your site out of all citations shown across all sites for that same grounding query".

Four properties make that different in kind from every vendor metric. The denominator is real, because Microsoft can count every citation it served for a grounding query — it served them, so this is a census of its own output rather than a sample of somebody's guess about demand. The unit is a grounding query, the retrieval query the AI system actually issued rather than the words a person typed, which places the measurement at the level where the selection decision was made. The surface set is stated: citations across Microsoft Copilot, Bing and select partner AI experiences. And Microsoft refuses the competitive reading outright, stating that Citation Share "does not expose competitor domains" and is not a "competitive scoreboard".

That last constraint is why Citation Share will not displace the commercial products, and it is worth being blunt about the reason. What buyers want from share of voice is a named leaderboard. The only party in a position to produce an accurate one has declined to, so the market is served instead by measurements that can produce a leaderboard precisely because they are not censuses of anything.

How a vendor figure is assembled

Every commercial AI share-of-voice product follows the same sequence, and each step embeds a decision the buyer usually never sees.

• Choose a prompt set. Typically dozens to a few hundred prompts thought to represent category demand. This is the single largest determinant of the output: change the prompts and the score changes. No AI operator publishes query data the way Search Console publishes queries, so prompt sets are built from keyword research, client input or model generation, and they are proprietary.
• Choose surfaces. ChatGPT, Gemini, AI Mode, AI Overviews, Perplexity, Copilot. Coverage differs by vendor, and because the surfaces overlap so little — 13.7% citation overlap between AI Overviews and AI Mode by Ahrefs' December 2025 measurement — surface selection alone moves the number substantially.
• Run each prompt some number of times. Rarely disclosed and entirely decisive, for reasons the next section sets out.
• Parse for the brand. Mentions require string and entity matching, with all the ambiguity that implies for brands whose names are ordinary words. Citations require domain matching, which is clean. A blended visibility score combines a noisy measure with a precise one at an undisclosed weight.
• Divide by something. All competitor mentions, or all prompts, or only prompts where some brand appeared. The choice moves the result by large factors and is very often unstated.

None of these steps is illegitimate. All of them are invisible on the dashboard, which is what turns a defensible internal measurement into an unauditable external number.

Mentions and citations are two quantities, not one

The numerator problem deserves its own treatment, because the two candidates come apart in both directions and both directions have been named and measured.

Cited but not mentioned is what Seer Interactive called the ghost citation in March 2026: the answer uses your page as a source and links it without ever speaking your brand name (Seer Interactive). Mentioned but not cited is the reverse and is at least as common: the answer recommends you by name while linking somewhere else entirely.

Semrush's June 2026 index keeps the two apart on purpose, defining mentions as "how often a company appears in an answer" and citations as "which domains and pages AI platforms use as evidence", and reports the two sets diverging sharply on Gemini. A single share-of-voice number that silently averages them is reporting a quantity with no referent, and the first question to ask of any movement in such a score is which of the two moved.

There is a further temptation to resist here. Weighting citations by where they appear in an answer looks like a refinement and is not one. Microsoft states that its citation data "does not indicate ranking, authority, or the role of any page within an individual answer". Google states that all links in an AI Overview are assigned the same position. Two platforms have separately said there is no meaningful ordering of sources inside an answer, so a position-weighted share is a vendor's invention presented as a measurement.

The base rate nobody accounts for

Share of voice assumes there is a voice to share. Frequently there is not, and the denominator quietly absorbs the difference.

Semrush's July 2025 study of 5,000 keywords found that 1.7% of AI Mode responses contained no links at all, that 92% carried a sidebar of links and 7% placed links below the response. More consequentially, a large share of answers name no brand whatsoever — a factual or instructional answer has no brand share to divide, and no vendor publishes what proportion of its prompt set produces answers of that kind.

This is where two vendors measuring the same brand over the same prompts can diverge before either has parsed a single answer. A tool that computes share only across prompts where some brand appeared reports a materially higher number than one dividing by all prompts, and the entire gap is an artifact of the denominator rather than anything about the brand. It is rarely stated either way, and it is a fair question to put to any vendor in writing.

Why two vendors can both be internally correct and disagree by a multiple

Share of voice is a ratio of two counts, and in this field both counts are noisy. The variance has been measured. SE Ranking parsed the same 10,000 keywords three times on the same day in Google AI Mode and found 9.2% of URLs present in all three runs, with 14.7% domain overlap across the three (SE Ranking, 29 August 2025). Ahrefs reported 45% of AI Overview citations changing between generations. A share computed from one run per prompt therefore carries an error term larger than most of the quarter-on-quarter movements a report would highlight, and both the numerator and the denominator carry it, so the ratio can move without any brand's underlying position changing at all.

Layer that on top of divergent prompt sets, different surface coverage, different numerator definitions and different denominators, and the conclusion is unavoidable: there is no conversion factor between two vendors' share figures. They are not the same measurement expressed on different scales. They are different measurements sharing a name.

Trend lines from a single vendor deserve the same scrutiny for a narrower reason. Semrush's own index moved from 2,500 prompts to 126 million between successive versions. A series whose denominator changes by five orders of magnitude is not a series, and the question to ask before comparing any two dates is whether the prompt set and the sampling depth were identical on both. A changed prompt set invalidates the comparison however the chart is drawn.

Reading the 36-brand finding correctly

Semrush's expanded AI Visibility Index covered 126 million US AI search prompts collected between January and April 2026, across ChatGPT, Gemini, AI Mode and AI Overviews, in 22 industries. Its most-quoted result is that only 36 brands maintained visibility across every platform studied (Semrush, June 2026).

That figure is routinely repeated as evidence that AI search is extraordinarily concentrated, with 36 brands dominating the field. Read carefully, it says something different and more interesting. It is a statement about cross-platform consistency: 36 brands held visibility on every surface at once. Given that AI Overviews and AI Mode share only 13.7% of citations at all, low cross-platform consistency is the expected result rather than a surprise, and the finding describes surface divergence at least as much as market concentration.

The distinction matters operationally. If the finding were about concentration, the strategic response would be to accept that a handful of brands own the category. If it is about divergence — and the overlap figures suggest it largely is — the response is that visibility has to be earned surface by surface, because being cited well on one says remarkably little about the others.

What to ask before accepting any share figure

Six questions separate a measurement from a number: what the prompt set is and whether it changed between the dates being compared; how many times each prompt was run and how the runs were aggregated; whether the numerator counts mentions, citations, or a blend, and at what weight; what the denominator includes, specifically whether prompts producing no brand at all are in it; which surfaces were queried; and whether the consumer product was queried or an API standing in for it.

Two absences should be stated alongside any answer. No vendor was found publishing its formula, prompt set or runs per prompt in a form that would allow independent replication, which means no commercial AI share-of-voice figure on the market as of 29 August 2026 could be reproduced from published information. That is recorded as an absence rather than an accusation, but it is what unauditable means in practice.

And no study was found linking any AI share-of-voice score to traffic or revenue outcomes with a stated method. The claim that share of voice drives commercial results is made in vendor marketing and is not supported by any primary work located here. What would settle it is a study correlating a disclosed share metric against measured AI referral traffic or conversions across many brands, with the prompt set published. Until that exists, the metric is best read as an internal diagnostic on a fixed prompt set, and never as a market position.

Frequently asked questions