AI share of voice measures the proportion of LLM-generated answers (ChatGPT, Perplexity, Gemini, AI Overviews) in which your brand is cited, across a given universe of queries. It is GEO's core KPI: it replaces keyword rank tracking, which has become unreadable in a world where the answer is synthesized rather than listed. Concretely, you query a basket of representative requests, you count the answers that mention you, and you divide that total by the number of queries tested. A 30% share of voice means you appear in one answer out of three. This indicator breaks down by platform, because citation mechanics differ: ChatGPT relies heavily on off-site mentions, Perplexity favors fresh sources, AI Overviews stays anchored to the organic top 10. Measuring, segmenting and growing this share of voice is now the core work of any serious generative visibility strategy.
Definition: what AI share of voice is
AI share of voice is the proportion of LLM-generated answers in which your brand is cited, across a defined universe of queries. It is the generative equivalent of advertising share of voice or rank tracking: an indicator of relative presence, measured at the scale of a market rather than a page.
The shift is profound. In classic SEO, you track the rank of a URL on a keyword. The logic is binary and positional: you are 3rd, then 1st, then 2nd. In a ChatGPT or Perplexity answer, there is no longer a list of ten links. There is a synthesis that cites three to six sources, sometimes fewer. Either you are in it, or you are not. Keyword ranking becomes unreadable because the object being observed has changed in nature.
Why this KPI is becoming central
More than half of Google queries now trigger an AI Overview, and ChatGPT exceeds 900 million weekly users. A growing share of search intent produces no click: the answer is consumed on the spot. In this world, the only honest measure of your visibility is not the traffic received but the frequency at which you are cited in the answer. That is exactly what AI share of voice captures, and that is why a GEO agency makes it the heart of its reporting rather than a secondary line.
The calculation formula
In its simplest form, AI share of voice is calculated as: number of answers where the brand is cited, divided by the total number of queries tested. On a basket of 100 queries where you are cited 28 times, your raw share of voice is 28%.
This raw version is enough for a first diagnosis, but it flattens two useful pieces of information: not all citations are equal, and not all queries carry the same commercial weight.
Weight by citation quality
Being cited as the first source of an answer does not carry the same value as being mentioned in passing in the last paragraph. You therefore assign a coefficient to each citation according to its position and nature: explicit citation with a link, brand mention without a link, mere presence in a list. This weighting turns a binary count into a measure of influence.
Weight by query volume
A high-volume query carries more weight than a niche request. By multiplying each presence by the associated volume, you obtain a weighted share of voice that reflects commercial reality rather than a flat arithmetic average.
| Criterion | Raw share of voice | Weighted share of voice |
|---|---|---|
| Calculation | Citations / queries | Citations x position x volume |
| Reading | Fast, binary | Granular, business-oriented |
| Use | Initial diagnosis | Steering and arbitration |
| Limit | Ignores citation value | Requires reliable volume data |
How to measure it in practice
Measuring your AI share of voice rests on three steps: build a representative query basket, query each platform repeatedly, then count and weight the citations. The rigor of the measurement depends entirely on the quality of these three steps.
Select 50 to 200 queries covering your key intents: informational, comparative, transactional. The basket must reflect your prospects' real journey, not just your favorite keywords. It is the foundation of the whole measurement.
LLMs are non-deterministic: the same query gives different answers from one call to the next. Query each request at least three to five times per platform, then average. A single measurement is a misleading snapshot.
Automatically spot mentions of your brand and your URLs in each answer, handling spelling variants. A brand cited without a link counts, but differently from a source citation.
Calculate share of voice per platform, because mechanics differ, before consolidating. A global score that masks a collapse on Perplexity is useless.
Share of voice only makes sense in relative terms. Measure the same queries for your three to five direct competitors to place your position on the market.
This last step deserves dedicated work: that is the whole point of a competitive benchmark on LLMs, which turns an isolated share of voice into an actionable market ranking. And counting citations benefits from automation: to miss no appearance, rely on a real method of tracking brand mentions in ChatGPT rather than on intermittent manual readings.
According to the Ahrefs analysis of 200,000 domains (December 2025), off-site brand mentions correlate more strongly with AI citations (YouTube 0.737, Reddit, Wikipedia) than Domain Rating (0.266). Share of voice measurement must therefore integrate external signals, not just your own site.
Optimization levers by platform
Growing your AI share of voice requires acting differently per platform, because each cites according to its own logic. Only one cross-cutting lever works everywhere: the technical citability of your pages.
The common foundation: making the page citable
LLMs do not execute JavaScript at collection time. A client-rendered page is invisible to them. SSR or static HTML is therefore non-negotiable. Beyond rendering, structure each answer into self-sufficient passages: the optimal citable passage is between 134 and 167 words, answers a question directly and is understandable out of context. FAQPage schema remains a strong signal for AI Overviews.
ChatGPT: bet on off-site mentions
ChatGPT relies heavily on external signals. Your presence on Wikipedia, Reddit, YouTube and sector directories weighs more than your Domain Rating. The lever here is a distributed brand-mention strategy, not an isolated on-page optimization. To target precisely the queries where you must appear, start from a diagnosis of your brand visibility in ChatGPT.
Perplexity: freshness and the primary source
Perplexity favors recent content and sources that hold authority on a precise subject. Publishing original research, clearly dating your content and updating it regularly increases your eligibility.
AI Overviews: stay anchored to the top 10
92% of AI Overviews citations come from the organic top 10, but 47% come from positions 5 to 10. In other words, classic SEO remains the entry ticket: you don't need to be first, but you need to be on the first page. And only 11% of domains are cited by both ChatGPT and AI Overviews, which confirms that you must work each platform separately.
Want a quantified starting point? The AI Visibility Score freely assesses your current presence in generative answers and identifies your first levers, platform by platform.
Share of voice vs SEO rank tracking
AI share of voice does not eliminate rank tracking: it doubles it on the terrain where ranking no longer says anything. On a query that triggers a generative answer, your position in the blue links only explains part of your real presence; what the model cites in its synthesis escapes the classic rank tracker.
The two indicators answer different questions. Rank tracking tells you where a URL sits in an ordered list. AI share of voice tells you how often your brand is selected in an answer that no longer looks like a list. As long as a significant share of queries keep active blue links, rank tracking remains useful for organic traffic. But on intents now absorbed by AI Overviews and assistants, share of voice is what steers.
The mistake would be to pick a side. Best practice is to make the two coexist: positions on still-"blue" queries, share of voice on generative queries, with an overlap zone you watch closely because that is what tips over first.
Common mistakes and measurement traps
Most AI share-of-voice measurements are distorted by the same mistakes: a biased query basket, a single unaveraged measurement, and the absence of competitive comparison.
The complacent basket
The most common mistake: testing only the queries where you already know you appear. The basket must include queries where you are absent, otherwise your share of voice is artificially inflated and steers nothing. A good basket is a little painful to read.
The single measurement
Querying a request once and drawing a conclusion from it is the classic statistical mistake. LLMs vary. Without repetition or averaging, you measure noise, not a signal. This is also why reading too frequently on a small basket is counterproductive: a monthly pace on a solid basket beats an unstable daily tracking.
Confusing citation and link
A brand mention without a link counts in share of voice: it proves the model knows your brand. Treating it as nonexistent underestimates your real presence and makes you miss a major off-site lever.
Forgetting the competitive context
A 30% share of voice is excellent if the leader peaks at 35%, mediocre if it is at 70%. The absolute value says nothing without comparison. This is the final trap, and the most frequent.
Tools and measurement frequency
Beyond a small exploratory basket, measuring AI share of voice by hand does not hold: you must query four platforms, repeat each request several times, handle brand variants and consolidate. That is a volume that demands dedicated tooling.
The GEO tools market exploded in 2025-2026, with mention-tracking, query-basket and citability-scoring solutions. They do not all measure the same thing or with the same statistical rigor: some settle for a single call per query, which reintroduces precisely the single-measurement bias. Before committing, compare the approaches in this overview of 2026 GEO tools, checking three points: call repetition, platform coverage and brand-variant handling.
On frequency, the rule is simple: basket stability beats reading frequency. A monthly reading on a wide, well-built basket produces a usable steering signal. Daily tracking on ten queries produces only noise. Reserve close readings for moments when you test a precise change (new content, off-site mention campaign) and want to measure its effect.
Request a free GEO audit: we measure your real AI share of voice, platform by platform, against your competitors.
Questions fréquentes
What is AI share of voice?+
AI share of voice is the proportion of LLM-generated answers (ChatGPT, Perplexity, Gemini, AI Overviews) in which your brand is cited or mentioned, across a given set of queries. It replaces keyword rank tracking in a context where the answer is synthesized. A 25% share of voice means you appear in one answer out of four.
How do you calculate AI share of voice?+
You divide the number of answers where the brand is cited by the total number of queries tested, across a representative query basket. You can refine it by weighting each citation by its position in the answer and by the associated search volume. The calculation is done per platform, then consolidated.
Does AI share of voice replace SEO rank tracking?+
It complements it rather than fully replacing it. Rank tracking remains relevant for classic organic traffic. But on queries that trigger a generative answer, your position in the blue links only explains part of your presence: AI share of voice captures what ranking no longer sees.
How often should you measure AI share of voice?+
A monthly reading is a good pace for most brands. LLM answers vary from one call to the next, so you must query each request several times and average. Measuring too frequently on a small basket mostly produces statistical noise.
Do you need a dedicated tool or can you measure manually?+
You can start manually on a small basket to understand the mechanics, but serious tracking requires a tool that queries platforms repeatedly, detects brand variants and consolidates scores. Manual measurement does not scale to a basket of 100 to 200 queries tested every month across four platforms.
