Blog · Measurement

Share of voice in AI answers, and its limits

Published 29 April 2026 · Updated 12 October 2026 · 9 min read

Share of voice in AI answers, and its limits
Photograph: Umar ben / unsplash

The short answer: Share of voice in AI answers is the fraction of brand slots you occupy across a fixed set of prompts, run on fixed engines, against a fixed competitor list. It is worth tracking because it is comparative, and it misleads in four known ways: the denominator is something you chose, the sample is far smaller than it looks, a mention is not a recommendation, and the number does not convert to traffic.

Every part of that definition carries a decision somebody made. Change the prompt list and the number moves. Add a competitor and the number falls without anything happening in the world. That is not a reason to abandon the metric. It is a reason to write the decisions down next to the score.

What the number is actually counting

On the brand questions in our audit prompt set, the assistants we run do not return a ranked list of ten results. They write a paragraph and name a handful of companies inside it. Your share of voice is your count of those names divided by every brand name across the answers you collected, times a hundred. Mention share and citation share are different quantities and people conflate them constantly, which is why citation share vs mention share is worth reading before you present either one.

Pew Research Center's analysis of 68,879 Google searches by 900 US adults, collected in March 2025 and published on 22 July 2025, found that 18 per cent of searches produced an AI summary, and 88 per cent of those summaries cited three or more sources. That is the shape of a Google AI Overview in that sample, on Google Search, in that month. We have not seen an equivalent public dataset for the other assistants, so treat three or four cited slots as a Google figure rather than a property of answer engines in general. Where an answer does carry only a handful of named vendors, the ceiling on anyone's share is low, the floor is zero, and most of the distribution sits in a band narrow enough that small absolute changes look dramatic in percentage terms.

The denominator is a choice, and it drifts

Here is the first failure mode, and it is the one that quietly ruins quarterly reporting. Your share of voice has a denominator made of two things you control: which prompts you run, and which names you count as brands.

Prompt drift is the obvious half. Someone adds eight prompts about a new feature, your brand wins most of them because you launched it, and the headline score rises four points. Nothing changed about how engines see you on the forty prompts you had before. The honest fix is a frozen core set, versioned, with additions tracked in a separate basket and reported separately for at least one full cycle.

Entity drift is the subtle half. Engines name open source projects, marketplaces, review sites and categories alongside vendors. Whether a community project counts as a competitor in your denominator is a judgement call, and whoever makes it can move the score by several points without touching a page on your site. Write the inclusion rule down, and apply it to the raw answers with a script rather than by eye, so it is auditable later.

The sample is smaller than the dashboard suggests

A prompt set of forty five prompts across five engines looks like 225 observations. It is nearer forty five, because the prompt is the independent unit and the five engines are not independent draws: they answer the same question and in several cases retrieve from overlapping web indexes. We cannot see those indexes, so we cannot put a figure on the overlap. The safe reading is that your effective sample is closer to the prompt count than to the prompt count times the engine count, and the error bars should be drawn accordingly.

Run the arithmetic yourself. If your measured share is 20 per cent on 45 prompts, the standard error on a simple proportion is the square root of 0.2 times 0.8 divided by 45, which is about 6 percentage points. Two standard errors either side puts the honest interval somewhere around 8 to 32 per cent. A move from 20 to 24 in a month is inside the noise. Teams present that move as progress every day of the week.

There is a second source of variance on top of sampling. Google's own documentation on AI features describes a query fan-out technique, where one question becomes several related searches across subtopics and data sources, and states plainly that AI Mode and AI Overviews may use different models, so the responses and links they show will vary. One prompt is not one retrieval. Repeat the same prompt on consecutive days and you are sampling a distribution, not reading a value. The practical consequence is in how to measure AI visibility without fooling yourself: three runs per prompt is a minimum, and the spread across runs belongs in the report next to the mean.

A check you can run this week

Split-half reliability. Divide your prompt set into two random halves and compute share of voice separately for each half on the same day. Then compute it for the whole set on two different days. If the gap between your two halves is larger than the gap between your two days, the metric is being driven by which prompts you picked rather than by anything the engines did. It takes about ten minutes in a spreadsheet, and in audits we have run it has changed what a dashboard was believed to be showing. We have not published those runs, so read that as our experience rather than as evidence you can check.

A mention is not a recommendation

Share of voice counts presence. It does not read the sentence. A brand named as the expensive option, the one with the steep learning curve, or the one a buyer should avoid for small teams, scores exactly the same as a brand named as the default choice. Sentiment and framing need a separate column, graded on the actual sentence, because the two can move in opposite directions at once.

The research literature on this is older than most dashboards. The KDD 2024 paper *GEO: Generative Engine Optimization* (arXiv 2311.09735) reports visibility gains of up to 40 per cent, measured on the authors' own GEO-bench benchmark rather than on live consumer assistants, so read it as evidence that the measure matters and not as a gain you should expect. What is portable is the definition: the paper scores visibility with position-aware and word-count-aware measures rather than a binary present-or-absent flag, because being named last in a trailing clause is not the same as being the subject of the first sentence. Neither that paper nor anything we have measured shows what prominence does to a reader's behaviour, so the argument for keeping it is that a yes or no throws information away, not that it predicts a click.

Engines differ here too. In our own runs a brand has been prominent in one assistant and absent in another on the same question in the same week, which is why how Gemini answers brand questions is a separate post rather than a paragraph. Reporting a blended cross-engine score hides exactly the variance you need to act on.

The number does not convert to traffic

This is the limit people find most uncomfortable. The same Pew analysis found users clicked a traditional result in 8 per cent of visits where an AI summary appeared, against 15 per cent where none did, and clicked a source cited inside the summary in about 1 per cent of visits. That is Google Search, desktop and mobile browsing, one month, one panel. It is the best public evidence we know of, and it does not cover the standalone assistants. What it does establish is that on Google, more slots in an answer is not a mechanism for proportionally more sessions.

So treat share of voice as a leading indicator of consideration, not as a demand metric. It has no impression volume behind it, unlike paid share of voice, so a 30 per cent share across prompts nobody asks is worth less than a 10 per cent share on the three questions that precede a purchase. Weight your prompt set by how often the question gets asked and by how close it sits to a decision, and say in the report how you weighted it.

Expect lag as well. In the audits we have run, access fixes have landed within a crawl cycle, identity work has taken longer, and third party corroboration has been slowest of all, which how long AEO takes to show up in answers covers in detail. A share of voice reading taken two weeks after a change is usually measuring the engines' refresh schedule rather than your work.

Honest limits

This post does not establish how answers are shaped outside Google. The click and citation figures here come from one Pew panel of 900 US adults browsing Google Search in March 2025, and we have not found comparable public data for the standalone assistants, so nothing here should be read as a measurement of them. The claims drawn from our own work, the split-half test, the engine-to-engine disagreement and the order in which fixes land, come from client audits we have not published, and are experience rather than evidence you can check. We have no defensible weighting for position inside an answer, because we have no click data from most of these surfaces to calibrate against, so any position-weighted score we publish is a convention. And we cannot say how much of the run-to-run variance comes from retrieval versus generation, which matters because only one of those two responds to anything you do to your site.

If you want to see how we present this alongside the evidence rather than as a single figure, the sample audit report shows the format, including the spread across runs and the prompt list the score was computed on.

FAQ

Is share of voice in AI answers worth tracking at all? Yes, as a comparative trend on a frozen prompt set, reported with its interval. It is a poor headline number and a decent directional one. The value is in the gap between you and the named alternatives, not in the absolute figure.

How many prompts do I need before the number means anything? More than most teams run. At 45 prompts a measured share near 20 per cent carries an interval of roughly plus or minus 12 points at two standard errors. Doubling the prompt count narrows that by about a third, which is often cheaper than arguing about a four point move.

Should I report one blended score across engines? Report per engine and show the blend underneath. In our runs engines disagree on the same question, and a blend can stay flat while one assistant drops you entirely, which is the event you most need to catch.

Does a higher share of voice mean more traffic? Not reliably. In Pew Research Center's March 2025 analysis of Google searches, click rates on pages carrying an AI summary were roughly half those without, and clicks on the summary's own cited sources were rare. That is Google only. Treat share of voice as consideration, and attribute revenue through other means.

Sources

All three were opened and checked against the sentences that cite them on 12 October 2026.

Related reading

How to measure AI visibility without fooling yourselfCitation share vs mention shareHow long AEO takes to show up in answersHow Gemini answers brand questionsWhat AI visibility tools can and cannot seeVariance, sample size and the honest error bar