Distrify

Brand Share of Voice in AI Answers: What It Measures

August 9, 2026·14 min read
A founder around thirty in a knit sweater compares two columns of handwritten notes in a notebook next to her laptop, in a wooden studio corner lit by a window

Brand Share of Voice in AI Answers: What It Measures

Brand share of voice in AI answers is the proportion of replies, across a fixed list of questions, where an assistant names your brand rather than someone else. Nobody publishes this number for you. It is a percentage you build yourself, out of your questions, your competitor list and your sampling schedule.

That is what makes it both useful and fragile. Useful, because a relative number says something a raw citation count cannot: you are not simply present or absent, you are named in one answer out of five while someone else is named in three. Fragile, because every choice that feeds the calculation can move the result by tens of points while nothing has changed in the real world. Here is how the number is built, what it is worth, and the five places where it misleads you.

TL;DR

  • Brand share of voice in AI answers does not exist outside the protocol that produced it: your questions, your competitor list, your number of samples.
  • You can build it by hand in an hour: a frozen question list, several runs per question, one spreadsheet row per observation.
  • It answers something a citation count cannot: are you losing ground, or is the whole topic being served less this month.
  • Its five known weaknesses: answer variability, sample size, question phrasing, the competitive set you pick yourself, and the gap between being named and being recommended.
  • Verdict: a decent trend instrument over months, a poor level instrument at any single moment. A few points of difference mean nothing.

Table of contents

What brand share of voice in AI answers actually means

The idea comes from advertising and press coverage: out of all the talk in a market, how much of it is yours. Moved to assistants, the definition looks simple. You ask a hundred buying questions to ChatGPT, Perplexity, Gemini or Copilot. Your brand is named in eighteen of the answers. Your share of voice is eighteen percent.

Three details change everything in that calculation.

First, the denominator is your questions, not the market's. Nobody will hand you the real list of what buyers ask an assistant, because those conversations are private. You are working from a list you wrote, and that list is a hypothesis about your market.

Second, an answer is not a ranking. An assistant that names four brands does not order them the way a results page does. Treating a first-sentence mention as equal to a closing aside, or not, is a decision you make and should write down.

Third, the same prompt does not return the same answer twice. That is what makes a one-off reading unusable, and it is the main weakness, covered below.

If you have never recorded anything yet, start with the base routine in AI visibility tracking. Share of voice is a calculation laid on top of those readings. It does not replace them.

Why the number appeals, and what it actually promises

Counting citations has a known flaw: the count does not read. Knowing that your brand appeared seven times this month tells you nothing until you know whether seven is a lot. Seven citations out of ten questions asked is excellent. Seven out of two hundred is absence.

Share of voice fixes that by giving you a denominator, and fixes a second problem by giving you a reference point. A drop in your citations can mean two opposite things: either you are losing ground, or the whole topic is being served less this month. Without the other brands in the calculation, you cannot tell those apart. With them, if you fall while everyone falls, most likely nothing happened at your end.

That is the real promise, and it is more modest than the one usually attached to it. Share of voice is not a thermometer for your reputation. It is a way to know whether a change you observed came from you or from the channel.

How to build the number, step by step

Five steps, no tooling, one spreadsheet.

  1. Freeze a list of twenty to thirty questions. Buyer questions, written the way a person would ask them, not keywords. They stay unchanged for at least three months, otherwise you are no longer measuring the same thing. If you are stuck, take the questions your support inbox actually receives.
  2. Freeze a competitor list. Five to eight names, chosen because buyers mention them, not because they annoy you. Keep a spare column for unexpected names that keep coming back: they are usually the instructive part.
  3. Run every question at least five times, on each assistant you follow, hours apart. This is the only protection against answer variability, and it is what turns an anecdote into a measurement.
  4. Record one row per reading: date, assistant, question, brands named, where your brand sits in the answer, and whether the mention came with a link. Nothing else.
  5. Compute two numbers, not one: your share, and the share of each competitor you track. Yours alone is useless.

Twenty five questions, five runs, two assistants comes to two hundred and fifty readings. Expect ninety minutes the first time and under an hour afterwards. The real cost is not the reading, it is the regularity: a measurement taken once produces no information at all.

If that regularity is exactly what you do not have, it is the problem distrify is built for: one campaign, written, published and reshaped across channels without you keeping the rhythm by hand. The product is not open yet, and the engine behind it already publishes daily for live sites.

Join the waitlist

What the measure says and what it does not

What share of voice measures What it does not measure
How often your brand appears across your questions How often it appears across the questions you did not write
How that frequency moves month to month How many people ask those questions at all
The gap between you and the brands you listed Your standing against brands you did not list
Being named in an answer Being recommended, preferred or chosen
A market wide move, if you track the others too The revenue any citation produced

The right hand column is the honest part of the subject. Most of the disappointment around this metric comes from people reading a percentage as market share. It is not market share. It is a share of your own questions.

The five places where the measure misleads you

1. Answer variability. An assistant does not return the same answer twice, even minutes apart. On a single run, your share of voice is a draw, not a measurement. This is the worst weakness because it produces the most confident conclusions: one check that happens to name you creates certainty without information.

2. Sample size. Twenty questions in one pass gives twenty observations. On twenty observations, two extra citations take you from ten to twenty percent. You would call that a doubling. It is noise.

3. Phrasing. Ask for the best tool to do X, then for an alternative to a well known player, then which provider a beginner should pick: three different answer universes. Your share of voice is inseparable from the style of your questions. Two companies in the same market with two different lists get two numbers that cannot be compared.

4. The competitive set you choose. This is the most comfortable bias, so the most dangerous one. Drop the two most cited brands from the list and your share rises mechanically. Nobody will stop you. Write the list once, date it, and only change it on the record.

5. Being named is not being recommended. An answer can name you as the expensive option, as a counter example, or in a list where you are the one that does not fit the need. Counting cannot tell the difference. Read ten full answers every month, or your percentage will climb while your reputation slides.

That last point connects to something covered in getting mentioned in ChatGPT: what you can work on is the material that makes a brand citable, never the citation itself, which no method can guarantee.

Comparing yourself to competitors without cheating

Comparison is the entire point of the metric, and it is also where protocols quietly loosen. Three rules are enough.

Same protocol for everyone. Competitors are measured on your questions, your runs, your assistants, on the same day. A number collected for you midweek and for them at the weekend is not a comparison.

Same status for every citation. If a closing mention counts as a citation for you, it counts as one for them. The urge to grade your own appearances carefully and theirs roughly is real.

A market category, not a list of enemies. The brands that show up in answers are not always the ones you track commercially. This is the most valuable intelligence in the exercise: an assistant is telling you which neighbourhood it considers yours. If an unexpected name appears in a third of the answers, it belongs to your market on this channel whether you like it or not.

One note for catalogues: on buying questions, assistants often cite comparison pages, buying guides and marketplaces before brand sites. A low share can be a format problem rather than a reputation problem, which is the angle of AI search optimization for ecommerce.

How large a gap has to be before it means anything

Here is the reading rule we apply, deliberately conservative.

  • Under one hundred readings in a month: nothing is reportable, even internally. You record, you do not interpret.
  • Between one hundred and three hundred readings: read an order of magnitude, such as we are named in fewer than one answer in five. Not a decimal.
  • A gap under five points between two months: treat it as zero, unless the same direction repeats three months running.
  • A gap over fifteen points: go and read the answers. A move that size almost always has a findable cause, often one question whose answer changed in nature.
  • A market wide move: if competitors move the same way by the same amount, that is not your performance, that is the channel.

Three months is the first horizon where the series becomes readable, and comparing two consecutive months is nearly always misleading. None of this is specific to assistants: it mirrors classic search, with the same temptation to conclude early, described in the generative engine optimization guide.

Reporting the number to people who did not read this page

A percentage travels faster than its caveats. Once eighteen percent is written in a deck, it becomes a fact, and six weeks later somebody asks why it fell to fourteen when nothing was done wrong.

Three habits prevent that. Report the number with its protocol attached in the same sentence, questions and runs included, so the figure never travels alone. Report your competitors' numbers on the same slide, because a shared decline is the single most useful thing the metric produces. And give a range rather than a point, since the underlying answers are variable by nature.

If someone asks for weekly reporting, the honest answer is that the series will not carry information at that pace. Producing a number every week that nobody can act on is a way of looking busy, not a measurement programme, and the same reasoning applies to which parts of a workflow are worth automating at all, covered in SEO automation.

When it is worth the hour, and when it is not

It is worth it when you already have some presence on the topic and want to know whether your content work moves anything. It is worth it when you have to choose between two topics and want the one with the least crowded neighbourhood. And it is worth it to prevent a bad decision: seeing the whole market drop stops you from dismantling something that was working.

It is not worth it in three cases. If you are starting from zero citations, your share is zero and will stay there for months: the number cannot measure progress that is happening upstream, in the material you publish. If your market is tiny, the number of readings needed to escape noise is more than you can sustain. And if you lack the regularity, one reading per quarter gives a series too short to separate a trend from an accident.

In those three cases the effort belongs on production, not on measurement. That is the unglamorous version of the advice, and it is rarely given, because measuring sells better than writing. The production side is what content distribution strategy and repurposing one post into social formats are about, and the rest of the writing lives in the blog.

FAQ

Is brand share of voice in AI answers an official metric anywhere? No. No assistant publishes how often it names a brand. Every number you will see, including your own, comes from a sampling protocol. The question to ask in front of any percentage is always the same: which questions, how many runs, which competitors.

How many questions make the number credible? Twenty to thirty questions run five times per assistant gives a usable base, roughly two hundred to three hundred readings a month. Below one hundred readings, a month to month gap is noise and should not trigger a decision.

Can I measure this without paid tooling? Yes. A spreadsheet, a frozen question list and an hour a month are enough. Tools save data entry time and keep the protocol stable. They do not create any data you could not collect yourself.

Does a high share of voice guarantee sales? No, and nobody can promise it does. Being named is neither a recommendation nor a visit: between the citation and the purchase there is still a click, a page and a decision. Track sessions and conversions alongside, without mechanically attributing one to the other.

Should I measure on every assistant? Start with two, the ones your customers mention. Adding an assistant multiplies reading time without changing your decisions in the early months, and cross assistant comparison rarely matters to a company that is starting out.

Does publishing more articles raise share of voice? Not mechanically, and writing for the sake of writing does nothing: the pages that get cited are the ones that answer a question precisely. The question of mass produced content and its quality is covered in AI content and Google penalties.

Conclusion

Brand share of voice in AI answers is a good trend instrument and a poor level instrument. It deserves an hour a month, a written protocol that does not move, and cautious reading: no decision under a hundred readings, no conclusion under five points, no comparison across two different protocols.

What actually moves the number is not the measurement, it is what you publish between two measurements, and above all publishing it regularly, everywhere your buyers look. That is the work distrify automates: one campaign, written, published and reshaped across SEO, social and AI answers, aligned on the same keywords. The product is not open yet, and a seven day free trial is planned at launch.

Join the waitlist