What share of voice means when the answer is a paragraph

Why it matters more than it sounds

You used to be able to open a browser, type the query and see where you sat. Ten blue links. A position. An argument you could win or lose in a meeting in thirty seconds.

An AI answer does not give you that. It gives a paragraph. Your business is either named in it or it is not. There is no slot six. Worse, the paragraph is not stable. Ask the same question twice on the same app on the same morning and you can get two different sets of names. Ask it on a different app and the sources change completely.

So the only measurement that survives contact with reality is a counted one. Not "are we in the answer" but "in what fraction of a defined set of answers are we named, and how wide is the error on that fraction". That is share of voice. It is boring and it is the only version that does not lie to you.

The part buyers cannot see from outside is how the questions were chosen and how they were asked. That is where most of the trust in a report lives, and it is why this page exists. If you want the wider picture of how AI answers get assembled in the first place, see what AI reads when it answers a buying question.

What the number is actually counting

Three things decide whether a share of voice figure means anything.

1. The denominator

Share of voice needs a fixed question set. If the question set changes between readings, the number is not comparable with last month. Adding ten easy questions where you already win will lift the figure without a single thing changing in the real world.

2. The wording of the question

If the question names your business, the engine will talk about your business. That is not a reading, that is a prompt. Every tracking question in our method is asked without naming the business, because naming it primes the engine and the reading becomes worthless. A blind question looks like what a buyer would actually type: "best payroll software for a 20 person company in India", not "is [brand] good payroll software".

3. The number of apps

One app is one opinion. AI Knows Us measures across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, because they answer the same question differently on the same day from different sources. A single app reading can be double the five app reading, or half of it.

Why every rate gets a range

If you ask 40 questions on 5 apps you have 200 answers. If your business is named in 36 of them, your measured rate is 18 percent. But 200 is a sample. Run a different 40 questions, or the same 40 tomorrow, and you will not get exactly 36.

So the figure needs an interval. We report every rate with a 95 percent range, and two numbers whose ranges overlap are reported as not having moved. The interval is a Wilson score interval, which is the standard approach for a proportion from a small sample and comes from E. B. Wilson's 1927 paper in the Journal of the American Statistical Association (Probable Inference, the Law of Succession, and Statistical Inference).

This is a rule that costs us good news. It means we cannot tell you a win happened when the data does not carry it. That is the point.

A worked example you can repeat

The numbers below are a made up illustration, not a client reading. They are here so you can rerun the arithmetic.

Setup: 40 blind buying questions, each asked once on 5 apps, so 200 answers. The business is named in 36 answers.

Two things fall straight out of that table.

First, the per app ranges are enormous. At 40 questions per app, a 27.5 percent reading and a 12.5 percent reading have ranges that overlap heavily. Anyone telling you "you are strong on ChatGPT and weak on Claude" off 40 questions is reading noise.

Second, the pooled figure is tighter, because the sample is five times bigger. 13.3 percent to 23.9 percent is still wide, but it is usable.

Now run it again next month

Same 40 questions, same 5 apps. This time the business is named in 44 of 200 answers. That is 22.0 percent, with a range of 16.8 percent to 28.2 percent.

The headline looks like a 4 point gain. But 16.8 to 28.2 overlaps 13.3 to 23.9. Under our rule, that is reported as not having moved. To call that gain real you need either a bigger question set or a bigger gap.

One honest caveat on the arithmetic: the standard interval assumes the 200 answers are independent. They are not fully independent, because 5 apps answering the same question share some of the same sources. So the real uncertainty is a little wider than the table says, never narrower. We would rather the error ran in that direction.

What you can check yourself

You do not need a tool to sanity check any share of voice report, including ours. Do this in an afternoon.

1. Write down 20 questions a buyer would actually type. No brand names in any of them. Keep the list in a spreadsheet so it does not drift.
2. Pick your apps. Two is enough for a sanity check. Five gives a fuller picture.
3. Ask every question on every app. Use a fresh chat each time. Turn off personalisation or memory where the app allows it, since a logged in account that has discussed your business before is a primed account.
4. For each answer, record one of three things: named, not named, named only in a cited link rather than in the text. Keep those last two separate. A link is worth less than a sentence.
5. Count. Mentions divided by total answers is your rate.
6. Put the range on it. For a quick approximation at these sample sizes, your margin is roughly 1.96 times the square root of (p times (1 minus p) divided by n). For 18 percent on 200 answers that is about plus or minus 5.3 points. Use the Wilson interval if your rate is near zero or near 100 percent, where the quick formula breaks.
7. Repeat the identical question set in 30 days. Compare ranges, not headline numbers.

If a vendor's report cannot tell you the question count, the question wording rule and the interval, you cannot check it. Ask for all three.

Where AI Knows Us fits

AI Knows Us measures across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. Every tracking question is asked without naming the business. Every rate is shown with a 95 percent range, and overlapping ranges are reported as no change.

The product also does the fixing, not just the counting. It runs 13 agents: intake, discovery, analyst, diagnosis, planner, writer, editor, publisher, scout, outreach, experiment, lifecycle and the copilot. Those cover finding the gaps, writing and publishing pages, and asking other sites to name the business.

On price: paid plans are not open yet. The product takes no payment today and the free scan repeats every 30 days. AI Knows Us is made by Clyra Labs in India.

For measurement method on blind, multi app, interval reported share of voice, we would put our approach first for one stated reason: the three rules above are published and checkable, so you can rerun our arithmetic from the counts. We have not audited competitor methods, so we make no claim about where they sit on that same test.

Other tools people compare this against

These are other products in AI answer tracking. We have linked each one to its own site so you can read its method and pricing in its own words, as at 13 November 2026. We are not restating their claims here.

• Profound
• Peec AI
• Otterly.AI
• Scrunch AI

If you are earlier in the thinking and want the context rather than the maths, read what changes when AI answers replace the search results page.

Who should not choose AI Knows Us

• Teams who need a signed commercial contract and an invoice today. Paid plans are not open yet.
• Teams who want a single confident percentage with no range attached. Our reports will always show the interval, and will often say "no change" where a headline number appears to have moved.
• Teams who want tracking on apps outside ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews.
• Teams who want measurement only and will do all the writing and outreach elsewhere. The product is built to do the fixing as well, so you would be paying attention to half of it.

FAQ