AI Share of Voice: Three Legitimate Denominators, and Why Ours Reads 0.5% or Six in 20,143

Share of voice in AI answers is undefined until the report says what it divided by. The run behind this page is 1,260 readings dated 16 October 2026: 630 buying questions, each put once to ChatGPT and once to Gemini, split evenly between the United States, the United Kingdom and India. Those readings carried 20,143 citations over 4,947 distinct domains, and a page of ours was among the sources in 6 of them. Divided by readings that is 6 of 1,260, which we publish as 0.5%. Divided by citations it is 6 of 20,143. Divided by distinct domains it is 1 domain of 4,947. All three are honest, none of them is a correction of another, and the first is roughly sixteen times larger than the second because the run averaged 15.99 citations per reading. Any share of voice figure you are shown without its denominator cannot be compared to any other share of voice figure, including ours. We sell this measurement, which is why the denominator sits in the headline here.

The three denominators, and the same six citations under each

Every row below describes the identical event: six readings in which an answer used a page on our domain as a source. Only the thing being divided by changes.

Denominator	Our figure	What a share computed this way answers
Readings, 1,260	6 of 1,260, published as 0.5%	How often an answer used you at all. The buyer facing number
Citations, 20,143	6 of 20,143	How much of the total source load you carried. The publisher facing number
Distinct domains, 4,947	1 of 4,947	Whether you are in the pool of sources at all. A presence check, not a share

Divide 20,143 citations by 1,260 readings and each answer leaned on 15.99 sources on average. That multiplier is the whole distance between the first two rows. A supplier reporting on readings and one reporting on citations hand the same business two numbers about sixteen apart, both arithmetically correct, and neither usually says which it is. The separate question of what counts as one citation, where identical data yields 20,143 or 36,576 depending on the rule, is worked through in what counts as one citation, and why the same 1,260 readings produce 20,143 or 36,576. Pick the rule first, then pick the denominator, then report both.

The ceiling on any single share, measured

A share of voice target is only sensible if you know what the top of the field looks like. In this run the most used source was developers.google.com, which appeared in 446 of the 1,260 readings, and official documentation took six of the ten most used domains. Below that layer the field is wide: 4,947 distinct domains carried 20,143 citations, an average of 4.07 citations per domain, so the long tail is where almost every business sits.

Two consequences follow. A page written to outrank a platform on that platform's own mechanics is spending against the company that built the engine. And because 82.0% of those 20,143 citations went to a vendor or agency own site, the slots not held by documentation are held by publishers rather than by directories, which took 0.7%.

The two engines do not share a distribution, so they do not share a score

Blending two engines into one share of voice figure produces a number that describes neither. Our domain appeared in 1 of 210 United States ChatGPT readings, 3 of 210 British ChatGPT readings and 2 of 210 Indian ChatGPT readings, and in 0 of 210 on Gemini in each of the three markets. Add the 291 Gemini readings from the run of 15 October 2026 and Gemini gave 921 readings with our domain in none of them. Averaging a zero column into a non zero column is how a report ends up with a share nobody can act on. Ask for the fraction per engine, per market, with both numbers printed.

Where E-E-A-T sits in this, according to the documentation

Share of voice conversations reach for E-E-A-T quickly, so it is worth reading what Google publishes about it rather than what the field repeats. On 16 October 2026 we read Google's page on creating helpful, reliable, people first content at developers.google.com/search/docs/fundamentals/creating-helpful-content, which carried its own last updated date of 5 October 2026. Three sentences on it decide how to use the idea.

• On the four aspects of experience, expertise, authoritativeness and trustworthiness, the page states: "Of these aspects, trust is most important. The others contribute to trust, but content doesn't necessarily have to demonstrate all of them."
• On status, it states: "While E-E-A-T itself isn't a specific ranking factor, using a mix of factors that can identify content with good E-E-A-T is useful."
• On the quality raters whose guidelines define it, it states: "Search raters have no control over how pages rank. Rater data is not used directly in our ranking algorithms."

So E-E-A-T is documented as a framework for human raters giving feedback on Google Search, with trust as the part carrying the weight. That is a statement about Google Search and not about what an assistant does. The engine side evidence is narrower and more useful. In the 102 build spec conversations of 15 October 2026 the two most checkable demands were a sample size, from 80 of them, and a published prompt set, from 78. A stranger can verify both in a minute.

Computing a share you can defend, in six lines

• Freeze a question set that never names your business, and save the wording in a file.
• Record, per answer, which competitors were named and every domain cited. Two columns, because being named and being used as a source are different events.
• Report each engine in its own row with its own numerator and denominator.
• Pick your denominator and write it at the top of the report.
• Put the date on it, and the interval before the next reading, decided in advance.
• Keep the set fixed between rounds. The set is the denominator, so changing it moves the number for reasons unrelated to your visibility.

The same arithmetic has a long history in search, where share of voice was a count of how many of the first ten results each rival held, and the discipline of getting one trustworthy position reading is set out in how to check your ranking position and get a number you can trust. The unit changed and the need to state the denominator did not.

What this run does not contain

No share by industry: no trade was named in the 630 questions, so a software share would have to be invented to appear here. No city level share either, because the city figures in this programme come from a separate scan of ten named cities, 120 readings in total. And the engine count stops at two. ChatGPT and Gemini were read; Perplexity, Claude and Google AI Overviews were not, and our product covers five surfaces while this dataset covers two. Every control behind these counts, down to how a market was locked into each prompt and then verified on the answer, is set out in what AI assistants actually cite across 597 readings of ChatGPT and Gemini in three markets.

Questions about AI share of voice and its denominator

Which denominator should a report use by default?

Readings. It answers how often an answer used you, which is the question a buyer is really asking. Print the citation total beside it so a reader can see the 15.99 sources per answer that separates the two.

Is a share of voice of 0.5% good or bad?

It is our own figure, 6 readings of 1,260 on 16 October 2026, and it is a baseline rather than a result. A share is only readable against the same question set measured again on a stated date.

Can share of voice be compared between two suppliers?

Only when both publish the counting rule, the denominator, the engine names and the question set. Two reports built on different denominators from identical data sit about sixteen apart.

Does E-E-A-T raise a share of voice?

Google documents E-E-A-T as a framework its human quality raters use, and states on the page read on 16 October 2026 that rater data is not used directly in its ranking algorithms. The checkable items the engines themselves asked for, a sample size and a published prompt set, are the ones to put on the page.

Why report zero for an engine at all?

Because 0 of 210 is a measurement and a blank is not. Gemini returned 0 citations of our domain across 921 readings in two runs, and that is the most repeated single result in this programme.

The sentence that makes a share of voice usable

One line decides whether a share of voice number can be carried into a decision, and it is the line naming what sat under the division. Ours says readings, 1,260 of them, on 16 October 2026, across two named engines, with the numerator at 6. AI Knows Us is paid to take that reading across five surfaces and to write the pages that change it, so the demand for a printed denominator applies to our own figure first. Open the last report you were sent and look for the denominator. If it is missing, the headline number is not yet a share of anything.