What Does AI Share of Voice Count, and How Do You Calculate It?

AI share of voice is how often AI answers name your brand compared with your rivals. It works like a guest list. Ask an AI assistant a buying question, and it names a few brands and leaves the rest off.

The tricky part is that two tools can give you different numbers for the same brand on the same day. The reason is simple: they divide by different things. Below, one small example shows how.

Key takeaways

• AI share of voice counts the answers that name your brand, out of all the answers collected.
• Named and linked are two separate counts, and a link alone isn't a mention.
• Divide by total mentions instead of total answers, and the same data gives a different number.
• Always show the base and the 95% range, or the percentage says less than it seems.

What is AI share of voice?

AI share of voice is the share of AI answers that name your brand, next to the share that name each rival you track. Firemark counts the answers that name each brand, out of all the answers it collected, and then ranks the brands.

Old-style share of voice is about media. It's a brand's share of the ad spend in its category. A blog post on the IPA's site describes excess share of voice as buying a bigger share of the ads than your share of the market, in order to grow that market share.

The AI version is simpler and more direct. It asks one thing: when a buyer asks, does the answer say your name? It makes no promise about sales.

Why does that matter? An AI answer works like a short list from a helpful friend. It names a few brands and skips the rest, and a ranking report can't tell you which side of the list you're on.

What does it count?

AI share of voice counts answers, not mentions elsewhere on the web. For every answer, Firemark checks two things: is your brand named, and is your site linked. It keeps the two counts apart, as Firemark's method page explains.

Named means your brand's name is in the words of the answer. Linked means your site is one of the sources listed with it, like the Sources view ChatGPT shows when it searches the web, as OpenAI's help page describes.

That makes four outcomes: named and linked, named only, linked but not named, and absent. A name that appears only in the source list doesn't count as named, because the answer never said it.

If an engine fails to answer, Firemark doesn't count that as your brand being absent. It marks that engine as short for the run.

How do you calculate it?

Divide the answers that name your brand by all the answers collected, then multiply by 100. Do the same for each rival, and do it for each engine too, since each one answers a little differently.

Here's a small example. Picture 100 answers from one engine, and four brands.

Brand A is named in 50 answers, B in 40, C in 30 and D in 10. These numbers are sample data.

One answer can name several brands, so the shares of answers add up to 130%. Think of a dinner party where each of 100 guests names a few favorite restaurants. You end up with 130 names said, even though only 100 people spoke.

The last column shares out those 130 names, so it adds up to 100%. Check the math: 50 out of 100 answers is 50%. But 50 out of 130 names is 38.5%.

Brand A leads in both columns, but the number you'd report is different. That's the whole problem. The demo report shows a share of answers for five engines, using sample data.

Why do tools report different numbers?

Tools report different numbers because they divide by different things. Picture a pizza night.

"How many of the 100 guests mentioned your pizza place?" is one question. "What slice of all the pizza places mentioned is yours?" is another. Both are fair, and they give different numbers.

Published guides and tools use at least three kinds of divisor. Some divide by all the answers, like the first column. Some divide by all the brand mentions, like the second. Others give each mention points for position or tone, then divide by the total points.

When two reports don't match, check the prompts, the engines and the divisor before you compare the percentages.

Firemark divides by the answers collected, for three reasons:

• Steady: with a share of mentions, your number can drop when a rival is named more often, even though your own count hasn't changed.
• Visible: the base is on the page, so "named in 40 of 100 answers" can be checked against the raw answers.
• Simple: a count needs no point scale that someone has to defend.

Because of this, shares can add up to more than 100% across brands. That's fine, not a bug.

The rank follows the counts. A rival named in fewer than 10 answers of a run shows as "fewer than 10" and gets no rank. Firemark tracks 1 to 10 rivals per brand on every plan, listed on its pricing page.

How much can you trust the number?

A share is a sample, much like a poll. If you ask 100 people, you wouldn't bet on the exact figure. You'd want a margin of error.

In the example, 50 of 100 answers is 50%. The 95% range, using the Wilson interval, runs from 40.4% to 59.6%. A rival at 40 of 100 has a range of 30.9% to 49.8%.

Those two ranges overlap, so the lead isn't settled. A one-point gap is within the noise, not a result.

More answers make the range narrower. At a 40% share, the 95% range is 23% to 59% on 25 prompts and 31% to 50% on 100. A bigger set gives a sharper picture, but it takes more prompts to build and read.

Firemark shows a percentage only from 20 prompts up. It reports a movement only after three comparable runs and a change of 5 points or more, and it draws a 95% range for each engine. That range covers sampling only, so it can't cover an engine changing how it answers.

Answers also change from run to run. SparkToro and Gumshoe found that AI tools rarely give the same list of brands twice. Even so, the author called visibility percent across dozens to hundreds of prompts, run more than once, a reasonable measure.

The post rejects the order of brands inside one answer as a measure. SparkToro ran the study with Gumshoe, an AI tracking vendor, and says so in the post.

When does a weighted score fit better?

A weighted score is like judging a talent show by how loud the applause is, not just how many people clapped. A first-place recommendation might earn 1 point and a mention in a list 0.6 points, and a warning might lose points.

It fits when you need one number for a board that says whether mentions help or hurt. But the point scale is a choice, so two tools with different scales can't be compared.

Firemark records which third of the answer holds your first mention, and whether the mention reads positive, neutral or negative. It reports the share as a plain count, though. If you want one weighted figure, a platform built to score that way fits better.

If you want a number you can check against the raw answers, count over a base. Either way, say which one you used.

How do you read the number?

Read it next to your rivals, across runs and per engine. A 50% share means half the answers name you, which says little until you know a rival's share on the same prompts.

Split the count by funnel stage too. A brand named when buyers compare but missing when they research has a different problem from one named early and dropped at the decision. One share hides both.

There's no universal good score. It depends on your prompts, your market and who you compete with. One blended number can also hide gaps. In Firemark's sample demo data, ChatGPT names the example brand in 62% of answers and Gemini in 40%, so the blended 50% hides a 22-point gap.

Connect Firemark's MCP server to Claude or ChatGPT and you can ask which engine names you least. Start with at least 20 prompts, count over a base and keep the range next to the percentage. Firemark has a demo report, and a new account gets one free run of 10 prompts on 4 engines.