AI share of voice: how to measure it with position weighting

In brief

AI share of voice is your share of all brand mentions in AI answers to a fixed set of questions, compared with your competitors. The fairest version weights each mention by its position, so being named first counts more than being named last. A common method gives each mention a weight of one divided by its rank.

Key points

AI share of voice compares your mentions in AI answers with those of competitors.
Simple counts treat a first mention and a last mention as equal, which flatters brands named late.
1/rank weighting gives 1 point for first, 0.5 for second, about 0.33 for third, and so on.
Each brand's share is its weighted total divided by the weighted total for all brands.
Share of voice should be split by engine and by market, because each can tell a different story.
It is a sample, so read it as a trend over several periods.

AI share of voice is the share of brand mentions you win in AI answers, compared with your competitors. You run a fixed set of buyer questions through AI engines, list every brand each answer names, and work out what share belongs to you.

The idea comes from advertising, where share of voice meant your share of all the ad space in a market. In AI search, the "space" is the answer itself. It is usually short, often names only a handful of brands, and the order matters.

Because a mention near the top of an answer is worth more than one at the bottom. When ChatGPT lists five firms, a buyer skimming the answer is more likely to act on the first name than the fifth. Many answers also put the strongest recommendation first.

A simple count ignores this. It gives the same credit to a brand named first with a clear reason as to one tacked on at the end. Over a whole prompt set, that can make a brand that is always named last look as strong as the one that leads.

Position weighting is not a new idea. The original GEO research paper measured visibility with a position-adjusted word count, which gives more weight to sources cited earlier in an answer.

How does 1/rank weighting work?

Each mention gets a weight of one divided by its position in the answer. The first brand named scores 1. The second scores 0.5. The third scores about 0.33, the fourth 0.25, and so on.

The steps are simple.

For each answer, list the brands in the order they first appear.
Give each brand a weight of 1 divided by its position.
Add up each brand's weights across all answers.
Add up the weights for all brands together.
Divide each brand's total by the grand total. That is its share of voice.

The weights fall quickly at first and then level off. So the gap between first and second is large, while the gap between sixth and seventh is small. That matches how people read a list.

A worked example with made-up brands

Here is a small example. Four made-up window companies, Alderpane, Brightsash, Clearholm and Duneview, compete in one town. We ask an AI engine five questions about replacement windows and note the order each brand is named.

Answer	1st	2nd	3rd	4th
Answer 1	Brightsash	Alderpane	Clearholm
Answer 2	Alderpane	Brightsash
Answer 3	Clearholm	Brightsash	Duneview	Alderpane
Answer 4	Brightsash	Duneview
Answer 5	Clearholm	Alderpane	Brightsash

Now add up the weights. Alderpane scores 0.5 in answer 1, 1 in answer 2, 0.25 in answer 3 and 0.5 in answer 5, a total of 2.25. Doing the same for the others gives the table below. The grand total for all four brands is 8.75.

Brand	Mentions	Simple share	Weighted total	Weighted share
Brightsash	5	35.7%	3.33	38.1%
Clearholm	3	21.4%	2.33	26.7%
Alderpane	4	28.6%	2.25	25.7%
Duneview	2	14.3%	0.83	9.5%

The weighting changes the story. On a simple count, Alderpane is ahead of Clearholm, with four mentions against three. Once position is counted, Clearholm moves ahead, because it was named first twice. Alderpane is named often but rarely first. That points to a clear goal for Alderpane, which is to move up the answer rather than simply appear in more of them.

In real tracking you would use many more questions and several runs of each. Five answers are far too few to trust. The maths stays the same.

Mention rate looks at you alone. It is the share of answers that name you. In the example, Alderpane appears in four of five answers, a mention rate of 80%.

Share of voice looks at you against everyone else. Alderpane's weighted share is only 25.7%, because other brands were named too, and often earlier. Both numbers are useful. Mention rate tells you about reach. Share of voice tells you about competition. The guide to measuring AI visibility covers how the two fit with other measures.

Share of voice counts and ranks mentions. It does not say whether a mention helped. Three extra measures fill that gap.

Sentiment. A mention that says "some customers report delays" is not the same as one that praises you.
Recommendation. Whether the answer actually suggests choosing you, and why.
Citations. Whether the engine linked to your own pages, which shows your content is being used as a source.

Some tools fold these into a single prominence score. That is handy, but keep the plain share of voice figure too, so you can see what is driving any change.

Split it by engine, by market and by question group. Each split can tell a different story.

By engine. ChatGPT and Perplexity search the web and cite pages. Gemini and Claude may answer from what the model already knows. A brand can lead in one and be missing in another.

By market. A firm with branches in several towns may lead in one and trail in another. Local AI visibility matters for any business that serves more than one area.

By question group. Group questions by stage, such as early research, comparison and choosing a supplier. Leading on research questions but trailing on supplier questions is a common and fixable pattern.

You improve it by appearing more often, and earlier, on the pages AI engines rely on. Start by looking at the answers where competitors are named ahead of you, and at the pages those answers cite.

If the same comparison article keeps appearing, make sure you are included and described fairly.
If review sites dominate, work on a steady flow of detailed, recent reviews.
If your own pages are cited but you still come low, make the answer on the page clearer and more specific.
If one engine leaves you out entirely, check whether its crawler can reach your site.

There is more on this in how AI engines choose sources.

More than most people expect. A share of voice built on a handful of answers can swing by several points just because one answer changed. A few hundred answers per period gives a much steadier figure.

For one brand in one market, 20 to 50 questions is a sensible start. Run each one across the engines you track, and more than once per period if you can. That quickly adds up to a few hundred answers. Businesses with many locations or product lines will need more, because each split needs enough answers of its own.

Keep the core questions fixed from month to month. You can add new ones, but report them separately until they have some history. The tracking guide explains how to build and run the set.

Share of voice is a sample, not a census. AI answers vary between runs, and consumer apps can use a person's location and history. OpenAI notes that ChatGPT may use location and saved memories when it rewrites a question into searches. So small moves between weeks are often noise.

The prompt set also shapes the result. A set full of questions that suit one brand will favour it. Keep the set balanced, keep it stable, and read the trend over several periods.

How Axiom GEO helps

Axiom GEO reports share of voice weighted by position, using 1/rank, for every brand named across your prompt set. It also gives a 0 to 100 prominence score that folds in recommendation, citations and sentiment. Competitor share of voice reads every stored answer, records every brand named, and tracks whether each was recommended and why, over time and per engine. AI visibility tracking runs the prompts on a schedule, and markets let you split share of voice by country or region.

Sources

These are the external pages referred to in this guide.

Frequently asked questions

How do you calculate AI share of voice?

Run a fixed set of questions through AI engines and list every brand named in each answer, in order. Give each mention a weight, such as one divided by its position. Add up each brand's weights, then divide by the total for all brands. The result is each brand's position-weighted share of voice.

What is 1/rank weighting?

It is a way of scoring mentions by where they appear. The first brand named in an answer scores 1, the second scores 0.5, the third about 0.33 and the fourth 0.25. It reflects the fact that readers pay more attention to the first names in a list, without ignoring later ones.

What is a good AI share of voice?

There is no fixed figure. It depends on how many competitors appear in your answers and how the prompt set is built. A useful test is whether you are ahead of the brands you actually lose work to, on the same questions, and whether your share is rising over several periods.

Is AI share of voice the same as mention rate?

No. Mention rate is the share of answers that name you, looked at on its own. Share of voice compares you with every other brand named. You can have a high mention rate but a low share of voice if competitors are named more often or earlier than you.

Should I track share of voice per engine?

Yes. ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews draw on different sources and behave differently. A brand can lead in one and be absent in another. An overall figure is handy for reports, but the per-engine split shows where to act.

Does share of voice include sentiment?

Not in its basic form, which counts mentions and weights them by position. Some tools add a separate score that folds in whether the brand was recommended, whether it was cited and the tone of the mention. It helps to read share of voice alongside sentiment, so a negative mention is not counted as a win.

AI share of voice, and how to weight it by position