What is AI share of voice?

Share of voice comes from classic media planning. There it measures what portion of a category's ad spend or impressions a brand captures compared with its direct competitors. GEO tools borrowed the name for a similar but different problem: how often a brand shows up in an AI assistant's answers, compared with the rival brands the same tool tracks in the same batch of prompts.

The common mistake is treating it as a synonym for AI Visibility. They are not the same. AI Visibility answers a binary question per prompt: does the brand show up or not. Share of voice answers a relative question: of all the mentions produced across that prompt set, how many belong to this brand and how many to the rest. A brand can have high AI Visibility and low share of voice at the same time, if its competitors show up even more often in those same answers.

No standards body defines share of voice for generative AI. Unlike a metric with a fixed formula, such as a search engine's click-through rate, each GEO provider builds its own numerator and its own denominator, and calls the result by the same name its competitors use. That freedom isn't necessarily bad, but it requires whoever receives the figure to know they're reading that tool's own calculation, not a measurement the whole industry recognizes.

The calculation always follows the same general shape: count the brand's own mentions in a set of AI answers, count the mentions of every tracked brand in that same set, and divide the first figure by the second. The result is expressed as a percentage.

What changes from tool to tool are the decisions behind those two figures, and almost none of them publish those decisions in detail. The first decision is which brands enter the denominator: if the client picks three competitors by hand, or if the tool auto-detects every brand that appears, the number moves without the real market having changed. The second is what counts as a mention: an exact brand string, a reference to the entity even if the name doesn't match letter for letter, or only mentions that carry a link to the brand's site. The third is whether position gets weighted: a mention in the first sentence of an answer doesn't carry the same weight for a user as a mention at the end of a list of ten brands, but few tools make that distinction. The fourth is the set of prompts and models: querying one model once a week with fifty fixed prompts produces a different number than querying four models daily with five hundred rotating prompts.

One example shows the effect: if one tool tracks ten brands in a category and another tracks twenty-five in the same category, the brand measured by the first has better odds of a high percentage, simply because it competes against fewer rivals in the denominator, not because its real presence in AI answers improved.

None of these decisions is wrong on its own. The problem shows up when two tools publish a share of voice percentage without explaining how they got there, and someone compares the two numbers as if they measured exactly the same thing.

Share of voice gets used to justify budget decisions: if a brand climbs five points in a quarter, someone will ask to repeat whatever was done that quarter. That decision only makes sense if the number measures the same thing month over month.

Before presenting the figure to a client or to management, it's worth asking the provider in writing whether the denominator changed from the prior period. The question takes a minute and prevents crediting a campaign for a move that actually came from a change in methodology.

The question worth answering before acting on the figure is simple: did the market change, or did the sample change? If the tool added a new competitor to the denominator, or expanded the prompt set from a hundred to three hundred, a rise or drop of several points may have nothing to do with the SEO or content work done that quarter.

This also affects comparisons between agencies or between tools. A report that shows 42% share of voice and another that shows 28% for the same brand in the same month don't contradict or correct each other: they're almost certainly measuring with a different denominator. Treating that gap as one tool's error, instead of a difference in method, leads to decisions based on the wrong number.

• Ask the GEO tool to document which brands enter the denominator and how many there are.
• Fix the set of competitor brands for the whole period you want to compare, and note any change.
• Ask how a mention is counted: exact brand text, entity reference, or link-only.
• Compare share of voice only within the same tool and the same method, never across different providers.
• Check the prompt sample size before reading a rise as a trend: with few prompts, the number is noise.
• Keep the actual AI answers behind the percentage, not just the final figure, so it can be audited later.

• Comparing the percentage from two different GEO tools as if it were the same metric.
• Changing the list of tracked competitors without noting the change in the report.
• Reporting share of voice without saying how many prompts and which models back it.
• Confusing share of voice with AI Visibility, which answers a different question.
• Reading a one-month swing on a small prompt sample as a real trend, instead of checking that sample size first.