AI Share of Voice

How it is calculated

There is no standard formula owned by any one body. The common method works like this:

1. Fix a set of questions your buyers ask, in the shape "best [category] for [type of buyer]."
2. Run each question on each engine you care about, several times.
3. In every answer, record which brands are named, using a fixed list of your brand and the competitors you track.
4. Divide your brand's count by the total count of all tracked brands across those answers.

If your brand is named 30 times and all tracked brands together are named 120 times, your share of voice is 25%. Calculate it per engine first. ChatGPT, Gemini and AI Overviews behave differently, and a blended number hides that.

Two details change the result more than most teams expect. The competitor list sets the denominator, so adding or dropping a rival moves your share even if no answer changed. And the counting unit must be written down: one mention per answer, or every repetition of the name. This is a different number from mention rate, which divides answers that name you by all answers collected and ignores competitors entirely.

Mention, citation or recommendation

Share of voice is only as meaningful as the outcome it counts. As we set out in Cited, Mentioned, Recommended, these are three different outcomes:

• Citation: the answer links to your page as a source.
• Mention: your brand is named in the answer text.
• Recommendation: the answer presents your brand as a suitable option for the stated need, or favors it.

They diverge sharply. In a Semrush study published June 9, 2026, covering 115 prompts across 14 countries and four engines, Gemini named brands in 83.7% of appearances but linked to them only 21.4% of the time, while ChatGPT linked 87% of the time and named the brand in 20.7% (Semrush). A share of voice built on citations and one built on mentions can tell opposite stories about the same engine.

Most teams should calculate share of voice on mentions, then a second, stricter version on recommendations, counted only on selection-type questions ("best X for Y," "alternatives to Z"). The recommendation share sits closest to the buyer's shortlist. Report which one you mean every time.

Why answers vary, and why that shapes the method

AI answers are not fixed. Researchers testing five language models configured to be deterministic, on eight tasks over 10 runs each, found accuracy varied by up to 15% between runs, and none of the models delivered repeatable accuracy across all tasks (Atil et al., arXiv).

Brand lists behave the same way. In a study published in January 2026, SparkToro and Gumshoe.ai had 600 volunteers run 12 prompts on ChatGPT, Claude and Google's AI Overviews or AI Mode, for 2,961 runs in total. The tools returned the same list of brands fewer than 1 in 100 times and the same list in the same order fewer than 1 in 1,000 times. Some brands still appeared in most answers: one hospital showed up in 69 of 71 responses (SparkToro).

Retrieval adds more movement. In repeat-run ChatGPT data reported by Search Engine Land on July 8, 2026, switching between hidden retrieval pipelines cut cited-URL overlap by roughly 45% (Search Engine Land). We cover that research in Why Your AI Citations Keep Changing.

The practical consequence: a single run of a question is an anecdote. Share of voice is a frequency, so it needs repeated sampling of a fixed prompt set, with the same design every period you intend to compare.

Choosing the question set

The question set is a sample you design, and every share of voice is conditional on it.

• Start from buyer questions. Use sales call notes, search queries, review site language and support tickets, phrased the way a buyer would type them.
• Cover the buying stages. Problem questions, category questions, comparison questions and vendor-specific questions. Recommendations only occur on selection-type questions, so include enough of them.
• Keep branded prompts separate. If the question names your brand, the answer naming you proves nothing.
• Include several phrasings per intent. Buyers phrase the same need many ways, and a single phrasing can flatter or punish you by chance.
• Fix the set and version it. Changing questions mid-quarter breaks comparison. If you must change them, rebaseline and note it in the report.

Reporting pitfalls

• Small samples. A share calculated from a few dozen answers swings widely from week to week. Report the count behind every percentage, such as "named 30 times in 240 answers," and treat changes inside normal run-to-run variation as noise until they persist.
• Mixing engines. Pooling ChatGPT, Gemini and AI Overviews into one share lets the engine with the most answers dominate. Report per engine, then a clearly labeled blend if needed.
• Vanity prompts. Questions chosen because you already win them, or that contain your brand name, inflate share and say nothing about the market.
• Silent denominator changes. Adding a competitor or switching from mentions to recommendations changes the metric. Note it, or the trend line is fiction.
• Position as precision. Rank-in-answer is the least stable part of the output, so weight by position only with many runs behind it.

Our checklist in How to Read an AI Visibility Statistic applies to vendor share of voice claims too: ask what the denominator is and what was excluded. Used carefully, share of voice is a solid competitive benchmarking metric, and it starts with tracking the questions buyers actually ask, across engines, run after run.