Measuring Share of Voice in AI Answers
The formulas behind mention rate, citation rate and AI share of voice — and the sampling pitfalls that make benchmarks lie.
Once you can measure mentions at all, the next question is the competitive one: compared to whom? Share of voice in AI answers is the number that turns tracking into strategy — it prices your visibility against the competitors you actually fight — but only if the prompt set and the sampling are honest. This article gives the definitions, the formulas and the pitfalls. The raw tracking method it builds on is in How to Track Your Brand Mentions in ChatGPT, Gemini and Perplexity.
The prompt set comes first
Share of voice is meaningless until the denominator is defined. Three prompt roles make a defensible set: category prompts ("best tool for X"), which measure who gets recommended when buyers are shopping; competitor prompts (prompts naming each competitor by name), which measure who shows up when someone else is the starting point; and brand prompts, which measure what is said about you once you are in the conversation.
Freeze the competitor list the way you froze the prompts — the brands you genuinely lose deals to, not the ones that flatter the chart. And keep the set stable: a share-of-voice number computed on one set cannot be compared to a number computed on another, which is the most common way these metrics get abused.
The three formulas
- Mention rate — prompts where your brand appears in the answer ÷ total prompts. Read it per engine; averages hide the surfaces that matter.
- Citation rate — prompts where at least one of your URLs is cited ÷ total prompts. It measures the retrieval half of the pipeline, which mention rate alone cannot see.
- AI share of voice — your mentions ÷ all brand mentions in the set across the run, competitors included. It is the competitive number, and the one to trend monthly once the first two are stable.
Variants exist — weighted by answer position, restricted to category prompts, split by engine — and each is legitimate if you pick one definition and keep it. Position weighting is worth considering: being named first in the answer and being named last are different outcomes, and many teams in the space treat early mentions as more valuable, though how much more is not a settled question. Consistency beats sophistication: a simple formula tracked for six months outperforms a clever one you redefine quarterly.
Sampling noise: runs, dates and versions
One run of one prompt is an anecdote. Grounded, browsing answers vary from run to run because the retrieval step returns different pages, so run each prompt several times per cycle — three to five is a common working range — and treat a prompt as "mentioned" on a majority basis or track the fraction explicitly. Log the date and the model version with every run: providers ship model updates frequently, and an update can move every number overnight for reasons unrelated to your work. Compare like-for-like windows, same prompts, same engine versions, same runs-per-prompt.
Common mistakes that make benchmarks lie
- Comparing your number to a public benchmark built on a different prompt set — the denominators make the comparison meaningless.
- Presenting a single run as a trend; without repeats, you are graphing noise.
- Counting citations as mentions or merging the two, which hides which half of the pipeline is broken.
- Ignoring sentiment and position — a last-line mention with wrong pricing is not a win.
- Cherry-picking the engine where the number looks best and dropping the others.
What "good" looks like
There is no universal target, because share of voice depends entirely on your niche and your named competitor set. The honest benchmark is internal: trend your own line month over month, and set the first real goal as beating the median of your competitor set on category prompts. When you get there, the useful question becomes distribution — which prompts, which engines, which pages earn the mentions — and that points straight back at the fix lists in The AI Answer Audit Checklist. For the tooling that automates the sampling, see The AEO Tool Landscape: A Roundup; for the underlying definitions, What Is Answer Engine Optimization (AEO)? remains the starting point.
Keep reading
- How to Track Your Brand Mentions — the data collection these formulas consume.
- The AEO Tool Landscape: A Roundup — trackers that compute these numbers on a schedule.
- ChatGPT Recommendation Optimization — the playbook for the prompts you are losing.