AI Share of Voice: How to Benchmark Against Competitors

AI share of voice (AI SOV) is the percentage of AI answers to your category's buyer questions in which your brand appears — and, in its stricter form, the percentage in which you're the recommended pick. It's the closest thing generative engines have to a rank tracker, and it's the metric that makes "how are we doing in AI answers?" answerable with a number instead of anecdotes.

This guide defines the metrics precisely, then walks through a benchmarking method that survives AI's non-determinism — the property that breaks most naive measurement attempts.

The Metrics, Defined

Teams use "share of voice" loosely, which makes benchmarks incomparable. Fix the definitions first:

Metric	Definition	What it tells you
Mention rate	% of answers (across your question set) that name your brand	Are you in the conversation at all?
Recommendation rate	% of answers where you are the explicit pick	Are you winning the conversation?
First-mention share	% of list-style answers where you're named first	Position inside answers buyers skim
Citation share	% of cited sources that are your domains	How much of the answer's evidence you own
Relative SOV	Your mention rate ÷ (sum of mention rates of tracked brands)	Your slice of the category's total AI presence

Mention rate flatters you; recommendation rate pays the bills. A brand can be mentioned in 80% of answers and recommended in 10% — "in the room" but never the pick. Track both, and lead reports with recommendation rate.

The Benchmarking Method

Fix the question set. Benchmarks require a stable panel: 20–50 buyer questions covering category picks, comparisons, alternatives, and use-case-specific asks. Changing the questions changes the number, so version the set — a documented prompt pack — and keep it constant within a benchmarking period.
Fix the engines and modes. ChatGPT with and without browsing produce different answers. Decide which engines and modes are in scope and hold them constant.
Sample, don't spot-check. The same question to the same engine can produce different answers across runs. A single run per question is a coin flip wearing a lab coat. Run each question multiple times per period (or at minimum across multiple days) and compute rates over all runs.
Count with consistent rules. Decide up front what counts as a mention (name in the answer body — not only inside a URL), and what counts as a recommendation (an explicit pick or "best for your case is..." — not mere presence in a list). Write the rules down; the person counting in Q4 won't remember Q2's judgment calls.
Score every tracked competitor the same way. The same runs, the same counting rules, applied to each rival. Your recommendation rate only means something next to theirs.
Report the spread, not just the average. "Recommended in 40% of runs" reads very differently if it's stable across engines versus 80% on Gemini and 0% on ChatGPT. Per-engine breakdowns tell you where to work.

How many runs is enough? There's no magic number, but the intuition is simple: the closer two brands' rates are, the more runs you need before the gap means anything. Three runs per question per engine per month is a practical minimum for trend reporting; if you're trying to detect a 5-point shift between near-tied rivals, you need substantially more — or you should report the pair as "contested" rather than pretending to resolution the sample can't support. Weight engines by your buyers, too: if your prospects overwhelmingly use ChatGPT, a blended average that dilutes it with engines they don't use is a vanity aggregate.

A Worked Example (Hypothetical)

Say you monitor 20 questions on two engines, four runs each per month: 160 answers. Your brand appears in 96 of them — mention rate 60%. It's the explicit pick in 40 — recommendation rate 25%. Your nearest rival is mentioned in 120 (75%) and picked in 56 (35%). Relative SOV across the three brands you track: you 33%, rival A 41%, rival B 26%.

Now the reading. The headline isn't the 60% — it's the 10-point recommendation gap to rival A. Cut by question type, you find your recommendation rate is 45% on head-to-heads but 8% on open category questions: buyers who already know you win the comparison, but you're losing the "best tool for..." shortlists where rival A's content dominates the citations. That's a precise, workable diagnosis — and none of it was visible in the mention rate. This is the analytical payoff of keeping metrics separated and question-level detail intact.

Setting Targets Without Fooling Yourself

Baseline first, targets second. Run two full periods before committing to any goal; the first period's number always moves once your sampling stabilizes.
Target the gap, not the absolute. "Close the recommendation gap to rival A from 10 points to 5 in two quarters" survives model refreshes better than "reach 40% SOV," because refreshes move everyone's absolute numbers at once.
Pair every SOV target with an accuracy floor. Being recommended more while engines misquote your pricing is winning the wrong game; track both, and treat a false-claim spike as overriding any SOV win.
Expect step changes. SOV moves in steps (a model update, a big citation shift), not smooth curves. Judge trends across quarters, not weeks.

Benchmarks Worth Having

You vs. category leader: the gap in recommendation rate on category questions.
You vs. nearest rival: head-to-head win rate on direct comparison questions.
You vs. yourself: the trend line — the benchmark that turns SOV into a program metric. Same panel, same method, every period.
Segment cuts: SOV on enterprise-flavored questions vs. SMB-flavored ones often diverges sharply and reveals positioning problems no aggregate shows.

Pitfalls That Invalidate Benchmarks

Non-determinism ignored. One run per question produces numbers that swing week to week for no real reason. If your SOV moves 15 points in a week, check sample size before celebrating or panicking.
Personalized sessions. Logged-in history skews answers. Measure from clean sessions.
Leading prompts. "Why is [You] the best?" inflates SOV and measures nothing. Questions must be neutral, phrased the way buyers actually phrase them — see how buyers actually ask AI.
Counting rule drift. If "mentioned" quietly becomes "mentioned or cited," your trend line is fiction.
Averaging away the story. A stable aggregate can hide a critical question flipping to a rival. Keep question-level detail underneath the headline number — displacement hides in averages (our displacement detection playbook covers catching it).

Cadence and Reporting

Weekly scans feed monitoring; monthly aggregates make good benchmarks; quarterly deltas belong in leadership decks. Present three numbers — mention rate, recommendation rate, and the top rival's recommendation rate — plus the question-level wins and losses behind them. Resist the composite-index temptation: a single blended "AI score" that mixes mentions, recommendations, and citations feels executive-friendly but hides which lever moved, and the first question anyone asks about a composite is what's inside it anyway. For how SOV fits into a broader measurement stack alongside citation and accuracy metrics, see how to measure generative engine optimization.

Running this by hand is doable but tedious: the sampling requirement is what breaks spreadsheets. Perciva runs the panel across engines on schedule and computes these rates per question and per rival — you can see the output format in our sample report.