Why mention-counting alone misleads
"We're mentioned in 6 of 10 answers" sounds strong — until you learn a competitor is in 9 of 10, first every time, while your six mentions are trailing caveats. Share of voice forces the comparison into the metric: everyone's mentions are counted, so the number reflects the race, not just your lane.
- Share: your mentions ÷ all brand mentions, per question set and period.
- Position: named first vs named last inside answers — first-mentions carry the recommendation weight.
- Citation share: whose URLs the engines actually link — presence in text and presence as a source diverge, and the source list is where authority shows.
Measuring it honestly
- Fix the question set BEFORE looking at results — picking questions you already win is the classic way teams fool themselves.
- Sample repeatedly: answers vary run to run, so a share computed from single runs is noise wearing a percentage sign. Multiple samples per question, per engine.
- Count competitors by the same matching rules as yourself (aliases included: "HubSpot" and "HubSpot CRM" are one brand — and so are your own nicknames).
- Keep engines separate before averaging: 60% on Perplexity and 10% on ChatGPT is strategy-relevant detail a blended 35% hides.
- Track the trend, not the level: the level depends on your question mix; the trend under a FIXED mix is the real signal.
Reading the number: what moves it
Share of voice responds to two different machines, on two different clocks. On-site fixes (answer-first passages, sourced claims, crawler access) are re-crawled quickly on live-search engines. The consensus layer — reviews, roundups, community presence — changes far more slowly, and it is where durable positions tend to sit. Your own re-checks are what report the timing. A share that moves with no on-site change is a reason to look at off-site activity — yours or a competitor's — rather than proof that either caused it.
Promvia computes share of voice from your tracked questions across seven engines — with the alias matching, per-engine breakdowns and multi-sample checks above built in — and its history chart shows whether your slice is growing. The definition matters more than the tool: however you measure, measure by these rules or the number will comfort you instead of informing you.
Frequently asked questions
What's a good AI share of voice?
There's no universal benchmark — categories differ in how many brands answers name. The useful comparisons are your own trend under a fixed question set, and your share versus the category leader's.
How many questions do I need for a meaningful number?
Enough that one question can't swing the total: 15–30 tracked questions is a practical floor for a small brand; below that, treat the number as directional only.
Should sentiment be part of share of voice?
Keep them separate: share tells you how often you're in the conversation, sentiment tells you how the conversation treats you. A rising share of negative mentions is a different emergency than absence.