Marketing wants a number that means “are we winning in AI?” Product wants a number that survives a follow-up question. LLM share of voice is the rare metric that can satisfy both, provided you define it precisely and refuse the two shortcuts that make it meaningless.
Definition and formula
LLM share of voice is your brand’s share of the addressable AI visibility in a defined topic set, over a defined period, across a defined set of engines. In its simplest defensible form:
SOV = your citations ÷ (your citations + tracked competitor citations)
Every word in the definition is load-bearing. Change the topic set and the number changes. Add an engine and the number changes. Report it without those three parameters, and you have published a mood, not a metric.
Where naive share-of-voice goes wrong
- Mixing populations. Counting simulated-prompt mentions and real referred sessions in the same numerator produces a number with no unit. Pick one population per metric.
- Unweighted engines. Treating a citation in an engine that sends you meaningful traffic as equal to one in an engine that sends none flatters whichever engine is easiest to win.
- Elastic competitor sets. Adding or dropping competitors silently changes the denominator. Freeze the set for a quarter and version it.
- Query sets written by the vendor. If the prompt list is chosen to flatter you, the number will. Write it from real demand: support tickets, sales objections, search console queries.
How to compute a defensible cross-engine SOV
- Step 1 — freeze the topic set. Twenty to fifty questions your buyers actually ask, written down and versioned.
- Step 2 — freeze the competitor set. Five to ten names, reviewed quarterly, never mid-period.
- Step 3 — collect per engine. ChatGPT, Perplexity, Claude, Gemini and Copilot separately. Never blend before you have looked at the split.
- Step 4 — weight by outcome. Weight each engine by the referred sessions it actually sends you, so the composite tracks business reality rather than engine count.
- Step 5 — publish the parameters with the number. Topic set version, competitor set version, engines, date range. Every time.
A sample dashboard sketch
| Panel | Metric | Owner |
|---|---|---|
| Composite SOV | Weighted share across five engines, weekly | CMO |
| Per-engine SOV | Unweighted share per engine, weekly | Content lead |
| Cited pages | Landing pages receiving AI referrals | Content lead |
| Crawl coverage | Share of published URLs fetched in 30 days | Engineering |
| AI-referred conversion | Conversion rate vs organic search | Growth |
The bottom two rows are what make the top row credible. A share-of-voice line that goes up while crawl coverage and AI-referred conversion stay flat is a measurement artefact, and somebody in the room will eventually notice.
Getting the traffic half right
The referral and cited-page inputs come from your own server, not from a model API. That is the part Citealytics handles: per-engine AI referrals, cited landing pages, and an AI Visibility Score per page, all classified against the public crawler registry. Pair it with a prompt-simulation tool from the ranked tools list and you have both halves of a defensible number. The vocabulary trap to avoid before you start is covered in mentions vs citations vs recommendations.