Metrics and intelligence
Canonical metric contract
Every metric definition declares target/component applicability, unit, numerator, denominator, operation, version and low-sample threshold. Every snapshot declares sample size, confidence, low-sample state, dimensions, evidence and calculation time.
Core intelligence includes mention coverage, share of voice, citation presence, position/rank, source influence, prompt coverage, entity accuracy, narrative/sentiment, product visibility, campaign lift, cost, latency and success rate.
Representative metric contracts
| Metric | Numerator | Denominator | Unit | Required context |
|---|---|---|---|---|
| Mention coverage | Eligible answers mentioning resolved target | Eligible completed answers | Percent | Provider/model, prompt set, market, language, window |
| Citation presence | Eligible answers citing an approved/relevant source | Eligible completed answers | Percent | Citation extraction state and source policy |
| Competitive SOV | Weighted mentions for one eligible target | Weighted mentions for all eligible compared targets | Percent | Competitor set and identical sampling scope |
| Prompt coverage | Covered eligible prompt clusters | Eligible demand/prompt clusters | Percent | Cluster version and demand source |
| Entity accuracy | Correctly resolved/represented target samples | Reviewed eligible entity samples | Percent | Resolution policy and ambiguous exclusions |
| Run success rate | Completed requested samples | Requested eligible samples | Percent | Retry/exclusion policy |
Different metric versions must not be plotted as one continuous series without an explicit compatibility rule or visible boundary.
Share of voice
Share of voice is meaningful only when competitor targets and a representative sample exist. A result can sum to 100% and still be unreliable. Always read it with sample size, confidence, prompt coverage, providers, markets and time range.
Intelligence areas
- Visibility: mention and citation coverage by target, provider, prompt, market and language.
- Competition: owned-versus-competitor share of voice and evidence.
- Entity & Demand: canonical entity resolution, aliases, prompt clusters and unmet demand.
- Narratives: claims, sentiment, recurring themes and factual consistency.
- Safety: misinformation, unsupported claims, brand-risk patterns and anomalies.
- Citation graph: source domains and relationships between answers, citations and targets.
Interpretation guardrails
- Low-sample results are directional, not definitive.
- Missing citations and negative sentiment are findings, not proof of causal business impact.
- Simulation output cannot replace observed provider answers.
- Attribution confidence is capped by tracking coverage.
- Metric-version changes must be visible in comparisons and exports.