Every metric must identify its measurement universe and evidence class.

Define the measurement universe before interpreting the number.

A valid result should identify the entity or brand, competitor set where relevant, prompt/intent universe, AI platform set, markets, languages, time period, methodology version and evidence class. Changing any of those can change the meaning of the metric even if the label stays the same.

Six evidence classes keep unlike evidence separate.

External AI measurement uses defined prompt panels and repeatable sampling.

Prompt and intent design

Panels can combine category research, customer-priority questions, search/customer demand signals, competitor language, product/category taxonomies, local-language research and emerging topics. Prompts are grouped into intent families rather than selected only because they favor a brand.

Sampling

Record the exact or normalized prompt, platform/model, date/time, geography/language, repeat number and relevant environment/configuration. Where stochasticity is material, repeated samples are preferable to one answer.

Entity extraction and recommendation

Resolve aliases, subsidiaries, products, spelling variants and local-language names to canonical entities. A recommendation is not the same as a mention: classifications can distinguish recommended/preferred, included as an option, neutral mention, negative recommendation/avoid, competitor-only and absent. Rules should be versioned and testable.

AI demand is estimated unless authorized platform data proves otherwise.

SIGNAL does not claim direct knowledge of all private AI prompts. Estimated demand can combine approved inputs such as search demand, first-party site search, customer query data, prompt panels, trend data, authorized platform telemetry and category models. Every demand number should identify whether it is estimated or platform-provided.

Citation evidence, source influence and narratives require different claims.

A citation is directly observable when an AI answer exposes a source, link or reference. Source Influence can instead rely on direct citation evidence, platform retrieval evidence, temporal/source correlation or a modeled relationship; “cited” must not be equated with “caused the answer.”

Narrative analysis can characterize dimensions such as reliability, value, innovation, safety, quality, service, trust and sustainability. For factual accuracy, compare recurring claims against an approved source-of-truth dataset.

First-party Analytics prioritizes server-side evidence.

AI crawler and agent analytics should prioritize server, CDN or edge logs because automated agents may not execute browser JavaScript. Requests can be classified as human, known crawler, retrieval agent, authenticated agent, likely automation or unknown using documented combinations of identity signals.

AI referral classification can use referrer information, URL parameters, platform link patterns, customer tags or authorized platform data.

Attribution is not automatically incrementality.

Depending on data and campaign design, incrementality can use randomized holdouts, geo holdouts, matched markets, synthetic controls, controlled time series or another approved design. Reports must disclose the design, assumptions, exclusions and confidence.

Verify starts from a claim and a frozen audit scope.

1. Record the claim.
2. Freeze scope and metrics where practical.
3. Acquire evidence.
4. Normalize time and scope.
5. Run the methodology.
6. Test sensitivity and confounders.
7. Assign support status.
8. Deliver the evidence package.

Support status is one of: Supported, Partially supported, Unsupported, Inconclusive, Not measurable.

Baselines should represent normal variation while remaining relevant to the campaign period. Changes to platform set, prompt universe, competitor set, market or methodology must be annotated rather than silently absorbed into an uplift number.

“Confidence” must name what kind of confidence it means.

Confidence may refer to a statistical interval, classification confidence, evidence completeness or methodological reliability. SIGNAL must not display a generic confidence percentage without identifying the underlying meaning.

Normalize high-level concepts without erasing platform-native evidence.

Different AI systems produce different answer structures. SIGNAL can normalize concepts such as mention and recommendation while preserving platform-native evidence. Paid-media definitions remain platform-specific where direct comparison is imperfect.

Model/platform updates can materially change observed answers; known measurement discontinuities should be annotated rather than automatically attributed to a campaign. Multilingual measurement should use native-language prompts and local source environments rather than translation-only panels.

Reproducibility requires enough metadata to reconstruct the conclusion.

Enterprise reports should retain scope, data sources, methodology version, result, exclusions and uncertainty. Material scoring or classification changes should be reflected in a public change log once metrics are used commercially at scale.

See Metrics for metric definitions and Verify for the independent audit layer.