A score is a conditional index

The AEO score combines several recorded-answer signals on a 0–100 scale. It is not a probability of being recommended, a search ranking, a conversion forecast, or market share. Component weights are provisional product-policy choices rather than externally calibrated revenue effects. The general V2 family and its retailer/business-objective revisions are versioned; inspect the version and objective on a saved result before applying a formula to an older snapshot.

Eligible observations and question clusters

Baseline scoring uses valid, assessable observations aligned to the relevant objective. Neutral baseline questions are distinct from branded confirmation, exact-product discovery, persona, preference, or stability experiments. Repeats of one question are averaged within its cluster so repeating an easy question does not manufacture independent coverage. The recorded provider, model, country, language, configuration, and question version remain part of the measurement context.

Intent and provider balancing

For the general AEO V2 baseline, informational questions receive 20% of the intent mix, comparison questions 40%, and commercial questions 40%. Questions are clustered and providers are balanced within the aggregation. Missing intent groups are not redistributed to the remaining groups. As a result, a portfolio containing only high-performing informational questions cannot stand in for a complete three-intent AEO portfolio.

Example
General AEO V2 intent mix
intent-balanced value =
  0.20 × informational value
+ 0.40 × comparison value
+ 0.40 × commercial value

Repeated observations stay inside their question cluster.
A missing required intent is not treated as a zero or renormalised away.

General AEO V2 component weights

The table describes the general baseline definition. Objective-specific retailer and business revisions must be read using their saved component definitions. Citation weight is applicable to the AI-search definition, not bolted onto model tests with unavailable search metadata.

Mode and tracked competitionPresenceRecommendationProminenceCompetitionOwn citations
Model test, with competitors30402010—
Model test, without competitors354520——
AI search, with competitors2535151015
AI search, without competitors304015—15

Prominence and competitive utility

Prominence uses the supported recommendation rank, not mention order: utility equals 1 / log2(rank + 1). Rank one contributes 1, rank two approximately 0.631, and rank three 0.5. A confirmed non-recommendation contributes zero; an unknown rank remains missing. Competitive utility is 1 for a decidable win, 0.5 for a tie, and 0 for a loss. An indeterminate comparison is not a loss.

Commercial score and business type

For the general retailer/marketplace/mixed commercial definition, component weights are shop recommendations 50, first retailer position 20, distinct purchase-question coverage 15, competitor position 10, and verified shop links conditional on recommendations 5. Brand, manufacturer, service, and B2B objectives use their applicable purchase recommendation, prominence, and coverage definitions instead; they must not be judged as though every business were a retailer. Publishing has no automatic purchase objective merely to fill a card.

Publication gates and missing components

The reviewed general implementation requires at least 80 points of available component weight, at least three distinct eligible questions, and at least six eligible observations. General AEO also needs its required intent groups. Missing components are not assigned invented values and their weights are not silently redistributed. A partial score can therefore omit a contribution and must expose that fact. Full, partial, provisional, and unavailable results communicate different evidence conditions.

Intervals, history, and interpretation

With sufficient distinct questions, a full eligible result can receive an indicative question-cluster bootstrap interval; the reviewed engine uses 240 deterministic resamples and requires at least ten question clusters for that interval. It describes variability of this recorded panel, not every possible real-world prompt. Saved score snapshots preserve formula and cohort context. Do not subtract a legacy score from a new objective revision and label the result a genuine performance improvement.

AEO rates, retailer wins, and unclear answers explainedComparable measurements, repeated sampling, and uncertaintyCatalogue matching, coverage buckets, and the CI Score