Questions define the observed market

Commerce Intelligence evaluates recorded AI answers to a configured buying-question panel. Scenarios, country, language, providers, models, and the selected period determine which observations enter an analysis. It does not exhaustively inspect every store, search query, or consumer transaction. A product absent from the panel's answers may still be important to your customers.

Collection is separate from extraction

A provider response first becomes a saved answer. Extraction identifies the products, brands, retailers, categories, attributes, supported recommendation positions, and associated evidence in that answer. A collected response is not necessarily fully classifiable. Check collection and extraction status separately, particularly when a provider fails, output is truncated, or an entity cannot be identified reliably.

Products, brands, retailers, and citations

These entity types must stay distinct. A manufacturer can be recommended without a retailer being recommended. A retailer can be named as a source without being endorsed as a place to buy. A product heading can establish product order without establishing retailer order. Multiple entities can be recommended in a single answer, so entity totals can exceed the number of answers.

Normalisation and identity

Normalisation reduces harmless differences in case, spacing, punctuation, aliases, and canonical names. Product generation, size, connector, material, or bundle differences can be material and must not be erased just to make a match. A similar product name is a candidate identity rather than proof. Ambiguous references remain reviewable instead of being automatically promoted to a confident catalogue match.

Recommendation evidence and deduplication

An entity is counted as recommended only when the relevant assessment supports a positive recommendation. Repeating its name several times in one answer does not create several independent recommendation answers. Coverage uses unique answer–product pairs, while other views use answer-level or entity-unit denominators. Question, provider, run, and passage provenance make these counts inspectable.

Catalogue comparison is a second decision

After identifying an observed product, the system compares it with the connected catalogue. This is a separate assessment from whether the model recommended it. A high recommendation rate can coexist with an unresolved catalogue identity. Review variants and the current catalogue snapshot before interpreting 'no catalogue match' as a real assortment gap.

Optional external observations

Where the workspace supports shopping or product enrichment, an external observation can add an offer URL, observed price, retailer, or freshness information. Availability depends on the configured service and workflow. An enrichment timestamp is not an ongoing stock guarantee. A missing shopping observation is not a zero price, a proof of discontinuation, or permission to compare different variants as identical offers.

A complete evidence chain

For a useful product opportunity, inspect the question and original answer, the positive product passage, exact identity and variant, eligible sample, catalogue state, and any external observation date. Keep recommendation evidence separate from freshness and commercial suitability. A persuasive summary without that chain is weaker than a smaller finding with inspectable evidence.

Commerce metrics: answers, recommendation units, and sharesCatalogue matching, coverage buckets, and the CI ScoreHow market opportunities are prioritised and reviewed