Customers and catalogs speak differently
People search through occasions, needs, style, slang, and context. Catalogs rely on controlled merchant attributes. AI can translate between them, but only if generated language remains separate from authoritative product facts.
Constrain the output
Generate structured attributes inside an allowed schema, validate unsupported claims, preserve source and model lineage, and make rollback possible. The goal is richer retrieval—not permission for a model to redefine the product.
Measure relevance and harm
Evaluate with human judgments, search behavior, zero-result rates, conversion signals, and error reviews. Better discovery must be balanced against irrelevant keywords, misleading claims, and taxonomy drift.