Frame the work
The situation I was solving
Customers search in human language while catalogs rely on merchant attributes and taxonomy. Uncontrolled AI enrichment risks unsupported claims, irrelevant keywords, and damage to source data.
Read the underlying principle: The opportunity comes first. The technology comes second. →Define what changes
What becomes better
A controlled pipeline can improve discoverability while preserving catalog authority, lineage, rollback, and measurable relevance.
Design the system
How I work through it
I constrain generation to structured attributes, validate against an allowed schema, preserve source and model lineage, and evaluate with both human judgments and search behaviour.
Go deeper: Most teams optimise the engine before rethinking growth. →Carry the learning
What I carry forward
Language models are most useful in search when they expand how products can be understood without gaining authority to rewrite what the product is.
Explore all educational field notes →