A home-shopping store, a beauty retailer and a streaming service share customers but not a single product code or category name. Labelling every item with the same set of interests — the step done with an LLM in production — makes one business's customers readable in the others.
Interests such as skin health, home cooking or true crime, each with a one-line definition, grouped under a dozen themes. The definitions are what keep the labelling consistent.
Product and content metadata — name, category, brand, synopsis — goes to an LLM with the taxonomy; the answer must use labels from the list only. Tens of thousands of items become comparable in one pass.
Each customer's purchases and views roll up into an interest profile, turned into percentiles. Pick a cohort anywhere; compare its profile with everyone else's in every business, with a test for whether the difference is real.