Problem
Shoppers want to know which items go with the one they are viewing. At Target the catalog is huge, and what matters differs by product: color, kid-friendliness, flavor, material, brand or a hundred other attributes. A human expert cannot work out the important attributes for every pairing. The Product Recommendations Team had to build a "related accessory" algorithm for Electronics and Home; the article describes the Home model (GRAM).
Example from the article: for pillowcases recommended with a sheet set, color and material matter most. For a book recommended with a kids' craft kit, the intended audience (infant, kids, adults) matters most.
Solution in brief
An LLM reads the product data and assigns importance weights to attributes for each pair of core (seed) and accessory item types. An accessory's score is the sum of the weights of the attributes whose values match the seed item. The LLM also supplies "harmonious" attribute values, and merchants add a list of commonly co-purchased accessory types to diversify results.
Questions this case answers
- How to recommend complementary items across a very large catalog when the attributes that signal fit differ for every pair of item types and manual rule curation does not scale
- How to score item pairs by attribute compatibility when exact value matching leaves too few candidates
- How to recommend accessories that go with a product when the catalog has hundreds of thousands of items and the relevant attributes differ by product type?
- How to decide which attributes matter for each pair of product types without a human expert reviewing every pair?
Tools and methods
- LLM-generated attribute weights — importance per pair of item types
- Attribute-match scoring — sum of matched weights, sales rank breaks ties
- LLM relative value matching — values that coordinate, not only equal
- Human-in-the-loop (HITL) co-purchase list — merchant-curated accessory types
- A/B test — add-to-cart flyout, Home category
Google or a one-time email code. No password.