Context
Koushik and colleagues augmented a curated Amazon ESCI subset of 1,887 query-product pairs from 266 queries with LLM-generated attributes and human-reviewed explanations. Their controlled reranking experiments report nDCG@10 of 0.9382 for a human-feature oracle, 0.9258 for a human-free Q+P configuration, and 0.9150 for synthetic approximations.
What changed
The September 20, 2026 preprint tests richer, LLM-assisted product and query attributes plus human post-edited explanations, separating directly intent-satisfying products from peripheral complements and substitutes.
Why it matters
Shopping discovery depends on whether a product actually meets the query's primary requirements, not just textual similarity. Structured attributes and explanations offer measurable signals for product selection.
Visibility OS interpretation
Visibility OS interpretation: Intent-satisfying product attributes may be more informative than raw similarity alone. Human-oracle results are an upper-bound diagnostic, not a deployable production result; no live AI shopping platform was tested.
What remains uncertain
Do the improvements persist at full-catalog scale, on non-English queries and in live shopping assistant recommendations? How do annotator disagreement and synthetic attribute errors affect downstream results?