Operator glossary
COSMO
Amazon's common-sense knowledge model that maps queries to products by inferred intent — why 'gift for new dad' can rank items whose listings never contain those words.
COSMO is Amazon's semantic layer: a knowledge model that connects what shoppers mean to what products are for. It lets search rank a camping chair for 'concert seating' or a white-noise machine for 'newborn sleep help' even when the listing never used those phrases, because the model has learned the use-case relationships from behavior at scale.
For listings this changes the job from covering keywords to declaring use cases. State who the product is for, what situation it serves, and what problem it ends — in plain sentences COSMO can parse. Listings that only enumerate specs give the model nothing to connect, and they are the ones losing impressions as semantic matching takes share from lexical matching.
The numbers
- Published by Amazon Science (SIGMOD 2024) as a large-scale common-sense knowledge system; production experiments reported measurable engagement gains when COSMO relationships fed search.
- It maps products to inferred use-cases and audiences ('camping chair' → 'concert seating'), which is why intent-declaring copy wins impressions that keyword lists miss.
Primary sources: COSMO paper (Amazon Science)