COSMO is the name Amazon's researchers gave to a framework that builds a large knowledge graph of commonsense relationships between products, uses and shopper intents, and applies it to search and recommendations. Where classic Amazon search matched the words in a query to the words in a listing, COSMO-style reasoning can connect a query such as "shoes for a pregnant woman" to products with the attributes that intent implies, such as slip-on and cushioned, even when the listing never uses the word "pregnant."
Why it matters
Amazon has said publicly that this kind of reasoning is used in its search, and sellers see the effect: listings appear for queries they never targeted, and queries that used to be won by exact keyword matches are now shared with products that fit the intent better. For listing work that means two things. Keyword density matters less than it did, and being explicit about who the product is for, what it is used with and what problem it solves matters more, because those are the relationships the graph is built from.
COSMO and Rufus are related but different. COSMO is about how search interprets queries and products; Rufus is the shopper-facing assistant. A listing built for one tends to serve the other.
What moves it
- Use-case and audience language in bullets and A+ Content: the situations, people and companion products the item goes with.
- Structured attributes, which give the graph unambiguous facts to link.
- Consistency with behavior. Amazon's model learns from what shoppers buy after a query, so a listing that converts for a query reinforces the association.
- Category placement, which anchors the product in the right neighborhood of the graph.
How we use it
We write listing copy in two layers: the terms shoppers type, verified through indexing checks, and the intents behind them, expressed as plain statements about use, audience and compatibility. We then watch Search Query Performance for queries the listing has started to receive that it does not literally contain, which is the practical sign that the semantic layer is working.