Key takeaways: COSMO is a system Amazon researchers described in a 2024 paper: a knowledge graph of commonsense relationships between what shoppers search for, why they want it and which products serve that purpose. It lets search connect "shoes for a pregnant woman" with slip-resistant footwear without the listing containing those words. For sellers it means use cases, audiences and attributes now matter alongside keywords. It does not mean keywords stopped mattering.
What COSMO is, according to Amazon
In 2024, Amazon scientists published a paper describing COSMO, a framework for building a large-scale commonsense knowledge graph for e-commerce. The idea is to capture relationships that a human shopper takes for granted and a keyword engine does not: that a camping trip implies a need for something waterproof, that a "gift for a new dad" implies certain product types, that "shoes for pregnant women" implies comfort and slip resistance rather than a search for the literal words.
The paper describes generating these relationships with large language models from Amazon's search and purchase data, filtering them with human review and machine judgment, and using the resulting graph to improve search relevance and product recommendations. Amazon reported that the system improved relevance in its experiments.
That is what is publicly known. Amazon has not published how heavily COSMO weighs in live ranking, which marketplaces or categories it applies to, or how it interacts with the existing relevance and sales-based signals. Claims about exact mechanics beyond the paper are speculation, and this article avoids them.
Lexical matching versus intent matching
Read this flowchart as text
- Shopper query: shoes for pregnant women
- Lexical engine: match words in listing fields
- COSMO layer: infer intent: comfort, support, slip resistance, easy on
- Listings containing those words
- Listings whose attributes and copy serve that intent
- Ranked with sales, CTR, CVR and other signals
A purely lexical engine ranks the listings that contain the query's words, then orders them by performance signals. An intent layer adds candidates that serve the purpose behind the query even when they lack the words, and can demote candidates that contain the words but do not fit the purpose. The performance signals still decide the order among relevant candidates.
For sellers the practical shift is this: a listing used to become a candidate for a query by containing its words. Now it can also become a candidate by clearly being the right kind of product for the need behind the query. "Clearly" is the operative word, and it is determined by what your listing says about use, audience and attributes.
What changes for listings
Use cases become indexable in a new way. Stating that a bottle is for hiking, that a lamp is for reading in bed, or that a chair suits people over six feet tall connects the product to intents that shoppers express in a hundred different phrasings you could never enumerate as keywords.
Audience statements carry weight. "For beginners", "for small apartments", "for sensitive skin" are the kind of relationships a commonsense graph encodes. If your listing is silent about who it is for, the graph has to infer it from weaker signals.
Structured attributes matter more. Material, item form, special features, target audience, recommended uses and the category-specific fields are the cleanest inputs to any semantic system, because they are machine-readable values rather than prose. A listing with rich attributes gives the graph facts to connect; one with blanks gives it guesses.
Consistency across fields matters. A system reasoning about your product across title, bullets, attributes and A+ is confused by contradictions in a way a keyword matcher was not.
Reviews become a relevance source. Amazon's data about what shoppers say about the product after buying it feeds the same understanding. A product reviewed as "great for camping" by many buyers is connected to camping whether or not the listing says so; a product reviewed as "too heavy to carry" is connected to the opposite.
What does not change
Keywords still matter. The paper describes COSMO as an addition to search relevance, not a replacement. Shoppers still type product-type phrases and attributes, and a listing still needs to contain the words for the primary phrases in its title, bullets and backend field. The keyword process in Amazon Keyword Research: How to Find the Terms That Sell and the backend rules in Amazon Backend Search Terms: Best Practice for the 250-Byte Field are unchanged.
Performance signals still order the results. Sales velocity, click-through and conversion rates remain the way one relevant listing outranks another. A semantically perfect listing that converts poorly does not hold page one.
Amazon's listing policies still apply. Adding use cases and audiences is writing better copy, not stuffing. A bullet that lists twenty scenarios is no more credible to a semantic system than to a human.
What we do differently because of it
The changes are modest and they were good practice before COSMO:
- Add one explicit "who and what for" statement to the bullets and A+ Content for every ASIN, drawn from the use cases buyers actually name in reviews and Search Query Performance.
- Complete every structured attribute, using Amazon's valid values where offered.
- Check the listing for contradictions across fields after every edit.
- Read Search Query Performance for queries that describe a need rather than a product, and check whether the listing appears for them. Those long-tail intent queries are where semantic matching shows first.
- Treat the review cluster analysis as a relevance input as well as a product-quality input.
The Rufus side of the same shift, the assistant that answers questions from the same fields, has its own checklist in Amazon Rufus Listing Checklist: 15 Things the AI Assistant Reads. How we prepare a listing for both is below.
Frequently asked questions
Is COSMO live in Amazon search now?
Amazon's 2024 paper describes it as deployed in experiments that improved search relevance. Amazon has not published a detailed account of how widely it runs in production or how much weight it carries. The safe assumption is that intent-based relevance is part of search and growing, without knowing the exact share.
Do I still need keywords in my title if COSMO understands intent?
Yes. COSMO adds intent understanding on top of word matching, and Amazon's own paper frames it that way. The primary product-type phrase and defining attribute belong in the title as before. What changes is that use-case and audience statements now also earn relevance.
Can I optimize directly for COSMO?
Not in the sense of a specific tactic. You can give the system accurate material to reason with: filled attributes, explicit use cases and audiences, consistent fields, and a product whose reviews confirm what the listing says. Anything beyond that is guesswork about an unpublished system.
How is COSMO different from Rufus?
COSMO is a knowledge system Amazon uses inside search and recommendations to understand intent. Rufus is the shopper-facing AI assistant that answers questions. They draw on overlapping data about your listing, so the same listing discipline serves both.