What Is Amazon COSMO? The Knowledge Graph and What to Change
By Gulfam Haidar, Founder, GrowEventure. Last updated September 2026.
Amazon Cosmo: quick summary
In short, this guide covers amazon cosmo in plain terms. Below you will find the key points on amazon cosmo, a breakdown you can act on today, and clear answers to the most common questions about amazon cosmo.
Amazon COSMO is a common-sense knowledge graph, introduced in an ACM SIGMOD 2024 paper, that infers what products are used for, who uses them, and what they pair with. It was built from about 1.87 million search-buy pairs and 3.14 million co-purchase instances, forming roughly 6.3 million nodes and 29 million edges across 18 categories. It learns from aggregate shopper behavior, not from your listing text directly. That last point is the one most guides miss, and it changes how you optimize.
COSMO in plain English
A knowledge graph is a web of facts. It stores things as points, called nodes, and the connections between them, called edges. COSMO is Amazon’s knowledge graph built to give its search common sense about products. It knows, for example, that a headlamp is used for camping, that a sippy cup is for toddlers, and that batteries pair with a flashlight.
Regular search matches words. COSMO matches meaning. When a shopper types “camping trip essentials,” the words alone will not surface half the right products, because few listings say “camping trip essentials.” COSMO can connect the query to products it has learned are used for camping, even when the exact phrase never appears in the listing.
The research behind COSMO
Amazon researchers published COSMO in a paper at ACM SIGMOD 2024, a top database and data-systems conference. The system was not built by hand. Amazon mined it from real shopper behavior at scale.
Per the paper, COSMO drew on roughly 1.87 million search-buy pairs, which are cases where a shopper searched a term and then bought a product, and about 3.14 million co-purchase instances, which are cases where two products were bought together. From that behavior, Amazon extracted common-sense relationships and checked them for quality before adding them to the graph.
The build process is worth understanding, because it explains why quality control matters. Amazon used large language models to generate candidate relationships from the behavior data, then filtered those candidates through both human review and automatic checks. Weak or nonsensical links were dropped. Only the relationships that passed made it into the graph. So COSMO is not a raw dump of everything shoppers do. It is a filtered set of connections judged to reflect real common sense. That filtering is why the graph can say a headlamp is used for camping with confidence, rather than linking every product to every search that happened to occur nearby.
The scale of the graph
The numbers describe a large structure. Per the ACM SIGMOD 2024 paper, COSMO holds roughly 6.3 million nodes and about 29 million edges, spread across 18 product categories, using 15 relation types.
Nodes are the things: products, uses, audiences, and attributes. Edges are the labeled links between them, like “used for” or “used with.” So a single product node can connect through many edges to many use nodes and audience nodes. That is how one flashlight can be linked to camping, power outages, car emergencies, and dog walks at night, all at once.
The scale is what makes it powerful. With millions of nodes and tens of millions of edges, COSMO holds a dense web of everyday connections that no team could write by hand. Across 18 categories, it covers a wide slice of what Amazon sells, from home goods to electronics to outdoor gear. For a seller, the takeaway is simple. Your product is already a node in a graph that connects it to uses and audiences based on how shoppers behave. The question is whether those connections match the customers you actually want. You steer them by naming your true uses clearly and earning sales from the shoppers who search them.
(alt: “Amazon COSMO knowledge graph scale statistics.”)
The relation types that matter
Of the 15 relation types, four map cleanly to things you control on a listing.
- Used_For_Func: what the product does or the task it completes. A blender is used for making smoothies.
- Used_For_Audience: who the product is for. A booster seat is used for toddlers.
- Used_With: what the product pairs with. A grill is used with charcoal.
- Capable_Of: what the product can do. A jacket is capable of blocking wind and rain.
Each of these is a signal you can state in plain words on your listing, so shoppers arrive for the right need and buy, which is the behavior COSMO learns from.
The other relation types cover things like where a product is used, what problem it solves, and what event it suits, but these four are the ones most sellers can act on today. Take a stainless steel travel mug as an example. Its Used_For_Func links might include keeping coffee hot and preventing spills in a car. Its Used_For_Audience links might include commuters and office workers. Its Used_With links might include a car cup holder or a single-serve coffee maker. Its Capable_Of links might include maintaining temperature for six hours and fitting standard cup holders. Write your listing so each of those true connections appears in plain language, and you make it easy for both shoppers and Amazon’s systems to place your product correctly.
How COSMO learns (and why text isn’t enough)
Here is the insight that changes your strategy. COSMO learns from aggregate search-and-buy behavior, not directly from your keywords. Writing “used for camping” in your title does not force COSMO to file you under camping. What teaches COSMO is a pattern of shoppers who search camping terms, click your listing, and buy.
So you influence COSMO indirectly. Your listing text still matters, because it gets you found and drives the click. But the sale is what confirms the relationship. Relevant conversions are the currency. Name your use cases so the right shoppers find you, then earn the purchase so the behavior backs up the claim. This is the same principle behind how Amazon ranks products across A9, A10, and COSMO.
Does COSMO actually move sales?
Amazon reported results from live tests of COSMO. Per the ACM SIGMOD 2024 paper, applying COSMO produced roughly a 0.7 percent sales lift in online experiments. Treat that as directional. It is a modest number, but at Amazon’s scale it is meaningful, and more importantly it tells you COSMO is not a lab experiment. It is in use, shaping what shoppers see.
Put that 0.7 percent in perspective. Amazon moves hundreds of billions of dollars in sales a year. A lift of even a fraction of a percent across that base is enormous in absolute terms. For an individual seller, the point is not the headline number. It is that Amazon judged COSMO good enough to run in production. When a company invests in a system and reports a real lift, that system is not going away. It shapes the searches your buyers make right now, whether or not you have optimized for it.
What COSMO is not
It helps to be clear about the limits. COSMO is not a dial you can turn. You cannot log into Seller Central and edit the graph. You cannot pay to add a relationship. It is also not a replacement for the ranking signals that still decide order, like conversion rate and stock. And it is not a reason to write false use cases into your listing hoping to trick it, since the system learns from what shoppers actually do, and bounces teach it the opposite. Understanding what COSMO cannot be pushed to do keeps you focused on the one lever that works: clear, true content plus relevant traffic that converts.
What to change on your listing for COSMO
You cannot edit the graph. You can feed it clear signals and relevant traffic. Do this:
- Name explicit use cases in your title and bullets, in plain words shoppers actually search.
- Name your audience where it is true, like “for large-breed dogs” or “for first-time buyers.”
- Name complementary products in your description and A+ content, so pairings are clear.
- Drive relevant traffic that converts, since the sale is what teaches COSMO the match is real.
Do not stuff false uses. If shoppers arrive for a use your product does not serve, they bounce, and that teaches COSMO the opposite of what you want.
A short before-and-after shows the difference. A weak description reads: “Our premium mug is the perfect choice for anyone who loves a great beverage experience.” It names no use, no audience, no pairing. A COSMO-aware version reads: “Commuters keep coffee hot for the full drive with a leak-proof lid that fits standard car cup holders. It also works for tea at a home office desk and cold drinks on a summer hike.” Now the function, the audience, and the pairing are all stated in words shoppers search and Amazon can connect. The second version does not just describe the product. It tells the systems what the product is for.
Mapping relation types to listing sections
| Relation type | What it captures | Where to signal it on your listing |
|---|---|---|
| Used_For_Func | The task the product completes | Title lead and first bullet |
| Used_For_Audience | Who the product is for | Bullets and A+ content |
| Used_With | Products it pairs with | Description and A+ comparison module |
| Capable_Of | What the product can do | Bullets and image spec callouts |
(alt: “Mapping COSMO relation types to Amazon listing sections.”)
COSMO, Rufus, and the future of Amazon search
COSMO and Rufus work as a pair. COSMO supplies the common-sense connections, and Rufus, Amazon’s AI shopping assistant, uses that context to answer buyer questions and recommend products. When a shopper asks Rufus “what do I need for a first camping trip,” the answer leans on the kind of use-case knowledge COSMO holds. So the same work pays off twice: it feeds the knowledge graph and it gives Rufus clear material. Our Rufus optimization playbook turns this into step-by-step actions.
The direction is clear. Amazon is moving from pure keyword matching toward understanding intent. A shopper no longer has to know the exact product name. They can describe a need, a person, or an occasion, and expect the right products to surface. For sellers, that raises the value of honesty and clarity. A listing that plainly states what the product does, who it serves, and what it pairs with is easier for the knowledge graph to place and for the assistant to recommend. The listings that win the next few years will read less like keyword lists and more like clear, truthful answers to real needs.
Frequently asked questions
What is Amazon COSMO?
COSMO is a common-sense knowledge graph that infers what products are used for, who uses them, and what they pair with.
When was COSMO introduced?
Amazon researchers introduced it in a paper at ACM SIGMOD 2024.
How big is COSMO?
Per the paper, roughly 6.3 million nodes, about 29 million edges, across 18 categories, using 15 relation types.
Does COSMO read my listing text?
Not directly. It learns from aggregate shopper behavior, so your text gets you found while the sale confirms the match.
What relation types does COSMO use?
Fifteen in total, including Used_For_Func, Used_For_Audience, Used_With, and Capable_Of.
Did COSMO increase sales?
Amazon reported roughly a 0.7 percent sales lift in live tests. Treat the figure as directional.
How do I optimize for COSMO?
Signal real use cases and audiences in your content, then drive relevant traffic that converts.
Want the connections mapped for your catalog? Sharpen your term list with Amazon keyword research, see the full range of our Amazon SEO services, or book a free audit.
Sources: COSMO paper, ACM SIGMOD 2024 (primary source for all statistics); Amazon Science blog coverage of COSMO where available.
Rather have experts handle this?
GrowEventure is a marketplace account-management agency that can optimize and rank your listings for your brand. Book a free store audit and we will point out the quickest wins.
