CoSMo is Amazon's commonsense knowledge graph. It learns why people buy products, not just what products are, by mapping each product to the situations, people and needs it serves: who it is for, what it does, where and when it is used, what it is used with. Amazon's researchers define 15 types of these relations. Amazon does not show sellers a CoSMo score. What you control is whether your listing states those relations in plain words, so your product matches the intents CoSMo has learned from shoppers. Repeating keywords does not help, because CoSMo is not counting them.
Why this matters
Unlike most of what sellers are told about Amazon's AI, CoSMo has a public paper behind it. Amazon researchers presented it at SIGMOD 2024, the ACM data management conference. In it they report an online A/B test on roughly 10% of Amazon's US traffic, run over several months, where one CoSMo feature in search navigation lifted product sales by 0.7% relative to the control. The authors put that at hundreds of millions of dollars a year. Navigation engagement rose 8%.
That was one small feature on the search page. Extended to all navigation traffic, the authors project a revenue increase in the billions.
What CoSMo is, and what it is not
CoSMo stands for Common Sense Knowledge Generation and Serving System. Amazon writes it COSMO. It is a knowledge graph: a very large map of statements such as "slip-resistant shoes are used by pregnant women" or "a camera case and a screen protector are capable of protecting a camera".
Three things it is often confused with:
| What it is | The question it asks | Documented by Amazon? | |
|---|---|---|---|
| A9 / "A10" | Amazon's search ranking. "A10" is a seller nickname, never an official Amazon name. | Does this listing match the words typed? | Partly, through Seller Central guidance |
| CoSMo | A knowledge graph of why people buy things | What need, person or situation does this product serve? | Yes. SIGMOD 2024 paper, Amazon Science blog |
| Alexa for Shopping (formerly Rufus) | The AI assistant in the search bar and on product pages | What is the best answer to this shopper's question? | Partly, through launch announcements |
CoSMo is not a content score. The paper describes building and serving a knowledge graph. It does not describe grading individual listings, a 0 to 100 scale, or a penalty for low scores. Guides that describe it that way are describing their own models, not Amazon's.
CoSMo does not replace search ranking. The paper deploys it alongside search: to understand broad queries, to suggest refinements, and in research experiments on relevance and recommendations.
How CoSMo learns why people buy
The paper describes a pipeline built on two shopper behaviors.
→ Search-buy: a shopper searches a query, then buys a product within a short session.
→ Co-buy: a shopper buys two products together.
A large language model is asked to explain each behavior. Why would someone who searched "winter coat" buy a long-sleeve puffer coat? Because it is capable of providing high-level warmth. Why buy a camera case with a screen protector? Because together they protect the camera.
Human annotators then judge a sample of those explanations for plausibility and typicality. Generic ones are thrown out. The paper names two it rejects: "customers bought them together because they like them", and "customers bought an Apple watch because it is a type of watch". Neither tells a system anything useful.
The approved examples train a smaller model, COSMO-LM, on about 30,000 annotated instructions. That model generates millions of relations across 18 major Amazon categories, from Clothing, Shoes & Jewelry to Pet Supplies. It is refreshed daily from new behavior logs.
Two consequences for sellers follow directly.
→ The relations come from shopper behavior, not from your listing. CoSMo learns that shoppers who want to "walk the dog" buy certain things. Your listing's job is to make it obvious your product is one of those things.
→ Shopper language wins. The knowledge is built from the words people search and the reasons behind their purchases. A listing written in brand or regulatory language describes the product in terms CoSMo did not learn from.
The 15 relations, in plain language
Table 2 of the paper lists 15 relation types. Each is a question a shopper is implicitly asking. The examples are the paper's own.
| Relation | The question behind it | Example from the paper |
|---|---|---|
| used_for_func | What job does it do? | dry face |
| used_for_eve | What activity or event is it for? | walk the dog |
| used_for_aud | Which audience is it made for? | daycare worker |
| capable_of | What can it do beyond its main job? | hold snacks |
| used_to | What task does it help complete? | build a fence |
| used_as | What else can it serve as? | smart watch |
| is_a | What kind of thing is it? | normal suit |
| used_on | When, or in what season? | late winter |
| used_in_loc | Where is it used? | bedroom |
| used_in_body | Which part of the body? | sensitive skin |
| used_with | What is it used together with? | surface cover |
| used_by | Who uses it? | cat owner |
| xInterested_in | What is the buyer interested in? | herbal medicine |
| xIs_a | Who is the buyer? | pregnant women |
| xWant | What does the buyer want to do? | play tennis |
The last three start with an "x" because they describe the buyer rather than the product. Add used_for_aud and used_by, and five of the 15 relations are about people. Those are the ones listings most often leave blank.
Not every relation applies to every product. A tumbler has no body part. A fence post has no season. The aim is to answer the relations that do apply, explicitly.
A real listing, relation by relation
Here is a worked reading of one public listing: the Stanley Quencher ProTour Flipstraw Tumbler, 30 oz, ASIN B0H1YLJZFF, as it appeared on amazon.com on 23 September 2026. The source is its title, highlights line, five bullets and description.
This is our reading of the text against the paper's relations. Amazon does not publish how CoSMo sees any listing. You can repeat the exercise on your own product.
| Relation | What the listing says | Stated? |
|---|---|---|
| used_for_func | "Leakproof hydration", keeps drinks ice cold | Yes |
| used_for_eve | "at the gym, on a hike, or at your desk" | Yes |
| used_in_loc | Car cup holder, gym bag, desk | Yes |
| is_a | Tumbler, insulated stainless steel cup | Yes |
| capable_of | Fits most car cup holders, dishwasher safe | Yes |
| xInterested_in | "Recycled stainless steel", "sustainable" | Yes |
| xWant | "Your hydration routine" | Partly |
| used_on | "All day". No season, no time of year. | Partly |
| used_as | Nothing. Reviews call it a thermos and a water bottle. | No |
| used_with | Nothing. No mention of ice, lids or straws sold separately. | No |
| used_for_aud | Nothing | No |
| used_by | Nothing | No |
| xIs_a | Nothing | No |
| used_to, used_in_body | Not meaningful for a tumbler | n/a |
Six relations clearly stated. Two partly. Five missing. Two that do not apply.
The pattern is typical of a well-written listing. Function, activity and place are strong, because copywriters naturally describe what a product does. Every relation about the person is empty. The listing never says who this tumbler is for.
The reviews say it for them. On the same page, buyers describe it as a commuter's car cup, a gift, and a cup for the whole family. Amazon's own review summary mentions one buyer calling it a perfect toddler-sized cup. Each of those is an audience relation the brand could state in its own words and has not.
The three gaps to check first
The Stanley reading is one listing, not a study. Still, there is a structural reason to expect the same three gaps elsewhere: copy is written about the product, and three families of relation are about something else.
1. The buyer. used_for_aud, used_by and xIs_a all ask who the product is for. Naming an audience can feel like excluding other buyers. It does not. "Designed for nurses on 12-hour shifts" does not stop a teacher buying it. It gives the knowledge graph a person to connect it to.
2. Time. used_on asks when. Season, occasion, time of day. "For summer road trips" or "a back-to-school gift" are relations. "All day" is barely one.
3. What it goes with. used_with asks about complements. CoSMo learns these from co-buy behavior, so shoppers are already connecting your product to others. Stating the pairing in your copy ("fits standard replacement straws", "pairs with our 20 oz lid") reinforces a connection the graph may already hold.
Why keyword stuffing does not help
Keyword stuffing was built for a system that counts matching words. CoSMo does not count words. It connects products to intents it learned from behavior.
A listing that repeats "insulated water bottle" six times gives CoSMo one fact six times. A listing that says who it is for, where it goes and what it pairs with gives CoSMo several different relations to match.
Here is the difference in a single description line for a generic insulated bottle, as an illustration:
→ Before: "Premium stainless steel water bottle 32 oz BPA free insulated water bottle for gym travel hiking outdoor sports"
→ After: "Keeps drinks cold through a full workday or a day on the trail. Built for commuters and hikers. Fits standard car cup holders and most bike bottle cages."
The keywords are still there: stainless, 32 oz, insulated, gym, hiking. But the second version states an audience (commuters, hikers), two contexts (workday, trail), a location (car) and a complement (bike bottle cages). The first states a product type and a list.
Amazon's title rules point the same way. Since 27 July 2026, titles are capped at 75 characters in every category except media. There is no room left to stack variants, so the relations have to live in highlights, bullets, the description and attribute fields.
Can you see your CoSMo score?
No. Amazon does not show sellers a CoSMo score, and its paper does not describe one.
What you can see is the direction of travel. Search refinements that group products by need rather than by category, which is what the paper documents. An assistant that answers "what should I get for a long hike" rather than matching the word "hike". Both need to know what your product is for, and only your listing can say it in your words.
Tools that report a CoSMo score, Cosmy included, are measuring something specific: how well your listing's text covers the 15 relations. That is a useful proxy, because it is the part you control. It is not a number Amazon exposes, and any tool implying otherwise is overstating what it can see.
How CoSMo and Alexa for Shopping relate
Amazon has not published how the two connect. Guides that describe a fixed pipeline, where one system filters and the other recommends, are describing a theory, not a documented architecture.
What is documented is that both were built to close the same gap: between how products are described and how people describe what they need. CoSMo does it at the level of the whole catalog. Alexa for Shopping does it one question at a time, reading your listing to answer a shopper directly.
They reward the same work. A listing that states its audience, its contexts and its complements gives CoSMo relations to match, and gives the assistant sentences to quote. The field-by-field side of that, from the 75-character title to item highlights and attributes, is covered in the Alexa for Shopping optimization guide.
What to do next
- Write two sentences about one ASIN: who it is for, and what problem it solves. Use the words a customer would use.
- Search your listing for both. Title, highlights, bullets, description, attribute fields. Usually neither appears.
- Walk the 15 relations. Mark each stated, partly stated, missing, or not applicable. The buyer relations, used_for_aud, used_by and xIs_a, are the likeliest gaps.
- Close the gaps in plain words. Name the audience, the season or occasion, and what it is used with. Do not imply them.
- Read your reviews for relations you missed. Every "I use this for..." and "bought it for my..." is a relation your customers stated and your copy did not.
Where to start
Walking 15 relations by hand works for one ASIN. It does not work for a catalog. A free Cosmy account at cosmy.ai scores a listing from 0 to 100 on how well its content covers the 15 CoSMo dimensions, shows which ones are missing, and tracks the score each time you re-run the ASIN. It is our measure of your content, built on the relations Amazon published, not a number Amazon shows.
