sellershortsHow Amazon Handles Vague Searches (COSMO)
Plenty of searches do not name what the shopper actually needs. Amazon built a commonsense knowledge graph to fill in the gap, and it changes what belongs in your bullets.
On this page
Quick Answer
COSMO is Amazon's commonsense knowledge graph for shopping, published at SIGMOD 2024. It exists because many searches need reasoning rather than keyword matching. A search for shoes for pregnant women implies slip resistance matters, even though nobody typed slip resistance. Amazon built COSMO from customer behavior, using search then buy pairs and co-buy pairs, with language models generating candidate relationships that were then filtered. The practical lesson is to state your use case, audience, occasion, setting, and what your product pairs with, because broad searches get resolved by intent and not just by words.
- Many searches imply a requirement the shopper never typed
- COSMO stores everyday relationships between products and reasons to buy
- It was built from real customer behavior, then filtered
- Fix: say the use case, audience, occasion, and pairings out loud
Most listing advice assumes the shopper tells you exactly what they want. Often they do not. They describe a situation and expect the store to work out the rest, the same way you would if you asked a shop assistant for something for a long flight.
Amazon has published research on how it handles that kind of search, and understanding it tells you which sentences in your listing are pulling weight and which are filler.
Reviewed by the SellerShorts editorial bench. SellerShorts runs an AI tool marketplace for Amazon teams.
Our Amazon Listing Optimizer takes an ASIN and returns a full optimized listing, covering title, bullets, description, and backend keywords, in one run. Push live to Seller Central in one click.
The vague search problem
A vague search is one where the words on screen do not fully describe what the shopper needs. The requirement is real, it is just implied.
Amazon's own example in the COSMO research is a search for shoes for pregnant women. The shopper did not say non slip. But a good result almost certainly needs to be easy to put on without bending down and safe to walk in. Keyword matching cannot get there, because the important words are not in the search.
- Situation searches. For a long flight, for a rainy commute, for a small kitchen.
- Audience searches. For toddlers, for seniors, for beginners.
- Occasion searches. For a housewarming, for a wedding guest, for camping.
- Problem searches. For sensitive skin, for back pain, for shedding dogs.
None of those name a product feature. All of them imply several.
What COSMO is
COSMO is a commonsense knowledge graph built by Amazon for shopping, described in a paper presented at SIGMOD 2024. In plain terms, it is a large structured store of everyday facts about products and the reasons people buy them.
The word commonsense is the key part. This is not technical specification data like voltage or thread count. It is the kind of thing a person knows without looking it up, such as a picnic blanket suits an outdoor event, or that a nightlight is for a child's bedroom.
Amazon built it because that knowledge is exactly what is missing when a shopper searches by situation instead of by product.
How Amazon built it
The build process matters, because it tells you the knowledge came from real shoppers rather than from someone's opinion.
- Start from behavior. Amazon used customer behavior signals, including search then buy pairs, where a shopper searched one thing and bought a particular product, and co-buy pairs, where two products get bought together.
- Use language models to phrase the relationship. Those behavior pairs were fed to large language models to generate candidate statements about why the pairing makes sense.
- Filter hard. Candidates were then filtered so only the reliable ones survived, since a generated statement is a hypothesis, not a fact.
That last step is why this is not just a model guessing. The behavior is the evidence, and the language model is the translator.
The kinds of relationships it holds
The relationships COSMO deals in are practical. Here they are with the sort of listing sentence that maps to each one.
| Relationship | Question it answers | Listing sentence that covers it |
|---|---|---|
| Used for | What job does it do? | Keeps drinks cold for up to 24 hours on site. |
| Used by | Who buys it? | Sized for toddlers aged 2 to 4. |
| Occasion | When is it used? | A tidy gift for a housewarming or a new office. |
| Location | Where is it used? | Fits a narrow hallway or a small entryway. |
| Pairs with | What goes with it? | Works with standard 12 inch cast iron pans. |
Notice that none of those sentences are keyword strings. They are normal statements a buyer would find useful, which is the point.
A worked example
Take a search for a lamp for reading in bed. Two listings sell almost the same product.
| Listing A | Listing B | |
|---|---|---|
| Title | LED clip lamp, 3 color modes | LED clip lamp, 3 color modes |
| Use case stated | No | Reading in bed without waking a partner |
| Audience stated | No | Late night readers and students |
| Location stated | No | Clips to a headboard or bunk rail |
| Pairs with stated | No | Fits headboards up to 2 inches thick |
Listing A answers what the product is. Listing B answers what the product is for, who it is for, and where it goes. For a search phrased as a situation, B has something to match on and A does not.
What to put in your copy
The practical advice is short. Say the things a shopper would otherwise have to infer.
- The use case. The actual job the product does, in the words a buyer would use.
- The audience. Who it suits, by age, skill level, body type, or situation.
- The occasion. Gift, travel, holiday, back to school, daily use, whatever fits.
- The setting. Kitchen, car, garden, office, dorm room, gym bag.
- What it pairs with. Products, models, or systems it works alongside.
- The problem it solves. The complaint that makes someone go looking in the first place.
Write them as normal sentences. A list of nouns does not communicate a situation.
Where each piece belongs
| Field | What goes there |
|---|---|
| Title | Product type plus hard specs. Keep it scannable. |
| Bullets | Use case, audience, occasion, and the problem solved. |
| Description | Setting, pairings, and the fuller story. |
| Attribute fields | Target audience, recommended uses, compatibility, where your category offers them. |
| Backend search terms | Phrases you could not fit naturally elsewhere, once each. |
Common mistakes
- Only describing features. Feature lists answer what it is and never answer what it is for.
- Assuming the use case is obvious. It is obvious to you because you sourced the product. It is not on the page.
- Claiming every audience. Great for everyone tells nobody anything and reads as filler.
- Stuffing situation phrases as keywords. Writing for camping for travel for gym for office in a row is not a sentence and does not describe a use.
- Skipping pairings on accessories. If you sell something that goes with another product, name that product and its models.
What nobody can tell you about COSMO
Being straight about the limits here matters, because this is a topic where confident nonsense spreads fast.
- There is no COSMO optimization checklist from Amazon. Amazon published research on how the graph was built. It has not published how a specific listing is scored or selected.
- You cannot see your entries. There is no Seller Central report showing what relationships exist for your ASIN.
- Treat size claims carefully. Specific counts of nodes and relationships circulate in secondary write ups and are not reliable, so we do not quote them.
- The advice is defensible anyway. Stating your use case, audience, occasion, and pairings makes your listing better for human shoppers regardless of what any system does with it.
Conclusion
A lot of shopping searches describe a situation rather than a product. Amazon built COSMO, a commonsense knowledge graph, to bridge that gap, and it was built from real customer behavior rather than from guesses.
You cannot optimize for it directly, but you can stop making people infer things. Say what your product is for, who it suits, when it gets used, where it goes, and what it works with. For a full listing built along those lines, try our Amazon Listing Optimizer. Next reads: what Amazon means by relevant, how to show up in Amazon AI shopping answers, and why attributes matter more than keywords.
References
Frequently asked questions
What is COSMO on Amazon?
COSMO is Amazon's commonsense knowledge graph for shopping, published at SIGMOD 2024. It stores everyday relationships between products and the reasons people buy them, such as what a product is used for, who uses it, and what occasion it suits. It exists to help Amazon understand searches that need a bit of reasoning rather than plain keyword matching.
Why does Amazon need a knowledge graph at all?
Because a lot of searches do not name what the shopper actually needs. Someone searching for shoes for pregnant women is implicitly asking for things like slip resistance and easy fit, but they never typed those words. Keyword matching alone cannot bridge that gap, so Amazon built a way to represent the implied requirement.
How was COSMO built?
From customer behavior, refined with language models. Amazon's paper describes starting from behavior signals such as search then buy pairs and co-buy pairs, using large language models to generate candidate relationships from that behavior, then filtering the candidates so only the reliable ones remain. It is behavior first, with the model used to put the behavior into words.
What kinds of relationships does COSMO cover?
Practical, everyday ones. What a product is used for, who typically uses it, what occasion or event it fits, what location or setting it suits, and what other products it goes with. These are the sort of facts a helpful shop assistant would know without being told.
How do I optimize my listing for COSMO?
You cannot optimize for it directly, and nobody outside Amazon can tell you how it scores a listing. What you can do is make sure your copy states the things it deals in: the use case, the audience, the occasion, the setting, and what your product pairs with. If a shopper has to infer it, so does the system.
Does this mean keywords no longer matter?
No. Broad and vague searches are where intent understanding earns its keep, but plenty of searches are still specific attribute strings. Keywords and attributes get you correctly matched on the literal searches. Covering use case and audience helps you on the vague ones. You want both, not one instead of the other.
Where should I write use case and audience details?
In the bullets and description mainly, plus the relevant attribute fields where your category offers them, such as target audience or recommended uses. Keep the title focused on the product and its hard specs, since the title also has to be scannable in a results grid.
Is COSMO the same as Amazon's AI shopping assistant?
No, they are different pieces of work. COSMO is a knowledge graph published as research in 2024. Amazon's AI shopping assistant is a customer facing product. They both deal with understanding what shoppers want, but they are separate things and Amazon has not published a spec tying your listing to either one.
AI Tools You Can Try
Write copy that covers use case, audience, and occasion.
Drop your ASIN. Get a title, bullets, description, and backend keywords that answer why someone buys, not just what the product is.
Try the Amazon Listing Optimizer →