- ProvenStated in Amazon's paper or on an official Amazon page. That is Amazon reporting on itself; nobody outside Amazon has reproduced it.
- PlausibleFollows from these sources, but Amazon doesn't say it.
- UnprovenNo Amazon source found, even if many blogs repeat it.
What Amazon COSMO is and how it works
A product catalog knows facts: size, color, material. It doesn't know that someone searching for “winter clothes” wants to stay warm, or that “camping” can mean winter camping and therefore winter boots. COSMO is Amazon's attempt to capture this common sense about shopping. Amazon's own researchers describe it in a paper at SIGMOD 2024, a database conference, and in an Amazon Science blog post from May 10, 2024.
According to the paper, COSMO is built in five steps Proven:
- Start with behavior. Two kinds of pairs: a search followed by a purchase in a short session (search-buy), and two products bought together (co-buy).
- Ask a language model why. The model gets the search query and the product title and completes an explanation, for example what the product is capable of that matches the search.
- Filter. Rules and a similarity check remove incomplete answers and answers that only repeat the query or the title. People then rate 30,000 answers for plausibility and typicality.
- Train a smaller model. COSMO-LM, fine-tuned from LLaMA 7B and 13B on these ratings, writes this knowledge at scale.
- Serve it. The results go into a cache and a feature store that other systems read. Frequent searches are loaded in advance, the rest is processed in daily batches, and the model is refreshed daily.
The result is a graph of products, search queries and intentions, linked by 15 relation types such as used for a function, used for an event, used for an audience, used in a location and capable of. The paper also names a limit of its own: because the model is refreshed daily, short events like flash sales are not reflected right away.
What the paper proves: the numbers
Every number in this table is in the paper, so each one is Proven in the sense of this page.
| Topic | Number in the paper | What it means |
|---|---|---|
| Size | 6.3 million nodes, 29 million links, 15 relation types | Counts of the research version, not of the system Amazon runs today |
| Coverage | 18 product domains, from Clothing, Shoes & Jewelry and Home & Kitchen to Electronics and Pet Supplies | Says nothing about which marketplaces use it |
| Input | about 3.1 million co-buy pairs and 1.9 million search-buy pairs | Everything starts from customer behavior |
| Human ratings | 30,000 rated answers; 44.3% of the search-buy and 14.5% of the co-buy explanations rated plausible | Raw language model output is often wrong, which is why COSMO filters it |
| Search relevance, offline | Macro F1 up 60% (28.44 to 45.52) with frozen models, up 28% (57.49 to 73.48) with trained models | Measured on Amazon's public shopping queries data from KDD Cup 2022: a benchmark, not a live ranking test |
| Recommendations, offline | Hits@10 up 4.05% (clothing) and 5.82% (electronics); one of six measures fell by 3.19% | One week of session data in two categories |
| Live test | Search navigation, several months, about 10% of US traffic: product sales up 0.7% (relative), navigation engagement up 8% | The only live result with numbers. The authors put it at hundreds of millions of dollars a year; that is their own extrapolation |
| Markets | Live test in the US; offline data from the US, Canada, the UK and India | Nothing on Europe |
Two things to keep in mind. First, the live result comes from search navigation: the refinement suggestions that lead from a broad search like “camping” to narrower ones like a camping air mattress for four people. The paper calls it a single, relatively minor feature on the search page. It is not the order of the search results. Second, the paper describes the state of 2024. What Amazon runs in 2026 it hasn't published.
What the paper does not prove
Most COSMO articles go further than the paper. These are the claims that come up again and again, checked against Amazon's sources:
| Claim in seller blogs | Evidence | What the sources say |
|---|---|---|
| “COSMO replaced A9 and A10 as Amazon's ranking algorithm.” | Unproven | The paper calls COSMO a system that supplies knowledge to search applications. It doesn't mention A9 or A10 and names nothing it replaces. |
| “COSMO ranks listings by intent instead of keywords.” | Unproven | The live result comes from search navigation, not from the order of results. The paper names no ranking factors and no weights. |
| “Write intent phrases and COSMO shows you to more buyers.” | Unproven | Relations start from customer behavior. No Amazon source says that a phrase in your copy creates a relation or improves your rank. |
| “Backend keywords no longer matter because of COSMO.” | Unproven | Nothing in the paper. Amazon's seller blog from September 2025 still names back-end search terms as a place for keywords. |
| “COSMO powers Rufus.” | Unproven | Neither Amazon's Rufus announcement nor its article on the technology behind Rufus mentions COSMO. |
| “COSMO weights conversion rate, reviews or price at a set percentage.” | Unproven | Amazon publishes no weights, and the paper contains none. |
| “COSMO is live in every marketplace.” | Unproven | The live test ran on about 10% of US traffic. Offline data came from the US, Canada, the UK and India. |
Plausible COSMO or a successor still feeds Amazon search in 2026, possibly in more places than navigation: the paper's introduction speaks of “various Amazon search applications”. Only navigation comes with numbers.
The paper is also clear that your listing text is part of the input, just not the starting point Proven. The prompt it shows contains the search query and the product title, and in the offline relevance test the model read title, description and attributes together with the COSMO knowledge. So “COSMO ignores your listing” is as wrong as “COSMO ranks your listing”.
COSMO, A9 and A10: what the names mean
- A9 Proven: A9.com was an Amazon subsidiary that researched and built search technology. Amazon announced its web search site in a press release in September 2004. Sellers still use “A9” as a name for Amazon's product search.
- A10 Unproven: there is no Amazon source for “A10”. The name comes from service providers.
- How Amazon describes its search Proven: in its SIGIR 2016 paper “Amazon Search: The Joy of Ranking Products”, Amazon describes machine-learned ranking models for categories instead of hand-tuned rules, plus language processing to match queries and products. That is a set of methods, not one versioned algorithm. The paper is from 2016 and says nothing about today's setup.
- Where COSMO fits Proven: according to its paper, COSMO's knowledge goes into a feature store that other systems read, and the paper tests it in search relevance, recommendations and navigation. No source describes it as one ranking formula, so “the COSMO algorithm” is a seller term, not Amazon's.
How retrieval and ordering work, and how to check at which stage your listing fails, is in the Academy lesson Amazon ranking explained. More claims that don't hold up are collected in 45 listing myths.
COSMO, Rufus and Alexa for Shopping
Rufus is Amazon's AI shopping assistant. Amazon announced it on February 1, 2024. On the product page it answers from the listing details, customer reviews and community Q&A, and it also uses information from across the web Proven.
On May 13, 2026, Amazon brought Rufus and Alexa+ together as Alexa for Shopping. In its update of November 13, 2026, Amazon lists 13 countries with full availability, among them the US, the UK, Canada, Australia and Germany. According to Amazon, Rufus helped over 300 million customers in 2025 Proven.
Amazon's article on the technology behind Rufus (October 2024) describes how the assistant first collects evidence from the catalog, customer reviews and community questions and answers, then writes the answer, and learns from customer feedback. It does not mention COSMO Proven.
COSMO and Rufus share a goal: understanding what a buyer means, not just which words they type. That they share data or models is Unproven. Amazon hasn't said so, and a page that claims it should show you where.
What Amazon does tell sellers Proven: in its seller blog of April 29, 2026, Amazon advises answering frequently asked questions in the product description so Rufus can recommend the product, and naming use cases and attributes. Amazon gives no advice for COSMO. This advice is for Rufus, and it points the same way. How to write for the assistant is in the Academy lesson Rufus and COSMO, how to structure the description in the Amazon product description guide.
What sellers can do with it
None of this needs a COSMO trick. It needs a listing that says plainly what the product is, who it is for and what it is used for. The table shows how well each step is backed.
| What you do | Why | Evidence |
|---|---|---|
| Name the use, the occasion and the audience in words buyers type: “for trail running on wet ground”, not “for every adventure”. | COSMO's relations are of exactly this kind: used for a function, an event, an audience, a location. Amazon's seller blog advises naming use cases. | Plausible that it changes your rank: Unproven |
| Answer the questions buyers ask in the description: size, fit, care, what's in the box, what the product is not made for. | Amazon advises it for Rufus, and Rufus answers from listing details, reviews and Q&A. | Proven |
| Fill in the attribute fields in Seller Central: material, dimensions, color, compatibility. | In the paper, the last step of COSMO's navigation narrows the results by attributes. | Plausible |
| Keep the title clear about what the product is. | In the paper's prompt, the model sees the search query and the product title. | Plausible |
| Leave out keyword lists and phrases written only for an algorithm. | Amazon writes that keyword stuffing can create a poor customer experience and hurt your rankings. | Proven |
| After each change, check whether the right buyers find and buy the product. | No source tells you what COSMO does with your listing. Your own numbers do. | Your data |
An example with running shoes. “Breathable running shoes, high quality” gives buyers and models nothing to work with. “Trail running shoes for wet ground, for wide feet, 10 mm drop, 9.5 oz in US size 9” names a use, an audience and facts a buyer can check. The second version is better for buyers whether COSMO reads it or not. Which attribute fields count for search and filters is in the Academy lesson Attributes, browse nodes and filters.
Listimo checks the part you control: your copy. The free Rufus checker shows in your browser which of nine common question types your title, bullet points and description answer; nothing is sent. The Rufus check in the toolbox goes further: by ASIN, an AI asks the twelve questions buyers have about your product, looks for each answer in the listing and writes a ready-made sentence for every gap. The preview is free with a free account, up to three times a day. The free listing checker scores your live listing from 0 to 100 across seven areas and shows which search terms it is missing. None of these tools can see what COSMO or Rufus does inside Amazon.
Frequently asked questions
What is Amazon COSMO?
COSMO is a knowledge graph of shopping intent that Amazon researchers described in a paper at SIGMOD 2024. It is built from what customers search and buy, links products to purposes like “used for winter camping” and covers 18 product domains. The paper reports its use in search navigation in the US.
Is COSMO Amazon's new ranking algorithm?
No Amazon source says so. The paper describes a knowledge graph that supplies intent knowledge to search applications, and its only live result with numbers comes from search navigation. It names no ranking factors and no weights, and it doesn't say it replaced A9. “A10” and “COSMO algorithm” are names from seller blogs, not from Amazon.
Does COSMO read my listing?
As context, yes. COSMO starts from customer behavior: a search followed by a purchase, or two products bought together. In the prompt shown in the paper, the language model sees the search query and the product title and explains the purchase. The paper does not say that text in your listing alone creates a relation or improves your rank.
Is Rufus or Alexa for Shopping built on COSMO?
Amazon doesn't say. Neither the Rufus announcement nor Amazon's article on the technology behind Rufus mentions COSMO. On the product page, Rufus answers from listing details, customer reviews and community questions and answers. Since May 13, 2026, Rufus has been part of Alexa for Shopping, first in the US; in its update of November 13, 2026, Amazon lists 13 countries with full availability, the UK and Germany among them.
How do I optimize my listing for Rufus and COSMO?
Write for buyers, not for a model: say what the product is, who it is for and what it is used for, answer the questions buyers ask in the description and fill in the attribute fields. Amazon itself advises answering frequent questions in the description so Rufus can recommend the product. The free Rufus checker shows which of nine common question types your copy answers.
More free tools: check which Amazon reviews you can get removed, Rufus checker for the questions buyers ask Amazon's AI, Amazon FBA calculator for the UK and Europe, Amazon FBA calculator for amazon.com, FNSKU label generator, UPC code generator with how to get real UPCs, EAN and UPC check digit calculator, Amazon image requirements checker, A+ content image sizes for every module, AI image label checker for Amazon, GPSR checklist for Amazon, Amazon Brand Registry checklist with trademark fees, unit price calculator for the German Grundpreis, ACOS and ROAS calculator for your ads, search terms byte counter for backend keywords and titles, product description formatter with line breaks, listing checker with a score from 0 to 100. All in one place: free Amazon seller tools. Numbers from over 1,200 checked listings: Amazon Listing Study 2026.
Sources
All sources checked on November 21, 2026. The numbers come from Amazon's own paper and have not been reproduced outside Amazon. Listimo is not affiliated with Amazon; this page explains public sources and is not an Amazon statement.
- Amazon Science: COSMO: A large-scale e-commerce common sense knowledge generation and serving system at Amazon (SIGMOD 2024, with the full paper as PDF)
- ACM Digital Library: the same paper in the SIGMOD-Companion '24 proceedings (DOI 10.1145/3626246.3653398)
- Amazon Science blog, May 10, 2024: Building commonsense knowledge graphs to aid product recommendation
- Amazon: Amazon announces Rufus, February 1, 2024 (with the note on the renaming on May 13, 2026)
- Amazon: Meet Alexa for Shopping, May 13, 2026, updated November 13, 2026
- Amazon Science, October 4, 2024: The technology behind Amazon's GenAI-powered shopping assistant, Rufus
- Amazon seller blog, April 29, 2026: How to boost Amazon listings to drive more sales
- Amazon seller blog, September 3, 2025: Amazon SEO, 7 ways to improve your product's search rankings
- Amazon press release, September 2004: A9.com launches new web site to make Internet search more effective
- Amazon Science: Amazon Search: The joy of ranking products (SIGIR 2016)