Amazon COSMO vs Keyword Search (A9/A10): What Is Confirmed and What Sellers Should Do

COSMO is a commonsense knowledge system that Amazon's researchers described in a paper at the SIGMOD 2024 conference; it uses a large language model to learn why shoppers buy things, such as which product suits which audience or use. Amazon has not published that COSMO sets rankings on Amazon.in, and it has never announced an "A10" algorithm. As of November 2026, anything more specific is the opinion of seller blogs, not Amazon documentation.

What is Amazon COSMO?

Amazon's science blog, published on 10 May 2024, describes COSMO as a framework that uses large language models (LLMs) to find commonsense relationships in customer behaviour and organise them into a knowledge graph. The graph links products to human contexts: what a product is used for, who uses it and where. Amazon's own example links slip-resistant shoes to pregnant women through a "used for audience" relationship.

The data it learns from is customer behaviour: queries paired with purchases, and products bought together in the same shopping session. The LLM proposes explanations. Heuristics and a classifier trained on human ratings then filter out weak ones.

The research paper is titled "COSMO: A Large-Scale E-commerce Common Sense Knowledge Generation and Serving System at Amazon". It was presented in the industry track at SIGMOD 2024.

Is COSMO the Amazon.in ranking algorithm?

Not confirmed. In Amazon's blog, the tests that showed improvement were relevance-ranking experiments on a public research data set from KDD Cup 2022, not a live Amazon store. The page reports that a model given COSMO knowledge beat the best baseline by about 28% in macro F1 and 22% in micro F1 when the encoders were fine-tuned. The same page says the graph is meant to help the recommendation engine make commonsense inferences. It does not describe how, or whether, it runs on Amazon.in.

So the honest reading is: COSMO shows the direction Amazon's research is heading, towards understanding intent instead of matching words. It is not a published list of ranking factors.

What about A9 and A10?

"A9" was the name of Amazon's search subsidiary and has long been used by sellers as shorthand for Amazon search. "A10" is different. We found no Amazon announcement, help page or documentation that says A10 replaced A9. Seller blogs and agencies use the term freely, and some write as if it were official. A Helium 10 podcast episode is even titled "No Such Thing As The Amazon A10 Algorithm", arguing the name was adopted by sellers.

For you, this means:

• Do not pay for an "A10 audit" that lists ranking factors with weights. Nobody outside Amazon has those.
• Do not rebuild listings around claims such as "external traffic boosts A10 rank" unless the person shows you an Amazon source.
• Do keep a record of what you change and what happened. Your own data is the only reliable evidence.

Keyword search vs semantic search: what changes

The table separates what Amazon has documented from what is common seller advice.

Both still matter. A shopper who types a product name still needs your title to contain it. Intent-based search adds a second job for the listing: say plainly who it is for.

What can an Indian seller do now?

These steps are low risk because they improve the listing for shoppers whether or not COSMO matters to ranking.

1. Fill every attribute field you can answer truthfully. Material, size, compatibility, age group, occasion. Structured data is the clearest way to state a use.
2. Write bullets as uses, then features. "Fits 15-inch laptops, for daily commuting" tells a buyer and a machine more than "premium quality bag".
3. Name the audience in the copy where it is true: "for new runners", "for toddlers aged 2 to 4".
4. Use A+ Content to show the use. Images with real alt text and short comparison charts help. See the glossary entry on A+ Content.
5. Answer real customer questions in the Q&A and in your copy, based on what buyers ask you. Do not invent questions.
6. Change in batches and measure. Pick five ASINs, change them, and compare sessions and conversion against five untouched ASINs for a few weeks.

Our full process is in the Amazon listing optimisation guide. For the shopping assistant side, read optimising Amazon India listings for Rufus.

How to test a listing change without fooling yourself

Seller blogs often claim a result after one edit. A fair test is slower but believable.

1. Choose ASINs with steady sales over the last four weeks, so you are not measuring a festival spike.
2. Change only the copy and attributes. Do not change price, images and ads in the same week.
3. Record sessions, unit session percentage (conversion) and organic orders before the change.
4. Compare against similar ASINs you left alone. If both groups moved together, the change was the season, not your edit.
5. Keep the edit if conversion or organic orders improve across a few weeks, and roll it out to the next group.

If you have Brand Registry, Amazon's Manage Your Experiments tool can split-test content. Amazon runs it inside Seller Central, so check the current eligibility rules there.

How is this different from Rufus?

Rufus is the shopping assistant that customers see. Amazon says it answers using Amazon.in and information from across the web, and on product pages it draws on listing details, customer reviews and community Q&As. COSMO is a research system for product knowledge. Some seller blogs say COSMO powers Rufus. We could not find Amazon saying so, so treat that as unconfirmed.

A related note: on 13 May 2026 Amazon renamed Rufus to Alexa for Shopping in the United States. We could not confirm whether the Amazon.in assistant has been renamed or when it might be.

Does this change how I run Amazon ads?

Ads still work on your bids and targeting. Better listing relevance can improve click-through and conversion, which feed into ad efficiency, but Amazon does not publish a formula. Keep your Amazon ads structure and test listing changes separately from bid changes, or you will not know which one worked.

FAQ

Is COSMO the same as Rufus? No. Rufus is the shopping assistant customers talk to. COSMO is a commonsense knowledge system described in Amazon's research papers. Amazon's published material does not say COSMO runs Rufus, and claims that it does come from seller blogs, not Amazon.

Did Amazon replace A9 with A10? We found no Amazon statement saying so. "A10" is a name used by sellers and agencies. Amazon has not published a document with that name, and ranking-factor lists attributed to it are unconfirmed.

Do I need to rewrite all my Indian listings for COSMO? No. Amazon has not told sellers that COSMO affects ranking. Improve the listings that sell most, using clear use-case copy and complete attributes, and measure the result. That work helps shoppers regardless of how the algorithm changes.

Where did COSMO get its data from? Amazon's blog says it uses queries paired with purchases and items bought together in the same session. An LLM generates explanations, and automated filters plus human-rated examples remove weak ones.

Does keyword research still matter? Yes. Shoppers still type product names and attributes, and your title and backend terms still need to contain them. Semantic search adds a second requirement: say who the product is for and what it does.

Can a seller buy COSMO optimisation? You can buy listing work that improves clarity and completeness. You cannot buy access to a COSMO score, because none is published. Be careful with any service that promises ranking gains from COSMO.

Next step

If you want listings rewritten around use cases and tested properly on a sample of ASINs, Blooprint's Amazon account management team does this for Amazon.in sellers. This page was checked in November 2026, and Amazon has not announced a change to search ranking since the sources above.