Amazon COSMO Algorithm Explained: How It Works and What Sellers Should Do (2026)

Short version: COSMO is a knowledge graph of shopping intent that Amazon built with a language model and described in a 2024 research paper. It is not a replacement for A9/A10. Amazon has published that it runs in search navigation. It has not published that COSMO sets your rank. The seller takeaway does not depend on that gap: write the use, the audience, and the occasion into the listing in plain words.

"COSMO" shows up in every Amazon SEO thread now, usually with claims no one sources. This guide sticks to what Amazon's own paper says, marks what is unconfirmed, and ends with a checklist you can run on one listing today. OptimalCentral is the operating system for your Amazon business; the intent grading in the call-out below is one of its tools.

What is Amazon COSMO?

COSMO is a research system from Amazon. The paper, "COSMO: A Large-Scale E-commerce Common Sense Knowledge Generation and Serving System at Amazon", appeared at SIGMOD 2024 (DOI 10.1145/3626246.3653398) and is listed on Amazon Science.

Its problem: existing e-commerce knowledge graphs hold product attributes and concepts, but not why people buy. The paper's own example: a shopper searches "shoes for pregnant women" and buys a slip-resistant shoe. COSMO records the link as (pregnant, requires, slip-resistant).

How COSMO builds its knowledge

1. Start from behavior. The inputs are search-then-buy pairs and co-purchase pairs, across 18 major product categories.
2. Ask a language model why. A large language model proposes an explanation for each behavior, such as "used for", "used by", or "capable of".
3. Filter hard. Heuristic rules and classifiers trained on human-labeled examples remove implausible or generic answers. The paper reports about 30,000 annotated instructions.
4. Fine-tune a smaller model. The result, COSMO-LM, generates intent knowledge at scale. The paper lists 15 relation types, including used-for-function, used-for-event, used-for-audience, capable-of, used-as, and is-a.
5. Serve it from a cache. Responses are stored as features. The model is refreshed daily, and the paper says it cannot react to real-time events such as flash sales.

Read the relation types as a list of what the system wants to know about your product: its function, its event, its audience.

What Amazon has said COSMO does in search

The paper evaluates COSMO on three tasks: search relevance, session-based recommendation, and search navigation. Two results matter to sellers.

• Relevance. On a public relevance benchmark (ESCI), adding COSMO-generated intent knowledge raised macro F1 by about 60% with a fixed encoder and about 25% with a tuned one. That is a research benchmark, not a ranking weight.
• Deployment. COSMO is live in search navigation. In A/B tests over several months on about 10% of US traffic, Amazon reports a 0.7% relative sales increase in that segment and an 8% rise in navigation engagement.

The paper says the deployed version targeted search navigation. It does not say COSMO re-orders the main results page.

COSMO vs. A9/A10 and Alexa for Shopping

The commonly used picture has three layers: A9/A10 keyword matching and performance signals at the base, COSMO's intent knowledge on top, and the conversational assistant above that. That picture is an industry summary, not an Amazon diagram. Two points hold either way:

• COSMO did not replace A9/A10. A listing that keyword matching never indexes gives intent knowledge nothing to work with. Start with the basics in the A9/A10 guide.
• The assistant link is unconfirmed. The paper does not mention Rufus, and Amazon has not published a document saying COSMO powers Alexa for Shopping. Many third-party guides assert it without a source. For what Amazon does say about the assistant, see what Rufus was and became and Alexa for Shopping vs. classic search.

For the full stack in one place, read how Amazon search works in 2026.

What this means for your listing

You cannot see or tune COSMO. You can control one input: whether your listing states the intent a shopper has in mind. A specs-only listing leaves the reader, human or model, to guess. Six things to state when they are true:

• Use case: what job it does ("keeps drinks cold for 24 hours on a hike").
• Audience: who it is for ("for travelers with a carry-on only").
• Occasion: when it is used ("camping", "office gift").
• Compatibility: what it fits or works with, by model or size.
• Attributes: material, size, capacity, certifications.
• Problem solved: the pain it removes ("no spills in a bag").

Three rules keep this safe:

• No keyword stuffing. Natural sentences carry intent. Repeated terms do not.
• Mind the title limit. Titles are capped at 75 characters from 2026-07-27 (media products are exempt). Put the most specific use or audience there, and the rest in bullets. See title optimization and bullet points.
• Do not seed your own Q&A. Amazon's community guidelines bar people with a conflict of interest from posting questions. Put common answers in bullets or A+ Content, and answer real questions as the seller.

Intent-phrased searches are also how shoppers talk to voice and chat assistants. The same copy supports them: see listing optimization for voice search and why a product may not appear in Alexa answers.

COSMO readiness checklist

• ☐ Title names the main use or audience within 75 characters
• ☐ Each bullet states a benefit and the situation it applies to, not only a spec
• ☐ Compatibility is stated by model, size, or standard where it matters
• ☐ Occasion and audience appear at least once in the listing
• ☐ Every claim is true and provable
• ☐ A+ Content repeats the use cases with images that show them
• ☐ Five intent-style searches tested against the copy ("gift for a runner who travels")

Frequently asked questions

What is Amazon COSMO?

COSMO is a system Amazon described in a SIGMOD 2024 research paper. It builds a knowledge graph of shopping intent: why people search for and buy things, such as the use, the audience, or the occasion. Amazon trained a language model on search and purchase behavior to generate this knowledge, then filtered it with human-labeled data.

Is COSMO the same as the A9 or A10 algorithm?

No. A9 is the name for Amazon's keyword-matching search ranking, and A10 is an industry label Amazon has not confirmed. COSMO is a separate system that produces intent knowledge. Industry write-ups describe it as a layer on top of A9/A10, not a replacement. Amazon's paper does not describe how COSMO output combines with the ranking function.

Does COSMO rank products in Amazon search results?

Amazon has not said so. The paper evaluates COSMO on search relevance, session-based recommendation, and search navigation, and reports a live deployment in search navigation. It does not state that COSMO sets the order of the results page. Treat claims that it does as unconfirmed.

Does COSMO power Alexa for Shopping (formerly Rufus)?

Amazon has not published that link. The COSMO paper predates Alexa for Shopping and does not mention Rufus. Many third-party guides say COSMO feeds the assistant, but they cite no Amazon source. Writing listings that state intent explicitly helps in either case.

How do I optimize my listing for COSMO?

State intent in plain words: the use case, the audience, the occasion, compatibility, key attributes, and the problem the product solves. Put it in the title, bullets, description, and A+ Content. Do not stuff keywords, and do not claim anything the product does not do.

Can I see how my listing scores on intent coverage?

Yes. The OptimalCentral AI Intent Audit grades a listing on six dimensions of intent coverage and writes the copy that closes each gap. It is included in every plan, and the free trial includes one audit.