What Is Amazon COSMO?
COSMO is the knowledge graph behind Amazon's shift from keyword matching to understanding intent. There's a lot of confident nonsense written about it. This is the honest version — the real numbers from Amazon's own paper, what COSMO does and doesn't do, and a straight answer on the question everyone gets wrong: does it actually power Rufus and Alexa for Shopping?
COSMO is a large-scale common-sense knowledge graph Amazon built to connect products to the intents and situations behind searches. Per Amazon's COSMO paper (SIGMOD 2024), it holds roughly 6.3M nodes and 29M knowledge relationships across about 18 categories, derived largely from real shopper behavior. It helps Amazon understand meaning, not just keywords. The big caveat almost no one states: Amazon has not confirmed that COSMO powers Rufus or Alexa for Shopping — that link is a community deduction, not an official fact.
Most COSMO content treats "COSMO runs the AI assistant" as established fact. It isn't. The COSMO research paper and Amazon's shopping-assistant materials do not cite each other, and Amazon has published no statement connecting them. The connection is a reasonable architectural inference the SEO community shares — and we'll act on it — but we won't dress a deduction up as a disclosure. If you've been sold "the official COSMO algorithm," you were sold something that doesn't exist publicly. Accuracy is the whole point of this page.
What COSMO actually is
COSMO stands for a common-sense knowledge generation and serving system. In plain terms: a giant map of how products relate to needs, uses, situations, and each other — built so Amazon can answer "what fits this intent?" rather than only "what contains this keyword?"
Figures as reported in Amazon's COSMO paper (Amazon Science, SIGMOD 2024). They describe the system at the time of publication and may have grown since.
Built from behavior, not guesses
What makes COSMO interesting is its source. Much of the knowledge is derived from real shopper behavior — patterns like search-then-buy (people who searched X ended up buying Y) and co-purchase (people who bought X also bought Z). Amazon's system mines these patterns, uses large language models to propose the "common-sense" relationships that explain them, and applies human annotation to keep the knowledge reliable.
The practical consequence: COSMO reflects how shoppers actually behave, aggregated at massive scale. That's powerful — and it's also why it moves slowly. A graph built from aggregated behavior refreshes in batches, so changes to your listing take time to register. Think weeks, not hours.
COSMO vs A9: meaning vs matching
These aren't rivals — they're different jobs. Confusing them is where most advice goes wrong.
A9-style ranking still decides whether you're eligible. A semantic layer like COSMO informs how well your product is understood and placed against intent. You optimize for both — keyword fundamentals and semantic clarity.
What the relationships look like
The "edges" in the graph are typed relationships — the common-sense links between a product and the world. The paper describes roughly fifteen relation types; here are the kinds of connections they capture (illustrative examples):
Why this matters to you: these are exactly the connections a complete, well-structured listing makes explicit. Vague, keyword-stuffed copy gives a knowledge graph little to map. Clear statements of use, audience, properties, and occasion give it strong, unambiguous edges to draw.
Does COSMO power Rufus and Alexa for Shopping?
Short answer: probably, in spirit — but Amazon hasn't said so.
It's an attractive story: Amazon builds an intent knowledge graph (COSMO), then launches a conversational assistant that answers by intent (Rufus, now Alexa for Shopping), so the graph must feed the assistant. Architecturally, that's plausible — they solve the same problem from two directions.
But here's what's actually true: the COSMO paper doesn't mention Rufus, the assistant's materials don't mention COSMO, and Amazon has published no statement linking them. So the connection is a reasonable deduction the community shares, not a documented fact. We build on it as a working model — because the optimization advice it implies (clarity, structure, intent coverage) is sound either way — but we flag it honestly, and you should too when you talk to clients or partners.
Whether or not COSMO literally feeds the assistant, the takeaway is identical: Amazon is modeling products by intent and relationships, and clear, complete, structured listings win in that world. You don't need the deduction to be confirmed for the work to be right.
What COSMO means for your listing
State meaning explicitly
Say what your product is for, who it's for, and where it's used — in plain, declarative terms a graph can map to intent.
Complete the structured data
Fill your backend attributes. They're the cleanest signals a semantic system can read about your product.
Cover real use-cases
Answer the situations and questions buyers actually have, so your product connects to more intents.
Give it time
Batch refresh means changes register slowly. Make the change, then wait a week or two before judging.
Is your listing legible to a semantic system?
Four quick questions. No email required.
Amazon COSMO — straight answers
A large-scale common-sense knowledge graph Amazon built to connect products to the intents and situations behind searches. Per Amazon's COSMO paper (SIGMOD 2024), roughly 6.3M nodes and 29M relationships across about 18 categories, derived largely from real shopper behavior. It helps Amazon understand meaning and intent, not just keyword matches.
A9 is keyword-based search ranking — it matches query terms to listings. COSMO is a semantic knowledge layer capturing relationships and intent: why a product fits a need. They're complementary; A9-style ranking decides eligibility, a knowledge graph informs intent understanding.
Widely assumed, not confirmed by Amazon. The COSMO paper and the assistant's materials don't cite each other, and Amazon has published no linking statement. It's a reasonable community deduction — treat it as a working model, not an official fact.
A graph built from aggregated behavioral data refreshes in batches, so listing changes propagate slowly — weeks, not hours. Give optimization work aimed at semantic understanding time before judging results.
Amazon increasingly models products by intent and relationships, not just keywords. Complete structured attributes, clear use-case coverage, and consistent, accurate content help any semantic system place your product correctly — regardless of exactly how COSMO connects to the assistant.
Why trust this explainer?
From the concept to the practice
COSMO is the "why." These pages are the "how" — start wherever your question is.
See how legible your listing is to a semantic system.
Our free self-scorecard scores your listing on the clarity and structure that semantic systems like COSMO rely on — about 10 minutes, no email required.
Get Your Free Scorecard