Amazon's COSMO Algorithm: What It Actually Is and Why Your Listings Need to Change

Amazon's Search Engine Got a Brain. Most Sellers Haven't Noticed.

Something fundamental changed in how Amazon decides which products to show shoppers, and the majority of sellers are still optimizing like it didn't happen. Amazon built a system called COSMO — short for Common Sense Knowledge Generation — that doesn't just match keywords to listings. It understands why a customer is searching and connects products to the intent behind the query.

This isn't a press release claim. Amazon published the research paper at SIGMOD 2024, one of the top database conferences in the world. The paper details a knowledge graph with over 6.3 million nodes and 29 million relationship edges spanning 18 product categories. In A/B testing on 10% of US traffic, COSMO drove a 0.7% increase in product sales — which at Amazon's scale translates to hundreds of millions of dollars in annual revenue. When an algorithm moves that kind of money, Amazon doesn't treat it as an experiment. It becomes the infrastructure.

What COSMO Actually Is (Without the Buzzwords)

Most articles about COSMO describe it as "intent-based search" and leave it there. That's vague enough to be useless. Here's what's actually happening under the hood.

COSMO is a knowledge graph — a structured database of relationships between products, attributes, and the reasons people buy things. It's built by feeding two types of real shopping data into a large language model:

• Query-purchase pairs: What someone searched for and what they actually bought afterward
• Co-purchase pairs: What products people bought in the same shopping session

The LLM analyzes these patterns and extracts relationships using four seed categories: usedFor, capableOf, isA, and cause. From those four seeds, Amazon's system generates 15 finer-grained relationship types that describe exactly how products connect to human needs.

Here's a concrete example. Amazon's system sees that thousands of customers who search "shoes for pregnant women" end up buying slip-resistant shoes. The LLM recognizes the commonsense connection: pregnant women need slip-resistant shoes for safety. COSMO encodes this as a used_for_audience relationship — linking slip-resistant shoes to pregnant women as a target audience.

The critical thing to understand: COSMO doesn't need the listing to contain the phrase "pregnant women" to make this connection. It infers the relationship from shopping behavior and commonsense reasoning. But — and this is the part most articles skip — your listing still needs to provide enough context for COSMO to confirm the match. If your slip-resistant shoes listing says nothing about safety, stability, or the types of people who benefit from non-slip footwear, COSMO has less signal to work with.

The 15 Questions COSMO Asks About Every Product

COSMO standardizes all intent understanding into 15 relationship types. Each one maps to a question the algorithm is trying to answer about your product. Based on Amazon's published research, the highest-value ones include:

• used_for_func — What function is this product used for? (e.g., "used for arch support")
• used_for_eve — What event or activity is this for? (e.g., "used for hiking," "used for weddings")
• used_for_audience — Who is the target user? (e.g., "for pregnant women," "for left-handed people")
• capable_of — What is this product capable of doing? (e.g., "capable of insulating in cold weather")
• used_to — What task does this accomplish? (e.g., "used to build a fence")
• is_a — What category or type is this? (e.g., "is a formal suit," "is a camping stove")
• used_as — What does this function as? (e.g., "used as a smart watch," "used as a nightlight")

The remaining relations cover properties like materials, styles, locations of use, and compatibility — but these seven are the ones that matter most for broad and vague searches, which is where COSMO has the biggest impact on which products appear.

Why this matters for you: Every one of these questions is something your listing content should answer clearly. Not with keywords stuffed into a title, but with natural statements in your bullets, description, and backend attributes that explicitly state function, audience, events, and capabilities.

How COSMO, A9, and Rufus Work Together

There's a lot of confusion about whether COSMO "replaces" the A9 algorithm. It doesn't. Here's the actual architecture:

A9 is the search engine. It handles keyword indexing, relevance scoring, and the basics of matching search queries to product listings. A9 still matters. Keywords still matter. If the word "insulated" doesn't appear anywhere in your listing, A9 won't index you for "insulated water bottle." That hasn't changed.

COSMO is the intelligence layer underneath A9. It's the knowledge graph that tells the search system what products are contextually relevant to a query — even when the query doesn't perfectly match the listing's keywords. COSMO is what allows Amazon to show a camping stove when someone searches "cooking gear for backpacking" even if the listing never uses the word "backpacking."

Rufus is the customer-facing AI shopping assistant that sits on top of both systems. Rufus reads your listing content — title, bullets, description, reviews, Q&A — and uses it to generate natural language answers to shopper questions. Amazon CEO Andy Jassy reported that 250 million shoppers used Rufus in 2025, with monthly active users growing 140% year over year. During Black Friday 2025, Rufus was active in 38% of shopping sessions.

The three systems feed each other. COSMO builds the knowledge graph that helps Rufus understand products. Rufus reads your listing to answer questions. A9 handles the keyword matching and indexing. If your listing is optimized for only one of these systems, you're losing ground to sellers who are covering all three.

We covered how to write bullet points specifically for Rufus in our post on writing bullets that Rufus actually recommends. This post is about the COSMO layer — the knowledge graph that decides whether Rufus even considers your product when answering a question.

What This Means If You Sell 5-50 Products

Here's where I need to push back on most COSMO content online. Almost everything written about COSMO is aimed at enterprise brands and agencies managing thousands of ASINs. The advice tends to be "use our $500/month tool to audit your COSMO knowledge graph alignment" or "hire us to do a semantic content overhaul of your catalog."

That's overkill if you're a solo seller with 20 SKUs.

The good news: COSMO actually helps smaller sellers. The old A9 system heavily favored products with massive sales velocity and deep review counts — the rich got richer. COSMO adds a new dimension. If your product genuinely fits a specific intent better than a bestseller does, COSMO can surface you for that intent-based query. Niche sellers who clearly communicate what their product is for, who it's for, and what it's capable of can now compete on relevance rather than just volume.

The practical translation: you don't need a "COSMO strategy." You need listings that clearly answer questions about your product in plain language. That's it. The sellers who overthink this are the same ones who used to obsess over exact-match keyword density in 2018.

The Specific Changes to Make in Your Listings

Enough theory. Here's what to actually do.

Fill Out Every Attribute Field in Seller Central

This is the single highest-impact, lowest-effort change you can make for COSMO. And almost nobody does it.

In Seller Central, go to your listing, click "Edit," and look at the full set of attribute fields. Most sellers fill out the required ones — title, bullets, description, images — and skip everything else. But Amazon exposes dozens of additional fields depending on your category:

• Subject Keywords — Up to 5 fields, 50 characters each. These index within 20 minutes and directly feed COSMO's category mapping.
• Intended Use — Describe activities, events, or conditions the product is designed for. "Hiking, camping, gym, travel" gives COSMO explicit intent signals.
• Target Audience — Be specific. "Left-handed chefs" is useful. "Everyone" is useless.
• Other Attributes — Material, color, size, compatible devices, age range, occasion type.

These fields map directly to COSMO's relationship types. When you fill in "Intended Use: hiking, camping, fishing," you're explicitly telling COSMO's knowledge graph that your product has a used_for_eve relationship with those activities. When you specify "Target Audience: pregnant women, elderly, people with mobility issues," you're creating used_for_audience connections.

Most of these fields are static — fill them out once, and they don't need regular updates. Spend 15 minutes per listing doing this, and you've done more for COSMO optimization than most enterprise brands paying agencies thousands per month.

Rewrite Bullets as Intent Answers, Not Feature Lists

This overlaps with Rufus optimization, but it's worth restating specifically through the COSMO lens.

COSMO builds knowledge by understanding why products are purchased. Your bullets need to communicate the "why" — not just the "what."

The second version still contains the keywords COSMO and A9 need ("camping," "cooking," "backpacking" are implied by "backcountry cooking," "alpine," "cook kit"). But it also tells COSMO's knowledge graph specific things: this product is used_for_eve (backcountry cooking), it's capable_of (working at high altitude, resisting wind), and it's used_to (boil water quickly). The first version just lists words.

Add Use-Case Language to Your Description

Your product description (2,000-character limit) is the most underused field for COSMO optimization. Most sellers either leave it blank or duplicate their bullets. COSMO indexes the description and uses it to build deeper context about your product.

This is where you can describe scenarios: "Whether you're car camping with the family or solo backpacking the PCT, this stove handles both without taking up half your pack." That single sentence creates multiple COSMO connections — family camping, solo hiking, the PCT (a specific trail Rufus can reference), and pack-size considerations.

Don't keyword-stuff the description. Write it like you're explaining the product to a friend who's trying to decide what to buy. Natural language is exactly what both COSMO and Rufus are optimized to parse.

Use Backend Search Terms for Intent Phrases

We covered the mechanics of backend search terms in a separate post — the 250-byte limit, no commas, no repeated words. But COSMO changes what you should put in that field.

Old approach: pack in keyword synonyms and misspellings.

COSMO-era approach: still include misspellings and synonyms, but also include intent-based phrases that don't fit naturally in your visible listing.

For a camping stove, old backend terms might be: camp stove portable butane gas burner outdoor cooking campstove

Better backend terms for COSMO: backpacking cooking ultralight thru hike trail cooking emergency preparedness power outage camp kitchen picnic tailgate

Those aren't product features — they're contexts where your product gets used. COSMO maps these intent signals to its knowledge graph, connecting your stove to searches like "what do I need for a thru-hike" or "cooking during power outage" — queries your listing would otherwise never match.

What You Don't Need to Do

Let me save you some time and money.

You don't need a "COSMO audit tool." Several companies now sell tools that claim to score your listing's "COSMO alignment" or show your "position in the knowledge graph." COSMO's knowledge graph isn't publicly accessible. These tools are making inferences about what COSMO might know based on search results — which is useful data, but it's not the same as seeing your product's actual node in the graph. If you have 10 products and a limited budget, your money is better spent on listing content improvements than on analytics tools telling you what to improve.

You don't need to abandon keyword research. COSMO sits on top of A9, not in place of it. Keywords still drive indexing. If "portable camping stove" doesn't appear in your listing, A9 won't show you for that search, and COSMO can't fix that. The shift is from keyword-only to keyword-plus-intent. Do your keyword research. Then make sure your listing also answers the intent questions.

You don't need to rewrite every listing overnight. Start with your top 5 sellers. Fill out their attribute fields. Revise bullet 1-3 to answer questions instead of listing features. Update the description. Check backend terms. Then watch your Unit Session Percentage (conversion rate) in Business Reports for 2-4 weeks before doing the rest.

The Opportunity Most Sellers Are Missing

Here's what I keep coming back to: COSMO is a system that rewards clarity. It rewards listings that plainly state what a product does, who it's for, when to use it, and what it's capable of. That's not a complicated optimization strategy. That's just good product communication.

The sellers who will struggle with COSMO are the ones whose listings are keyword soup — all the right words in the wrong order, saying nothing a human or AI can actually parse. The sellers who will benefit are the ones who write like they're talking to a real person about a real product.

Most solo sellers I know already think about their products this way. They know their customer. They know the use cases. They just weren't translating that knowledge into their listing copy because the old playbook said to focus on keywords. COSMO is the algorithm catching up to what good sellers have always known: people don't search for "stainless steel double wall vacuum insulated 32oz BPA free." They search for "water bottle that keeps drinks cold all day."

SellScope builds COSMO-era intent signals into every listing it generates. When you input your product details, it produces titles, bullets, and descriptions that answer the intent questions COSMO maps — function, audience, events, capabilities — while still hitting the keyword targets A9 needs for indexing. It's not magic, and it won't replace knowing your product and your customer. But it handles the structural work of encoding both keyword relevance and intent clarity into Amazon's required format, which is the part that trips most sellers up.

The Bottom Line

COSMO is a knowledge graph that maps products to the reasons people buy them. It uses 15 relationship types — function, event, audience, capability, and others — to decide which products are relevant to a shopper's intent, not just their search terms. It works alongside A9 (keyword matching) and Rufus (AI shopping assistant), and together they determine what shoppers see.

The single most important action you can take: open your top-selling listing in Seller Central right now, scroll through every attribute field, and fill in anything that's blank. Subject keywords, intended use, target audience, other attributes. That's 15 minutes of work that feeds COSMO exactly the structured data it was designed to consume. Then reread your first three bullet points. If they list features without explaining who benefits and why, rewrite them as answers to the questions your customers actually ask.