COSMO vs A9 vs A10: The Real Amazon Algorithm in 2026.

Why "A10" Became the Seller Community's Biggest Myth

Walk into any Amazon seller Facebook group, YouTube channel, or conference in 2025 and you'd hear sellers talking about "A10" like it was a confirmed algorithm update. Entire coaching programs were built around ranking on A10. Sellers blamed A10 for ranking drops. A10 was cited as the reason keyword stuffing stopped working. A10 was the new boogeyman.

The problem: A10 has never been confirmed by Amazon. Not in any official statement, developer documentation, patent filing, or peer-reviewed paper. A10 is a community-invented label for perceived algorithm changes - a placeholder name that stuck because sellers needed something to point at when their listings behaved differently than they used to.

This guide is the myth-busting breakdown: the real 3-layer algorithm stack Amazon runs in 2026, what COSMO actually does, how it differs from both A9 and Rufus, and exactly how to optimize listings for the current system. For the broader ranking strategy this algorithm work enables, see How to Rank on ChatGPT in 2026. For the parallel AI ecosystem shift happening outside Amazon, UCP vs ACP for Shopify Merchants covers the agentic commerce protocol war.

The Truth: A9 Never Got Replaced - COSMO Was Built on Top

A9 is Amazon's core ranking infrastructure. It handles keyword matching, sales velocity, conversion rate signals, inventory health, and the foundational ranking factors every seller has known about for years. A9 is still running. A9 has not been deprecated or retired. Amazon has made no announcement suggesting otherwise.

What happened starting in 2024 is that Amazon deployed COSMO as an AI layer operating alongside A9 - specifically focused on search relevance, recommendations, and search navigation. COSMO doesn't replace A9's core ranking math. It adds intent understanding on top of it.

The Practical Difference

Under pure A9, the question was: "Does this listing contain the words the customer searched for?" Keyword matching, with sales velocity and conversion amplifying the match.

Under COSMO plus A9, the question becomes: "Does this product actually solve the problem the customer described?" Semantic understanding, layered on top of keyword matching. A listing can match keywords perfectly and still lose ranking if COSMO decides the listing doesn't actually answer the intent behind the query.

Why Sellers Got Confused

The confusion came from Amazon's public silence. Amazon never announced COSMO with the fanfare that would have made sellers aware. The SIGMOD paper was technical, not consumer-facing. The Rufus launch got the headlines. Meanwhile, COSMO was quietly reshaping retrieval in the background. Sellers observed the effects (keyword stuffing less effective, intent-based listings winning) and needed a name for what they were seeing. "A10" was easier to say than "a commonsense knowledge graph deployed on top of A9."

What COSMO Actually Is (From the SIGMOD 2024 Paper)

COSMO stands for COmmon Sense MOdeling. It's Amazon's system for understanding the real-world intent behind shopper queries and mapping that intent to products.

How COSMO Builds Its Understanding

COSMO doesn't analyze listings in isolation. It builds a knowledge graph from two primary data sources:

• Query-purchase pairs - what people search for and then buy. If shoppers search "quiet vacuum for small apartment" and buy specific vacuum models, COSMO learns that those models are associated with "quiet" and "small space" attributes even if the listings don't explicitly use those words.
• Co-purchase data - what products people buy together in the same session. If buyers of vacuum X also frequently buy HEPA filter refills, COSMO builds relationships between vacuum X and HEPA filtration, allergen concerns, and filter-maintenance intent.

COSMO uses large language models to infer the commonsense relationships behind these behavior patterns, then validates those inferences through human-in-the-loop annotation. The result is a structured knowledge graph mapping products to real-world use cases, audiences, problems, and complementary products.

Where COSMO Is Deployed

The SIGMOD paper specifically documents deployment in Amazon's search navigation feature - the refinement tiles that appear on search results pages. The paper's authors note that COSMO's architecture is designed for broader deployment, and the parallel growth of Rufus (which serves intent understanding on the customer-facing side) strongly suggests systematic expansion toward semantic discovery across Amazon's search stack.

For the Rufus-facing implementation of these principles, see Amazon Rufus Optimization Guide. For how this translates to actual listing rewrites, Noun Phrase Optimization for Amazon is the tactical framework.

The 3-Layer Algorithm Stack: A9 + COSMO + Rufus

Amazon's 2026 discovery system is not one algorithm. It's three layers operating together.

How COSMO Builds Its Knowledge Graph

The mechanics matter because they tell you what actually signals relevance to COSMO (and therefore what to build into your listings).

The Data Pipeline

1. Raw behavioral dataAmazon collects every search query, every click, every purchase, every session-level co-purchase pattern. Billions of data points per day.
2. Pattern extractionCOSMO's LLM layer identifies recurring patterns. "Shoppers who search X tend to buy Y together with Z." "Shoppers who describe themselves as A tend to choose products labeled B."
3. Commonsense inferenceThe LLM generates hypotheses about why these patterns exist. "Shoppers buying vacuums with HEPA filters probably have allergies or pets, which explains the co-purchase of pet hair tools."
4. Human validationHuman annotators review LLM-generated inferences and flag false positives. Only validated relationships enter the production knowledge graph.
5. Daily refreshNew session data feeds back in. Patterns strengthen or weaken. Relationships get added or pruned.
6. Downstream applicationThe knowledge graph feeds search relevance, refinement tiles, recommendations, and (probably) Rufus retrieval.

What This Means for Your Listings

COSMO learns product attributes from shopper behavior, not just from your listing text. If shoppers consistently describe your product category with specific use cases or audience terms in their queries, COSMO picks up those patterns even if your listing doesn't explicitly use those words. Conversely, if your listing claims an attribute but shoppers never buy based on that attribute, COSMO deprioritizes that claim.

The practical implication: align your listing language with how shoppers actually describe their needs. Don't use internal product-team jargon. Use the natural language shoppers use in Brand Analytics search query reports. For a complete listing optimization framework, see Amazon Listing Checklist.

The 6 Commonsense Relation Types COSMO Uses

The SIGMOD paper identifies specific relation types that COSMO encodes in its knowledge graph. Understanding these relation types tells you exactly what signals to build into your listings.

Why Keyword-Only Listings Are Losing Ranking in 2026

Here's what's actually happening when a keyword-stuffed listing loses ranking without the keywords changing.

The Decay Pattern

Under pure A9, a keyword-stuffed title like "Garlic Press Kitchen Gadget Tool Cooking Utensil Stainless Steel Professional Home" would match a wide range of queries. Sales would reinforce ranking. The listing would hold position.

Under A9 plus COSMO, the same title matches queries but fails COSMO's intent evaluation. The title doesn't clearly communicate what problem it solves, who it's for, or what it does well. When shoppers land on the listing from a query, COSMO observes whether they convert at expected rates. A COSMO-incoherent listing tends to underperform expected conversion because shoppers land, don't find their intent confirmed, and bounce.

Bounce signals feed back into COSMO. COSMO downgrades the listing's semantic relevance. A9's keyword match is still there - but the AI layer is actively suppressing the listing's visibility because it's not answering shopper intent. Over 30-90 days, the listing's ranking decays even though nothing about its keyword strategy has changed.

What Sellers Observe

• Traditional keyword rankings remain stable
• Sessions drop
• Conversion rate drops
• PPC becomes more expensive (quality score erodes)
• "The algorithm changed" becomes the explanation

The algorithm didn't change. The AI layer on top of the algorithm noticed your listing doesn't answer intent and started routing traffic away. Fixing it requires rewriting for semantic coherence, not just adding more keywords.

COSMO's Search Navigation Deployment vs the Broader Stack

One important nuance from the SIGMOD paper: COSMO's documented deployment is in Amazon's search navigation feature - the refinement tiles that appear on search results pages suggesting narrower queries. The paper does not describe COSMO as a replacement for A9's core ranking algorithm.

What the Paper Documents

• COSMO powers the "refinement" suggestions you see on search results pages ("for saltwater fishing," "for small kitchens," "under $50")
• COSMO's knowledge graph informs which product types are most relevant to specific refinements
• COSMO adjusts its graph daily based on new behavioral data

What's Widely Believed But Not Formally Documented

• COSMO likely powers Rufus's retrieval layer, though Amazon has not officially confirmed
• COSMO likely influences broader organic ranking beyond just refinement tiles
• COSMO likely feeds into Sponsored Ad placement decisions, especially in Rufus-facing ad slots

Propagation Timing: Why Listing Changes Take 7-14 Days

Understanding propagation timing matters for measuring your optimization work. If you rewrite a listing today and check rankings tomorrow, you'll see nothing. COSMO doesn't update in real time.

The Cache Architecture

COSMO uses a two-layer cache strategy:

• Yearly pre-loaded patterns - Stable relationship patterns that rarely change (common product categories, well-established use cases) are cached long-term
• Daily batch processing - New behavioral signals are batched and processed daily. Listing changes don't enter the knowledge graph until this daily batch includes them, and they need enough associated behavioral data (queries, clicks, purchases) to build statistical confidence

The Real-World Implication

Expect 7-14 days before listing changes fully reflect in COSMO's evaluation. For major title or bullet rewrites, allow 30 days for full ranking stabilization. Flash sales and real-time events cannot be processed due to the daily batch cycle - which is why Rufus rarely surfaces flash deal content.

Measurement Cadence

If you're rewriting listings for COSMO optimization:

• Day 0: Document baseline rankings, sessions, conversion rate
• Day 7: Light check - unlikely to show meaningful change yet
• Day 14: Initial measurement window - early signals visible
• Day 30: Full measurement - the rewrite's impact is now stable
• Day 60: Second measurement - confirms sustained improvement

This is why the "rewrite and wait" approach beats the "constant tweaking" approach for listing optimization in 2026. For how this fits into a broader weekly optimization cadence, see Amazon Listing Optimization Services.

Auditing a Listing Against All 3 Layers

Here's the audit framework we run on client listings. Run this on your top 10 revenue products.

A9 Audit (Foundation)

• Primary keyword in title first 80 characters?
• Primary keyword in at least 3 of 5 bullets?
• Sales velocity meeting category expectations?
• Conversion rate within category benchmark?
• Inventory in-stock with healthy coverage?
• Price competitive with top 3 category competitors?

COSMO Audit (Intent Layer)

• Does the title communicate what the product IS (not just keywords)?
• Does the title communicate who the product is FOR?
• Do bullets cover at least 4 of the 6 relation types (used_for_activity, used_for_audience, has_capability, complementary_product, use_context, shopper_intent)?
• Does the listing answer the 3 most common questions in the category?
• Does A+ content reinforce intent signals rather than repeat title keywords?
• Are all product attribute fields filled in Seller Central backend?

Rufus Audit (Customer Interface)

• Ask Rufus 5 specific questions about the product. Does Rufus answer correctly?
• Does Rufus cite positive review sentiment? (Check review language for clear, detailed mentions of key features)
• Does Rufus flag any listing gaps or missing information?
• Does the listing appear when a shopper asks for the product's specific use case?

A listing passing all three layers is COSMO-ready. Listings failing 3+ items in any layer need rewrite prioritization. For the specific rewrite framework, see Noun Phrase Optimization for Amazon.

The 10-Point COSMO Optimization Checklist

Run this on every listing you're optimizing. Each item scored yes/no.

1. Title communicates product identity clearlyNot just keywords - what the product actually is, in natural language.
2. Title includes primary audience or use case"for home cooks," "for professional kitchens," "for outdoor adventurers."
3. Bullets cover all 6 relation typesActivity, audience, capability, complementary, context, intent - each represented.
4. Primary keywords included naturallyNot stuffed - woven into intent-communicating prose.
5. Specific measurable claimsMaterial, dimensions, capacity, wattage - not subjective adjectives.
6. A+ content reinforces intent signalsModules that explain use case, audience, problem-solved - not just feature lists.
7. Every product attribute field filledBackend Seller Central attributes are direct COSMO inputs.
8. Review sentiment aligns with listing claimsIf listing claims "durable" but reviews complain about durability, COSMO catches the mismatch.
9. Q&A section addresses common questionsIf shoppers repeatedly ask the same question, the listing has an intent gap.
10. Images show use case, not just product-on-whiteLifestyle and in-use imagery reinforce use_context relations.

Scoring 8+ out of 10 indicates a COSMO-ready listing. 5-7 indicates needed optimization. Under 5 indicates urgent rewrite required. For the complete Amazon review strategy that reinforces COSMO signals, see Amazon Review Strategy 2026.

What to Expect Through 2027 as the Stack Matures

COSMO and Rufus are still early. Here's what operators should expect to shift through 2027.

Near-Term (Q2-Q4 2026)

• COSMO's influence on broader organic ranking expands beyond search navigation
• Rufus usage grows from the current 2-3% of purchase interactions toward 10-15%
• Sponsored Ads in Rufus conversations become more prominent and better-measured
• Amazon rolls out more transparent reporting on Rufus-attributed sales
• Interests AI (the mood-based curation feature) gets broader user adoption

Mid-Term (2027)

• Keyword stuffing becomes actively penalizing, not just less effective
• Backend attribute completeness becomes a major ranking signal
• Review quality (length, specificity, helpful votes) outweighs review count
• External traffic that converts on Amazon becomes the single biggest ranking factor (continuing a 2025-2026 trend)
• Multi-modal signals (images, videos, A+ content) weight higher as Amazon deploys visual AI

The Operator Response

Sellers who rewrite listings for COSMO in 2026 will be ahead of 80% of the market going into 2027. The compound effect is significant: a COSMO-optimized listing gets better ranking, which drives more sessions, which strengthens COSMO's positive signals, which drives further ranking improvement. A9-only listings decay in the opposite direction.