Disclosure: This article contains affiliate links — we may earn a commission at no extra cost to you. This does not affect our rankings or conclusions.
Amazon COSMO is a machine-learning system Amazon built to mine “common sense” knowledge about shoppers from their behavior, so search can infer the intent behind a query instead of matching keywords alone. It was published by Amazon’s own researchers at the ACM SIGMOD 2024 conference. For sellers, COSMO matters because it pushes ranking further toward “does this product solve the shopper’s underlying need?” — but a lot of what circulates online about it is interpretation, not confirmed fact, so this guide separates the two carefully.
This page is the deep dive on COSMO. If you want the broader picture of how ranking works end to end, start with the Amazon SEO guide; COSMO is one layer inside that system, not a replacement for it.
What COSMO Stands For — and What It Is
COSMO is short for “Common Sense knowledge generation and serving at Amazon” — a name taken from the paper’s title rather than an official, letter-by-letter acronym Amazon has published. The problem it addresses is a real gap in older e-commerce search: a shopper who types “shoes for a nurse” isn’t naming a product category, they’re describing a need — long shifts, comfort, slip resistance. Keyword matching alone struggles with that leap. COSMO’s job is to build the bridge between the query and the reason behind it.
Think of it less as “the new ranking algorithm” and more as a knowledge layer that feeds intent signals into search relevance, recommendations, and search navigation. That distinction matters, and we come back to it, because “COSMO replaced A9” is one of the most common overstatements you’ll read.
Layer 1 — What Amazon’s Research Actually Says
Everything in this section comes from Amazon’s own SIGMOD 2024 paper and the accompanying Amazon Science blog post (sources checked 2026-07-21). These are the load-bearing facts; treat them as the confirmed baseline.
COSMO builds a commonsense knowledge graph from shopper behavior. The pipeline is recursive: a large language model generates hypotheses about the commonsense implications of query-to-purchase and co-purchase data; a combination of human annotation and machine-learning “critic” classifiers filters out the low-quality hypotheses; human reviewers extract guiding principles from what survives; and instructions based on those principles are used to prompt the LLM again. Amazon calls the fine-tuned model COSMO-LM.
It is scaled with surprisingly little hand annotation. The paper reports expanding the knowledge graph to 18 major product categories at Amazon, producing millions of knowledge assertions from only about 30,000 annotated instructions.
The relation types started narrow and got finer. The Amazon Science blog notes the system began with four broad commonsense relations — usedFor, capableOf, isA, and cause — and later refined these into a finer-grained set of frequently recurring relationships.
The headline accuracy number is a search-relevance benchmark, not a sales number. On the public Amazon Shopping Queries Data Set (built for the KDD Cup 2022), COSMO-enhanced cross-encoder models improved relevance-classification F1 over the best baseline. How large the gain is depends on whether the encoders were frozen or fine-tuned:
| Setup | COSMO F1 gain over baseline |
|---|---|
| Encoders frozen | +60% macro F1 |
| Encoders fine-tuned on a test-set subset | +28% macro F1 / +22% micro F1 |
In plain terms: COSMO’s intent knowledge helped the model judge query-to-product relevance much more accurately.
Amazon’s own paper reports positive online A/B results, including a specific sales figure. In its deployment section, the paper states that months-long online A/B tests covering roughly 10% of Amazon’s U.S. traffic produced a 0.7% relative increase in product sales in that segment — which the paper says translates to hundreds of millions of dollars in annual revenue — alongside an 8% increase in navigation engagement rate in the same segment. Two caveats matter and are easy to lose: this is a relative lift, measured in a controlled A/B test on a traffic segment, in the search-navigation context COSMO is deployed in — not a claim that any single listing’s sales rise 0.7%. (Amazon’s shorter Amazon Science blog post highlights the offline relevance metrics and does not repeat these online figures — which is why a lot of coverage assumes Amazon never published a sales number.)
That gap between the detailed paper and the shorter blog post is important for the next section.
Layer 2 — What the Industry Reads Into It (Clearly Marked)
Around these confirmed facts, seller-facing blogs have built an interpretation. Some of it is reasonable extrapolation; some of it hardens Amazon’s cautious language into precise-sounding claims. Here is where the two diverge.
- The “0.7% sales uplift” is real — but third-party retellings often strip its qualifiers. The figure comes straight from Amazon’s paper (see Layer 1), so it is confirmed, not invented by seller blogs. The interpretation problem is what happens next: blogs that repeat it (ZonGuru and others) frequently drop the relative, traffic-segment, and search-navigation context, leaving the impression that COSMO lifts any listing’s sales by 0.7%. It is a system-level result on a segment of traffic, not a listing-level lever you can pull. Trust the number; distrust the flattened version of it.
- “COSMO replaced A9” overstates it. Amazon has never announced retiring A9. The accurate reading is that COSMO adds an intent-and-context layer that works alongside keyword relevance and performance signals. Ranking is still relevance plus conversion; COSMO sharpens the relevance half.
- Specific “COSMO ranking factors” are inferred, not disclosed. Amazon does not publish a ranking formula, and nothing in the COSMO paper is a list of “do X to rank.” Any checklist you see labeled “COSMO ranking factors” — including the one below — is reasoned seller practice, not a leaked spec.
None of this means the interpretation is useless. It means you should know which claims you can build a strategy on (Layer 1) and which are directional (Layer 2).
Layer 3 — What Sellers Should Actually Do
Here’s the practical part. These recommendations follow logically from how COSMO works, and — usefully — they align with what already made a listing strong before anyone said the word “COSMO.” That’s the tell that they’re sound: an intent-understanding layer rewards listings that genuinely communicate what a product is, who it’s for, and how it’s used.
1. Write for the need, not just the keyword
If COSMO connects “shoes for a nurse” to comfort and slip resistance, then a listing that spells out the use case — “all-day comfort for 12-hour shifts, slip-resistant sole” — gives the model the exact signals it’s looking for. This is a shift in emphasis, not a new tactic: your keyword research should now explicitly capture problem-aware and use-case phrases, not only head terms and volume.
2. Fill in product attributes and structured fields completely
COSMO reasons over structured data, so the discovery attributes in your product template (material, use case, audience, occasion, compatibility) do real work. Leaving them blank leaves intent on the table. This is one of the lowest-effort, highest-consistency wins available.
3. Let reviews and Q&A carry natural language
Shopper reviews are part of the behavioral raw material intent models learn from. Reviews that describe how and why people use a product reinforce the associations you want. The mechanics of earning them compliantly are covered in the broader SEO workflow — the point here is that authentic, descriptive reviews are an intent asset, not just social proof.
4. Don’t abandon keyword fundamentals
An intent layer does not delete relevance. Your title, bullets, and backend search terms still need the right words in the right places — COSMO can only surface a product it can first match. The correct mental model is relevance + intent + conversion, and the workflow that ties them together lives in listing optimization.
For the research itself — mining competitor listings for the attributes and intent phrases they rank on, then checking your own coverage — a keyword and reverse-ASIN tool speeds up steps that are painful by hand. Data checked 2026-07-21: Helium 10 offers a capped free plan plus paid tiers with reverse-ASIN and rank tracking; you can start on the Helium 10 free plan and only upgrade when you hit the caps.
Common Misconceptions About Amazon COSMO
- “Keywords don’t matter anymore.” False. Relevance still gates whether you’re eligible to rank at all. COSMO changes what “relevant” is measured against, not whether it’s required.
- “COSMO is a ranking penalty for keyword stuffing.” There is no confirmed COSMO penalty. Keyword stuffing has always hurt click-through and conversion, and modern relevance modeling rewards clear intent over density — but “COSMO penalizes stuffing” states as fact something Amazon hasn’t published.
- “I need to optimize for COSMO separately.” You don’t optimize for COSMO the way you’d chase a specific algorithm. You optimize for shopper intent and complete data; COSMO is the mechanism that rewards that.
- “The 60% improvement means my sales go up 60%.” No. That number is a relevance-model accuracy benchmark (macro F1) on a research dataset, not a promise about any individual listing’s traffic or sales.
COSMO Readiness Checklist
Use this as a listing audit. It’s grounded in how COSMO works, not in disclosed ranking rules.
- Title and bullets state the use case and target buyer, not just the product noun
- Problem-aware and “product for [situation]” phrases are captured in your keyword map
- Every relevant structured attribute in the product template is filled (material, audience, use case, compatibility)
- Backend search terms still cover synonyms and long-tail (relevance foundation intact)
- Reviews and Q&A describe real usage; acquisition stays within Amazon’s rules
- Conversion elements (main image, price, rating) are strong — intent gets you matched, conversion still decides rank
- You’re measuring rank and conversion per keyword, changing one variable at a time
Frequently Asked Questions
What is Amazon COSMO?
Amazon COSMO is a system Amazon’s researchers published at SIGMOD 2024 that mines commonsense knowledge from shopper behavior to infer the intent behind a search query. It helps Amazon search understand why someone searches, not just the literal keywords, and feeds into search relevance, recommendations, and navigation.
Did COSMO replace the A9 algorithm?
No. Amazon has not announced retiring A9. COSMO is best understood as an intent-and-context layer that works alongside the existing relevance and performance signals. Ranking is still relevance plus conversion — COSMO sharpens how relevance is judged.
Does COSMO mean keywords no longer matter?
No. Keyword relevance still determines whether your listing can be matched to a query at all. What changes is the emphasis: alongside the right keywords, your listing needs to clearly communicate use case, audience, and the problem it solves.
Is the “60% improvement” a sales increase?
No. The 60% figure is an increase in macro F1 score — a search-relevance accuracy metric — measured on a public research dataset with frozen encoders. It is not a sales or traffic promise for any individual listing.
How do I optimize my listing for COSMO?
You don’t optimize for COSMO as a separate task. You write listings around genuine shopper intent — spelling out use cases and target buyers — and fill in structured product attributes completely. See the listing optimization guide for the full workflow.
Where can I read the original COSMO research?
It was presented at ACM SIGMOD 2024 as “COSMO: A large-scale e-commerce common sense knowledge generation and serving system at Amazon,” and Amazon Science published a companion blog post. Those are the authoritative primary sources; most seller-facing articles are interpretation on top of them.
Conclusion
COSMO is a genuine step in how Amazon search reasons — from matching words to inferring needs. But the honest summary is narrower than most headlines: Amazon’s paper confirms both a strong relevance-accuracy improvement and a measurable online sales lift (0.7% relative, on a ~10% U.S. traffic segment, in search navigation) — the catch is that this is a controlled, segment-level result rather than a per-listing guarantee, and third-party retellings often blur that distinction. For sellers, the takeaway is reassuring rather than disruptive. Write listings that clearly state what your product does and who it’s for, fill in every structured attribute, keep your keyword foundation intact, and let conversion do the rest. That was good practice before COSMO — and COSMO just raised the reward for doing it well. Continue with the Amazon SEO guide and keyword research guide to put it into a full workflow.