Amazon listing optimization for Rufus and COSMO: what replaces keyword stuffing

The question has changed, not the need for keywords

Amazon SEO used to be about how many times and where a keyword appears. It is now about how well the system understands what the product is, who it suits, which job it does, under which conditions — and why the seller’s claims deserve to be believed.

That does not make keywords useless. Amazon’s own SEO guidance still tells sellers to use relevant search terms in the listing. The same guidance defines keyword stuffing — repeating or overusing keywords — as bad for the customer experience and potentially bad for search performance.

So the task is not “how do I insert keywords for Rufus”. It is this:

Make every ASIN unambiguous, semantically complete, structured, evidenced and convertible — for a stack that combines lexical, semantic, behavioural, catalog and graph/LLM signals.

The practical priority order that follows from it:

semantic coverage → accurate catalog attributes → evidence for claims → a clear title and bullets → images and A+ → consistency with reviews and Q&A → click and conversion rate → continuous A/B tests.

A keyword is still how you name the entity. Repeating it is no longer a strategy.

What Rufus and COSMO actually are

Two things need separating before any tactics, because most of the advice in circulation mixes them up.

Rufus: an answer layer, not a ranker

Amazon describes Rufus as a generative-AI shopping assistant built on a shopping-specialised large language model. It was trained mainly on shopping data — the Amazon catalog, customer reviews and Community Q&A — supplemented by public web information. When it answers, it uses retrieval-augmented generation: it first retrieves the most relevant information from those sources, then the model writes the answer.

That matters for anyone writing listings. The old mindset was simple: the query is insulated water bottle, so write insulated water bottle several times. A Rufus-style system has to answer different questions:

• Will this bottle fit a bike cage?
• Does it fit a car cup holder?
• Does it have a straw?
• Can it go into a backpack without leaking?
• How long will water stay cold?

Repeating water bottle answers none of them. A diameter, a leakproof lid, a straw, an insulation duration and cup-holder compatibility answer all of them. In other words, the information gain of one concrete fact is potentially worth more than the tenth repetition of the head keyword. That is an inference from how Rufus is built, not a ranking weight Amazon has disclosed.

COSMO: product — function — audience — context

COSMO is even more telling. Amazon’s example is a query like shoes for pregnant women. The shopper never types “slip-resistant outsole”, yet commonsense reasoning can connect the audience, the situation and the product feature that matters.

COSMO is built roughly like this: candidate relationships are mined from query–purchase and co-purchase behaviour, an LLM proposes relationships between entities, and filtering, human annotation and ML classifiers clean the result. Among the relationship types Amazon describes are used_for_function, used_for_event and used_for_audience.

In Amazon’s experiments, adding COSMO knowledge to a model that matched queries against product descriptions produced large gains on the Shopping Queries benchmark — around +60% macro-F1 with a frozen encoder, and roughly +28% macro-F1 and +22% micro-F1 after fine-tuning against a strong baseline. That is a research result, not a promise of similar sales growth for any ASIN.

How the pieces fit

Amazon’s public material supports the following picture, though it is a reconstruction, not a published production diagram:

Amazon confirms individual elements — lexical, semantic and behavioural matching, ensembles of language and graph models, COSMO and Rufus retrieval — but not the full list of production signals or their weights. Anyone claiming to know “the Rufus ranking formula” is guessing.

Why keyword stuffing stopped paying off

Redundancy. Once the system has confidently recognised stainless steel water bottle, five more identical mentions add almost no meaning. 32 oz, double-wall insulation, straw lid, cup-holder fit, for hiking, for commuting add a lot. Amazon’s published retrieval research combines language models with graph neural networks precisely to judge relevance beyond literal word overlap.

Intent is richer than one phrase. COSMO explicitly models product-to-function, product-to-audience and product-to-context relationships. A listing that covers real use cases potentially gives the model more useful links than a listing with high keyword density.

Amazon itself advises against it. Its current SEO guidance asks sellers to write naturally and use relevant word variants in context, and names stuffing as a problem.

Content has to serve retrieval and conversion at once. Amazon’s research models use graph relationships built partly from co-view, co-purchase and query-to-purchase data. Amazon has not published a “CTR × CVR = rank” formula. But turning the title into unreadable keyword soup that lowers clicks and purchases is a poor bet in a system that feeds behavioural data into machine learning.

There is now competition for answers. When a shopper asks “which of these is better for travel?”, the assistant needs facts to justify a recommendation — and it takes them from the catalog, reviews and Q&A.

The unit of Amazon SEO is therefore no longer just a keyword. It is entity + attribute + relationship + evidence.

A working model for optimizing a listing

Start not with a keyword list but with a semantic map of the product. Describe six layers for every ASIN:

what it is → specifications → function → problem it solves → audience → situation of use.

For a bottle: insulated bottle → 32 oz, stainless steel → keeps drinks cold → fewer refills → commuters and hikers → office, gym, car, trail. The listing is assembled from this map.

Turn keyword research into intent clusters

Keyword research stays. What changes is the output: not a list of words but clusters of intent. Instead of water bottle, bottle, insulated bottle, stainless steel bottle, water bottles…, build:

• product — insulated stainless-steel bottle
• function — keeps drinks cold
• audience — commuters, hikers, gym users
• occasion — travel, office, outdoors
• feature — leakproof, straw, wide mouth
• compatibility — cup holder, backpack pocket
• constraint — capacity, dimensions, weight

Brand Analytics gives you the raw material: Search Query Performance, Top Search Terms and Search Catalog Performance show search terms with impressions, clicks, add-to-carts and purchases.

Use E-E-A-T as a method, not as an “Amazon factor”

E-E-A-T is Google’s term, not a declared Amazon ranking signal — and even Google says it is not a single ranking factor. As a content-quality framework, though, it is useful:

• Experience — show the product in real use.
• Expertise — give exact dimensions, materials, compatibility and limitations.
• Authoritativeness — keep brand and product information consistent everywhere.
• Trust — no unverified specifications, no invented certifications, no contradictions between bullets, images and Q&A.

For Rufus this is concrete, because reviews and Q&A are part of what it reads. “Leakproof” in the title next to dozens of reviews about leaks is not an advantage — it is a conflict in the evidence.

Catalog attributes before polished prose

On Amazon you do not control the page’s Schema.org markup, so “structured data” here means catalog attributes and product type definitions: material, dimensions, capacity, compatibility, colour, intended use and the other fields of your product type. The Selling Partner API exposes these definitions, including required and conditional attributes.

Amazon does not promise “+X positions” for a filled field. The value is different: the system gets an exact, machine-readable fact instead of having to guess it from marketing copy. The rule:

If a fact exists as a catalog attribute, fill the attribute first — then explain the benefit in the visible copy.

Diagnose by funnel stage

Optimize the whole funnel, not rank in isolation. Search Catalog Performance shows impressions, clicks, add-to-carts, purchases and conversion rates, which lets you find the stage that is actually failing:

That is far more useful than reacting to a dropped position by adding three more copies of the same keyword.

Before and after: two listings

Both examples are hypothetical and deliberately simplified. Every claim in a real listing must match the real product.

Insulated bottle

Before:

Stainless Steel Water Bottle 32 oz Water Bottle Insulated Water Bottle Leakproof Bottle Sports Water Bottle Hiking Bottle Travel Bottle

The product is recognisable, but almost the whole title is spent on synonyms. Neither a shopper nor a semantic system learns anything new.

After:

Brand X Insulated Stainless-Steel Bottle, 32 oz — Leakproof Straw Lid, Cup-Holder Friendly, for Travel & Hiking

And the bullets, instead of “HIGH QUALITY WATER BOTTLE — our quality water bottle is the best water bottle…”, become:

• Leak-resistant straw lid — designed to be carried upright in a backpack or gym bag.
• 32 oz capacity — fewer refills during work, travel or a hike.
• Cup-holder fit — state the exact base diameter and what it fits.
• Double-wall insulation — state a tested duration, only if it has actually been tested.
• Cleaning — list which parts are dishwasher safe, only if the manufacturer confirms it.

Each line adds a new potential product → function, product → audience or product → context relationship, or a verifiable attribute — exactly the kind of relationship COSMO works with.

USB-C hub

Before:

USB C Hub USB C Adapter Multiport USB C Dongle USB Adapter Laptop Adapter 7 in 1

After:

Brand X 7-in-1 USB-C Hub — HDMI 4K, 100W Power Delivery, SD/microSD, 3× USB, for Compatible USB-C Laptops

The bullets then carry semantic coverage rather than more USB C: supported display modes, maximum power-delivery passthrough, operating-system compatibility, port limitations, and the real scenarios — home office, travel, external monitor.

Limitations are often the most valuable lines on the page. They reduce wrong purchases and raise trust, and an assistant can answer “will this run two displays on my laptop?” only if the listing states compatibility plainly instead of hiding the gap behind another USB-C hub.

Test hypotheses, not word order

Manage Your Experiments randomly shows shoppers different versions of titles, images, descriptions, bullets and A+ Content, and reports conversion, units sold, sales, units per unique visitor, sample size and the probability that one version is better. Amazon itself recommends making versions different enough for the test to detect a real effect — swapping two words rarely qualifies.

Metrics and rollout

The most common mistake is a single KPI such as “organic rank on 20 keywords”. The new model needs a query-funnel scorecard:

Best Sellers Rank is not a substitute for a search KPI: Amazon defines it as a sales rank relative to the same category, separate from search ranking.

For a portfolio of ASINs we run this as a loop:

1. Baseline. Record Search Query Performance and Search Catalog Performance, and split queries into intent clusters: category, feature, problem, audience, occasion, compatibility.
2. Semantic map. Describe the six layers of the product.
3. Attribute audit. Fill missing catalog facts and remove contradictory ones.
4. Rewrite. Title, bullets and A+ — one clear primary entity, wider semantic coverage.
5. Evidence matrix. For every claim: source of proof → where it is shown → where it could be contradicted. For example: dishwasher safe → manual or test → bullet and image → watch reviews and Q&A.
6. Experiment. On high-traffic ASINs test the title and value proposition first, then bullets, then images and A+.
7. Analyse and repeat. Click rate, add-to-cart and conversion by query cluster; roll out the winner or return to the semantic map.

ASINs without enough traffic are not eligible for Manage Your Experiments — Amazon limits it to listings that can produce valid results. For those, compare sequential periods carefully and treat price, stock, ads, promotions and seasonality as confounders.

Risks worth planning for

Optimizing “for the AI”. The most dangerous outcome is replacing keyword stuffing with semantic stuffing: listing dozens of audiences and use cases. COSMO is built to capture plausible commonsense relationships, so writing that a product is “for camping, pregnancy, seniors, children, office, military…” without an objective basis is the same mistake in a new form.

Hallucinated claims. Copy generated by an LLM easily adds specifications the product does not have. Keep a source of truth: every size, material, certification, compatibility and performance claim needs internal proof.

Conflicting sources. Rufus can read the catalog, reviews and Q&A together. Look not only for missing keywords but for contradictions: the title says quiet, reviews keep saying loud; a bullet promises compatibility, a Q&A answer denies it.

Clicks without conversion. An aggressive title or image can lift clicks and attract the wrong buyers. Never optimize click rate without conversion, returns and review sentiment as guardrails.

False confidence in COSMO. Public data shows COSMO is used in Amazon search applications and improves relevance models in research. It does not show that COSMO is the main production ranker, just as Rufus is not a published replacement for search ranking. “We optimized for COSMO” should mean “we improved semantic relationships and catalog clarity” — not “we know Amazon’s secret algorithm”.

Checklist before you publish an updated ASIN

• Entity. The first words make it obvious what is sold; the primary keyword appears naturally.
• Attributes. Product type, dimensions, material, capacity, compatibility and other relevant fields are filled and consistent.
• Semantic coverage. Functions, problems, audiences and realistic use cases are described — without listing everything imaginable.
• Information gain. Every bullet adds a new fact instead of rephrasing the previous one.
• Evidence. Every quantitative, performance or certification claim is backed.
• Assistant readiness. Typical questions — will it fit, is it compatible, how is it different, what are the limits — can be answered from facts on the page.
• Consistency. Title, bullets, attributes, images, A+, manual, reviews and Q&A do not contradict each other.
• Readability. A person can read the title; images explain size, use and benefits; nothing important is buried in keyword soup.
• Measurement. Impressions → clicks → add-to-carts → purchases are recorded before the change.
• Testing. Each change has a hypothesis, a primary KPI and a guardrail defined in advance.

The formula in one line: keywords give the system a vocabulary, attributes give it facts, semantic relationships explain meaning, evidence creates trust, clarity turns relevance into clicks, and conversion confirms the right product was found.

So stop asking “how many times should the keyword appear?” and start asking: which question about this product can Amazon still not answer confidently from our catalog and our listing?