Rufus and COSMO change Amazon SEO in one main way: Amazon now matches products to what a shopper means, not only to the words they type. For Amazon Rufus COSMO SEO, that means listings that state plainly who a product is for, what it's used for and what it's made of give Amazon's AI something to match and quote. Keyword placement still decides whether you appear in regular search at all.
Two facts set the scene. COSMO is a knowledge system Amazon researchers presented at SIGMOD 2024, built to connect searches like "shoes for pregnant women" with products like slip-resistant shoes. Rufus, the shopping assistant Amazon launched in February 2024, was renamed Alexa for Shopping on May 13, 2026 and now answers questions inside the main Amazon search bar in the US.
This guide covers what Amazon has published about both systems, what changed for sellers, what stayed the same, and a before and after rewrite checked against Amazon's listing rules. It's written for sellers and agencies with live listings, not for people choosing what to sell.
Key Takeaways (TL;DR)
- COSMO maps products to uses, audiences and occasions. Amazon built it from search-then-buy and bought-together behavior.
- Rufus is now Alexa for Shopping (since May 13, 2026, US). Shoppers can type questions straight into the Amazon search bar.
- Amazon says the assistant draws on the product catalog, customer reviews, community Q&As and the web.
- Keyword search hasn't gone. A listing still has to be indexed for a phrase to appear in regular results, and titles are capped at 75 characters.
- State facts the AI can quote: materials, sizes, who it's for, where it's used. "Premium quality" gives it nothing.
- Writing for intent can drift into health or safety claims. Keep every use case inside what the product and spec table can back up.
- Amazon hasn't published how the assistant picks products, and it hasn't said COSMO powers Rufus. Treat any detailed "ranking factor" list with care.
Amazon Rufus COSMO SEO: at a Glance
| Topic | Rufus (now Alexa for Shopping) | COSMO |
|---|---|---|
| What it is | The AI assistant shoppers talk to | A knowledge graph of commonsense product relationships |
| Where shoppers see it | Search bar, chat window, AI overviews on search and product pages | They don't. It works inside search and recommendations |
| What it reads | Catalog, reviews, community Q&As, the web | Search-then-buy and bought-together behavior |
| First published | February 1, 2024 (beta) | SIGMOD conference, June 2024 |
| Status in 2026 | Renamed Alexa for Shopping, May 13, 2026 (US) | Deployed in Amazon search, per the research paper |
| What sellers control | Listing facts, A+ Content, images, review quality | The facts that let a product fit a use, audience or occasion |
| What sellers can't see | How products are picked for an answer | The graph itself |
What Is Amazon COSMO?
Amazon COSMO (Common Sense Knowledge Generation and Serving System) is a system that teaches Amazon's search and recommendations the kind of reasoning a shop assistant does without thinking. Amazon's researchers give the example of a shopper searching "shoes for pregnant women": the store should work out that slip-resistant shoes are a good match, even if no listing mentions pregnancy.
The research paper describes it as an industry-scale system deployed in Amazon search. It covers 18 major product categories.
How the COSMO algorithm builds its knowledge
The Amazon COSMO algorithm starts from shopper behavior, not from listing copy. Amazon's write-up names two inputs: query-purchase pairs (what someone searched, then bought) and co-purchase pairs (what was bought in the same session). A large language model is asked why each pair makes sense, and human reviewers and filters throw out weak answers like "people bought them together because they like them."
What survives becomes a set of relationships such as *used_for_function*, *used_for_event* and *used_for_audience*. A camera case and a screen protector bought together become "capable of protecting a camera." Slip-resistant shoes get linked to pregnant women as an audience.
What COSMO changed in Amazon's tests
In Amazon's experiments on its public Shopping Queries dataset, adding COSMO relationships to a relevance model raised macro F1 by 60% over the best baseline when the models weren't fine-tuned. After fine-tuning, the COSMO model still led by 28% on macro F1 and 22% on micro F1.
Those are lab numbers on a test dataset, not ranking gains for any seller. What they show is direction. Amazon's relevance models now have a second source of meaning beyond the words in a query and a listing.
From Rufus to Alexa for Shopping: What Changed in 2026
Rufus launched in beta on February 1, 2024, described by Amazon as an assistant trained on its product catalog, customer reviews, community Q&As and information from across the web. On a product page, Amazon said Rufus answers from listing details, reviews and Q&As. That's the part sellers can shape.
It grew fast. Amazon says Rufus helped over 300 million customers in 2025. On the Q1 2026 earnings call, Andy Jassy said monthly active users were up 115% and engagement up 400% year over year, and Amazon had earlier put Rufus at nearly $12 billion in incremental annualized sales.
What Alexa for Shopping added
On May 13, 2026, Amazon merged Rufus with Alexa+ into Alexa for Shopping. The changes that matter for listings, from Amazon's announcement:
- Shoppers can ask questions in the main search bar, which recognizes a question and hands it to the assistant.
- AI overviews appear at the top of search results in the app and on product detail pages.
- Shoppers can select products in search results and get a side-by-side comparison of features, prices and reviews.
- Price history now covers up to a full year.
- The assistant uses each shopper's history and Alexa conversations to personalize answers.
The launch covers US customers. Amazon's announcement doesn't give dates for other storefronts.
Why the search bar change matters most
Before May, Rufus sat in its own chat panel, and a shopper had to choose to open it. Now the same box a shopper has always typed keywords into also takes questions like "best work shoes for nurses." The assistant's answer and the regular results share the page.
That puts AI answers in front of shoppers who never meant to use an assistant. A listing that the assistant can't describe or compare loses ground on the first screen, whatever its organic rank.
How Rufus and COSMO Change Amazon SEO for Listings
Classic Amazon SEO asks one question: does my listing contain the words people type? Amazon AI search adds a second: can Amazon tell what my product is for, and who should buy it? Five changes follow from that.
1. Relevance now includes use, audience and occasion
COSMO's relationships are about context: what a product does, for whom, at what event. A query like "gifts for new moms" (about 1,084 Amazon searches a month in the US) or "running shoes for flat feet" (≈822) describes a need, not a product.
Your listing can't appear for those through the graph alone if Amazon can't tell what the product is. Clear attributes and plain statements of use give the system something to connect.
2. Answers draw from reviews, not only your copy
Amazon says the assistant answers product questions from the listing, reviews and Q&As. If 40 reviews say a backpack's straps dig in, a shopper asking "is this comfortable for long hikes?" can get that back, whatever your bullets claim.
That makes product problems an SEO problem. A recurring complaint in reviews is something to fix in the product or address honestly in the listing, since the assistant will surface it either way.
3. The shopper may never read your bullets
AI overviews and side-by-side comparisons summarize products before anyone opens a detail page. A comparison pulls the same fields from each product. If your listing leaves material, size or compatibility blank, your column is the one with gaps.
4. Vague copy has nothing to quote
"Premium quality" and "all-day comfort" match nothing a shopper asks. "1-inch heel" and "removable memory foam insole" answer specific questions. Listings written to fit a character count with adjectives give the assistant less to work with than listings written as facts.
5. Paid placement now reaches the assistant
Amazon has started selling ad placements inside AI conversations. Sponsored Prompts became billable on a cost-per-click basis on March 25, 2026, and Jassy said nearly 20% of shoppers who interact with a brand prompt keep talking to the assistant about that brand. Organic visibility in answers now competes with paid placements, the same way organic rank always has. Sponsored Prompts are bought in Amazon Ads; this guide covers the organic side, the listing content the assistant reads.
What Hasn't Changed: Keyword Search Still Decides Indexing
Most shoppers still type short phrases, and regular search results still depend on indexing. If a shopper types "non slip shoes for women" (≈1,570 searches a month) and none of those words sits in your title, Item Highlights, bullets, backend terms or attributes, you aren't in the pool. COSMO adds meaning on top of that match; it doesn't replace it.
The rules for those fields still apply in full. Titles in most categories are capped at 75 characters since July 27, 2026, Item Highlights adds 125 searchable characters, and backend search terms have their own byte limit. Our guide on how to rank for keywords on Amazon covers the indexing test and the field limits step by step.
So the job has two layers. Keywords get you indexed; facts about use and audience make you understandable to Amazon's AI. Dropping the first to chase the second loses regular search traffic you already have.
How to Optimize Your Amazon Listing for AI Search
Amazon Rufus optimization, now aimed at Alexa for Shopping, is mostly the listing work you already do, written with more facts and fewer adjectives. These six steps run in order of effort.
Fill every attribute with an accurate value
Attributes are the structured facts Amazon can read without interpreting prose: material, size, closure type, intended use, target audience. They feed search filters and give comparisons something to line up. Open the listing in Seller Central and fill every dropdown that applies, with the true value.
The spec table on your product page is where those values show up. Amazon SEO Ranking's AI rewrites use it as the source of truth, so a rewrite never says "leather" when the spec table says "synthetic". If a value is missing there, add it in Seller Central first; the copy can only state what the spec table backs.
Write bullets as answers to buyer questions
List the questions a buyer asks in a store. What's it made of? Will it fit my pan? Can I wash it? Then answer each in a bullet header and one plain sentence. A bullet like "Fits 13 x 18 inch half sheet pans" gives the assistant a quotable answer, and it still carries keywords for regular search.
Keep bullets under 200 characters so they show in full on mobile; Amazon allows up to 255. The free Amazon character limit checker shows where each field gets cut. If you rewrite with the AI Copy Rewriter, Ask AI can redo a single bullet, so you can turn one vague bullet into an answer without touching the rest.
Name the use, audience and occasion, within what you can prove
Say who the product is for and where it's used: "for nurses, servers and retail staff on their feet 8 to 12 hours" tells Amazon an audience and a context. This is the place to be careful. Use cases slide easily into claims, and "non-slip", "for pregnancy" or "relieves back pain" are safety and health claims that can get flagged or suppressed. The worked example below shows where that line fell.
Write your approved facts and banned claims down once, in a brand fact sheet. Amazon SEO Ranking applies yours to every rewrite, Ask AI edit and bulk run, and flags health claims and claims the spec table doesn't support as you edit. Then score the new copy as a draft audit, against the live listing's images, rating and Prime status, before anything goes live.
Use A+ Content and images for the facts that don't fit
Amazon's original Rufus announcement says answers come from listing details, and A+ Content is part of the detail page. Use it for comparison charts, dimensions and care instructions. The A+ Content generator drafts modules and a brand story on every plan, including Free.
Amazon's announcement also lists visual search among the assistant's features, so image text should be readable and the main image still has to follow Amazon's rules. Our page on Amazon product image requirements lists them with passing and failing examples. The AI Vision Image Audit on Pro reviews the gallery itself and flags unreadable infographic text and missing lifestyle or scale shots.
Read your reviews the way the assistant does
Sort reviews by most recent and note the three most common complaints. Each is a likely answer the assistant gives. Fix what you can in the product, and answer the rest in the listing with facts: if shoppers say the sizing runs small, a sizing note in the bullets helps both the shopper and the assistant's answer.
The audit's Reviews & Social Proof criterion fails below a 4.3 average rating and tells you to address recurring complaints. Treat that as a prompt to read the reviews, since the rating is the part a listing edit can't move directly.
Check each marketplace separately
Alexa for Shopping launched in the US, but shoppers in every storefront read your listing, and each has its own copy and its own language. Audit the listing in each storefront you sell in rather than copying one version across. Amazon SEO Ranking's audit covers 23 Amazon storefronts, and rewrites come back in the marketplace's language, so an amazon.de listing gets German copy and an amazon.com.be listing gets French.
For agencies and brand teams, the work is catalog-wide. Bulk Optimization audits and rewrites up to 100 ASINs per run on Pro and 500 on Agency, each row in its own marketplace language, with a claim check column on every AI draft. Comparing a client's or brand's audited listings side by side then shows which one scores best on AI-Readiness, so you can use it as the template for the rest.
Worked Example: Rewriting a Work Shoe Listing
Take a fictional listing, Halden women's slip-on work shoes. We ran each version through Amazon SEO Ranking's listing copy check on 28 October 2026. The length limits, consistency checks and risky backend terms are free; the AI compliance scan is Pro. Search volumes are US Amazon estimates from the same tool.
The original title, at 96 characters:
Halden Shoes Women BEST Comfortable Non Slip Shoes Premium Quality Work Shoe for All Day Comfort
The bullets opened with ALL CAPS headers like "PREMIUM QUALITY" and "100% SATISFACTION". The check found:
| Problem | What the check said | Found by | Why it matters for AI search |
|---|---|---|---|
| Title length | 96 characters against a 75 limit | Length check (free) | Amazon may rewrite it and choose which words survive |
| "100% SATISFACTION" | High severity: reads as a satisfaction guarantee | AI compliance scan (Pro) | Can trigger suppression, and it's a fact the assistant can't use |
| "BEST", "finest materials" | Medium: unsubstantiated superlatives | AI compliance scan (Pro) | Says nothing about the product a shopper could ask about |
| "Non Slip" in the title | Medium: a safety claim with no slip-resistance standard named | AI compliance scan (Pro) | A claim the listing can't back |
| ALL CAPS headers | Low: against formatting rules | AI compliance scan (Pro) | Readability for shoppers and AI alike |
First rewrite: too far toward intent
The first rewrite tried to write for COSMO's example audience. It said the slip-on design helps "in the third trimester" and the tread "channels water away" on wet floors.
The AI compliance scan flagged both at low severity. The pregnancy line read as a health benefit, and the water claim implied slip-resistance performance without a test standard. This is the trap in writing for intent. COSMO's research connects pregnant shoppers to slip-resistant shoes from what they bought, so the seller doesn't need to make that link in the copy.
It's also what a general AI assistant tends to write when you ask it to "rewrite this for Rufus". ChatGPT or Claude used on its own doesn't see the listing's spec table, your approved facts or Amazon's claim rules, so it adds the use case and the claim together. Through Amazon SEO Ranking's Claude connector on Pro and Agency, Claude can run these same checks on a draft from inside the chat.
Final rewrite: facts only
- Title (67 characters): Halden Women's Slip-On Work Shoes, Rubber Grooved Outsole, Wide Fit
- Item Highlights (85 characters): Low 1-inch heel and removable memory foam insole for long shifts standing and walking
- Bullets: grooved rubber outsole for kitchen, hospital and shop floors; memory foam insole and 1-inch heel for staff on their feet 8 to 12 hours; slip-on, no laces; medium and wide widths; wipe-clean synthetic upper
- Backend (111 bytes): nurse server kitchen restaurant hospital clog loafer comfortable standing all day memory foam wide width ladies
Every field passed its limit, and the AI scan returned no flags. The rewrite states heel height, insole, widths, audience and workplace, each one backed by the spec table, plus the outsole and upper materials the original never named. On a live listing, the next step is a draft audit: it scores this copy against the listing's real images, rating and Prime status, and shows the change from the last live score.
The keyword the rewrite left out
We first tried "non slip slip resistant" in the backend, since "non slip shoes for women" gets about 1,570 Amazon searches a month in the US. The AI scan flagged it there too, at low severity: the same safety claim, hidden from shoppers but still in the listing, with no slip-resistance rating to back it.
| Search phrase | Est. monthly US searches | Decision |
|---|---|---|
| non slip shoes for women | ≈1,570 | Left out until the shoe has a documented slip-resistance rating |
| slip resistant shoes women | ≈112 | Same |
| shoes for pregnant women | ≈74 | Left to Amazon's behavior data |
That's a real trade-off, and the right call depends on the product. A shoe tested to a slip-resistance standard should say so in the bullets and can target those phrases. One that isn't tested gives up the traffic rather than carry a claim it can't back. "Shoes for pregnant women", at about 74 typed searches a month, mostly reaches Amazon through behavior and conversation, which no keyword tool measures.
What Amazon Hasn't Said Yet
Amazon has published how COSMO is built and what Alexa for Shopping does. It hasn't published how the assistant chooses which products to recommend, and the COSMO paper doesn't mention Rufus. Many seller guides state that COSMO powers Rufus; that's a reasonable guess, not something Amazon has confirmed.
There's also no report in Seller Central that shows how often your product appears in AI answers. Until there is, judge changes by what you can measure: indexing for your keywords, sessions, conversion and the questions shoppers ask. Be wary of any tool or agency selling a precise "COSMO score" or a ranked list of Rufus factors without saying how it's calculated, since Amazon hasn't published one.
What an AI-readiness check can honestly measure
A tool can't see inside Amazon's assistant. It can check whether a listing gives the assistant something to work with, and it should tell you exactly what it checks. Amazon SEO Ranking's audit has two AI-Readiness criteria, out of 12 equally weighted criteria with published pass rules:
- Intent-driven natural language: does the copy say who the product is for and which problems it solves, in the words a shopper would use?
- Extractable factual content: does the copy state concrete facts an AI can quote, such as dimensions, materials, compatibility and model numbers, rather than vague claims like "higher quality"?
An audit of a Canon all-in-one printer on amazon.ie shows how that reads in practice. It scored 50 on AI-Readiness: its bullets gave print speeds and resolution, but the description said "more efficient" with no comparison, and nothing said who the printer suits or which connectivity, size or cartridge it uses. Those are facts a shopper asks an assistant, and the listing doesn't answer them.
Everything You Need to Know About Amazon Rufus COSMO SEO
| Topic | What to know | Rule or number |
|---|---|---|
| COSMO | Knowledge graph of uses, audiences, occasions | Built from search-then-buy and bought-together data |
| COSMO results | Better relevance in Amazon's own tests | +60% macro F1 (frozen), +28% (fine-tuned) |
| Rufus | Amazon's AI shopping assistant | Launched February 1, 2024 |
| Alexa for Shopping | Rufus merged with Alexa+ | May 13, 2026, US |
| Reach | Rufus users in 2025 | 300 million+ |
| Where it shows | Search bar, chat, AI overviews, comparisons | Search results and product pages |
| Its sources | Catalog, reviews, community Q&As, the web | Per Amazon's announcement |
| Paid placement | Sponsored Prompts | Billable since March 25, 2026 |
| Indexing | Still required for regular search | Words must sit in an indexed field |
| Title | Main keyword first, facts not adjectives | 75 characters |
| Bullets | Answers to buyer questions | Under 200 characters for mobile, 255 max |
| Claims | Use cases only where the product backs them | No health, safety or guarantee claims without proof |
| Unknowns | How products are picked for answers | Not published by Amazon |
Check Your Listing's AI-Readiness, Free
Writing for Amazon's AI is where generic AI copy goes wrong: it adds the use case and the unsupported claim in the same sentence. Amazon SEO Ranking's rewrites are grounded in the listing's own spec table and your brand fact sheet, then checked in code for Amazon's length limits, pack counts, health claims and claims the spec table doesn't support. On Pro, the AI compliance scan runs on every rewrite too, the scan that caught the pregnancy line and the "non slip" backend terms above. You can score the result as a draft before it goes live.
Every audit scores the listing against 12 published criteria, two of which measure AI-Readiness for Alexa for Shopping (formerly Rufus) and COSMO. You see which criterion fails and why, with the pass rules in the open.
It's built for sellers and agencies improving listings that are already live, and for brand teams managing a catalog across storefronts: the audit covers 23 Amazon marketplaces, rewrites come back in each marketplace's language, and Bulk Optimization handles up to 500 ASINs per run on Agency. It's trusted by eRetail Search specialists at Nestlé and eCommerce SEO agencies such as Reach AI.
The free plan includes 3 full audits a month with no credit card required. Run a free Amazon listing audit on the ASIN you want to check, try the AI Amazon listing generator on the parts that fail, or compare plans and pricing: Pro is $79 a month with a 7-day free trial and adds the AI compliance and image scans.
FAQ
What is Amazon COSMO?
Amazon COSMO is a commonsense knowledge system that links products to the uses, audiences and occasions shoppers have in mind. Amazon researchers presented it at the SIGMOD conference in June 2024 and describe it as deployed in Amazon search. It's built from search-then-buy and bought-together behavior, which a language model explains and human reviewers filter. In Amazon's tests, adding COSMO data raised a relevance model's macro F1 score by 60%.
Is Rufus still available on Amazon?
Rufus is no longer available under that name in the US, because Amazon renamed it Alexa for Shopping on May 13, 2026. The new assistant combines Rufus's product knowledge with Alexa+ and the shopper's history. Shoppers can now ask it questions in the main Amazon search bar as well as in a chat window. It's free for any signed-in US customer, with no Prime membership or Echo device needed.
How do I optimize my Amazon listing for Rufus?
You optimize an Amazon listing for Rufus, now Alexa for Shopping, by stating plain, checkable facts about the product in every field the assistant reads. Fill all attributes, write bullets that answer buyer questions, and name the audience and setting the product is made for. Keep the main keyword in the first words of a 75-character title so the listing stays indexed for regular search. Avoid health, safety and guarantee claims you can't prove, since compliance checks flag them.
Does COSMO replace the A9 algorithm?
COSMO doesn't replace the A9 algorithm; it adds commonsense relationships to the relevance models that Amazon search already uses. A listing still has to contain a query's words in an indexed field to appear in regular results for that query. Amazon's paper describes COSMO data as an extra input to a model that also reads the query and the product. Sellers who drop keywords to write only for intent lose regular search traffic.
What data does Alexa for Shopping use to recommend products?
Alexa for Shopping uses the product catalog, customer reviews, community Q&As and information from across the web, according to Amazon. It also uses each shopper's shopping history, browsing and Alexa conversations to personalize answers. On a product page, answers come from listing details, reviews and Q&As. Amazon hasn't published how it weighs these sources when picking which products to recommend.
Do keywords still matter on Amazon with AI search?
Keywords still matter on Amazon with AI search, because regular search results still depend on a listing being indexed for the words a shopper types. "Non slip shoes for women" alone gets about 1,570 Amazon searches a month in the US. Titles are capped at 75 characters, so the main keyword belongs at the front. AI search adds a second job, describing use and audience, on top of keyword placement.
Can I pay to appear in Alexa for Shopping answers?
You can pay to appear in Amazon's AI shopping conversations through Sponsored Prompts, which became billable per click on March 25, 2026. Andy Jassy said on the Q1 2026 earnings call that nearly 20% of shoppers who interact with a brand prompt keep talking with the assistant about that brand. Paid prompts sit alongside organic answers, which still come from listing facts and reviews. Ads don't fix a listing the assistant can't describe.
Isn't optimizing for Rufus and COSMO the same as normal listing work?
Optimizing for Rufus and COSMO is mostly the same listing work, with a heavier weight on facts and a lighter one on adjectives. The difference shows up in three places: attributes, bullets written as answers, and reviews, which the assistant quotes. In our worked example, the rewrite that scored clean with the AI compliance scan stated seven checkable facts, from heel height to widths, where the original relied on words like premium and comfortable. A free audit shows whether your listing passes the two AI-Readiness criteria, intent-driven language and extractable facts, before you change anything.
Audit the listing you plan to rewrite and see how it scores on the two AI-Readiness criteria, plus the other 10 published criteria. Free, no card required.
Run free audit
