Key takeaways
Rufus is Amazon's generative-AI shopping assistant, built on AWS Bedrock and grounded by the COSMO commonsense knowledge graph. It understands shopper intent, not just keywords, so listings increasingly compete on how well they match a real buying need. Amazon confirms the assistant and its semantic stack; the precise ranking impact remains industry inference. Write clear, human-readable, intent-rich listings.
- Rufus is Amazon's confirmed AI shopping assistant, powered by AWS Bedrock (Amazon Nova, Anthropic Claude Sonnet, and a custom Amazon model) plus retrieval over Amazon's catalog and reviews.
- COSMO is Amazon's real, peer-reviewed commonsense knowledge graph (6.3 million nodes, 29 million edges) that maps products to the contexts and intents they serve.
- The shift is from lexical keyword-match toward semantic intent-match, so clear human-readable listings can outperform old keyword-stuffed ones.
- Amazon confirms Rufus, Bedrock, and COSMO exist; the term "A10" and any precise ranking weights are seller-community inference, not official.
- In May 2026 Amazon began unifying Rufus and Alexa+ into "Alexa for Shopping," retaining Rufus's recommendation capabilities under the new brand.
What is Amazon Rufus?
Definition
Rufus is Amazon's generative-AI shopping assistant, built on AWS Bedrock and grounded by the COSMO commonsense knowledge graph, designed to understand a shopper's intent rather than only matching keywords. It answers shopping questions, compares products, surfaces recommendations, and explains product details and reviews inside the Amazon Shopping app and website.
For sellers, the significance is structural. Rufus reads your listing, your reviews, and your customer Q&A as source material it can quote, compare, and summarize back to a shopper. That moves the contest from how many keywords a listing contains toward how clearly it communicates what a product is and who it is for. This guide is part of our broader explainer on Amazon's ranking algorithms, and it complements our deep dives into the A9 algorithm and the A10 algorithm.
Rufus and Bedrock at a glance
- What Rufus is
- Amazon's generative-AI shopping assistant
- Launched
- Early 2024, then to all U.S. customers
- Runs on
- AWS Bedrock (Nova, Claude Sonnet, custom model)
- Semantic layer
- COSMO knowledge graph (SIGMOD 2024)
- COSMO scale
- 6.3M nodes, 29M edges, 18 categories
- Grounding method
- RAG over catalog, reviews, and Q&A
- Reported impact
- Users 60% more likely to complete a purchase
- 2026 status
- Unifying into "Alexa for Shopping"
What Amazon has actually confirmed
Amazon publicly confirms the following building blocks of its generative-AI shopping layer:
- Rufus is an AI shopping assistant in the Amazon Shopping app and on the website, introduced in early 2024 and rolled out to all U.S. customers, designed to answer shopping questions, compare products, surface recommendations, and explain product details and reviews.
- AWS Bedrock is Amazon's fully managed service for building generative-AI applications on top of foundation models. Amazon confirms Rufus runs on Bedrock and uses a mix of models including Amazon Nova, Anthropic's Claude Sonnet, and a custom Amazon model trained on store knowledge.
- COSMO is Amazon's commonsense knowledge graph, published by Amazon researchers at SIGMOD 2024. It contains 6.3 million nodes and 29 million edges mapping products to attributes, use cases, and shopper intents across 18 categories.
- Scale and behavior: Amazon reports more than 250 million customers used Rufus in a year, and states that customers who use Rufus during a shopping journey are 60% more likely to complete a purchase.
What Amazon does not publish is its product-search ranking algorithm. So while the existence of Rufus, Bedrock, and COSMO is fact, any claim about exactly how much they move a specific listing's rank is inference.
How Rufus is built on AWS Bedrock
Per Amazon Science and the AWS Machine Learning blog, Rufus is not a single model. It is a system that orchestrates several models through Bedrock:
- Hybrid model strategy: Amazon combines Amazon Nova, Anthropic's Claude Sonnet, and a custom in-house model. Bedrock lets Amazon select and scale these models flexibly across different query types.
- Grounding and retrieval: Rufus uses retrieval-augmented generation (RAG) over the Amazon catalog, customer reviews, and community Q&A, and calls relevant Stores APIs, so answers are grounded in real product data rather than only model memory.
- Custom training: Amazon trained a foundational Rufus model primarily on shopping data - the entire Amazon catalog, customer reviews, and community Q&A, plus public web information.
- Infrastructure: Rufus runs on AWS Trainium and Inferentia chips and uses techniques like continuous batching and token-by-token streaming to serve responses at low latency to hundreds of millions of users.
The practical takeaway for sellers: Rufus reads your listing, your reviews, and your Q&A as source material it can quote, compare, and summarize back to a shopper.
COSMO: the semantic understanding layer
Traditional Amazon search has long worked on lexical matching - indexing the words in a listing and matching them to the words in a query. COSMO adds a semantic layer on top of that.
COSMO (Common Sense Knowledge Generation and Serving System) builds a knowledge graph of entity-relation-entity triples such as "slip-resistant shoes, used_for_audience, pregnant women." It connects a product to the contexts and intents it actually serves, not just the keywords it contains. In Amazon's published evaluation, COSMO improved search relevance and produced a measurable sales uplift in A/B testing.
This matters because it changes what "relevance" means. A listing can now be matched to a query even when the exact keywords do not overlap, as long as the product clearly serves the intent behind the search. Conversely, a keyword-stuffed listing that does not clearly communicate what a product is for can lose ground to a clearer, more human-readable competitor.
From keyword-match to intent-match: what changed for listings
The combined direction of Rufus, Bedrock, and COSMO is a move toward understanding the shopper's underlying need. For sellers, this reframes listing optimization:
- Old model: cram every plausible keyword into the title and bullets so the lexical index matches as many queries as possible.
- Emerging model: write clear, specific, human-readable copy that states what the product is, who it is for, what problem it solves, and the contexts it fits - the exact signals a semantic system maps to intent.
This is why industry practitioners report that clean, well-structured listings can now outperform older keyword-stuffed ones. Important nuance: keywords still matter for the initial candidate pool - your listing still has to be indexed for a query to be considered. The shift is in what wins once you are in the pool. Treat this as a both/and (cover the relevant terms and read like a human wrote it), not an either/or.
A note on "A10" and other unofficial terms
You will see Amazon-SEO blogs reference the "A9" and "A10" algorithms. Be precise about what these labels mean:
- A9 originated as the name of the Amazon subsidiary that built early product search; it is widely used as shorthand for Amazon's lexical search ranking.
- A10 is seller-community terminology, not an Amazon product name. Amazon has never confirmed an algorithm called "A10." It is an informal label practitioners use to describe observed changes in ranking behavior, such as more weight on conversion and external traffic.
By contrast, Rufus, Bedrock, and COSMO are real, named Amazon systems with official documentation and a peer-reviewed paper. When you read claims that "the A10 algorithm now uses Rufus to rank you," treat the mechanics as practitioner inference layered on top of confirmed components - directionally reasonable, but not an official description of how ranking works. For the full breakdown of those two labels, see our guides to the A9 algorithm and the A10 algorithm.
The May 2026 shift: Rufus becomes Alexa for Shopping
In May 2026, Amazon announced it is unifying Rufus and Alexa+ into a single experience branded "Alexa for Shopping" across the Amazon Shopping app and website. Reporting at the time described the standalone Rufus chatbot brand being retired, with Amazon stating it would carry Rufus's recommendation capabilities and shopping context into the unified assistant.
For optimization purposes, the underlying generative-AI stack - Bedrock-served models grounded by catalog data and the COSMO knowledge graph - is what reads your listing, regardless of the consumer-facing brand name. New consumer capabilities announced alongside the rebrand include asking questions directly in the main search bar, AI-generated overviews on search and product pages, side-by-side comparisons, up to a year of price history, and automated actions like Auto-Buy and Scheduled Actions. The optimization principle is unchanged: clear, intent-rich, well-grounded listings are what these systems can confidently surface and recommend.
History of Amazon's AI shopping stack: a timeline
Amazon's discovery stack evolved from lexical "A9" search into a generative-AI layer built on Bedrock and grounded by the COSMO knowledge graph, and is now being unified under a single shopping assistant.
-
2003
A9 founded
Amazon launches A9.com, the subsidiary that built early product search. "A9" became shorthand for Amazon's lexical search ranking.
-
2024
COSMO published
Amazon researchers present COSMO, a commonsense knowledge graph (6.3M nodes, 29M edges), at SIGMOD 2024, formalizing semantic intent mapping for shopping.
-
2024
Rufus launches
Amazon introduces Rufus, a generative-AI shopping assistant built on AWS Bedrock, initially in beta and then to all U.S. customers.
-
2025
Rufus scales
Amazon reports Rufus helped over 300 million customers and that users are 60% more likely to complete a purchase; the assistant gains agentic actions.
-
2026
Alexa for Shopping
In May 2026 Amazon unifies Rufus and Alexa+ into "Alexa for Shopping," retaining Rufus's recommendation capabilities under a new brand.
The listing signals that matter for intent-match
Where lexical search rewarded keyword coverage, Amazon's semantic stack rewards how clearly a listing communicates what a product is, who it serves, and the problem it solves. These are the signals that feed intent-match.
| Signal | What it captures |
|---|---|
| Intent clarity | How obviously the listing states what the product is, who it is for, and the problem it solves - the inputs a semantic system maps to shopper intent. |
| Use-case coverage | Whether the copy names real contexts and audiences (e.g., "for travel," "for sensitive skin") that COSMO-style graphs connect to relevant queries. |
| Review and Q&A richness | Rufus uses RAG over reviews and community Q&A; richer, specific customer language gives the assistant more grounded material to quote. |
| Attribute completeness | Structured fields and bullet specifics give both lexical indexing and semantic matching concrete, machine-readable signals. |
| Human readability | Clean, natural copy that an LLM can summarize and compare tends to be surfaced more confidently than keyword-stuffed text. |
The practical takeaway is that covering the right terms still matters for indexing, but clarity is what wins once your listing is in the running. Pairing those listing fundamentals with Amazon Sponsored Ads can accelerate the conversion signals these systems read.
How to optimize for Rufus and intent-match
To optimize for Amazon's generative-AI layer, write clear intent-rich copy, name concrete use cases and audiences, complete every structured attribute, and enrich the reviews and Q&A that Rufus retrieves and quotes.
-
Lead each listing with a plain-language statement of what the product is and the core problem it solves
Semantic systems map listings to intent; a clear value statement is the strongest intent signal you can give.
-
Name concrete use cases and audiences in titles, bullets, and A+ content (occasions, environments, skill levels, who it is for)
COSMO-style graphs connect products to the contexts they serve, so explicit context expands the intents you can match.
-
Replace keyword stuffing with natural, scannable copy that still covers the relevant terms
Practitioner evidence and Amazon's semantic direction both favor human-readable listings; keywords still feed indexing but no longer win on volume alone.
-
Strengthen reviews and community Q&A with specific, real customer language
Rufus retrieves and quotes reviews and Q&A via RAG, so this content directly shapes what the assistant tells shoppers.
-
Complete every structured attribute field accurately
Filled attributes give both lexical search and semantic matching unambiguous, machine-readable facts to rank and recommend on.
-
Anticipate the questions Rufus will be asked (comparisons, compatibility, fit) and answer them on the page
Listings that pre-answer buying questions give the assistant grounded material to recommend you confidently.
Rufus and Bedrock myths vs. reality
Few Amazon topics generate as much confident misinformation as the generative-AI stack. Here are the most common myths and what is actually true.
Myth Amazon runs an official algorithm called "A10."
Reality "A10" is seller-community terminology. Amazon has never confirmed an algorithm by that name; it is an informal label for observed ranking behavior, not an official system.
Myth Rufus replaces keyword optimization entirely.
Reality Keywords still feed the initial candidate pool through lexical indexing. The shift is that intent-match increasingly decides what wins once you are in the pool - so do both.
Myth More keywords always mean more visibility.
Reality With semantic understanding via COSMO, clear human-readable listings can outperform keyword-stuffed ones. Coverage still matters, but clarity now competes with volume.
Myth Rufus and COSMO are the same thing.
Reality Rufus is the consumer-facing AI assistant built on Bedrock. COSMO is a separate commonsense knowledge graph that provides semantic understanding. They are complementary, not identical.
Myth We know exactly how much Rufus and COSMO move a listing's rank.
Reality Amazon does not publish its ranking algorithm. The existence of these systems is confirmed; their precise ranking weight on any given listing is industry inference.
Frequently asked questions
Rufus is Amazon's generative-AI shopping assistant in the Amazon Shopping app and website, introduced in 2024. It answers shopping questions, compares products, surfaces recommendations, and explains product details and reviews. Amazon built it on AWS Bedrock and grounds its answers in catalog data, customer reviews, and community Q&A.
AWS Bedrock is Amazon's managed service for building generative-AI applications on foundation models. Amazon confirms Rufus runs on Bedrock and uses a mix of models, including Amazon Nova, Anthropic's Claude Sonnet, and a custom Amazon model, selected flexibly across query types to serve hundreds of millions of shoppers.
COSMO is Amazon's commonsense knowledge graph, published by Amazon researchers at SIGMOD 2024. It maps products to the contexts, attributes, and intents they serve using roughly 6.3 million nodes and 29 million edges, helping Amazon match listings to shopper intent rather than only matching keywords.
Amazon does not publish its ranking algorithm, so any precise ranking effect is industry inference. What is confirmed is the direction: Amazon's generative-AI stack understands intent and reads your listing, reviews, and Q&A as source material, which rewards clear, human-readable, intent-rich content.
No. "A10" is seller-community terminology, not an official Amazon product name. Amazon has never confirmed an algorithm called A10. It is an informal label practitioners use to describe observed ranking behavior. By contrast, Rufus, Bedrock, and COSMO are real, named, documented Amazon systems.
No. Keywords still feed the initial candidate pool through lexical indexing, so your listing must still be indexed for relevant terms. The change is that semantic intent-match increasingly decides what wins once you are in the pool, so write naturally and cover relevant terms - do both, not one or the other.
In May 2026 Amazon announced it is unifying Rufus and Alexa+ into a single experience called "Alexa for Shopping" across the Amazon Shopping app and website. Reporting described the standalone Rufus brand being retired while Amazon retained Rufus's recommendation capabilities and shopping context within the new assistant.
Lead with a plain statement of what the product is and the problem it solves, name concrete use cases and audiences, complete every structured attribute, and enrich reviews and Q&A with specific customer language. These give semantic systems the grounded, intent-rich signals they need to surface you.
The bottom line
Bottom line
Rufus, Bedrock, and COSMO are real, documented Amazon systems that push product discovery from keyword-match toward intent-match. The exact ranking weights stay private, but the direction is clear and the optimization response is durable: cover the relevant terms for indexing, then write listings, reviews, and Q&A that read like a human explaining what the product is and who it is for. Clarity is what these systems can confidently surface and recommend.
References
- How to use Amazon Rufus - About Amazon
- How Rufus scales conversational shopping with Amazon Bedrock - AWS Machine Learning Blog
- The technology behind Amazon's GenAI-powered shopping assistant, Rufus - Amazon Science
- COSMO: A Large-Scale E-commerce Common Sense Knowledge Generation and Serving System at Amazon (SIGMOD 2024)
- Amazon Bedrock - AWS
- Alexa for Shopping: Amazon's AI assistant for personalized shopping - About Amazon
- Amazon ditches Rufus chatbot, launches Alexa shopping agent - CNBC (May 2026)