How Alexa for Shopping works: what Amazon has confirmed and what it hasn't
Alexa for Shopping is not one model: Amazon says a router picks between Claude Sonnet, Amazon Nova and a custom store model that reads the catalog, reviews, community Q&A and the web. Amazon has not published how it chooses products.
The short version
Alexa for Shopping is Amazon's AI shopping assistant. It started as Rufus, in beta in the US in February 2024, and Amazon renamed it Alexa for Shopping on 13 May 2026. In Mexico it is called Alexa para Compras, and it launched there on 11 November 2026.
This guide separates three things, because they get mixed up all the time: what Amazon has said, what Amazon has not said, and what sellers repeat as fact without a source. Every "Amazon says" below comes from an Amazon page we read on 12 November 2026. The sources are at the end.
It is a system, not a single model
Amazon's own description is that the assistant uses a real-time router over a mix of models accessed through Amazon Bedrock. The models Amazon names are:
- Anthropic's Claude Sonnet
- Amazon Nova
- A custom model based on Amazon's store knowledge
The router picks the model for each type of question, to balance capability, speed and answer quality.
Amazon Science, in a 2024 write-up by the scientist who leads the effort, adds how the custom model was built:
- A language model trained from scratch on shopping data. Mainly the whole Amazon catalog, customer reviews and community question and answer posts, plus public information from the web, with the contribution of each source curated.
- Retrieval before answering. The assistant pulls information from sources Amazon considers reliable, such as reviews, the catalog and community questions and answers, and calls Amazon's store systems. How much each source matters depends on the question.
- Learning from shoppers. Thumbs up and thumbs down on answers feed reinforcement learning.
- A streamed answer filled with data. The shopper gets the first part of the answer while the rest is still being written, and the answer is populated with data by querying Amazon's internal systems.
What it reads
Put together, Amazon names these sources:
| Source | What Amazon says |
|---|---|
| Product catalog and detail pages | The base of the custom model and of retrieval |
| Customer reviews | A named source. In Mexico, comparisons include "the highlights of the reviews" |
| Community questions and answers | A named source |
| Information from the web | The custom model draws on information from across the web. In the US, Amazon named The New York Times, USA Today, Good Housekeeping and Vogue as sources for product and trend questions |
| Price and price history | Mexico shows the last 30, 60 or 365 days. The US post of November 2025 describes a 30 and 90 day tracker |
| Data from Amazon's internal systems | The streamed answer is populated by querying them |
| The shopper's own history | US account memory: kids, pets, preferences. The shopper can see, correct and delete it by chatting |
In the Mexico announcement, Amazon's own examples also mention budgets, products that "arrive on time" and options "with the best ratings from other customers". Those are the situations the assistant is built for.
What it does, by country
Amazon describes the features differently in each market, so be careful not to assume that what you read about the US applies to your marketplace.
- Everywhere it has launched (US, UK, Mexico): AI summaries in search results and product pages, and comparison of products.
- Mexico: side-by-side comparison of features, prices and review highlights, price history of 30, 60 or 365 days, and a cart the assistant prepares. The shopper completes the purchase. It is free, needs no Prime and no Echo, and is separate from the Alexa+ subscription.
- US only: account memory, price alerts, auto-buy for Prime members, Buy for Me for products outside Amazon's store, and shopping by text message, by invitation.
What COSMO is, and what it is not
You will see COSMO named in seller posts. COSMO is a paper from Amazon Science, presented at SIGMOD 2024. It describes a system that generates common-sense knowledge about products and the reasons people buy them, expanded to 18 major categories, and deployed in Amazon search applications such as search navigation.
What Amazon has not said is that COSMO is what powers Alexa for Shopping. Many in the industry read it that way, because the assistant sits on top of Amazon's search and attribute knowledge. It is a reasonable reading. It is not a statement from Amazon, so treat it as one.
What Amazon has not published
This is the part that matters most for sellers:
- How it chooses which product to recommend. No criteria, no weights, no order.
- Whether advertising changes what the assistant recommends. We found no official statement.
- A report or tool that shows a seller how the assistant mentions their products. As of 12 November 2026, we found no announcement of one.
- The router's rules and the size of the custom models.
Things sellers repeat that Amazon has not said
Be careful with these. None is confirmed, and none should be stated as fact: that the assistant reads text inside your images and weighs it, that A+ content has a specific weight, that there is an "assistant score" for listings, or any ranking formula with percentages.
One practical reading of ours, labelled as ours: an assistant that prepares a cart has little reason to recommend a product the shopper cannot buy right now. Keeping stock, the Buy Box offer and the current price right is cheap insurance. It is not a rule Amazon has confirmed.
What you control
Amazon's own pages point at the same places, and they are all yours:
- Keywords first. Shoppers still type searches. The assistant sits on top of search; it does not replace it.
- Detailed, consistent product information. Title, bullets, attributes, description and A+ content that agree with each other and with your own site. On Amazon US, titles are capped at 75 characters including spaces (except media) since 27 July 2026, and the new Item Highlights field takes up to 125 characters.
- Reviews and questions. They are a named source, and they are also the best list of what shoppers really ask. Read them for patterns: what people praise, what they compare it to, who they buy it for and what they wish they had known. Group those into the prompts you test. Never quote a review and never turn one anecdote into a promise.
- Price, stock and delivery that match what you promise.
- A way to see what the assistant says. Amazon has not announced a report for sellers, so ask it yourself, with your shoppers' questions, write down the answer and the date, and repeat it over several days. One run proves very little.
For a field-by-field checklist, see how to get an ASIN ready for Alexa for Shopping.
How Geonos fits
Geonos is an Alexa for Shopping optimization platform that does the two jobs together, Amazon SEO and Alexa for Shopping, for sellers on Amazon US, Amazon UK and Amazon México, and it launches soon. In the dashboard, for each ASIN, you see your priority prompts checked every day with every answer saved and dated, whether you are named and which competitors are named instead, your mention rate and share of voice with the number of runs behind them, your keyword position in Amazon search, the fixes ranked by what to change first, a listing ready to paste and designed product images. Nothing changes on Amazon until you approve it. See how it works or leave your email to hear when it launches.
Nobody can promise that the assistant will recommend your product. Geonos shows you where you stand and what to change first.
Sources
- Amazon, Amazon Rufus: Amazon's AI shopping assistant gets smarter and more personal, 18 November 2025.
- Amazon Science, The technology behind Amazon's GenAI-powered shopping assistant, Rufus, 4 October 2024.
- Amazon México, Conoce Alexa para Compras, tu asistente personal de compras con IA en Amazon, 11 November 2026.
- Amazon Science, COSMO: A large-scale e-commerce common sense knowledge generation and serving system at Amazon, SIGMOD 2024.
- Last checked: 12 November 2026. Amazon has not published how Alexa for Shopping ranks and chooses products; where this guide goes beyond Amazon's words, it says so.
Frequently asked questions
How does Alexa for Shopping work?
Amazon says it is a system of several models on Amazon Bedrock, chosen by a real-time router, including Anthropic's Claude Sonnet, Amazon Nova and a custom model built on Amazon's store knowledge. The custom model draws on the catalog, customer reviews, community questions and answers and information from the web.
What data does Alexa for Shopping use?
Amazon names the product catalog and detail pages, customer reviews, community questions and answers, information from the web, price and price history, data from Amazon's internal systems, and, in the US, the shopper's own history and preferences.
Is Alexa for Shopping the same as Rufus?
Yes, it is the same assistant. It started as Rufus, and Amazon renamed it Alexa for Shopping on 13 May 2026. In Mexico it launched on 11 November 2026 as Alexa para Compras.
How does Alexa for Shopping choose which products to recommend?
Amazon has not published it. There are no official criteria or weights. Anyone who gives you a ranking formula is guessing.
Is COSMO the algorithm behind Alexa for Shopping?
Not as far as Amazon has said. COSMO is a published Amazon Science system that is deployed in Amazon search. Many in the industry think it is related to the assistant, but Amazon has not confirmed it.
Does advertising change what Alexa for Shopping recommends?
We found no official statement on this, so we do not claim either way.
Can sellers see how Alexa for Shopping mentions their products?
As of 12 November 2026, we found no announcement from Amazon of a report for sellers. The practical way is to ask the assistant your shoppers' questions yourself, write down the answers with their dates and repeat them over several days. Geonos does this for each ASIN.
Does Amazon SEO still matter?
Yes. Shoppers still type keywords, and the assistant sits on top of search. Your keywords stay the base, and clear answers to shoppers' questions go on top of them.
Can anyone guarantee that Alexa for Shopping will recommend my product?
No. Amazon decides what it shows, and answers change. A tool can show you where you stand and what to change first, but not promise a result.