Amazon retired the Rufus brand on May 13, 2026, folding its AI shopping assistant into a single system called Alexa for Shopping. The technology behind Rufus kept running; what changed is that it now shares memory with Alexa+, so a shopper who starts researching a product on their phone can pick the conversation back up on an Echo speaker. At the time of the relaunch, Amazon said the assistant had already reached more than 300 million customers, with monthly active users up over 115% year over year. For brands, the new name matters far less than the mechanics underneath it: what decides which products get named out loud in an answer, and which get left out entirely.
Why Amazon merged Rufus into Alexa for Shopping
Amazon merged Rufus with Alexa+ because the two systems didn't share context. A shopper could compare hiking boots inside the Rufus chat on Amazon's app, then open Alexa on an Echo device and find no record of that conversation. Alexa for Shopping fixes that by giving both surfaces one memory.
Rajiv Mehta, Amazon's vice president of conversational shopping, said the goal was that the customer doesn't have to think about where they started a discussion with Amazon, according to GeekWire's coverage of the May 13 announcement. The rollout began in the US that same week.
The timing lines up with a broader shift. ChatGPT, Gemini, and Perplexity have each added shopping-adjacent features over the past year, and Google now lets shoppers check out with retailers like Walmart and Wayfair without leaving a chat. OpenAI tried the same move first, then pulled back its in-chat checkout in March 2026 after adoption came in lower than expected. Amazon holds the largest product catalog of any of them, which is exactly why the mismatch between AI shopping agents and traditional product pages, covered in our guide to agentic commerce and product-page readiness, matters more here than almost anywhere else.
The five layers Alexa for Shopping reads before naming a product
Alexa for Shopping doesn't run one ranking algorithm. It stacks five data layers, each answering a different question about a product, then blends them before generating a response. Get one layer wrong, and a product can be in stock, priced right, and still never get named.
Those five layers are:
Large language models, Claude Sonnet, Amazon Nova, and a custom model trained on Amazon's catalog, generate the actual sentences a shopper reads.
Retrieval-augmented generation, pulling from Amazon's own product pages, reviews, and Q&A, plus outside sources. Amazon said in November 2025 that Rufus draws on popular sources like The New York Times, USA Today, Good Housekeeping, and Vogue when a question calls for outside expertise.
A proprietary knowledge graph Amazon calls COSMO, which maps products to the real-world situations shoppers describe, connecting a phrase like "fits under an airline seat" to specific luggage dimensions, according to Perpetua's June 2026 analysis of the system.
Review sentiment analysis, which extracts recurring themes, positive and negative, from customer reviews rather than just averaging star ratings.
Account memory, which factors a shopper's past purchases and browsing across Amazon, Kindle, Prime Video, and Audible into what gets suggested.
That's a narrower filter than most AI engines apply. Microsoft's Copilot, by contrast, still draws from Bing's full search index before narrowing to an answer, so a listing that ranks nowhere near the top of Bing rarely gets named by Copilot either.
Alexa for Shopping doesn't behave like a search results page. It typically names five products in an answer, against roughly fifty in a standard Amazon search grid according to Perpetua's June 2026 analysis, and once agentic purchasing is switched on, there's no browsing fallback at all.
What actually gets a product named in an answer
Getting recommended by Alexa for Shopping has almost nothing to do with keyword density in a title. It comes down to whether a product's backend attributes, bullet copy, and reviews answer the specific, often oddly phrased question a shopper actually typed.
Amazon's own structured data doubles as training data for the model deciding what to say, so filling in backend fields like material, intended use, dimensions, and compatibility feeds the knowledge graph directly, according to Perpetua's optimization guidance.
Bullet points written in a simple order, what the shopper gets, what feature enables it, and who it's for, give the retrieval layer something closer to a direct answer than a marketing line does. Tinuiti's guide to the assistant recommends the same shift inside the product description: answer constraint-based questions, like whether a suitcase fits under an airline seat, in plain declarative sentences rather than leaving a shopper to work it out from a dimensions table.
The same entity discipline that helps a brand surface correctly in ChatGPT or Google's Knowledge Graph carries over here. When the retrieval layer reaches outside Amazon.com, it's Amazonbot doing the crawling, the same way GPTBot or ClaudeBot index a page for other assistants. A JSON-LD product schema that's already correct for those engines doesn't need a separate version here.
Negative review themes are worth answering directly in the listing copy, too. If reviewers keep flagging short battery life, addressing it head-on in the description gives the sentiment layer a credible counter-signal instead of leaving the complaint as the only theme it can extract.
Sponsored Prompts: real ad dollars, with real limits
Amazon now sells ads inside Alexa for Shopping's answers, called Sponsored Prompts, but the format is small and narrowly targeted today. It's worth testing at low spend. It is not yet worth building a strategy around.
A Sponsored Prompt is a suggested question tied to a specific product or brand, appearing in the "Customers ask" section of a product page. Clicking it triggers a generated answer, followed by an ad with an add-to-cart button, according to Intentwise's rundown of the format, published in November 2025. Advertisers are opted in automatically and can turn off individual prompts.
The reach is still thin. One midsized seller tracked by eMarketer received 88 clicks from Sponsored Prompts in the first quarter of 2026, against 500,000 clicks from standard Amazon ads over the same period, under 1% of total click volume. Cost per click ran close to 31 cents, cheaper than the 50 to 70 cents typical of standard placements, though eMarketer cautioned that benchmark rests on too little data to trust yet.
The format also targets shoppers who have already named a brand, rather than those still comparing options, something eMarketer's analysts described as worth treating "as a research tool, not a sales driver" until Amazon expands it to unbranded, higher-funnel queries. That branded-only targeting echoes a pattern VizibleAI has tracked on other engines too: a gap between being mentioned and being cited as the actual source a buyer acts on.
How to know if it's working
Measuring performance in Alexa for Shopping means tracking two different things: whether a product gets named in answers to realistic shopper questions, and whether the Sponsored Prompts running against it show up on relevant, high-intent queries. Amazon Ads' Prompts report, added to the console in late 2025, is the only place that connects the two.
The Prompts report shows impressions, clicks, and purchases per prompt, but Intentwise noted in late 2025 that "few account managers appear to have much data in the Prompts tab" yet, so sample sizes stay thin for most sellers. Outside the ads console, the more reliable check is manual: run the specific comparison questions a category's shoppers actually ask, which blender is quietest, what running shoes work for flat feet, and see which brands the assistant names first, if at all.
The underlying discipline matches what already governs getting cited by Claude directly, since Claude Sonnet is one of the models generating these answers: test the exact phrasing a buyer would use, not the keyword a marketer would choose, and track how the answer shifts over weeks rather than checking once.
Amazon hasn't published how often the COSMO graph refreshes its read of a given listing, so sellers who update backend attributes should allow at least a few days before re-testing, based on the lag Perpetua's team has observed across client accounts.
Frequently asked questions
What is Alexa for Shopping?
Alexa for Shopping is Amazon's AI shopping assistant, formed by merging Rufus, the chatbot Amazon launched in beta in February 2024, with its broader Alexa+ assistant. It answers product questions, compares options, and can complete purchases across Amazon's app, website, and Echo devices, using large language models, a proprietary knowledge graph, and a shopper's own account history to generate its recommendations.
When did Amazon rename Rufus to Alexa for Shopping?
Amazon announced the change on May 13, 2026, and began rolling out the unified assistant in the US that same week. The Rufus brand name was retired, though the underlying technology continues to power the shopping experience behind the scenes.
What data does Alexa for Shopping use to decide which products to recommend?
It combines five sources: large language models including Claude Sonnet and Amazon Nova, retrieval-augmented generation from product pages and outside publications, a proprietary knowledge graph called COSMO that maps products to real-world use cases, sentiment analysis of customer reviews, and a shopper's own account and purchase history.
How can a brand improve its chances of being recommended?
Fill in every backend product attribute Amazon allows, since these feed the knowledge graph directly. Write bullet points and descriptions that answer specific, constraint-based questions in plain declarative language, and address recurring negative review themes directly in the listing copy rather than leaving them unanswered.
Are Sponsored Prompts, Amazon's ads inside the assistant, worth the spend?
Early data suggests caution. One seller tracked by eMarketer received under 1% of their total ad clicks from Sponsored Prompts in early 2026, and the format currently targets shoppers who have already named a brand rather than those still comparing options. It's reasonable to test at low spend, not yet to rely on.