Amazon Rufus, now Alexa for Shopping: how to optimise your listings for AI search

A growing share of Amazon shoppers no longer type a keyword and scroll. They ask a question, such as “which of these is quietest?” or “will this fit a small kitchen?”, and an AI assistant answers it using what your listing, your reviews and the wider web say about your product. In the UK that assistant is still called Rufus. In the US, since May 2026, it is part of Alexa for Shopping.

There is a lot of confident advice about how to “rank in Rufus”. Most of it goes further than anything Amazon has published, including some of what we wrote ourselves in 2025. This guide sticks to what Amazon has documented, says plainly where the documentation stops, and gives you a method you can run on your own listings.

What is Amazon Rufus?

Amazon Rufus is Amazon’s AI shopping assistant: shoppers ask it questions in plain language and it answers, compares and recommends products using the catalogue, detail pages, reviews, community Q&As and the wider web⁵. In the UK it is still Rufus. In the US it is now part of Alexa for Shopping.

Amazon launched Rufus in the US in February 2024, describing it as an assistant trained on “Amazon’s extensive product catalog, customer reviews, community Q&As, and information from across the web"². It reached the UK in September 2024⁴, and Amazon’s UK newsroom now says it is available to all UK customers in the Amazon Shopping app and on desktop³.

On 13 May 2026 Amazon brought Rufus together with Alexa+ to create Alexa for Shopping, “available to U.S. customers on the Amazon Shopping app and website, and Echo Show devices"¹. The original Rufus announcement now carries a note that Rufus was renamed Alexa for Shopping². Amazon’s US help page describes one assistant “with shared memory and personalized preferences” across the Amazon app, the website and Alexa devices, and warns that it “may not always get things right"¹².

We have found no Amazon announcement of Alexa for Shopping for the UK, and we would not plan around a date nobody has published. The US version is still worth watching if you sell there too, because it puts an AI-written overview on your own product page. What the assistant says about your product becomes part of the page itself.

How much does it matter? Amazon says Rufus helped over 300 million customers in 2025¹, and reported in October 2025 that shoppers using Rufus were 60% more likely to complete a purchase⁷. Read that second figure carefully. It compares shoppers who chose to use the assistant with those who didn’t, and people asking detailed questions are usually closer to buying. It shows where high-intent shoppers are spending time, not that the assistant adds 60% to anyone’s sales.

What the assistant reads on your listing

Amazon is more specific here than most sellers realise. Its UK Seller Central help page says Rufus is trained on “our extensive product catalogue, customer reviews, community Q&As, product detail pages and information from across the web”. For product-specific questions such as “What are the dimensions?” or “Is this machine washable?”, it says: “we use information found on your product’s detail page"⁵.

In a 2026 post to UK sellers, Amazon also listed why the assistant gets things wrong: detail page information that is incomplete or inconsistent across sections, listings not updated for current specifications, and conflicts such as bullet points that say one thing while the product description says another⁶. It asked sellers to keep bullet points, descriptions and A+ Content consistent and to “use all available content fields"⁶.

That gives a clear map, and a clear edge where Amazon stops saying anything:

Has Amazon retired customer questions and answers?

We could not find an Amazon notice retiring the feature. What has changed is visible on the page: on many listings the old list of customer questions has been replaced by a box that sends a question to the assistant. Sellers in Amazon’s own forums report they can no longer view or answer unanswered questions, and Amazon’s reply pointed them to the Rufus help page rather than to a new Q&A tool¹¹. Yet Amazon’s help still names community Q&As among the assistant’s sources⁵.

So existing Q&A may still inform answers, but you can’t count on answering new questions there. If shoppers keep asking whether your pan works on induction, the answer belongs in a bullet point and an attribute, not in a thread you may never see.

What COSMO is, and what it isn’t

COSMO is real, and unlike most of what is written about it, it has a primary source. In a paper at SIGMOD 2024, Amazon researchers describe it as “a scalable system to mine user-centric commonsense knowledge from massive behaviors” and turn it into knowledge graphs of shopping intent⁸. Large language models generate candidate facts about why people buy things, classifiers trained on human-annotated data filter them, and a fine-tuned model, COSMO-LM, extends the graph across 18 major product categories⁸. The paper says COSMO was deployed in search navigation, where online tests showed a 0.7% relative increase in product sales in the tested traffic segment and an 8% rise in navigation engagement⁸.

The part that is useful to a seller is the shape of the knowledge. COSMO links products to intentions through 15 relation types, including what a product is used for (a function, an event or an audience), what it is capable of, where it is used, when it is used, what it is used with and who uses it⁸. The paper’s own example: a broad search for camping air mattresses can be broken down into mattresses for lakeside camping, mountain camping or four people⁸.

What COSMO is not is a published ranking algorithm for Rufus. Amazon has not said that COSMO powers its shopping assistant, and the paper describes its deployment in search navigation, not in an assistant. Our older posts treated “optimise for COSMO” as a ranking tactic. We now treat the paper as a writing brief: evidence that Amazon models products by the situations people buy them for, not only by keywords.

What Amazon doesn’t tell you

Amazon does not publish how its assistant chooses which products to recommend, or in what order. We checked. (For what Amazon has said about ordinary search ranking, see our guide to Amazon SEO.) So when you read that the assistant “weights Brand Registry”, that listing changes “re-index in 48 to 72 hours”, that brands “surfacing in Rufus” have 90% of attributes filled, or that “an internal Amazon deck” puts a number on any of this, ask for the source. We removed each of those claims from our own older pages because we could not find one.

What Amazon has documented is enough to act on. Product answers come from your detail page. Incomplete, outdated or contradictory pages produce wrong answers. Reviews are summarised for shoppers. None of that needs guesswork.

A readiness method for each listing

This is the process we run on client listings, starting with the products that earn the most. It takes an hour or two per product family.

1. Ask the questions your customers ask. Open Rufus in the Amazon app and ask category questions (“What should I look for in a [product]?”, “Which [product] is best for [use]?”, “Which [product] lasts longest?”) and then product questions on your own page. Note whether your product appears and exactly what is said about it, with the date. Answers change.
2. Check every answer against the listing. If the assistant is wrong, fix the source content first. If the listing is right and the answer is still wrong, open a help centre case with the ASIN, a screenshot of the question and answer, the type of issue and evidence such as a manual or the manufacturer’s website⁵.
3. Close the gaps field by field. Titles now have 75 characters⁹, so the title says what the product is and the item highlights carry uses and materials. Bullets answer objections. Attributes are complete. A+ content states key facts as text and includes a comparison chart for your range. Our listing optimisation guide covers the rules for each field.
4. Read your reviews the way the assistant will. Go through the last 50 to 100. Praise that keeps recurring belongs in the listing. A complaint that keeps recurring needs a fix to the product, or a clear explanation if it is a misunderstanding, because the assistant will repeat it to shoppers either way.
5. Make the range answerable. “Which one do I need?” is one of the most common questions shoppers ask. A comparison chart, consistent naming and a sensible variation structure keep them inside your brand instead of sending them to a competitor.
6. Keep control of what Amazon changes. Amazon now rewrites titles longer than 75 characters with its own AI, and brand owners have 14 days to review those changes in Review Listing Changes¹⁰. Keep a record of your approved titles and attribute values, and check for changes monthly.

Then measure. Conversion is what the assistant’s answers ultimately move, and your Business Report shows it product by product as Unit Session Percentage. The listing optimisation guide explains how to read it, and the Profit X-Ray flags listings that convert well below your own average.

For a quick self-check, the listing checker tests a title, item highlights and bullets against Amazon’s rules, then walks you through a 10-point AI-readiness checklist: brand and product type first, item highlights, feature-and-benefit bullets, complete attributes, text-based A+, a comparison chart, recent reviews that match shoppers’ questions, Brand Registry, a compliant main image¹³ and backend search terms that add new words. Those ten points are our team’s practice, not Amazon rules, and the checker says so.

A worked example: five filters and one question

A home goods brand we work with sold five models of replacement water filter cartridge. Each listing was accurate and full of specifications: filtration technology, contaminant removal, flow rate. The brand knew its products inside out and assumed its customers did too.

When we asked Rufus about the category, the questions it suggested and answered were about situations, not specifications: which cartridges suit hard water, which work with well water, which are easy to install, whether a filter removes heavy metals or improves the taste. On the brand’s own pages, nothing told a shopper which of the five models matched their situation. Shoppers moved between listings, and many either left or bought the cheapest.

We rewrote each listing to open with who it is for and which water problem it solves, added an A+ comparison across all five models, and wrote the answers to the assistant’s recurring questions into bullets and A+ text. We don’t quote a conversion figure for this one, because several things changed at once and we couldn’t isolate the effect. The lesson was the gap itself: the brand had assumed shoppers knew the difference between the models, and the assistant’s questions showed they didn’t.

The same exercise on a competing filter showed the other side. Asked whether that filter was a good choice, the assistant said customers liked its performance but that some had reported an unpleasant smell after use. The listing never addressed it. Whether it was a bad batch or a missing instruction, every shopper who asked was hearing the objection with no reply from the brand.

Where this leaves UK sellers

For now, UK shoppers use Rufus and US shoppers use Alexa for Shopping. The names will keep changing. The job doesn’t: write a listing that states every fact a buyer needs, in plain text, the same way everywhere it appears, and check what Amazon’s assistant says about it.

That is also the work that makes a listing convert for people who never open the assistant, which is why we don’t treat AI readiness as a separate project. It is listing quality, checked by a new and very literal reader. Start with the listing checker, then the full listing optimisation guide and our A+ content blueprint.