How to Optimise Amazon Listings for Alexa for Shopping (Rufus) and COSMO in 2026

On 13 May 2026 Amazon retired the Rufus name and moved the assistant into the main search bar as Alexa for Shopping. Same model, same recommendation logic, far more prominent. It now writes an AI overview above the results, runs side-by-side comparisons inside the results page, and answers the product questions that used to live in the Q&A section. If your listing does not give it something to work with, it recommends the listing that does.

I have been rewriting my own supplement listings against this system since April, and this guide is the process I ended up with. The first version of this article was written for Rufus. Everything below has been updated for the rebrand, the removal of customer Q&A, and the tool pricing changes since then.

What changed when Rufus became Alexa for Shopping?

Amazon renamed Rufus to Alexa for Shopping on 13 May 2026 and made it the default for every signed-in US customer, no Prime or Echo required. The underlying model and how it evaluates listings did not change. What changed is placement: it now sits in the search bar, generates overviews above results, and compares products inline.

That placement shift matters more than the name. Under the Rufus brand, the assistant was a button most shoppers ignored. As Alexa for Shopping it intercepts the search itself. Amazon's Q4 2025 earnings put Rufus at nearly $12 billion in incremental annualised sales, up from the $10 billion pace Andy Jassy had projected a quarter earlier. By Q1 2026, monthly active users were up 115% year on year and engagement up 400%. Amazon also says shoppers who use it are 60% more likely to complete a purchase.

You will still see "Rufus" everywhere. Seller Central help pages, third-party tools and most seller forums use both names, and I do the same in this guide. For the wider picture on where AI shopping agents are heading, including the ones outside Amazon, see our AI shopping agents explainer and the Amazon AI agent policy breakdown.

How do A10, COSMO and Alexa for Shopping work together?

Amazon search runs three layers at once. A10 handles keyword indexing and basic relevance. COSMO, a common-sense knowledge graph, maps your product to buyer intent across 15 relation types. Alexa for Shopping sits on top, interprets the natural-language query, and pulls from COSMO to decide which products to recommend and how to describe them.

Keywords still get you indexed. A10 has not gone anywhere, and a product that is not indexed for "insulated water bottle" will not be considered for "best bottle for the gym that stays cold all day" either. But repetition stopped helping years ago and now actively costs you characters that could carry meaning.

COSMO is the layer most sellers have never heard of. Amazon published the architecture in a 2024 paper at SIGMOD, describing a system that extracts "what this product does, who it is for, when it is used and what it is used with" from listings, reviews and behaviour data. When a shopper types a question, Alexa for Shopping uses that graph to match intent, not text.

What are the 15 COSMO relation types?

COSMO organises product knowledge into 15 relation types covering function, audience, context, classification, complementary products and buyer interest. Most listings I audit cover two or three of these clearly. The listings Alexa for Shopping surfaces first tend to cover eight or more, spread across the title, bullets, attributes and A+ text rather than crammed into one field.

ZonGuru maintains a useful plain-English breakdown of each relation type if you want worked examples beyond the table.

How should you rewrite your product title?

The title is the single highest-weight field for COSMO. Cover at least four relation types in it: key capability, product type, audience, and use context, followed by specifications. Every word should carry both a keyword for A10 and a piece of structured knowledge for COSMO. Repeating the head term does neither.

Here is a title I see variations of every week: "Yoga Mat Thick Yoga Mat Non-Slip Yoga Mat Exercise Mat Workout Mat for Home Gym". Four instances of "yoga mat", six of "mat". A10 indexes it fine. COSMO learns one thing: it is a yoga mat.

The rewrite: "Extra-Thick Yoga Mat for Bad Knees, Non-Slip Cushioned Home Workout Mat (72 x 26 inches, 8mm)". Same character budget. Now it carries audience (bad knees), capability (non-slip, cushioned), function (home workouts), location (home) and specification (exact dimensions). When a shopper asks "which yoga mat is best if my knees hurt," this title has already answered.

This does not mean longer titles. My best-performing supplement title after the rewrite was 11 characters shorter than the original. It means every word has to justify its place.

How should you restructure bullet points around intent?

Bullets are where the most space gets wasted. The default approach lists features with keyword variations. COSMO needs each bullet to answer a question a shopper would actually ask: what problem it solves, who it is for, where it is used, how it compares, and what you get.

Feature dump: "Made of premium 304 stainless steel. BPA-free. Double-wall vacuum insulation. Leak-proof lid."

Intent version: "Built from 304 stainless steel that keeps drinks cold for 24 hours and hot for 12. Tested in gym bags, backpacks and car cup holders without denting or leaking."

Same facts. The second version adds temperature retention as a measurable capability, three location contexts, and a durability claim. It answers "what is the best bottle that will not leak in my gym bag" directly.

The five-bullet framework I use:

1. What problem does it solve? Lead with the pain point, then the mechanism.
2. Who is it specifically for? Name audiences by use case, not demographics.
3. Where and when is it used? Concrete scenarios and environments.
4. How does it beat the alternative? Specific advantages, no competitor names.
5. What do you get? Specifications, contents, guarantee.

If you are weighing whether to write these yourself or hand them to a model, our AI vs manual listing writing test covers where each approach falls down. For catalogues past 50 SKUs, the product listings at scale guide is the better starting point.

What should go in backend keywords and attributes?

Mention each important keyword once in its most relevant field, then spend the remaining search-term bytes on intent phrases. More important than search terms: fill every product attribute Amazon offers. Target audience, intended use, subject matter and the other discovery attributes feed COSMO directly, and most sellers leave them empty.

The old backend strategy was to pack every variation and misspelling into the 250-byte search-terms field. That still helps A10 indexing, so keep doing the first pass. But do not repeat anything already in your title or bullets, and use what is left for phrases that read like questions: "best for sensitive skin", "works with induction cooktops", "fits under a standing desk".

Then open the product template and look at the attribute fields you have been skipping for years. When I audited my own catalogue in April, five of six listings had target audience blank and all six had intended use blank. Filling them took under five minutes per listing. It was the single change with the fastest visible effect.

How does A+ Content feed the knowledge graph?

COSMO parses the text in your A+ modules and uses it to extend your product's entry in the graph. That makes A+ Content a discovery asset, not just a conversion one. Use it for the relation types your title and bullets could not fit: secondary audiences, seasonal context, comparative positioning and less obvious use scenarios.

Think gift givers, caregivers, and people replacing a broken version of the product. Think "holiday gift", "back-to-school", "marathon training" where they fit. Think capability comparisons without naming anyone: "holds 50% more than a standard bottle", "8mm versus the usual 4-6mm".

Image-only A+ modules give COSMO nothing. Every module needs real text.

I use Listing Optimization AI for the A+ layouts because it generates modules with structured text blocks rather than pure image banners. The free tier covers the basic layouts. For sellers without a designer it closes the gap between no A+ Content at all (a missed COSMO signal) and agency work at agency prices. Code SELLERSTACKED20 takes 20% off the paid plans.

What replaced the customer Q&A section?

Amazon removed the customer Q&A block from product pages during the Rufus rollout. Alexa for Shopping now answers those questions itself, drawing on your listing text, attributes and reviews. You can no longer seed questions or reliably see unanswered ones, so the influence you had through Q&A has moved to reviews and attribute completeness.

This is the biggest change from the first version of this guide, which had a whole section on proactively answering questions. Sellers on the Amazon forums noticed the questions disappearing in 2025, and the Seller Central view of unanswered questions went with them.

What still works:

Reviews carry the most weight. The assistant extracts specific claims from review text. "This mat saved my knees during daily yoga" gives it a data point for Used_For_Audience (knee problems) and Used_For_Event (daily yoga). Five hundred "great product" reviews give it nothing. You cannot write reviews, but you can influence what customers write about: policy-compliant product inserts that ask about a specific use case, Vine for new launches, and prompt replies to negative reviews that name a specific dimension, since negative sentiment gets extracted too. Our AI review management strategy covers the tooling.

Attributes answer the compatibility questions. "Does this fit a Peloton cup holder?" used to be a Q&A entry. Now the assistant looks for it in your Capable_Of and Used_With coverage. If the answer is in your bullets or attributes, it can say yes. If not, it moves to the listing where it is.

Your own brand content is the third source. The assistant reads A+ text and Brand Story modules. Anything you would once have put in a Q&A answer belongs there now.

Which tools actually help with COSMO optimisation?

No subscription tool automates the full COSMO audit yet. Helium 10's Cerebro remains the best source of question-style keywords, and its 2026 AI Listing Builder has a Rufus mode. ZonGuru's per-ASIN Helix service does score against all 15 relation types, but it is a $30-per-listing add-on, not part of the $29 plan.

I ran the tools below across my six-listing supplement catalogue. Prices are current as of September 2026 and match our monthly price index.

Helium 10

Platinum is $99 a month on annual billing or $129 monthly. Code SELLERSTACK20 takes 20% off for six months; SELLERSTACK10 takes 10% off every month.

Cerebro is still the strongest keyword research tool for Amazon. The change for Alexa for Shopping is how you filter it: sort for question-style and long-tail conversational phrases rather than raw volume. Those are the queries the assistant is answering. The 2026 AI Listing Builder now has a Rufus enhancement setting that writes bullets around shopper questions and use cases. It got me to a decent first draft on four of six listings. It does not score against COSMO relation types, so I still did that pass by hand. Full test in our Helium 10 review, and the newer AI features are covered in the Helium AI Amazon Analyst review.

DataDive

DataDive is not a COSMO tool. I include it because its rank tracking is the cleanest way I found to measure whether the rewrite worked. Set up tracking for both your head terms and a handful of question-style phrases before you change anything, then compare at day 14 and day 30. Code SELLERSTACKED takes 10% off for six months. See the DataDive review.

ZonGuru

ZonGuru has done more marketing around COSMO than anyone, and its content on the relation types is genuinely good. Be clear about what you are buying, though. The Researcher plan is $29 a month on annual billing, $49 monthly, and includes the standard Listing Optimizer. The COSMO-specific product, Helix, is sold separately at $30 per ASIN and is a service, not a feature. For a six-SKU catalogue that is $180 on top of the subscription. The free COSMO Readiness Report is worth running on one listing to see the scoring approach. Details in our ZonGuru review.

Jungle Scout

Jungle Scout pulls from a different keyword data set than Helium 10, so running both surfaces phrases either would miss alone. AI Assist drafts listing copy from your keyword list and handles the A10 foundation well. It has no COSMO awareness at all, so the intent pass is entirely manual. Our Helium 10 vs Jungle Scout comparison covers when the lower price is the right trade.

For the broader research-tool picture, see our Amazon product research tools comparison.

How do you measure whether it worked?

A10 changes show in 24-48 hours. COSMO and Alexa for Shopping take 7-14 days to reflect a rewrite, and conversion impact is not reliably measurable until day 14-30. Set a baseline before touching anything, make all changes in one pass, then do not edit the listing again for two weeks.

Track four things: rankings for head terms and for a set of question-style phrases, unit session percentage, sessions from sources Seller Central labels as "other" (assistant-driven traffic does not get its own line yet), and review content, since new reviews that mention specific use cases suggest the assistant is surfacing you to the right shoppers.

Amazon does not yet expose assistant-specific visibility data to sellers, and I have not found a third-party tool that reliably tracks it either. For visibility in AI systems outside Amazon (ChatGPT, Google AI Overviews, Perplexity) the approach is different and covered in our guide to getting products cited by AI answer engines.

What did 60 days of testing on my own listings show?

I rewrote six supplement listings between 14 April and 13 June 2026 using the process above, changing nothing else during the window. Average COSMO coverage went from 3 relation types to 9. Unit session percentage across the six rose from 11.2% to 13.9% by day 60. Two listings picked up first-page rankings for question phrases they had never ranked for.

Some honest caveats. Six listings in one category is a small sample, and the period overlapped with the Alexa for Shopping rollout, so I cannot fully separate the rewrite from the placement change. One listing went backwards on its head term for nine days before recovering. And the attribute fields did more than I expected: the listing where I only filled attributes and left the copy alone still improved, which tells you where to start if you only have an afternoon.

The rewrite itself took roughly 40 minutes per listing including the Cerebro pull. The attribute pass was five minutes each. I would not call the conversion lift dramatic. I would call it the cheapest 2.7 points I have picked up on Amazon.

What mistakes hurt visibility most?

Keyword-stuffed titles. The number one Amazon SEO tactic for a decade is now counterproductive. A title that repeats the head term four times communicates less structured knowledge than one that uses those characters to cover four relation types.

Empty attribute fields. Every blank is a missing node in the graph. Five minutes per listing.

Editing too often. COSMO needs 7-14 days. Change the listing every two days and it never settles. Change, wait two weeks, measure, then iterate.

Optimising only for A10. "Find high-volume keywords, put them everywhere" optimises one layer of a three-layer system. A10 gets you indexed. COSMO and Alexa for Shopping get you recommended.

Treating the Q&A removal as a loss and stopping there. The signal moved to reviews and attributes. Sellers still waiting for Q&A to come back are leaving that ground to competitors.

What should you do this week?

Pick your top three listings by revenue. Score each against the 15 relation types and count how many it covers clearly. Fill every backend attribute field first, since that is the fastest win with the least risk. Then rewrite the title using the four-dimension formula and restructure one bullet from feature to intent.

Set up rank tracking for two head terms and three question phrases per listing before you save anything. Then set a reminder for 14 days and do not touch the listings until it fires.

The assistant is now the search bar. That is not a temporary interface test. Listings that communicate structured knowledge will keep compounding their advantage as more of Amazon's discovery runs through it.