How To Optimize Amazon Listings For Rufus And COSMO In 2026

A shopper opens Amazon on their phone and types something like "which shampoo will not strip colour treated hair" into the search bar. Before they see a single product page, they read an answer. That answer was assembled out of listing text, customer reviews and community questions, and it named a handful of products. One of them belonged to a seller who never found out why.

Amazon says the assistant behind those answers helped more than 300 million customers during 2025, and that on 13 May 2026 it stopped being called Rufus. It is now Alexa for Shopping, it answers inside the main search bar instead of a separate panel, and it carries the shopper's own history and prior conversations with it. Source: Amazon, Amazon primary, 2026.

We are still writing Rufus here, because that is the name sellers use and the mechanism did not change with it.

The part that will genuinely annoy you is that none of this arrives as a report. No notification, no dashboard, no flag in Account Health. Sessions hold up, the ads keep running, and conversion drifts down on the pages that used to carry the account.

So the work here is not damage control. It is going back through the listing and asking whether it can answer a question, because that is now one of the jobs it has.

What Amazon actually built, in plain terms

There are two layers here, and they get talked about as one thing, which is where the confusion starts.

Rufus, now Alexa for Shopping, is the surface. Amazon describes it as a generative AI shopping assistant trained on its product catalogue, customer reviews, community Q&As and information from across the web, and says that on a product page it answers using the listing details, the reviews and the community questions for that specific product. Sources: Amazon and Amazon, both Amazon primary.

Amazon has also said the assistant runs on several large language models, including Anthropic's Claude Sonnet and Amazon Nova, with retrieval from outside sources on top. Source: Amazon, Amazon primary.

COSMO is the intent layer underneath. Amazon Science published it on 10 May 2024 and presented it at SIGMOD that year. It is a commonsense knowledge graph built by having large language models read query and purchase behaviour, then keeping only the relationships that human annotators and trained classifiers agreed were plausible.

The relationships are the ones a shopper would recognise: what a product is used for, what it is capable of, what kind of thing it is, and what it causes, breaking down further into function, event and audience. Amazon says it covers 18 major categories and has been deployed in its search applications. Source: Amazon Science, Amazon primary, 2024.

Read those two layers together and the shift becomes concrete. A keyword index answers "which pages contain these words". An intent layer answers "which product suits a person doing this thing, for this reason, in this situation". Those are different questions, and a listing written to win the first one often has nothing to say to the second.

Why keyword stuffing now works against you

For years the incentive ran one way. More phrases in more fields meant more chances to be indexed, and the cost of an ugly title was low because the title was mostly read by a machine.

Two things changed that. The first is that the same text is now read by a system that summarises rather than matches, and a title listing eleven phrases summarises badly. When an assistant has to say in one sentence what a product is and who it suits, a page that never says either has nothing to give it.

The second is Amazon's own guidance, which has quietly moved. Amazon's seller blog now says plainly that AI powered search tools use listing content to recommend products, points sellers at descriptions that address the questions customers keep asking, and frames the description as a short buying guide covering who the product is for, what problem it solves and the benefits rather than a feature list.

On titles, the character limit is 200 and Amazon's own recommendation is 80 or fewer. Source: Amazon, Amazon primary.

That last number surprises people, because the advice for a decade has been to use all 200 characters. It is a recommendation and not a rule, so nothing breaks with a long title, but it is worth noticing when the platform says what it prefers.

A9 and A10 have not gone anywhere

This is the part the AI search commentary keeps skipping, and skipping it costs sellers money.

Amazon does not publish its ranking algorithm, so anybody claiming to know the weighting is guessing in public.

What is not in dispute is the direction of the mechanics: a product has to be indexed for a phrase before it can be retrieved for it, and the products that hold position are the ones that convert the traffic they get, sell steadily and earn clicks in search results. An assistant recommending a product it cannot retrieve is not a scenario that exists.

So the honest framing is additive. Backend search terms still do the indexing job, and our backend keywords page covers how we treat that field. Velocity, conversion and click through rate still decide who holds page one. What has been added on top is a layer that reads the page as a document and needs something to read.

How to tell whether this is already costing you

Five checks, none of which need a tool you do not already have.

1. Compare sessions against unit session percentage over 90 days. If sessions hold and unit session percentage slides, the traffic still arrives and the page has stopped closing. That is usually the first commercial signal.
2. Read your last twenty reviews and list the questions inside them. Then look at whether the page answers any of them. Most pages answer none, because the copy was written before those reviews existed and nobody went back.
3. Read the page out loud as an answer to one shopper question. "Is this good for a beginner" or "will this fit a 2019 model". If nothing on the page can be read out as an answer, there is nothing for an assistant to summarise either.
4. Open the Listing Quality Dashboard and count the missing attributes. The path is Inventory, then Manage All Inventory, then Listing Quality Dashboard, and Amazon groups its recommendations by benefit, including improving search results.
5. Look at your ads report for clicks without conversions on long search terms. Question shaped terms that click and never convert usually mean the shopper arrived expecting an answer the page did not contain.

What to check before you rewrite anything

There is a strong urge to rewrite the whole catalogue in a weekend, and it is worth resisting for a week, because a rewrite with no baseline is a rewrite nobody can evaluate afterwards.

1. Baseline sessions, unit session percentage and units per ASIN for the last 30 and 90 days. The path is Reports, then Business Reports, then Detail Page Sales and Traffic by Child Item. Without this, every later argument about whether the change worked is just opinion.
2. Confirm the ASIN is actually indexed for its main phrases. A page that is not retrievable cannot be recommended, so indexing is the floor and not the ceiling.
3. Pull the attribute gaps per ASIN, not per account. Attributes are defined per product type, so a field that is critical for a monitor is irrelevant for a shampoo, and account level averages hide the ASINs that matter.
4. Collect the real questions in one place. Reviews, community questions, customer service emails and returns reasons. This list is the raw material for everything that follows, and it is free.
5. Note which elements are testable. Amazon's Manage Your Experiments tool covers titles, main images and A+ content for brand registered sellers with enough traffic to reach a result, so some of this can be measured rather than argued about.

How we optimize an Amazon listing for Rufus, in the order we do it

This is the sequence our team uses on an established ASIN, structural work first, because those fields are where effort converts most reliably into machine readable meaning.

1. Write the buyer's decision in one sentence before touching the listing. Who this is for, what they are trying to do, and the one thing they are worried about. If that sentence cannot be written, the research is not finished.
2. Fill the structured attributes to completion for that product type. Material, quantity, size, fit, compatibility, ingredients, age range, whatever the type defines. These fields feed search filters and they are the least ambiguous statements about the product on the entire page. Amazon's own generative listing tool is built to populate more than 70% of required attributes, which tells you how much of the catalogue is sitting empty. Source: Amazon, Amazon primary.
3. Rebuild the title around identification rather than coverage. Brand, what the thing is, the variant that identifies it, the quantity or size. Amazon's 80 character recommendation is a useful discipline even where the limit allows 200, because a title that reads as a sentence can be quoted back to a shopper and a keyword list cannot.
4. Turn the bullets into answers to the five questions from your list. One question per bullet, the answer first, the reassurance second. This is the single change that most often moves conversion, and it is also the change that gives an assistant something quotable.
5. Write the description as a short buying guide. Who it suits, what problem it solves, what is in the box, what it does not do. That last one feels counterintuitive and it reduces returns, because the shopper who would have been disappointed self selects out before buying.
6. Treat images as data rather than decoration. The main image follows Amazon's site standards, including a pure white background and the product filling the frame, per Amazon's product image guide. The secondary images are where scale, material, quantity and use case get stated in words, because that is where a shopper looks before reading anything.
7. Fill the image keywords field on every A+ module. It is the alt text field in A+ Content Manager, it exists for accessibility, and it is one more place where the page describes itself in plain language. Amazon says A+ content "can help increase sales by an average of 5.6%", which is a modest number honestly stated, and our A+ content page covers how we build the modules.
8. Treat reviews and questions as part of the listing, because Amazon does. Amazon states that the assistant answers product questions using the listing details, the reviews and the community Q&As. Answering community questions properly is unglamorous work that adds text written in the shopper's own words, and for products with fewer than 30 reviews, Amazon's Vine programme is the legitimate route to a review base. Nothing outside Amazon's own programmes is worth the risk to the listing.
9. Keep the backend search terms doing their one job. Indexing. Not persuasion, not repetition of the title, and not a dumping ground for every phrase a tool exported.
10. Change one element at a time on your top ASINs and measure it. Manage Your Experiments exists for exactly this, and on ASINs without the traffic to test, the fallback is a clean before and after window with the baseline you took in the previous section.

We cannot promise this moves a ranking, because Amazon does not publish how it weighs any of it, and because the listing is one input among several that includes price and delivery speed. What we can say is that these are the inputs Amazon has told sellers it reads.

What this looked like on a real catalogue

Our team rebuilt 12 SKUs for a haircare brand across the US and UK using this approach: the item builds done properly, attributes completed, and the copy built around use cases, material and quantity rather than a keyword list. The listings indexed for more than 25 keywords within two weeks, conversion moved from 8% to 17%, and two variants picked up the Amazon's Choice badge.

Keywords were part of that work and they were not the reason it converted. It converted because the pages answered the questions a shopper asks about haircare: what hair type it suits, what is in it, how much you get, what it will not do.

That is the position we keep arguing for, and it is not the popular one in this industry. A listing is a narrative, not a keyword container. The AI layer did not create that principle, it just made ignoring it more expensive. If it is the whole catalogue that needs this rather than one page, our listing optimization page explains how we work through it.

Keeping the catalogue current instead of rewriting it every year

Amazon changes the shopping surface more often than it changes the fundamentals, so the useful goal is a listing that does not depend on one interface staying still.

1. Harvest questions once a quarter. New reviews, new community questions, new return reasons. Products acquire new objections as they reach new buyers, and a page written in 2023 is answering a 2023 objection.
2. Close attribute gaps as a monthly routine rather than a project. Values take up to about 48 hours to appear after a change, and twenty minutes a month keeps the structured layer honest.
3. Use the reading test on any new listing before it goes live. If a person can read the bullets and correctly describe who should buy the product, the summarising layer has what it needs too.
4. Watch unit session percentage per child ASIN monthly. Parent level numbers hide the drift you are trying to catch, and the child that quietly loses two points of conversion is usually the one carrying the ad budget.
5. Treat the annual keyword refresh as maintenance rather than strategy. Adding phrases to a page that cannot answer a question changes nothing except its length.

What we would do first if this were our account

Baseline sessions and unit session percentage on the top ten ASINs, because in six weeks that number is the only way to know whether any of this helped.

Then close the attribute gaps on those same ASINs, since those are structured fields with defined answers and no judgement involved. It is the highest certainty work available in a listing.

Then rebuild the bullets on the top three ASINs around the questions your own reviews keep raising, one element at a time so the result is measurable rather than a story.

We cannot promise Amazon's assistant will start recommending your product, and we would be careful with anyone who does, because Amazon publishes neither the retrieval logic nor a report telling sellers when they were cited.

What we can tell you before you spend anything is which of your pages can answer a shopper question and which cannot, and which attributes are empty on the ASINs that carry your revenue. That is what the free, no-obligation audit covers, and if your listings are already in good shape, we will say so.

Related guides

• The product attributes AI search reads
• The nine checks for a low conversion rate
• Title limits, banned wording and truncation

Common questions about Rufus, COSMO and AI search