How Alexa for Shopping evaluates a listing: AI listing optimization after Rufus

What changed (updated October 22, 2026)

This article first explained how Amazon Rufus evaluated listings. On May 13, 2026, Amazon renamed Rufus to Alexa for Shopping and merged it with Alexa+ (Amazon), so this version covers the assistant as it works now. Two other changes since March affect Amazon listing optimization for AI directly:

• The assistant moved into the search bar. US shoppers can ask questions in the main Amazon search box, and Amazon is adding AI overviews to the top of search results and to product detail pages (Amazon).
• Titles got shorter. Since July 27, 2026, titles in non-media categories are capped at 75 characters, with a new 125-character Item Highlights field; titles left over the limit can be rewritten by Amazon's AI (PPC Land).

What did not change is the core idea. An AI assistant reads your listing as evidence, and it can only recommend what it can state and defend.

The short answer: what Alexa for Shopping reads

Amazon has been fairly specific about inputs. It said Rufus was trained on the Amazon catalog, customer reviews, community Q&A and information from across the web (Amazon). Amazon's science team described a retrieval step that, before each answer, pulls evidence from reviews, the catalog and community Q&A and calls internal store services, with the most useful source depending on the question (Amazon Science). Alexa for Shopping adds the shopper's own context: orders, lists, browsing and stated preferences (Amazon).

Translated into seller terms, an assistant evaluating your product can draw on:

• Your catalog data: title, Item Highlights, bullets, description, attributes and variation structure.
• What buyers said: review text and ratings.
• What shoppers asked: questions and answers on the detail page.
• The offer: price, availability and delivery, which come from Amazon's store systems.
• The shopper: what this person bought, browsed and told Alexa before.

What Amazon has not published is the weighting: which of these matters most, whether there is a rating floor, how many products an answer includes. Every "ranking factor" list you read for Alexa for Shopping is third-party analysis, including the one implied in this article. We mark which parts are which.

From question to recommendation

It helps to picture the steps an assistant like this has to go through. The sequence below is our simplified reading of what Amazon has described, not Amazon's published logic.

Two things follow from this for anyone doing Amazon listing optimization for AI.

First, the constraint check is where most products drop out. If a shopper asks for a blender "quiet enough for an apartment" and your listing never mentions noise, the assistant has no evidence either way. It can still mention you, but it cannot defend you, and the products that state a decibel level have an easier case.

Second, reviews act as a second opinion on your claims. A listing that says "leak-proof" next to thirty reviews about leaks gives the assistant a contradiction. The safe move for a cautious system is to recommend something else, or to quote the complaint.

What is the Amazon COSMO algorithm?

COSMO comes up in almost every Rufus optimization guide, often described as the hidden algorithm behind the assistant. Here is what is actually public.

COSMO is a research system Amazon described in May 2024 and presented at the SIGMOD 2024 database conference (Amazon Science). It uses large language models to build a commonsense knowledge graph from shopping behavior: pairs of search queries and the products bought after them, and products bought together in the same session. The model explains why each pair belongs together, and those explanations become relations such as "used for audience" or "used for function". Amazon's own example links a search for shoes for pregnant women to slip-resistant shoes, a connection no keyword match would make.

In Amazon's offline tests on a public shopping-queries dataset, adding COSMO's knowledge improved relevance scores substantially over the baseline models. That is a research result on a benchmark, not a sales figure.

What Amazon has not said, as of October 2026, is that COSMO powers Rufus or Alexa for Shopping, or how it is used in production ranking. Seller guides that describe a "COSMO layer" deciding which products the assistant considers are making an educated guess. The guess is reasonable, because the idea matches what the assistant visibly does, but it is not confirmed.

The practical lesson holds either way. Systems like COSMO connect products to who they are for, what they are used for and in what situation. If your listing states those things plainly ("for small kitchens", "for runners with wide feet", "safe for septic systems"), you give any such system something to connect. If your listing is a string of keywords, you give it nothing new.

An AI assistant does not reward the listing that mentions the most keywords. It rewards the listing it can quote without being wrong.

Kris Krokos

How each part of the listing is read

Here is the listing element by element, with what Amazon has confirmed and what is reasonable inference.

Title and Item Highlights

The title is now 75 characters at most. That makes it a definition, not a keyword container: brand, product type and the one or two specs shoppers filter on. Item Highlights carries the next layer, such as material, use case and compatibility, and Amazon indexes it for search (M2E Pro). Amazon has not said how the assistant weighs Item Highlights, but a field designed to hold short factual claims is exactly the kind of text an AI summary can lift.

Bullets and description

Bullets are where you state features and limits in sentences. An assistant summarizing your product will paraphrase them, so write bullets that survive paraphrase: one claim per bullet, a number where a number exists, and the limitation stated next to the feature ("fits laptops up to 15.6 inches"). Generic bullets produce generic summaries.

Attributes and variations

Attributes are the structured version of your claims, and they are what comparison tables are built from. Alexa for Shopping can compare products a shopper selects from search results on features, prices and reviews (Amazon). A blank attribute becomes a blank cell next to a competitor's value. Variation structure matters for the same reason: if size and color children are split across several parents, or carry inconsistent values, the assistant cannot say cleanly which version answers the question. We go deeper on this in why messy product catalogs fail in AI commerce and why AI depends on structured data.

Images and A+ Content

Amazon supports visual search from shopper photos, but it has not said how much the assistant reads from your listing images or A+ modules when it answers. Treat images as proof for the shopper and repeat every important fact in text. A dimension that exists only inside an infographic may be invisible to a text model.

Reviews and Q&A

These are confirmed inputs. The assistant reads review text, not just the star average, and it surfaces themes: what buyers praise, what they complain about, whether the product matched expectations. Questions on your detail page are a free list of the constraints shoppers care about. If the same question appears five times, the answer belongs in your listing.

Price, stock and delivery

Price history of up to a year is now one tap away, and shoppers can set Auto-Buy rules that fire when your price hits a target. Out-of-stock products cannot be bought by an assistant on the shopper's behalf. Neither fact is a "ranking factor" Amazon has announced, but both shape whether a recommendation turns into an order.

What third-party data says, and its limits

The most careful public data we know of comes from AMALYTIX, a listing analytics vendor. In November 2024 it ran 500 top US search terms through Rufus, logged more than 1,300 recommended products and analyzed their listings (AMALYTIX). Recommended products had a median rating of 4.5 and a median of 2,991 reviews; 92.1% were Prime eligible and 94.2% were fulfilled by Amazon; almost none were rated under 4.0.

Read that correctly. It describes what Rufus tended to recommend for broad, high-volume terms in late 2024. The authors say it shows correlation, not causation: popular, well-reviewed, FBA products also win in classic search. It was measured before the 75-character title rule and before Alexa for Shopping. Use it as a sense of the bar in competitive categories, not as a set of thresholds.

Most other numbers in circulation, such as "the assistant shows five products instead of fifty" or the share of sessions it handles, come from agency analysis with less transparent methods. They may be right. They are not Amazon's.

Rufus optimization in 2026: what still works, what to drop

Still works:

• Stating the facts shoppers ask about, in text, in the title, Item Highlights, bullets and attributes.
• Consistent values across every variant and every field.
• Fixing the product problems behind recurring complaints.
• Answering detail-page questions in the listing itself.
• Stable price and stock.

Drop:

• Keyword-stuffed titles. The 75-character limit ends them, and Amazon's AI will rewrite what you leave over the limit.
• Vague superlatives ("best quality", "premium") with no fact behind them. An assistant has nothing to quote.
• Facts that live only in images.
• Inflated reference prices. A year of price history makes them easy to spot.
• Any tool or service that promises a guaranteed "Rufus score" or "Alexa ranking". Amazon has not published one.

A practical order of work

If you manage more than a handful of products, do this in order, top sellers first:

1. Title and Item Highlights. Rewrite to the new limits yourself before Amazon's suggestions go live.
2. Attribute completeness. Fill every attribute a shopper would filter on or ask about, with the same values across variants.
3. The five most-asked questions. Pull them from reviews and Q&A, and answer each in the listing.
4. Review themes. List the top three complaints per product, fix the cause or correct the claim.
5. Offer hygiene. Check price history for awkward spikes, and make sure stock covers a price cut.

For the full step-by-step version, read Amazon listing optimization for AI shoppers. For what the Rufus rename means beyond the listing, see Amazon Rufus is now Alexa for Shopping, and for how the assistant differs from classic search, Alexa for Shopping vs Amazon search. More marketplace coverage lives under Retail Platforms & Marketplaces.

FAQ

What is Rufus optimization?

Rufus optimization means making an Amazon listing easy for Amazon's AI assistant to understand, quote and recommend. Since May 2026 the assistant is called Alexa for Shopping, but the work is the same: clear facts in text, complete attributes, answered questions and healthy reviews.

What is the Amazon COSMO algorithm?

COSMO is an Amazon research system, published in 2024, that uses large language models to build a commonsense knowledge graph linking products to the needs, audiences and situations they serve. Amazon has not said whether it powers Alexa for Shopping.

How do I optimize an Amazon listing for AI?

State the facts shoppers ask about in your title, Item Highlights, bullets and attributes; keep values consistent across variants; answer recurring questions in the listing; fix the causes of recurring complaints; and keep price and stock predictable.

Does Alexa for Shopping read reviews?

Yes. Amazon says its assistant draws on customer reviews along with the catalog, community Q&A and web information. It uses review text to summarize what buyers like and dislike, not just the star rating.

Do keywords still matter for Amazon's AI assistant?

They matter for classic search, which still runs in the same search bar. For the assistant, plain statements of fact matter more than keyword volume, and the 75-character title limit leaves little room for stuffing anyway.