How to optimize an Amazon listing for Alexa for Shopping (2026)

Short answer: there is increasingly an assistant sitting between your listing and the shopper. Amazon retired Rufus as a standalone product on 13 May 2026 and folded the capability into Alexa for Shopping — an agentic assistant in the main search bar, on by default for signed-in customers in the US. You do not optimize for that layer with keyword strings. You optimize by answering questions: what the product is for, who it is for, under which conditions — in plain language, with structured attributes filled in, and with claims that can be grounded.

One caveat up front: as of August 2026 the German marketplace is not in the state described for signed-in US customers, where the assistant is the default. A broader roll-out is expected; we do not know a date for amazon.de. The work below is still worth doing — it improves the listing for the people who read it too.

What happened on 13 May 2026

Rufus was Amazon's shopping assistant: a separate chat panel that answered product questions and pulled recommendations out of the catalogue. On 13 May 2026 Amazon shut it down as a standalone assistant and folded it into Alexa for Shopping. The difference is not cosmetic. The assistant no longer sits next to search — it sits in the main search bar, which is where purchase intent starts anyway. And it is designed to be agentic: it carries out steps rather than only producing text.

For scale, Amazon has published its own figures for Rufus: roughly 300 million customers used it, attributed sales of about $12 billion on an annualized basis, and interactions up 210% year over year. Those are Amazon's own numbers, not independently verified, and Amazon defines what counts as “attributed”. As evidence that a meaningful share of shoppers now asks rather than searches, they are enough.

What COSMO changes about matching

COSMO is Amazon's publicly documented semantic layer, described in a paper on Amazon Science. Put simply, it infers relations between products and the situations they are used in: what something is for and who it is for, beyond the words literally typed into the box. String matching cannot do that.

In practice this cuts two ways. A listing can be relevant to a query in which none of its keywords appear literally. And it can be passed over for a query whose every word sits in its title, because the context does not fit. “Which terms do I put in so I get found?” becomes “Can a reader tell from my listing which situation this product is the right choice for?”

1. State plainly what it is for and who it is for

An assistant answers in full sentences, and it can only summarize what the listing actually says. Purpose and audience therefore have to be written out, not implied. “28 cm stainless steel pan” is a description. “Works on induction hobs, oven-safe to 240 °C, sized for small kitchens” is an answer.

• Name the use case in a readable sentence — in bullet one or two, not buried in the description.
• Name the audience when it decides the purchase: beginner or professional, indoor or outdoor, household or trade.
• Name the limits. “Not dishwasher safe” is not a weakness in your copy; it is an answer the assistant would otherwise have to guess.

2. Cover the questions shoppers actually ask

In conversation, shoppers ask different questions than they type into a search box: compatibility, care, dimensions, what is in the box, durability, how this differs from the next model up. You do not have to invent those questions — they are already in your reviews and in the reviews of your three strongest competitors, along with a note about which answer is missing.

Put each answer where it belongs: the two most common in the bullet points, the rest in A+ content. A+ is the right home for them because it holds comparison tables and usage steps that five bullets are too short to carry.

3. Put intent terms in the backend keywords

Backend keywords are the one place for terms that would read badly in visible copy. That is exactly where use-case and intent terms belong: “gift for dad”, “camping”, “small apartment”, “for beginners”, along with colloquial names, spelling variants and common misspellings.

The field's hard limits do not change: 500 bytes, not 500 characters — accented characters count double — and nothing repeated from the title, the bullets or the description.

4. Fill the structured attributes completely

An assistant reads data, not vibes. Dimensions, material, compatibility, age rating, power supply, box contents: what sits in a structured attribute field is unambiguous, while what sits only in prose has to be interpreted first. Empty attribute fields are the cheapest avoidable disadvantage in a listing, because filling them costs neither new copy nor new photography, only time.

• Work through your category's full attribute list once, not just the required fields.
• Enter dimensions with units into the attribute, rather than only writing them into a bullet.
• Keep attributes and copy consistent. A contradiction between the two is worse than a gap.

5. Make every claim verifiable

An assistant that recommends products stakes its usefulness on the recommendation. Claims anchored to something — a standard, a measured value, a material spec, a certification — are sturdier for that purpose than superlatives. “Waterproof to IPX7” is a different sentence from “extremely waterproof”, even when both describe the same product.

The second reason is the less comfortable one: an AI layer passes your exaggeration along. The shopper then buys on a promise you cannot keep, and the bill arrives as a return.

What not to do

• Keyword stuffing as a response to the AI layer. A layer that weighs context gets nothing out of a chain of terms — and the human reading afterwards gets even less.
• Invented questions and answers: an FAQ block full of questions nobody asked, answered by nobody who checked. It is not evidence, and the moment it contradicts your attribute data it does damage twice over.
• Copying the same text into several fields. Title, bullets, description and backend keywords have different jobs; a duplicate costs space and buys nothing.
• Claims about competing products, or other brands' names inside your listing. That breaches the guidelines and adds legal exposure on top.

What nobody outside Amazon knows

How Alexa for Shopping selects products, and how much weight COSMO carries in the ordering, is not known outside Amazon. There is no formula and no checklist with a guaranteed effect; anyone selling you one made it up.

What is above are robust practices, not ranking mechanics. They follow from what Amazon has published itself, and they do no harm in any scenario: a listing that states its purpose clearly, carries complete data and makes no promise it cannot keep is the better listing with or without an AI layer.

Where to start

Take your three highest-revenue ASINs. Write down the ten questions shoppers really ask about them — from reviews, return reasons and support email. Check which of those the listing answers today. Then fill the attribute fields completely and replace every unsupported claim with a smaller one you can back up.

That last step is where most listings get stuck, and it is why Torch Listings ties every statement to a review, a competitor listing or a data point: what is evidenced is what an AI layer can ground and pass on.