Amazon Rufus: How Amazon's AI Shopping Assistant Picks Products

1. 01

   Rufus, now called Alexa for Shopping in the US, reads your listing the way a careful shopper does: title, bullets, backend attributes, A+ text, reviews, community Q&A.

2. 02

   In Workflow Labs' tests it answers with about five named products, where a results page shows around 50.

3. 03

   If a fact isn't written on your listing, Rufus can't use it to pick you.

4. 04

   Ratings work like a gate: in an Amalytix study of 1,300+ recommended products, none had under 4.0 stars.

We still say Amazon Rufus here because most sellers and most searches do. For the wider marketplace picture, see our guide to digital shelf optimization.

What Amazon has confirmed

Amazon trained Rufus on its catalog and customer reviews, plus community Q&As and information from across the web (About Amazon). Writing in IEEE Spectrum, Amazon's Trishul Chilimbi explained that it uses retrieval-augmented generation: it parses the question, works out which sources can answer it, then pulls from the catalog, reviews or Q&A posts, or calls Amazon Stores APIs (IEEE Spectrum).

The matching layer is COSMO, a knowledge graph Amazon built from queries and co-purchases that links products to relations like used_for_function or used_for_audience. In one test, it lifted the macro F1 score of relevance models by 60% over the best baseline (Amazon Science).

1. 1

   The shopper asks

   "Gooseneck kettle with temperature control for green tea, under $80."

2. 2

   Rufus parses the request

   It picks the sources that can answer, from catalog data to live prices.

3. 3

   It maps the need to products

   A graph like COSMO connects "green tea" to adjustable temperature. A listing that never states that fact is hard to connect.

4. 4

   It filters on facts and trust

   Budget and stated features, then ratings. Amalytix found no recommended product below 4.0 stars.

5. 5

   It answers with a shortlist

   A few named products, each with a reason from the listing or reviews.

What outside research adds

• Amalytix ran the top 500 generic US search terms through Rufus as conversational queries and logged more than 1,300 recommended ASINs. The median product had a 4.5 rating and 2,991 reviews, and 87.2% used A+ or A+ Premium content (Amalytix).
• An April 2026 review by Workflow Labs concluded that backend attributes matter more than copy or imagery, and that many brands leave about half of them empty (PPC Land).

Read both as patterns. Amalytix says its correlations don't prove cause, and in repeat runs only 2 to 3 products per keyword appeared every time. Any list of "the 10 Rufus ranking factors" is a guess.

One number deserves more attention: the median organic rank of a recommended product was 41. Rufus often picks products far below page one, so page-one rank isn't the gate most teams think it is.

ChatGPT reasons the same way (see how ChatGPT recommends products), so here is the same kettle request as an AI assistant would answer it:

ChatGPT

Illustrative demo

ChatGPT

ChatGPT

Each reason is a fact a listing stated. "Precise temperature" alone gives the assistant nothing to quote.

• Fill every backend attribute

  Start with best sellers, and match each value to the visible copy

• Rewrite bullets around uses

  Who it's for, when to use it, and the spec that proves it

• Answer the questions shoppers ask

  Mine reviews and Q&A for repeat questions

• Fix rating problems at the root

  Below 4.0 stars recommendations look unlikely; read low reviews for gaps the copy can close

• Keep facts consistent everywhere

  Capacity and materials should match on Amazon, on your own site and in Google

• Check what the live page shows

  On a shared ASIN other sellers can contribute data, so confirm the bullets shoppers see are yours

Don't skip the attributes. In our audits they're the most common gap, because nobody sees them and nobody owns them.

Measuring Rufus visibility

Amazon offers no public interface for querying Rufus (PPC Land). What works is a fixed set of 20 to 50 prompts per category, run monthly with identical wording, logging which products appear. Phrasing alone moves the answer, so never edit prompts between runs. Pair it with sessions and conversion per ASIN, and see measuring AI visibility for the method outside Amazon.

Ads inside the answer

Sponsored Prompts, ads inside the assistant's answers, became billable on March 25, 2026 (EcomCrew). We'd still fix the listing first. Follow-up answers come from your listing, so a paid prompt on a thin one buys a conversation your content can't finish.

The same facts should reach Google, where AI Mode shopping reads your Google Merchant Center feed. Our digital shelf optimization guide covers syndication to other retailers.