Amazon
genAI shopping assistant
In 2025, Amazon's genAI shopping assistant Rufus reached an annualized run rate of 10 billion USD in incremental sales and made its users 60% more likely to buy, according to the Q3 2025 earnings call.
Key points
- Conversational genAI shopping assistant on the Amazon app and site.
- In-house LLM plus RAG over the catalog, product pages, and reviews, with agentic functions.
- Run rate of about 10 billion USD in incremental sales, users 60% more likely to buy.
- Level A evidence, living status confirmed.
Objective
Turn product search into a dialogue to address purchase objections earlier in the journey and increase conversion across a catalog of hundreds of millions of items.
The deployment
Rufus is a conversational assistant built into the Amazon app and site. The customer asks a question in natural language (compare two models, check compatibility, find a gift within a budget) and Rufus answers by drawing on the catalog, the product pages, and the reviews. It has evolved toward agentic functions: recommending a single product via Help Me Decide, buying on the customer's behalf via Buy for Me, and searching beyond the Amazon catalog. Launched in beta in February 2024 in the US, extended to seven other markets in 2025.
Results Proof A
Figures announced by CEO Andy Jassy on Amazon's Q3 2025 earnings call, reported by Fortune and the business press. Result tied to financial results, with the attribution method (seven-day rolling window) specified.
How it works
Inferred typical approachThe internal detail is not public. Here is a proven approach that leads to the same result, to adapt to your stack.
The stack in detail
- llm LLM maison Amazon Language model built in-house and specialized for shopping (catalog, reviews, web); Amazon does not publish the exact architecture.
- outil RAG sur catalogue, fiches et avis Retrieval of the relevant product pages, specs, and reviews before generating the answer, to keep claims traceable.
- infra Infrastructure AWS Training and inference of the in-house foundation models on AWS infrastructure.
- outil Fonctions agentiques Help Me Decide et Buy for Me Recommendation of a single product with a generated justification, and purchase on the customer's behalf.
How it runs, concretely
For ops teams-
1Natural language question customer
The customer types or dictates a question in the search bar or on a product page.
-
2Retrieval and generation AI
The model retrieves the relevant product pages, specs, and reviews, then generates a synthetic answer with suggested products.
-
3Decision support AI
On a hesitant basket, Help Me Decide proposes a single product with a generated justification.
-
4Measure the downstream impact data team
Purchases following an interaction are attributed via a seven-day window to steer optimization.
The quality and freshness of the product pages and reviews. If a product has few reviews or incomplete specs, Rufus answers poorly and the conversion effect fades on that item.
How your customers perceive this type of use
Sourced studiesLes consommateurs n'acceptent pas les chatbots par defaut : 64% prefereraient que les entreprises n'utilisent pas d'IA dans leur service client (Gartner, 2024) et pres d'un utilisateur sur cinq du service client par IA n'en retire aucun benefice (Qualtrics, 2025). L'acceptation se construit sur trois conditions mesurees par Salesforce : savoir qu'on parle a une IA, pouvoir escalader vers un humain, comprendre la logique de l'agent.
Acceptance conditions
- Etre informe qu'on parle a une IA et non a un humain (pres de 75% le demandent, Salesforce 2024)
- Un chemin d'escalade clair vers un agent humain (45% plus enclins a utiliser l'agent IA, Salesforce 2024)
- Une logique de l'agent clairement expliquee (44% plus enclins, Salesforce 2024)
Red lines
- Rendre l'humain injoignable : c'est la premiere inquietude des consommateurs sur l'IA dans le service client (Gartner 2024) et 50% craignent que l'IA les coupe du contact humain (Qualtrics 2025)
- Remplacer le service client par l'IA sans alternative : 53% envisageraient de partir chez un concurrent (Gartner 2024)
Sources: Salesforce 2024 · Gartner 2024 · Qualtrics 2025
How to replicate
Inference, not sourcedData prerequisites
- a structured and up-to-date product catalog
- usable customer reviews
- first-party browsing history
Org prerequisites
- a data science team or an internal LLM platform
- guardrails against hallucinations on product specs
Possible stack
- a foundation LLM (via API or self-hosted)
- a RAG engine over the catalog
- a conversion attribution layer
The plan, step by step
- Step 1Scope the high-traffic product questions and structure access to the catalog, product pages, and reviews.Deliverable: v1 scope + RAG index over product pages and reviews
- Step 2Build the assistant (LLM + RAG) with guardrails against hallucinations on product specs.Deliverable: Prototype that answers within scope while citing its sources
- Step 3Integrate the assistant into the journey (search bar, product page) and set up purchase attribution on a rolling window.Deliverable: Assistant in beta on a share of traffic, measurement wired in
- Step 4Run the conversion A/B test users vs non-users and fix the categories where the assistant answers poorly.Deliverable: Conversion impact read and error rate per category
- Step 5Extend the scope and add decision-support functions.Deliverable: Assistant in production with an assisted-sales dashboard
First step: Scope a set of high-traffic product questions and wire a RAG onto the existing product pages and reviews.
The same pattern, proven in other industries
Booking.com
genAI trip planning assistant
Burberry
genAI copilot for consultants (search, image recognition, complete-the-look)
Homes.com
natural-language voice and text search, then conversational assistant
Sources
- S1 Amazon says its AI shopping assistant Rufus is so effective it's on pace to pull in an extra $10 billion in sales Established press archive pending
- S2 Amazon says its AI shopping assistant is gaining traction, with Rufus users up 115% Secondary archive pending
An error, newer info, a source?
This page lives on its accuracy. If a figure has moved, if the deployment has changed, or if you have a higher-quality source, tell us. Every sourced correction is verified before publication.