The Amazon recommendation system started with item-to-item collaborative filtering and now blends deep learning and generative AI. Agentic AI is the next step for retailers: agents that combine live behaviour, inventory, margin and context to choose the next best product, offer or message, within business rules, across web, app, store and service channels.
Updated October 2026
Who this is for: Chief digital officers, heads of e-commerce, merchandising and CRM, and retail data leaders in fashion, grocery, electronics, home and marketplaces.
The business problem: recommendations that ignore the business
Most retail recommenders still answer one question: "what did similar shoppers click?" They do not know a product is out of stock in the shopper's region, that its margin is negative after shipping, or that the customer returned the same item last month. The result is wasted placements, promoted items that cannot ship, and personalisation that feels generic. Every homepage slot, email and search result has a cost; poor recommendations waste it at scale.
What changed in 2026
- Amazon describes how generative AI now personalises product recommendations and descriptions in its store, as covered on About Amazon.
- Amazon Personalize remains AWS's managed recommender and can be combined with Amazon Bedrock foundation models for segmentation and content variations.
- McKinsey's State of AI 2025 found 23% of organisations scaling an agentic AI system somewhere, with no more than 10% in any one function: retail recommendation agents are still early.
- Gartner warns over 40% of agentic AI projects may be cancelled by 2027 because of cost, unclear value or weak risk controls, so ROI discipline matters.
How the Amazon recommendation system works
Amazon's approach, described in its own history of the recommendation algorithm, began with item-to-item collaborative filtering, published by Amazon engineers in IEEE Internet Computing in 2003. Instead of matching similar customers, it matches items that are often bought or viewed together, which scales to huge catalogues and works in real time. Over time, Amazon added deep learning, more signals and generative AI. The lesson for other retailers: start from relationships between products and behaviour, then layer in context.
From recommender to recommendation agent
| Classic recommender | Agentic recommendations | |
|---|---|---|
| Goal | Predict clicks or purchases | Optimise a business goal: margin, conversion, retention |
| Inputs | Behaviour history | Behaviour plus live inventory, price, margin, returns, weather, events |
| Output | Ranked product list | Product, offer, bundle or message, plus the channel and timing |
| Rules | Hard-coded filters | Policies the agent must follow (brand, pricing, fairness, privacy) |
| Learning | Periodic retraining | Continuous feedback and experiments |
Reference architecture
clicks, carts, orders, stock, price
real-time features, embeddings, PostgreSQL / pgvector
collaborative filtering, Personalize, embeddings
re-ranks with margin, stock and policy
web, app, email, store screens
Figure: Agentic retail recommendations. Live signals become features; models generate candidates; an agent re-ranks them against margin, stock and policy; outcomes feed back for learning.
Keep candidate generation fast and cheap. Use the agent for the last step, where business context matters. Fresh features are critical; see real-time feature engineering. Graph models also help with product relationships, as in our GNN applications guide.
Use cases retailers start with
- In-stock aware recommendations: never promote what cannot ship to that postcode.
- Margin-aware bundles: cross-sell accessories that lift basket value without heavy discounting.
- Returns-aware sizing: steer shoppers away from items they are likely to return.
- Next best action in CRM: choose product, offer and channel for each customer.
- In-store and clienteling: give associates suggestions on tablets, linked to autonomous store technology.
Are your recommendations ignoring stock and margin? Get a free retail recommendation assessment with a ranked list of quick wins.
How to start in 30 days
- Week 1: Pick one surface (product page, cart or email). Record current conversion, revenue per session and out-of-stock clicks.
- Week 2: Join behaviour, inventory and margin data; define the policies the agent must obey.
- Week 3: Add an agent re-ranking step on top of your current recommender, offline-tested on past sessions.
- Week 4: Run an A/B test on a small traffic share; decide on numbers, not opinions.
Where it's heading: 2020 → 2030
| Period | Retail recommendations |
|---|---|
| 2020 | Collaborative filtering and deep learning recommenders; batch updates. |
| 2026 | Generative AI descriptions and shopping assistants; agents start re-ranking with business context. |
| 2028 (expected) | Agents coordinate offers, pricing and inventory across channels within set policies. |
| 2030 (our view) | Shoppers' own AI agents negotiate with retailer agents; machine-readable catalogues and trust become key. |
How NexaStack helps
NexaStack provides a composable agent framework, a unified inference engine and an observability and evaluation layer, with policy and safety controls, deployable in your own cloud account, private cloud or on-premises. You keep your data platform and recommender; NexaStack runs and governs the agents on top. See also our AI agent infrastructure stack.
FAQ
How does the Amazon recommendation system work?
It started with item-to-item collaborative filtering, recommending products often bought or viewed together, and has added deep learning and generative AI over time.
Does Amazon use AI to recommend products?
Yes. Amazon uses machine learning and generative AI to personalise recommendations and product descriptions.
What are the latest retail recommendation trends in 2026?
Generative AI shopping assistants, agentic re-ranking with stock and margin, real-time features, and preparing catalogues for shoppers' own AI agents.
What does it cost, and how do we protect customer data?
Start with one surface and an A/B test, so spend tracks proven lift. Run models and agents inside your own cloud account or private environment, honour consent flags, and keep personal data out of prompts where it is not needed.
Want a stock- and margin-aware recommendation agent live on one surface in 30 days? Book a retail recommendation pilot with NexaStack.
