AI Product Recommendations: How Ecommerce Stores Can Use AI
How AI recommendation models work in ecommerce: model types, data needs, cold start, offline and online evaluation, build vs buy, LLMs and risks.
Quick answer
AI product recommendations use models trained on product data and shopper behavior to predict which products are relevant in a given context. Common approaches are co-purchase patterns, collaborative filtering, content similarity, embeddings and sequence models, usually combined and constrained by merchandising rules for stock and margin. Language models are now used to interpret needs and explain recommendations, grounded in the live catalog. Most stores should buy rather than build, evaluate models offline and then against a holdout, and watch for cold start, popularity bias and irrelevant suggestions.
Where This Fits
Recommendation types, placements and honest measurement are covered in ecommerce product recommendations; Shopify's tools are in Shopify product recommendations. This guide explains the models and how to judge them.
Model Families
The diagram above compares the main families. In practice, systems combine several: for example, embeddings for similar items, co-purchase for complements and a ranking model on top.
Data Requirements
| Data | Used for |
|---|---|
| Product views, adds, purchases with IDs | Behavior-based models |
| Orders | Complements and bought-together |
| Product attributes and categories | Similarity and cold start |
| Text and images | Embeddings |
| Stock, price, margin | Filtering and business rules |
| Consented user identifiers | Personalization across sessions |
Cold Start and Long Tail
New products have no interactions, and many catalog items have few. Content-based similarity and embeddings from product text and images help them appear; merchandising rules can seed new launches. Without this, recommendations concentrate on a small set of popular products, which limits discovery.
Are your recommendations actually adding revenue?
ZSpace evaluates recommendation models against holdouts and tunes them with your merchandising rules.
LLM-Assisted Recommendations
Language models can turn a request like “a gift for a runner under 80” into constraints, choose candidates from a retrieval system and explain why each fits. The key is grounding: the model should only recommend products retrieved from the live catalog, with real prices and stock, and should not invent specifications.
Evaluating Models
| Stage | Method | Watch |
|---|---|---|
| Offline | Held-out interaction data; precision, recall, coverage | Optimizing for clicks, not revenue |
| Online | A/B test against a holdout | Revenue per session, AOV, conversion |
| Guardrails | Monitor after launch | Returns, margin, stock-outs, diversity |
Build vs Buy
Platform features and vendors handle most needs, include tooling for rules and reporting, and improve over time. Building makes sense for large catalogs with unusual relationships (compatibility, configuration), strict data requirements or a strong in-house data team, and it means owning pipelines, retraining and monitoring. See AI ecommerce.
Placement Strategy by Page
The right recommendation depends on where the shopper is and what they're trying to do.
| Page | Shopper goal | Recommendation type |
|---|---|---|
| Homepage | Explore | Personalized picks, trending, recently viewed |
| Category page | Browse a range | Popular in category, personalized ordering |
| Product page (above add to cart) | Decide | Alternatives, similar items |
| Product page (below) | Complete the purchase | Frequently bought together, accessories |
| Cart | Finish | Low-cost add-ons, compatible items |
| Post-purchase email | Use and return | Complementary items, replenishment |
| Empty search results | Recover | Popular or related items |
Measuring Recommendations Properly
Recommendation widgets often report their own clicks and attributed revenue, which overstate their effect because shoppers might have found those products anyway. Test strategies with random assignment and measure whole-visit outcomes: revenue per visitor, conversion, average order value and returns. Keep placement the same when comparing algorithms. See personalization testing.
Merchandiser Controls
Merchandisers need to shape recommendations: exclude out-of-stock, low-margin or sensitive products, prevent recommending items from different gender or age ranges where inappropriate, boost strategic ranges, and set rules for specific products. Good tools combine model output with these controls and show why a product was recommended. See AI merchandising.
Recommendations Without Much Data
Small stores and new catalogs often lack the behaviour data that collaborative filtering needs. Content-based approaches work better here: recommend products with similar attributes, from the same collection or complementary categories defined by merchandisers. Rules such as "accessories for this product type" and bestsellers within a category are simple, transparent baselines. Move to learned models as data grows, and test each step against the simpler baseline.
| Data available | Suitable approach |
|---|---|
| Little behaviour data | Attribute similarity, merchandiser rules, category bestsellers |
| Moderate order history | Co-purchase (frequently bought together) |
| Rich behaviour data | Collaborative filtering, session-based models |
| Rich data plus text and images | Hybrid and embedding-based models |
Privacy and Consent in Recommendations
Personalized recommendations use browsing and purchase data, so they fall under privacy rules and consent settings in many markets. Non-personalized recommendations (similar items, frequently bought together across all customers) can work without individual profiles and are a sensible fallback when consent isn't given. Avoid recommendations that reveal sensitive purchases, for example in shared-device or email contexts. See ecommerce privacy and customer data.
Risks
- Popularity bias narrowing what shoppers see
- Recommending just-bought or out-of-stock items
- Irrelevant pairings that reduce trust
- Measuring clicks instead of incremental revenue
- Generated explanations that misstate facts
- Heavy scripts slowing product pages
Planning AI recommendations?
Talk to ZSpace about recommendation systems, data pipelines and testing.
Conclusion
AI recommendation models work when they're fed clean data, combined with merchandising rules, grounded in the live catalog and judged on incremental revenue. Start with platform or vendor models, measure honestly, and invest in data before model complexity. For cross-sell and upsell strategy, see cross-selling and upselling.
Common questions
Models learn from product data and shopper behavior, such as what's viewed and bought together, which shoppers behave alike, and which products are similar, then score products for each context.