How do I build product recommendations for an ecommerce site?

To build product recommendations for an ecommerce site, start simple and evolve toward personalization.

1) Pick the recommendation use cases

Common placements:

• Homepage: trending, popular, personalized picks
• Product page: similar items, frequently bought together
• Cart/checkout: add-ons, accessories, complementary items
• Email/push: recently viewed, back-in-stock, replenishment

2) Collect the right data

You’ll usually need:

• Catalog data: product ID, category, brand, price, attributes, tags, availability
• User behavior: views, clicks, add-to-cart, purchases, search queries
• Context: time, device, location, referral source
• Feedback signals: returns, ratings, dislikes, skips

3) Start with baseline recommendation methods

These are easy to implement and often strong enough initially:

Popularity-based

Recommend top-selling or most-viewed items.

• Good for new users and homepage
• Easy fallback when you lack history

Rule-based / editorial

Example:

• “If user bought a laptop, recommend a laptop sleeve”
• “Show items from same category and price band”

Content-based

Recommend products similar to the one being viewed based on attributes/text:

• Category, brand, price, color, material, description embeddings
• Great for “similar products”

Collaborative filtering

Recommend based on behavior of similar users or item co-occurrence:

• “Users who bought X also bought Y”
• Works well once you have enough interaction data

4) Add personalization

Use each user’s:

• Browsing history
• Purchase history
• Preferred categories/brands
• Price sensitivity
• Recency/frequency of interactions

A simple approach:

• Score products by similarity to the user’s past interactions
• Combine with business rules like stock, margin, or promotion priority

5) Build candidate generation + ranking

A practical recommender often has two stages:

Candidate generation

Pull a few hundred likely items from:

• Similar products
• Popular items in user’s category
• Co-purchase graph
• Search/query matching

Ranking

Sort candidates using a model or scoring function with features like:

• User-item affinity
• Category match
• Price proximity
• Recency
• Inventory status
• Margin / promotion flag

6) Handle cold start

New user

Use:

• Popular items
• Trending by region/device
• Contextual recommendations
• Session-based behavior

New item

Use:

• Content-based features
• Category/brand/price similarity
• Exploration to collect interactions

7) Evaluate recommendations

Offline metrics:

• Precision@K / Recall@K
• NDCG@K
• MAP
• Coverage and diversity

Online metrics:

• CTR
• Add-to-cart rate
• Conversion rate
• Revenue per session
• Average order value
• Return rate

Always A/B test against a control.

8) Balance relevance with business goals

You may want to optimize for:

• Revenue
• Conversion
• Margin
• Inventory clearance
• Customer satisfaction

Common constraints:

• Don’t recommend out-of-stock items
• Avoid over-recommending discounted low-margin items if that hurts profit
• Ensure diversity so the list isn’t repetitive

9) Production architecture

Typical pipeline:

1. Ingest events from site/app
2. Update user/item profiles
3. Generate candidates
4. Rank them
5. Filter by business rules
6. Cache results for fast serving
7. Log impressions/clicks/purchases for retraining

10) Recommended implementation path

If you’re starting from scratch:

1. Popularity + category-based rules
2. Co-purchase / “also bought”
3. Content-based similarity
4. Collaborative filtering
5. Learning-to-rank / deep personalized model

11) Useful model choices

• Baseline: popularity, association rules
• Lightweight: matrix factorization, item-item similarity
• Mid-level: gradient boosted rankers
• Advanced: two-tower retrieval models, deep ranking models, sequence/session models

12) Don’t forget safeguards

• Deduplicate recommendations
• Respect privacy and consent
• Avoid recommending already purchased items unless relevant
• Add explainability where useful:
  • “Because you viewed…”
  • “Frequently bought with…”

If you want, I can also give you:

1. a simple architecture diagram,
2. a Python/SQL implementation, or
3. a modeling approach for a small vs. large ecommerce site.