"Customers who bought this also bought…" is one of the most reliable revenue levers in e-commerce, and modern AI makes it far better than a static related-products widget. Done well, recommendations lift average order value and help shoppers find what they actually want. Done badly, they feel random or creepy. The difference is in the approach and the data.
The approaches that work
There's no single "AI recommendation" — there are a few proven techniques, and the best systems blend them:
- Collaborative filtering. "People similar to you liked these." It finds patterns across shoppers without needing to understand the products themselves. Powerful, but weak for brand-new items and new users.
- Content-based. Recommends items similar to what a shopper viewed, using product attributes — or, increasingly, embeddings that capture a product's description and image semantically.
- Behavioral / session-based. Reacts to what someone is doing right now — this session's clicks and cart — which often predicts intent better than months-old history.
Start with the data you already have
Recommendations are only as good as the signals feeding them. You almost certainly already collect what you need:
- Purchase history — the strongest signal of real preference.
- Browsing and click behavior — views, dwell time, add-to-cart, and abandons.
- Product catalog data — categories, attributes, and descriptions to reason about similarity.
Get this into a clean structure — a well-modeled PostgreSQL schema of events and catalog data — before reaching for any fancy model. Most recommendation projects fail on messy data, not weak algorithms.
Where semantic AI genuinely helps
Classic techniques struggle with cold start: a new product no one has bought yet, or a first-time visitor. This is where embeddings shine. By turning each product's title, description, and image into a vector, you can recommend semantically similar items on day one — no purchase history required. A pgvector similarity search finds "products like this one" even for an item added an hour ago.
For shoppers, the same technique powers "more like this" and natural-language search that understands intent, not just keywords.
Don't be creepy
Personalization crosses a line fast. A few rules keep it helpful rather than unsettling:
- Recommend, don't reveal. Show relevant products without spelling out how much you're tracking someone.
- Stay in context. Suggestions should relate to what the shopper is doing now, not surface something private from weeks ago.
- Respect privacy and consent. Handle data transparently and honor opt-outs; trust converts better than any algorithm.
Measure what matters
Vanity metrics like click-through are easy to game. Tie recommendations to the numbers that pay: average order value, conversion rate, and revenue per session. Run A/B tests — recommendations on versus off, or one strategy versus another — so you know the lift is real. And keep a fallback (best-sellers, or category top picks) for when personalized data is thin, so no shopper ever sees an empty slot.
Start with clean data and a simple, measurable approach, then layer in semantic AI where it clearly helps. If you want to build a recommendation engine that measurably lifts revenue rather than just looking smart, talk to us.
Originally published on the Doktouri Agency blog. We build web, mobile, SaaS, and AI products — let's talk.
Top comments (0)