- 1 Collect views, carts and purchases per user
- 2 Start with simple rules such as bought together
- 3 Place blocks on product page, cart and emails
- 4 Hold out a control group for each placement
- 5 Add model-based ranking once rules are measured
Personalised recommendations show each customer products they are likely to want, based on what they and similar customers viewed or bought. They appear as blocks on product pages, in the cart, on the home page, in emails and inside apps. They fit stores and apps with a catalogue big enough that customers cannot easily browse it all, and with enough traffic to learn from.
Why it works
Most shoppers see only a small part of a catalogue. Recommendations surface items they would not have found, at the moment they are already shopping. A matching accessory in the cart, or a refill a month after the first purchase, solves a real need instead of adding noise.
Relevance is what earns attention. A generic “you may also like” block full of unrelated best sellers gets ignored. A block that clearly connects to what the customer just did feels like help.
How to do it
- Get clean event data. Track product views, add-to-cart, purchases and searches with a consistent product ID across site, app and email tools.
- Start with rules. Build “frequently bought together” from orders, “viewed this also viewed” from sessions, and “complete the set” from your own product relationships. These are easy to check by hand.
- Choose placements with intent. Product pages suit alternatives and add-ons. The cart suits small add-ons that do not distract from checkout. Post-purchase emails suit complementary items and refills.
- Exclude the obvious misses. Remove out-of-stock items, products already bought that customers rarely buy twice, and items that do not fit their size or region.
- Use recommendations in email and push. Add a block based on recent browsing to your newsletters, and send a browse follow-up when someone views several products in one category.
- Measure with a control group. Hide recommendations for a small share of users, or show generic best sellers, and compare revenue per user. Click rates alone can mislead.
- Add smarter ranking once you have data. Many commerce platforms and email tools include recommendation engines. Test them against your rules before switching fully.
- Review regularly. Check samples of recommendations each month for odd or awkward pairings, and adjust rules.
Example
A D2C home goods brand shows “best sellers” in every recommendation block, so customers buying a duvet see random candles. The team switches the product page block to items often bought with the viewed product, and the cart block to small add-ons under a set price. Post-purchase emails a week after delivery show items that complete the room. They keep a holdout group that sees the old best sellers block. After a few weeks, they compare revenue per visitor between the two groups before rolling the change out.
A mobile recipe app uses saved recipes to recommend new ones on the home screen, and swaps out ingredients the user has marked as disliked.
Metrics to track
- Revenue per visitor or user in the personalised group compared with the control. This is the main measure of success.
- Average order value, which should rise if add-on recommendations are relevant.
- Share of orders containing a recommended item. Growth shows customers find the suggestions useful.
- Recommendation click-through rate by placement, used to compare blocks and formats.
- Repeat purchase rate for customers who receive personalised post-purchase emails.
When it backfires
- Creepy or sensitive suggestions. Recommending products that reveal health, pregnancy or other private details can upset customers. Exclude sensitive categories.
- Distracting checkout. Too many suggestions in the cart can reduce checkout completion. Keep the cart block small.
- Recommending what they just bought. Suggesting the same durable item again looks careless.
- Measuring clicks only. A block can get clicks while pulling sales away from items customers would have bought anyway.
- Bad data. Mismatched product IDs or missing events lead to irrelevant suggestions.
- Filter bubbles. Showing only items close to past purchases can hide new categories a customer might love. Leave some room for new arrivals and best sellers.
Growth Marketing Academy
Learn the framework behind this tactic
These lessons explain the math and the mental model, so you can adapt the tactic instead of copying it.
- Track 4 · Retention RFM Segmentation: Score Every Customer and Act on the Right Ones Score customers on recency, frequency and monetary value, name the segments, give each one a playbook, and build it in a spreadsheet or SQL. Includes the B2B usage version.
- Track 6 · Monetization Pricing and the Value Metric Pick the unit you charge for, package it into good/better/best tiers, research willingness to pay and localise prices for India and other PPP markets.
Questions people ask
Do I need machine learning to start?
No. Simple rules such as "frequently bought together" or "others who viewed this also viewed" often work well and are easy to explain. Move to models when you have enough data and a way to measure improvements.
What should I show customers with no history?
Show best sellers, items popular in their region or category, or products related to the page they are on. Personalise as soon as they view or add something.
How do I handle privacy?
Tell customers in your privacy policy how you use browsing and purchase data, follow consent rules for tracking in your market, and avoid recommendations that reveal sensitive information.