Product recommendations are everywhere — “customers also bought”, “you might like”, “complete the look” — and yet most retailers cannot say with confidence whether their recommendations add revenue or simply decorate the page. This guide is about getting that answer: where to place recommendations, which algorithm fits each placement, and how to measure the lift honestly.
It is written for the people who own the result, not the people who write the code.
What recommendations are actually for
Recommendations do three distinct jobs, and conflating them is the first mistake teams make:
- Discovery — helping shoppers find products they would not have searched for. Most valuable on the homepage and category pages.
- Consideration — offering relevant alternatives and comparisons when a shopper is evaluating a specific product. The job of the product page.
- Basket building — adding complementary items to increase order value at the cart and checkout.
Each job needs a different algorithm and a different success metric. A “complete the look” carousel on a product page and a “trending now” rail on the homepage are not the same feature wearing different labels.
Where to place recommendations
Placement determines value more than the cleverness of the algorithm. The placements that consistently earn their keep:
Homepage
Shoppers here often lack specific intent. Use trending, new arrivals, and personalised picks for returning visitors with history. Keep first-time-visitor experiences sensible — popularity-based fallbacks work well.
Category and listing pages
Surface top sellers and personalised re-ranking of the grid so the most relevant products for that shopper rise to the top. This overlaps with merchandising — see AI Merchandising Control.
Product pages
The richest opportunity. Two recommendation types belong here:
- Similar products (“you might also like”) for when this one is not quite right.
- Complementary products (“frequently bought together”, “complete the look”) to build the basket.
Keep them visually distinct so shoppers understand the difference.
Cart and checkout
Focus on low-friction, complementary add-ons — accessories, consumables, warranties. Avoid distracting, high-consideration recommendations that pull people back out of the funnel.
Post-purchase and email
Order confirmation pages and lifecycle emails are underused. Replenishment reminders and complementary follow-ups perform well because intent is already proven.
Which algorithm for which job
You do not need to understand the maths to choose well. The practical mapping:
- Popularity / best sellers — simple, robust, and a strong baseline. Use for new visitors and as a fallback everywhere.
- Collaborative filtering (“people who bought X also bought Y”) — powerful for discovery and basket building, but needs traffic and suffers with new products.
- Content-based (similarity by product attributes) — good for “similar items” and for new products with no purchase history.
- Hybrid — blends the above to cover each other’s weaknesses. This is what we recommend for most retailers.
We compare these approaches in depth, including their failure modes, in Recommendation Algorithms Compared. The headline: there is no single best algorithm, only the best fit for a placement and your data maturity.
The data you need underneath
Recommendations are only as good as the signals feeding them. The essentials:
- Clean product data with consistent attributes, so content-based similarity works.
- Behavioural events — views, add-to-carts, purchases — captured reliably across devices.
- Identity resolution so a shopper’s history follows them across sessions and channels. This is the hard part, and we cover it in The Unified Customer Data Model.
A common pitfall: launching recommendations on shaky data, getting poor results, and concluding the technique does not work. Fix the foundations first.
Personalisation without creepiness
Personalisation works best when it feels helpful, not surveilled. Some principles we hold to:
- Be relevant, not eerily specific. Recommending a category a shopper has browsed is helpful; referencing a single product they looked at once can unsettle people.
- Respect consent and privacy law. Personalisation must sit within your consent framework. See GDPR and AI in eCommerce.
- Give a graceful default. New or anonymous visitors should still get sensible, popularity-based recommendations.
- Avoid filter bubbles. Mix in discovery so you are not endlessly showing variations of one thing.
Measuring whether it actually works
This is where most programmes fall down. Recommendations almost always look like they are working — the carousels get clicks — but clicks are not lift. To know the truth:
- A/B test against a holdout. Show some shoppers no recommendations (or a simple baseline) and compare. Without a control, you are guessing. Our view on testing properly is in A/B Testing with AI.
- Measure incremental revenue, not attributed revenue. Attribution credits the recommendation for sales that may have happened anyway. The holdout tells you the genuine uplift.
- Watch the right metric per placement. Discovery placements: click-through and downstream conversion. Basket placements: attach rate and average order value. Product-page placements: assisted conversion.
- Look at revenue per visitor, not just conversion rate. A recommendation can lift order value without touching conversion. See Revenue per Visitor.
A realistic expectation
In our experience, well-implemented recommendations contribute a meaningful but not headline-grabbing share of revenue — typically a single-digit to low-double-digit percentage of orders touch a recommendation, with real incremental lift somewhat lower than the attributed figure suggests. Be sceptical of vendors quoting only attributed numbers.
An implementation checklist
Before you launch, work through:
- Have you chosen one job per placement and the matching algorithm?
- Is your product and behavioural data clean and connected?
- Do you have a fallback for new and anonymous visitors?
- Have you set up an A/B test with a holdout before rollout?
- Are your success metrics defined per placement and tied to revenue?
- Have you checked the experience on mobile, where carousels behave differently?
The bottom line
Effective recommendations come from matching the right algorithm to the right placement, feeding it clean and connected data, and measuring incremental lift through a proper holdout rather than trusting attributed clicks. Done that way, they are a dependable contributor to revenue. Done casually, they are page furniture.
If you want help designing a recommendations programme and proving its lift, our AI Search and Recommendations team can scope and measure it with you. Get in touch for a frank assessment of where the opportunity sits in your store.