Recommendation Engines for E-commerce: Boost Sales with Personalised AI

Walk into a good independent shop and the owner will notice what you pick up, remember what you bought last time and suggest something you had not thought of. Online, that role is played by a recommendation engine for ecommerce. Done well, it makes a large catalogue feel curated. Done badly, it shows you the kettle you bought yesterday for the next three weeks.

A recommendation engine is software that predicts which products a shopper is most likely to want, then displays them in the right place: the homepage, product pages, the basket, search results, emails and push notifications. It uses customer behaviour analytics (views, clicks, purchases, returns, searches) together with product data to rank items for each person or session.

The business case is well documented. McKinsey’s personalisation research found that personalisation most often drives a 10 to 15% revenue lift, and that 71% of consumers expect companies to deliver personalised interactions. For UK online retailers facing rising acquisition costs, selling more to existing visitors is often the cheapest growth available.

This guide is for ecommerce directors, founders, CTOs and marketing leaders at UK and US retailers and marketplaces. It explains how recommendation algorithms work, which approach suits your catalogue, where to place recommendations for maximum effect, what ASOS has done at scale, the common pitfalls and how to build or buy the right personalisation engine.

Why Product Recommendations Matter More Than Ever

Three pressures make ecommerce personalisation a priority rather than a nice extra:

• Acquisition is expensive: Paid search and social costs keep rising, so conversion rate and average order value have a bigger effect on profitability.
• Catalogues are growing: With thousands or millions of SKUs, customers cannot browse everything. Relevant suggestions help them find products they would otherwise never see.
• Expectations are set by the biggest platforms: Shoppers used to Amazon, Netflix and Spotify notice when a site treats them like a stranger.

A personalised shopping experience also reduces friction in mobile browsing, where screen space is limited and every scroll risks losing attention.

How Recommendation Algorithms Work

1. Collaborative Filtering

Collaborative filtering recommends products based on the behaviour of similar shoppers. If people who bought a particular running shoe also tended to buy a specific pair of socks, the system suggests those socks to the next shoe buyer. It needs no understanding of the products themselves, only interaction data.

Its weakness is the cold start problem: new products with no interactions, and new visitors with no history, give it nothing to work with.

2. Content-Based Filtering

Content-based filtering recommends items with similar attributes to those a shopper has engaged with, using category, brand, colour, material, price, text descriptions and increasingly image features.

It handles new products well because attributes exist from day one, but can become repetitive, showing endless variations of the same thing.

3. Hybrid and Deep Learning Approaches

Most production systems are hybrids. Modern recommendation engine development typically uses a two stage design: a fast retrieval model narrows millions of products to a few hundred candidates, then a ranking model orders them using rich features such as session behaviour, stock levels, margin and time of day.

Deep learning models, including two tower networks and sequence models that treat a browsing session like a sentence, now power many of the largest retailers’ recommendations.

Do, Checkout: Machine Learning vs Deep Learning: Which One Is Right for Your Business?

Where to Place Recommendations for the Biggest Impact

Placement matters as much as the algorithm. Each position serves a different shopper intent:

1. Homepage: For returning visitors, show items related to recent browsing. For new visitors, show trending products for their location or traffic source.
2. Product detail pages: “Similar items” helps undecided shoppers. “Frequently bought together” drives cross-sell.
3. Basket and checkout: Low cost add-ons and accessories (cross-sell upsell AI works best here with affordable, relevant items, not expensive distractions).
4. Search results: Personalised ranking puts the items a shopper is most likely to buy at the top.
5. Empty and zero result pages: Recommendations rescue sessions that would otherwise end.
6. Email and push: Back in stock alerts, price drops and post purchase suggestions timed to the product’s replenishment cycle.

Real-time product suggestions that react to what a shopper is doing in the current session typically outperform recommendations based only on long-term history, particularly for anonymous visitors.

Real Business Example: ASOS

The challenge.:

• ASOS, the UK online fashion retailer, sells tens of thousands of products to millions of customers across many markets, with new lines arriving constantly.
• Fashion adds extra difficulty: styles change quickly, products often have short lives and fit problems drive expensive returns.

The solution:

• ASOS built an in-house AI team to develop personalisation tools. Its product pages feature a “You Might Also Like” carousel driven by machine learning, alongside tools such as Style Match visual search, which lets shoppers upload a photo and find similar items, and Fit Assistant, which uses past purchases and returns to recommend sizes.
• ASOS data scientists have described using collaborative filtering, reasoning that products often bought by the same customers are similar in some way, and building models that generate complete outfits from a single seed product.

Implementation:

• Rather than a single engine, ASOS combined several models serving different moments: similar item recommendations, outfit completion, visual search and size recommendations.
• It also introduced a Profile Builder that asks new customers to react to a small sample of products, reducing the cold start problem.

Business outcome:

• These tools help customers navigate a vast catalogue, increase cross-sell through outfit suggestions and reduce returns through better size advice, addressing both revenue and cost.
• The broader lesson is that personalisation is most effective as a set of targeted tools, each solving a specific customer problem, rather than one generic widget.

Ready to personalise your store? Talk to IIH Global about a recommendation engine built around your catalogue and customers.

Measuring the Return on a Recommendation Engine

Click-through rate on a carousel is easy to report but can mislead, because some shoppers would have bought those products anyway. Stronger measures include:

• Incremental revenue per visitor measured through controlled A/B tests
• Conversion rate and average order value for exposed versus control groups
• Items per order and attach rate for accessories
• Return rate on recommended items (recommendations that cause returns destroy margin)
• Catalogue coverage: the share of your range that ever gets recommended

Run holdout groups continuously, not just at launch. That is the only reliable way to show ongoing ecommerce conversion optimisation value to the finance team.

Common Mistakes to Avoid

• Recommending what was just bought: Suppress purchased items unless they are consumables with a known repurchase cycle.
• Ignoring stock: Suggesting out of stock or low size availability items frustrates shoppers. Feed live inventory into ranking.
• Popularity bias: Algorithms naturally favour bestsellers, which starves new and long tail products of exposure. Add diversity and freshness controls.
• Ignoring margin and returns: The most clicked item is not always the most profitable one.
• Slow page performance: A recommendation widget that adds a second to load time can cost more conversions than it gains.
• Weak product data: Missing or inconsistent attributes cripple content-based models. Clean catalogue data first.

Privacy and Compliance

Personalisation relies on personal data, so UK GDPR and the Privacy and Electronic Communications Regulations apply. Be clear about tracking in your cookie consent, avoid sensitive inferences, give customers control over personalised marketing and keep data only as long as needed.

The Information Commissioner’s Office publishes practical guidance on cookies, profiling and automated decision making that every ecommerce team should read. Personalisation that feels invasive damages trust faster than it builds sales.

Build, Buy or Customise?

SaaS personalisation engines and platform plug-ins (for Shopify, Magento, BigCommerce and similar) offer a quick start and suit smaller stores. As traffic, catalogue size and ambition grow, their limits become visible: restricted control over ranking logic, difficulty combining business rules with models and costs that scale with traffic.

Custom engines make sense when you have a large or unusual catalogue, significant traffic, proprietary data such as loyalty or store sales, or business rules (margin, stock, brand agreements) that off the shelf tools cannot handle.

IIH Global’s machine learning development services cover recommendation models end to end, and our AI integration services connect them to your storefront, search, CRM and email platforms.

If you have an internal team that needs extra capacity, you can also hire AI developers to work alongside them.

Future Trends in Retail AI Recommendations

Generative AI is changing how recommendations are presented. Conversational shopping assistants can ask clarifying questions (“Is this for a wedding or everyday wear?”) and explain why they suggest a product. Multimodal models that understand images and text together are improving visual similarity and style matching.

Real-time personalisation is extending beyond product lists to page layout, content and promotions. At the same time, tighter privacy controls and the decline of third party cookies are pushing retailers to invest in first party data and consented personalisation.

Conclusion

A recommendation engine for ecommerce is one of the most direct ways AI can increase revenue, because it acts at the moment of purchase. The technology is mature, but results depend on good product data, sensible placement, attention to stock and margin, and disciplined testing.

Start by identifying the pages and moments where shoppers most often hesitate or leave, test a focused recommendation there and measure incremental impact honestly. If you want expert help designing a personalisation engine that fits your catalogue and customers, speak to IIH Global.

Frequently Asked Questions

What is a recommendation engine in ecommerce?

It is software that analyses shopper behaviour and product data to predict which items each customer is most likely to want, then displays them across the website, app and emails.

What is the difference between collaborative and content-based filtering?

Collaborative filtering uses behaviour of similar shoppers to suggest products. Content-based filtering recommends items with similar attributes to those a shopper already viewed or bought.

How much can product recommendations increase sales?

Results vary by retailer, but McKinsey research found personalisation most often lifts revenue by 10 to 15%. Controlled A/B testing is the best way to measure your own uplift.

Do small online stores need a recommendation engine?

Smaller stores can start with platform plug-ins or simple rules such as bestsellers and frequently bought together, then move to machine learning as traffic and catalogue size grow.

Are AI product recommendations GDPR compliant?

They can be. Businesses must obtain valid consent for tracking cookies, be transparent about profiling, avoid sensitive inferences and let customers opt out of personalised marketing.

Published by: UK Crecso