Personalised product recommendations: your route to more revenue

Personalised product recommendations are the most underrated revenue lever in ecommerce. Shops spend thousands on advertising to win new customers and ignore the potential in the visitors they already have.

The numbers are clear: personalised recommendations generate 20 per cent more revenue than generic suggestions (McKinsey). Smart recommendation engines lift the conversion rate by 26 per cent on average (Barilliance). Amazon generates around 35 per cent of its revenue through recommendations alone. That is no accident. It is the result of years of optimising algorithms, data and user behaviour.

The technology behind recommendations: how algorithms think

Modern recommendation systems rest on three core algorithms. Each has its strengths, weaknesses and ideal use cases.

Collaborative filtering: “Customers like you also bought…”

Collaborative filtering analyses the behaviour of similar users. If customer A and customer B both bought product X, and customer A also bought product Y, product Y is recommended to customer B as well. The advantage: you need no product data, only user behaviour. The drawback: the cold-start problem. New products without a purchase history do not get recommended.

It gets more effective as the data grows. From 10,000 transactions a month, collaborative filtering delivers significantly better results than manual recommendations.

Content-based filtering: “Products with similar attributes”

This algorithm analyses product attributes: category, colour, price, material, brand. A customer looking at red trainers is shown more red trainers or trainers from the same brand. The advantage: it works straight away, even without purchase history. The drawback: it creates filter bubbles. The customer only sees variations of what they already know.

Hybrid systems: combining both approaches

The most advanced recommendation engines combine both methods. Netflix and Amazon use hybrid systems that link user behaviour with product attributes. The results are more precise, more varied and less prone to the cold-start problem. Implementation is more complex: the return justifies the effort from 50,000 visitors a month.

GDPR, first-party data and the future of personalisation

With the end of third-party cookies, the data behind personalised recommendations changes fundamentally. Shops that do not switch to first-party data lose the ability to personalise. The answer: build up your own data systematically.

First-party data you should be collecting today:

• Purchase history: the most valuable data source. Which products does the customer buy, how often, at what time of year?
• Browsing behaviour: which categories do they visit, how long do they stay on a page, which filters do they use?
• Email interaction: which newsletters do they open, which links do they click?
• On-site search: what does the customer search for on your site? Search queries reveal needs the customer would not otherwise tell you.

The legal basis: Art. 6(1)(b) GDPR (performance of a contract) covers personalisation within an existing customer relationship. For prospects you need consent (Art. 6(1)(a) GDPR), which has to be obtained through a transparent cookie banner.

Behavioural triggers: the right recommendation at the right moment

Context decides how well a recommendation works. The same suggestion on the home page and in the basket produces different conversion rates. Leading shops use seven behavioural triggers to place recommendations at the right time and in the right context:

Browsing behaviour: a customer keeps visiting the “Running shoes” category without buying. A pop-up with “This week’s most popular running shoes” after the third session converts at 8 to 12 per cent.

Abandoned basket: 69.8 per cent of all baskets are abandoned (Baymard Institute). An email with personalised recommendations for similar products one hour after abandonment recovers 10 to 15 per cent of the lost revenue.

Purchase history: a customer buys protein powder every 6 weeks. An email after 5 weeks saying “Your protein powder may be running low” converts at 25 to 35 per cent.

Price tracking: a customer has viewed a product several times without buying. A price drop alert converts at 20 to 30 per cent.

Time on page: customers who stay on a product page for more than 3 minutes without buying get exit-intent pop-ups with complementary products or reviews.

Device switching: a customer browses on mobile but buys on desktop. Cross-device tracking makes consistent recommendations across all channels possible.

Seasonality: coats in winter, sunglasses in summer. Context-aware recommendations raise relevance by 25 to 35 per cent.

Recommendation engines: build or buy

For shops that want to go beyond manual recommendations, there are three options:

Specialist recommendation services: tools such as Nosto, Dynamic Yield (Mastercard) or Barilliance offer AI-driven personalisation. Cost: €200–2,000/month. The return typically turns positive within 3 months. Ideal for shops with 50,000 visitors a month or more.

In-house development: machine learning models in Python (scikit-learn, TensorFlow) on your own data. High upfront effort (€50,000–150,000), but full control and no ongoing licence fees. Only worth it for shops with their own development team and >100,000 visitors a month.

Measuring the return: what counts and what does not

Measuring the return on personalised recommendations takes more than a look at revenue. Four metrics are essential:

Important: always compare against a control group. Show recommendations to 50 per cent of visitors and none to the other 50 per cent. After 4 weeks, measure the difference. Otherwise you will not know whether the revenue increase came from the recommendations or from seasonal effects.

The most common mistakes with personalised recommendations

Five patterns come up again and again:

1. 1
   

   Automating too early

   Shops with fewer than 500 transactions a month invest in expensive AI tools, although manual recommendations would deliver the same results. There is not enough data for machine learning.

2. 2
   

   No context

   Recommendations on the home page that ignore what the user is doing right now are generic suggestions with an expensive label. Personalisation needs context.

3. 3
   

   Over-personalisation

   Too many personalised elements on one page overwhelm the user. The ideal density: 2 to 3 personalised modules per page. More than that creates a “Big Brother” effect.

4. 4
   

   No A/B tests

   Recommendation engines get implemented and never tested. Every algorithm, every placement, every number of suggestions should be tested. Without testing, you are guessing.

5. 5
   

   Neglecting the mobile experience

   60 to 70 per cent of traffic comes from mobile devices. Recommendation carousels that work on desktop are unusable on a smartphone. Mobile-first design is a must.

Real-time personalisation and predictive analytics

The next generation of personalisation does not happen in minutes but in milliseconds. Real-time personalisation analyses what the user is doing right now and adapts the page instantly.

Examples of real-time personalisation:

• Dynamic home page: a customer who keeps visiting sports shoes does not see general bestsellers on the home page but running shoes and fitness equipment.
• Prices for returning customers: regular customers automatically see VIP prices without entering a code.
• Content personalisation: product descriptions change based on the customer’s level of experience. A beginner sees explanations, an expert sees specifications.

Predictive analytics goes one step further: it predicts what the customer will buy next, before they know it themselves. Amazon uses predictive analytics to stock products before demand rises. In ecommerce, predictive analytics can be used to:

• identify churn risks (customers who might leave)
• predict next orders (for automated email triggers)
• calculate price elasticity per customer segment

The technical prerequisite: a data warehouse (such as Google BigQuery or Snowflake) and a machine learning model trained on historical data. The effort is high, the competitive advantage long-lasting.

Examples: what leading companies do differently

Amazon: “Customers who bought this item also bought…” is based on item-to-item collaborative filtering. The system compares products, not users. The result: it scales to billions of items and stays highly relevant.

Netflix personalises both recommendations and thumbnails. A user who prefers action films sees a different preview image for the same film than a user who prefers dramas. The click-through rate rises by 15 to 20 per cent.

Spotify: “Discover Weekly” uses collaborative filtering at playlist level. Every user gets 30 new songs every week, based on the listening patterns of similar users. The playlist has a retention rate of 60 per cent, higher than any other Spotify feature.

The implementation roadmap

1. Data analysis (weeks 1–2): export your order data for the last 12 months. Identify the top 50 products by revenue and the most common basket combinations.
2. Manual recommendations (weeks 3–4): define 2–3 complementary products for each of the top 50 products. Place them on the product page and in the basket.
3. Measurement (weeks 5–8): track clicks and conversion of the manual recommendations. Calculate the AOV impact.
4. Tool evaluation (weeks 9–10): if the manual impact is positive, evaluate tools such as Nosto, Barilliance or Dynamic Yield.
5. Automation (weeks 11–16): implement the chosen tool. A/B test algorithms, placements and design variants.
6. Optimisation (ongoing): review the metrics monthly. Adjust the strategy quarterly.

Personalised product recommendations are not a nice-to-have. They are a strategic competitive advantage that grows stronger every month. If you invest early, you collect data that competitors who start later will not have.