Personalised Product Recommendations in E-Commerce: Where AI Really Works and Where It Does Not

Product recommendations are often considered one of the fastest ways to deploy AI productively in e-commerce. The expectation: activate a system, let algorithms work, increase revenue. This notion is fuelled by success stories from large platforms like Amazon or Zalando, which allegedly generate a significant portion of their revenue through recommendations.

The reality for most companies looks different. Personalised product recommendations are neither easy to implement nor do they automatically guarantee better results. There is a considerable gap between perceived simplicity and operational complexity.

This article provides an analytical overview of AI-powered product recommendations. It is aimed at e-commerce managers, growth and CRM leads, and marketing leaders who want to understand what is realistically achievable, what prerequisites must be met, and where the typical pitfalls lie.

What Personalised Product Recommendations Actually Are

Personalisation as Probabilistic Relevance

Product recommendation systems do not understand customers. They calculate probabilities. The central question is not "What does this customer want to buy?" but rather "Which products have the highest statistical probability of being relevant to this user?"

This distinction is fundamental. Recommendation systems are based on patterns in historical data: purchasing behaviour, browsing activity, click paths, basket compositions. They identify correlations and derive predictions from them. However, they understand neither motivation nor context nor actual purchase intent.

A customer who has bought blue shirts multiple times receives recommendations for more blue shirts. Whether they are currently looking for a gift, want to change their style, or bought the blue shirts for someone else remains hidden from the system.

Personalisation, Segmentation, and Contextual Relevance

These three concepts are often confused but function differently:

Personalisation uses individual user data: purchase history, browsing behaviour, explicit preferences, interaction history. The more data points available, the more specific recommendations can become. The problem: for new customers or users with low activity, this data is missing.

Segmentation groups users by shared characteristics: demographic data, purchasing behaviour patterns, entry channels, price sensitivity. Recommendations are optimised at group level, not individually. The advantage: relevant recommendations are possible even with limited individual data.

Contextual relevance considers situational factors: time of day, device, entry page, current basket contents, weather, season. Context signals can supplement individual data or replace it when user data is missing.

In practice, effective systems combine all three approaches. Pure personalisation fails with new customers. Pure segmentation ignores individual differences within groups. Pure context does not capture longer-term preferences.

Recommendations Influence, They Do Not Create Purchase Intent

A common misconception: recommendations generate demand. This is incorrect in most cases. Recommendations influence product selection within an existing purchase intent. A customer already looking for shoes can be directed to a specific model through recommendations. A customer without purchase intent is rarely converted to a buyer through recommendations.

This distinction has practical consequences. Recommendation systems are most effective when customers are already in the decision process: on product pages, in the basket, after a search. They are less effective on home pages or in emails to inactive users.

Which KPIs Can Realistically Be Improved and Why

Conversion Rate

Relevant product recommendations reduce search effort. Customers find suitable products faster, which increases the probability of purchase completion. Typical improvements range from 5 to 15 percent, depending on baseline and product complexity.

Causal mechanism: Reduced decision complexity, increased relevance of displayed products, shorter customer journey.

Typical side effects: With overly aggressive personalisation, discovery can be restricted. Customers see only "more of the same" and potentially miss interesting products outside their previous pattern.

Average Order Value

Cross-selling and upselling recommendations can increase average order value. "Customers also bought," "Goes well with," and "Upgrade options" are established formats.

Causal mechanism: Increased visibility of complementary products, perception of value through bundles, anchoring to higher-value alternatives.

Typical side effects: With wrong recommendations, trust decreases. Overly aggressive cross-selling tactics can be perceived as pushy and worsen customer experience.

Typical improvements range from 5 to 20 percent, heavily dependent on product structure. Shops with natural complementary products benefit more than those with homogeneous assortments.

Revenue per Visitor

As a combination of conversion rate and order value, revenue per visitor is the aggregate success metric. Improvements of 10 to 25 percent are possible with well-implemented systems.

Time on Site

Personalised content increases engagement. Users spend more time exploring when displayed products appear relevant.

Bounce Rate

Relevant recommendations on landing pages and product pages reduce bounces. Users who immediately see relevant alternatives leave the site less often.

Repeat Purchase Rate

Personalised post-purchase recommendations and email campaigns promote repeat purchases. Replenishment reminders for consumable products and recommendations based on previous purchases can strengthen loyalty.

Customer Lifetime Value

As a long-term metric, CLV captures the total value of a customer relationship. Recommendation systems contribute through higher purchase frequency and larger baskets.

Return Rate

Better product fit through precise recommendations can reduce returns. Particularly size recommendations and style advice can help avoid wrong purchases. Typical reductions range from 3 to 10 percent.

Why Improvements Are Incremental Rather Than Exponential

All mentioned effects are incremental in nature. A recommendation system improves the baseline; it does not transform it. Shops with already good UX and navigation see smaller improvements than those with obvious optimisation potential.

The often-cited high revenue shares from recommendations at large platforms are the result of years of optimisation, massive data volumes, and specialised teams. They are not replicable through simple tool implementation.

Industry Relevance

Why Fashion, FMCG, and Beauty Particularly Benefit

Fashion is a natural use case: high SKU count, complex product attributes, visual dominance, frequent repeat purchases, and strong discovery behaviour.

FMCG benefits from replenishment logic: regular repeat purchases, clear preference patterns, high purchase frequency.

Beauty combines elements of both categories: strong loyalty to found products, complex attributes, potential for regime extension and cross-selling.

Where Recommendations Add Little Value

Very small assortments: With fewer than 100 products, manual curation is often more effective than algorithms.

Single-purchase products: For products with very low repeat purchase probability, recommendations for follow-up purchases are barely relevant.

Highly regulated products: Pharmaceuticals, financial products, or other heavily regulated categories have limited possibilities for algorithmic recommendations.

Maturity Levels of Recommendation Systems

Level 1: Rule-Based Recommendations

Recommendations based on manually defined rules. Typical results: 3 to 8 percent conversion increase. High manual maintenance effort for large assortments.

Level 2: Machine Learning-Based Recommendations

Algorithms learn from user data and identify patterns automatically. Typical results: 8 to 15 percent conversion increase.

Level 3: Context-Aware Real-Time Recommendations

Recommendations consider real-time signals. Typical results: 15 to 25 percent conversion increase.

Level 4: Cross-Channel Orchestration

Recommendations coordinated across all touchpoints. Measurable CLV increases and reduction of cannibalisation between channels.

Tools and Solutions in Practice

SMB and Mid-Market

Nosto, Klevu, Algolia Recommend, and Clerk.io are suitable for companies with 1,000 to 50,000 visitors per day.

Mid-Market and Enterprise

Bloomreach, Dynamic Yield, Salesforce Commerce Einstein, and Constructor are suitable for companies with more than 50,000 visitors per day.

Implementation Reality

Personalisation Is Not Plug and Play

Tool activation is not the same as successful implementation. The biggest investments lie in data preparation, integration, testing, and continuous optimisation.

Typical Implementation Timelines

SMB: 6 to 10 weeks

Mid-Market: 3 to 5 months

Common Misconceptions and Failure Modes

Frequent errors include: over-personalisation, blind trust in algorithms, ignoring consent and data gaps, treating recommendations as a marketing widget, and missing ownership between marketing, tech, and merchandising.

Readiness Framework

Decision criteria: data availability and quality, stable traffic volume, clear ownership, willingness to test and iterate.

Conclusion

Personalised product recommendations are not a quick win. They are a strategic capability that delivers measurable improvements when implemented correctly but requires continuous investment in data, technology, processes, and organisation.

Before such systems are scaled, organisational maturity should be assessed. This illustrates why AI is not just a tool question, but a decision-making question. For a well-founded assessment of strategic AI readiness, ThePromptMonk offers strategic AI consulting.

For operational support with implementation, find more information on our E-Commerce Services page and the AI in E-Commerce overview. For a complementary look at AI-driven pricing, see our article on dynamic price optimisation.

Back to blog overview