Customer journeys vary dramatically - from impulse purchases of guitar picks to year-long research cycles for grand pianos worth tens of thousands of euros. Some customers, such as guitarists, frequently purchase consumables and accessories, thereby building extensive product histories. Others, such as pianists, might make a single major purchase every decade or so. This diversity in both product types and purchasing behaviors meant traditional recommendation approaches fell short.

"We had everything from tiny harmonicas to grand pianos, plus professional lighting equipment and components to build entire stages," explains Falco Wittchen, Analytics and Data Science Lead at Thomann.io. "Some customers buy guitar strings every three months, while others spend a year researching a €35,000 piano purchase. The decision to buy can happen in a minute for a cable, or take nine months for a grand piano."

The existing in-house recommendation system had become difficult to maintain, in light of the rapidly advancing technological developments in recent decades. Therefore the decision was made to move from a rule-based system to a modern ML approach to reduce manual intervention and to adapt more flexibly to evolving customer behavior patterns and unlock more options in terms of personalization.

"When someone buys a guitar, we don't want to show them another identical guitar," explains Thiede. "Google's different models let us show complementary products like amplifiers or cases, creating a more relevant shopping experience."

The implementation leveraged multiple Google Cloud products working in concert:

Vertex AI Recommendations powers the core personalization engine

BigQuery handles massive product catalog and customer data processing

Bigtable manages real-time user session data and complex customer email relationships

Cloud Composer orchestrates the entire ML pipeline

Google Analytics provides behavioral data to feed the recommendation algorithms

The new recommendation system was developed in-house by Thomann.io’s Data Division with the support of its trusted partner and leading system integrator adesso, which also operates an office in Berlin. They had previously helped Thomann modernize its entire data platform, creating a scalable, cloud-native architecture that now powers not only recommendation services but also Thomann’s broader digital transformation, including the ongoing migration of its shop infrastructure.

"adesso helped us establish the necessary conditions for data products like this recommendation engine," explains Thiede. "They paved the way at different stages, helping us create a platform that's truly state-of-the-art for data processing speed and quality."

The team implemented a careful rollout strategy, starting with newsletters (lower risk), then the mobile app, and finally the website. At each touchpoint, they conducted rigorous A/B testing against the legacy system over specific time periods, ensuring accurate attribution models and comparable measurement frameworks.

The system handles complex business logic specific to Thomann's operations, including managing N-to-N relationships between users and email addresses, and filtering out dangerous goods (such as batteries and liquids) that cannot be shipped to specific locations. These customizations demonstrate the flexibility of Google Cloud's infrastructure in accommodating specialized business requirements, while building on the solid foundation established through our partnership with adesso.

The AI-powered system respects customer privacy by carefully managing consent. When users decline Google Analytics tracking, the system seamlessly falls back to non-personalized logic while maintaining the shopping experience. This approach strikes a balance between the benefits of personalization and strict GDPR compliance, ensuring that data governance never compromises user trust.

Thomann also addressed algorithmic fairness challenges that emerged during implementation. Early testing revealed that the system was particularly drawn to expensive vintage guitars (€150,000-€350,000), which attracted clicks but generated no sales. These "curiosity clicks" skewed recommendations toward unrealistic products. The team implemented business rules to ensure that recommendations prioritize genuine purchase intent over mere browsing curiosity, striking a balance between algorithmic optimization and practical business outcomes.

The results speak volumes about relevance and impact. Recommendations increase the average order value by 24% – an extraordinary figure for any e-commerce platform. Beyond revenue metrics, the team discovered compelling qualitative success stories: customers recommended a single snare drum, purchased entire drum kits, while others, shown blue guitars, bought the same model in red, demonstrating genuine purchase influence rather than mere browsing.

Thomann is expanding beyond basic recommendations, exploring how generative AI through Vertex AI's LLM capabilities could enhance product discovery and customer education, potentially creating AI-powered music advisors that combine product recommendations with educational content.