Show 3. juni 2026
October 10, 2021
Transforming Fashion PPC with AI-Powered Returns Prediction
Refyne
for
BESTSELLER
Kategori :
Type :
andet
Show 3. juni 2026
October 10, 2021
Refyne
for
BESTSELLER
Kategori :
Type :
andet
In fashion e-commerce, returns are not a side issue. They are one of the industry’s most persistent structural challenges, eroding margins, distorting performance signals, and making growth look stronger than the underlying economics actually is. For leading fashion brands, this creates a fundamental tension. How do you keep scaling demand when a significant share of reported revenue never becomes retained revenue?
That was the strategic challenge BESTSELLER chose to confront.
In 2025, BESTSELLER, the Danish fashion group behind brands such as Jack & Jones, ONLY, Vero Moda, Name It, Vila, and Selected, faced a structural profitability issue across its European direct-to-consumer business. Return rates above 50% in some markets made returns one of the most margin-destructive mechanisms in the business. At the same time, Google Ads bidding continued to optimise toward gross conversion value, rewarding transactions based on checkout value rather than actual retained revenue.
The insight was simple, but category-defining in its implications. The fashion industry had accepted a bidding model built for gross sales in a category shaped by net outcomes. BESTSELLER saw the opportunity to challenge that logic before the market did.
An initial idea, excluding frequent returners, proved operationally weak because more than 60% of traffic consisted of new or anonymous users without purchase history. Instead, BESTSELLER reframed the problem at a more fundamental level. Rather than trying to change customer behaviour, they chose to change the commercial intelligence behind every bid.
The strategic pivot was to move from gross revenue optimisation to a Net Revenue model, predicting the retained value of each order at the moment of purchase and letting that value inform bidding in real time. That made BESTSELLER an early mover in applying AI not just to media efficiency, but to one of fashion’s most important profitability challenges. The goal was clear. Improve cost-efficiency, prioritise high-value customers, and protect volume, all while making growth more economically intelligent.
To turn that strategy into a real operating advantage, BESTSELLER partnered with REFYNE to build a real-time AI prediction system integrated directly into the Google Ads bidding engine. This was not a reporting enhancement or a dashboard layer. It was a new way of competing in the auction, designed specifically for an industry where checkout value alone is an incomplete signal of commercial value.
The solution consisted of three connected layers.
First came the data foundation. A reverse ETL pipeline moved prepared datasets from BESTSELLER’s Snowflake environment into BigQuery, creating a unified layer across brands, categories, and markets.
Second came the prediction engine. Using Google Cloud Vertex AI, machine learning models were trained on historical transactions and return data to estimate the expected retained value of each incoming order. Importantly, the model did not rely on personal identifiers or historical user profiles. Instead, it focused on transaction-level signals available in the moment, including basket composition, size duplication, payment method, and basket value. This made the model usable at scale, including for anonymous and first-time customers, and enabled a 93% prediction accuracy across orders.
Third came the activation layer. An overnight batch setup was replaced with real-time scoring via server-side Google Tag Manager, so every conversion could be assigned a predicted net revenue value and passed into Performance Max instantly. In effect, BESTSELLER gave Google’s bidding engine a smarter commercial signal than the market standard.
To validate the impact, BESTSELLER used a GeoLift framework across matched test and control regions in the Netherlands and Sweden. This gave the business a rigorous way to prove that the model did not merely look innovative. It created measurable auction and profitability advantages in practice.
What makes this case powerful is not that BESTSELLER changed the existence of returns in fashion. It is that BESTSELLER found a smarter way to compete in a category defined by them. In an industry where many brands still optimize media toward gross demand, BESTSELLER used AI to build a commercially intelligent model, one that improved the quality of growth rather than simply chasing more of it.
The results were above expectations and significant.
In the Netherlands test market, Return on Ad Spend increased by 50%. Across test markets, Cost Per Click decreased by 21% to 24.5%. Most importantly, this was achieved without suppressing transaction volume. Rather than pulling back indiscriminately, the model reallocated spend toward orders and customers with stronger retained-value potential, allowing BESTSELLER to maintain, and in some cases improve, order counts while materially strengthening unit economics. Additionally the achieved savings of some 1.3 mio euro were then reinvested back into the continuous loop to spark further growth.
That is what makes the approach market-leading. BESTSELLER did not accept the traditional trade-off between efficiency and scale. It showed that more intelligent bidding can protect both.
The effect also extended beyond media. The same AI-generated forecasts are now used in collaboration with product, buying, and finance teams for reconciliation, forecasting, and assortment planning. What started as a paid media innovation has developed into broader commercial infrastructure. During H1 2026, the model is being rolled out globally across brands and markets, a strong signal that this is not a one-off test, but a scalable capability with first-mover potential in a category that has long struggled to operationalise returns intelligently.