Fleur Verwijs is performance marketing manager at Omoda, one of the Netherlands’ leading e-commerce retailers for footwear and clothing. Together with her team, she’s responsible for driving the project that revolutionised its approach to returns — and delivered more profit.
Omoda is an innovative, family-owned Dutch fashion retailer. Although we were established as a footwear retailer in 2000, our heritage stretches back as far as 1875, when our founder’s grandfather made shoes and offered a home delivery service… by horse.
By 2007, we’d begun selling online and overseas. Further diversification came when we began selling clothing, which now accounts for around 35% of our total revenue.
Like for many fashion retailers, the past few years have been challenging. The acceleration of online shopping, changes in payment methods, and the introduction of fast fashion have meant more and more returns. Some of the biggest Dutch online platforms see between 30% and 70% of their products sent back.
As clothing is the most returned product category, it’s something we knew we needed to address. Not only for financial reasons, but for environmental ones too.
That’s when we turned to our partners at Google and Dept agency to help us create an AI-powered solution to help us predict, and lower, returns.
Return to sender… no more
We had two major problems to tackle. The first one was strategic change. When we first launched clothing, the focus was more on growing our revenue. But in early 2023, our focus shifted to profit growth instead of revenue growth to help ensure our long-term success.
The second was the volume of returns. With 50% of our total revenue being sent back via returns, we needed to lower this.
At the time we were using Smart Bidding algorithms to power our advertising and reach customers based on total order value. We focused our marketing budget on customers who spent a lot, however it meant they potentially returned a lot too. For example, a person buying a pair of shoes in three different sizes to try on at home. These customers bring revenue, but also return costs.
With 50% of our revenue growth being sent back via returns, we needed to lower this.
That’s why we created a new model to help predict whether an order was going to be returned — and, if so, how much of the order might be returned.
“Everything is hosted within Google’s ecosystem,” explains Dept’s data scientist Maria Guerrero. “We have all the consented data from customers and orders in data warehouse BigQuery and we used that to train the models. To get the data in real-time, we also incorporated Google’s Firestore, Dataflow, and Cloud functions. Machine learning models were developed and deployed using Vertex AI.”
“Now, when an order comes in, we can mix historical data — what a customer ordered and returned before — and combine it with new real-time data, including variables like payment methods, to get a prediction.” Guerrero says.
Not every order has the same value to our business. Orders which are less likely to be returned are more valuable. This new process has allowed us to optimise our marketing budget, focusing our bidding on customers with lower return rates while still reaching those that are in the market for higher-margin products.
Delivering the goods with AI
We wanted our real-time solution to be as agile as possible. To make that work, we needed an infrastructure that could process the predictions in milliseconds.
Building all of this took around a year and then we needed to ensure our board’s buy-in. The board is made up of many innovative and creative people, including the founding family members. They were excited about what we could do, but also wanted to inject their own ideas. The CEO, for example, wanted to add a data layer to account for additional costs associated with an order, such as packaging and shipping costs, etc.
So we kept updating the models. When we were ready to launch, we analysed all the data with a fine-tooth comb and watched for any tweaks that were needed.
As a last, but very important step, we also wanted to activate this predictive data in Google Ads where we use AI-powered ads solutions such as Performance Max and Broad Match to get the most out of our investments.
Results that deliver
As the result came in, we couldn’t have been happier.
We were able to correctly predict whether or not any items would be returned for between 70-75% of orders by using this model. And with this knowledge we were able to decrease returns by 5% as the model focused on generating orders with high profit rather than simply high revenue.
Not only that, but we also increased profit margin by 14%.
This was very much a concept. Now we want to optimise the model. Other applications we are considering include modelling to reduce CO2 emissions and combining prediction models with the personal sizing recommendation we introduced last year. This should help reduce returns even further.
Whatever you do, if your challenge and strategy are clear — as ours were — and you have the right teams to work with, you really can solve any problem. We’ve gained an in-depth understanding of customers’ return behaviour, and can now use that to help prevent returns, maximise our profits, and deliver the best service possible.
