For even the most optimized e-commerce websites, one challenge remains: understanding the mind of the customer to predict their behavior. While in brick and mortar shops a well-trained sales person might pick up on the subtle behavior of a customer in doubt, this is much harder in an online environment. But what if online visitor behavior still reveals their intent to leave? To predict visitors exiting in real-time, OnMarc has leveraged Amazon Web Services (AWS) to build a new machine learning product: MarcSense Retain.
A customer in doubt might elicit subtle clues that indicate a state of doubt, sometimes without being conscious of it. Thanks to this behavior, one can take subtle actions that will stop a visitor from abandoning their shopping cart. With a combination of real-time prediction and well-timed personalization, MarcSense Retain can turn high volumes of data into a more relevant customer experience. With great success: for Lampdirect.nl the tool could reduce the drop-out in their checkout by 48%.
So how does it work? Retain is a machine learning model, using boosted strategies, to learn patterns in clicking and mouse movements that occur prior to a visitor leaving the website. It combines general movements that give away the intent of leaving the page with gradual changes in a particular user’s behavior to predict whether a user will leave the sales process. If this happens, the tool can produce personalized content or trigger the website’s Content Management System to do so. What separates this approach from more ‘standard’ content recommendation is that timing is of the essence: only when a customer is in doubt, you can present them more information or give out a discount to stop them from leaving. This advantage leads to a more relevant customer journey, and the results speak for themselves: after deploying MarcSense Retain, OnMarc’s customer websites showed up to 57% increase in conversion rates.
The next step in exit prevention
This might lead you to think: is machine learning actually needed here? Tools have been available on the market for years already, showing a pop-up the moment a user is at risk of leaving the page. Which additional value does machine learning bring to the table?
Kevin van Kalkeren, current Manager Product Management & Data Science at OnMarc with experience in data consultancy, says: “I always found myself wondering how effective and accurate those tools were. Of course, a visitor moving to the ‘X’ in the top corner of the browser means they are leaving, but what if you are too late? Once the conscious choice to leave the funnel has been made, you’d need more persuasive messages to get the potential customer back. This is ineffective at best, and very costly at worst, if you would start to hand out high discounts to all visitors leaving at that point.” OnMarc has been working with the collection software Celebrus for years, enabling real-time collection of clicking, scrolling and cursor movement. “Our customers already got real value from their data, but we had never looked into the potential of all that movement data combined at scale. So, with the challenge of the exiting visitor in our minds, we set out to unlock the potential of user’s movement data as well.”
Among multiple OnMarc customers that chose to deploy MarcSense Retain on their websites, Lampdirect.nl, a World Wide Lighting (WWL) brand, managed to achieve 48% reduction in drop-off, also thanks to the great people that are developing the solution. Jimmy van den Eerenbeemt is currently Manager Insights (Business & Marketing Intelligence) at WWL, but started his career in data at OnMarc. His first topic of research there was to predict exits through mouse movement.
“Mouse movements have a close relationship with emotion. From lab studies, we already knew different emotional states lead to different mouse paths.” says Van den Eerenbeemt. “And since purchasing behavior is also heavily influenced by emotion, it would only make sense if there would be some kind of link between how a visitor moves and what their mental state prior to purchasing is.” Yet, the gap between idea and product is never an easy one to bridge. The volume and complexity of data, the intelligence involved to come up with predictive features and training a performant model were the first hurdles to overcome. But running this as a SaaS-tool came with more challenges.
Machine learning at scale
“While originally this started as a project to see if a bespoke model could predict exits in time, we started thinking: what if exit behavior is universally detectable? That could make this from a consultancy project into a scalable product to help multiple companies at once.” Van Kalkeren explains their shift to product. “This change in thinking was helped by our migration to AWS. While collecting data itself has been easy with Celebrus, the management of environments we would need to deploy Retain at scale like that would be considerable on traditional data centers. And even before that, just the pioneering that is required to test machine learning products over and over is helped massively by services like Amazon SageMaker. Not only it provides best-in-class hardware, but it also allows us data scientists to experiment safely, easily and at scale.”
”Once the model was tested thoroughly and we felt confident we had a general model that could help reduce exit on different websites, we started developing the pipeline to do so in real-time.” says van Kalkeren. The solution leverages various microservices that transform real-time mouse- and clicking behavior to machine learning features. Deploying and managing applications in the fast and cost-effective way a SaaS-tool like MarcSense Retain requires would not be possible without CI/CD aids such as CloudFormation and CDK. Not just because the data is voluminous, but because it involves various steps that go beyond the standard data preparation steps such as normalization. To templatize the solution and to help streamline automation of the infrastructure, OnMarc decided to leverage infrastructure-as-code (IaaC) thanks to Amazon CloudFormation . This allows to onboard new customers easily and get started fast with training their custom exit detection models. “So even when we run into the limits of the general model, we can still rely on a scalable way to help our customers quickly. We just provision a model (re)training pipeline and can put it into production in no time” remarks Van Kalkeren.
Experimentation at the core
In addition to advanced customer insights, Lampdirect can now run more elaborated experiments. “At the start, Retain did not directly increase our conversions” says Van den Eerenbeemt. “But we found that the real difference is made through relevant personalization. We know our customers best, so combined with Retain’s real-time insight, we optimized the content to be more relevant.”
This also speaks to the culture within WWL. “We have a way of looking at our organization as Vikings and Nerds. The Vikings are fast-moving and creative, wanting to innovate and try out new things. The Nerds might take a calmer pace, but try their hardest to measure and support all the new concepts with data and technology. That creates a dynamic that is flexible and innovative.”
Conclusion
OnMarc has reached a new step in its business with a flexible architectural backbone to put data science into production. “We can now help smaller companies in leveraging their own data like enterprise companies would. And enterprise companies can use Retain to free up their analytical resources to focus on their own innovation and optimization.” Says Van Kalkeren.
———
Felipe Augusto Chies is a Sr. Business Development Manager for AI and ML at AWS. Felipe has spent a substantial part of his career working with Machine Learning. Prior experiences are: Co-founder and Head of Product of Axelera AI, Product Sales Manager at AAEON, Tech Lead and Industry Specialist at Intel.