Companies don’t often say it aloud, but not all customers are equal. The goal is to attract and keep the ones who will return regularly to shop, rather than those who might make one purchase and are never heard from again.
Figuring out which is which isn’t always easy. Many fashion retailers still base their predictions on a model developed decades ago by marketing academics to estimate how much a customer will spend with their brand – their lifetime value. The formula combines details about previous purchases with some basic demographic information to calculate the lifetime value of a shopper.
That model, created in a pre-digital world, is showing its age. Faced with rising customer acquisition costs, brands today are trying to figure out which customers to target before they’ve even thought about buying something. They’re enhancing their predictions of customer lifetime value by incorporating the massive amounts of data shoppers shed online, using advanced analytics such as AI to process the information into usable insights.
Germany’s Mytheresa, which built its profitable e-commerce business on a loyal clientele of high-spending customers, is among them.
“I will share a secret,” chief executive Michael Kliger told BoF. “The most important part is who you target as your new customers.” It’s “purely AI” now doing that work for the company, he added.
The technology has some advantages. “The beauty of AI is that the data that you’re using to create a lifetime value is much richer and more complex than the data that retailers and fashion companies have used in the past,” said Brandon Purcell, an analyst covering customer analytics and AI at Forrester, a research and consulting firm. “In retail and in fashion in particular, very few companies are using AI for that.”
Still, it’s not a silver bullet. AI models are only as good as the data they’re based on. But when they work well, they can let companies make better-informed decisions about where to direct their marketing and other resources to get the best return on investment.
Spotting Big Spenders
Mytheresa’s AI looks at a mix of what Kliger called obvious signals and not-so-obvious ones. Among the obvious: “If the first purchase is a high-value, ready-to-wear item, much higher correlation to future spend levels than if the first item is a €200 pair of sneakers,” he said.
But it also tracks how shoppers react to marketing messages, their browsing activity on Mytheresa’s site, and even data points such as which payment method they choose. The company wouldn’t elaborate on exactly how it uses these measures to predict loyalty, but Kliger added: “If you have many of those factors and constantly look at not only what they buy, but what they look at, what emails they open, and you really have a database, then we have a pretty high predictive power for first-time customers and whether they will end up becoming one of our best customers.”
The insights shape Mytheresa’s decisions about where to direct its marketing dollars and which relationships to nurture. Those shoppers flagged as high potential may get first access to new products, priority in the order in which Mytheresa ships purchases, or have their calls answered first when they phone customer service, Kliger explained. The data can tell Mytheresa which services a customer is likely to want, such as personal styling; allow it to better target emails, texts, or push notifications; and help it forecast who will eventually stop shopping, or ”churn” in business parlance.
The company introduced data analytics and algorithms — the sequence of instructions a computer follows to process data — as well as a new system for marketing across media channels in 2017. In a regulatory filing last year, it pointed to these factors as the reason for its falling customer acquisition costs, “a trend we believe is rare in the industry, despite growing our active customer base from 400,000 in fiscal 2019 to over 486,000 in fiscal 2020,” it said.
Mytheresa isn’t alone in looking to new varieties of predictive analytics to help it make better decisions. Nike’s spate of tech acquisitions in recent years began in 2016 with Zodiac, a firm specialised in forecasting customer behaviour and lifetime value. ASOS has looked to machine learning as a way to better understand shoppers, in part because the free shipping and returns it offers mean attracting the wrong customers can actually cost it money.
Emad Hasan, co-founder and chief executive of Retina, which recently introduced a product that uses AI to identify high-value customers, saw how prevalent negative lifetime values can be while working at Facebook and PayPal, where he was involved in analytics for merchants. There, he discovered many businesses spent more money to get customers in the door and keep them than they were ultimately worth.
“It was fascinating to me that, for most businesses, about 30 to 50 percent of their customer base was lifetime unprofitable,” he said.
Companies such as Google and Facebook have long offered brands a way to pinpoint the shoppers they want, but Hasan said from what he sees, more brands are starting to rely on the information they’re able to collect from customers themselves.
“What’s happened, at least in the digital world, is a lot more signals have become available early on in the customer journey,” he said. Brands are making sure it’s channeled into their own data warehouses. This information may also become increasingly valuable as privacy concerns limit the data companies are able to share about their users. Hasan’s company is effectively a bet that AI can be more effective at processing all these signals into useful insights than the old model.
The Limits of AI
It’s hard to say how much better AI is at predicting a customer’s lifetime value. The term “AI” may conjure notions of sentient software, but it’s essentially predictive math. For it to work well, it needs good data to base its predictions on, and more doesn’t necessarily mean better.
“If you don’t have data that’s representative of your entire customer base or that’s representative of reality — or you just have hygiene issues in your data — then the model’s going to inherit those issues,” Purcell said.
Measures of customer lifetime values are already just probabilities rather than guaranteed outcomes. When problems slip into the data used to make predictions, the model may simply not be very useful, or it can point a business in the wrong direction if it’s not careful.
“Doing diligence on your data, understanding your data, preparing it correctly — all that data hygiene stuff — it’s not sexy, but it is so, so important,” Purcell said.
Mytheresa, for one, feels the work is worth it.