Overstock.com Inc. has overhauled several key marketing systems to improve its ability to quickly and intelligently tailor web pages and emails to each shopper’s interests. Early returns are promising, says J.P. Knab, chief marketing officer at the web-only retailer.
The goal is to move away from static web pages that every visitor sees to a “pageless” approach that creates web content on the fly based on shopper behavior, including what the shopper is doing right now.
“When you enter Overstock.com, no matter what page, if we know what you’re looking for, that page should personalize to you,” Knab says. “It should not be the same type of page every other consumer is seeing. We had to strip out our entire marketing stack to do that.”
The two big pieces Overstock, No. 32 in the 2018 Internet Retailer Top 500, overhauled were a system that stores customer data and another that houses product images and other assets.
Working with technology provider mParticle, Overstock created what Knab calls a “customer data platform” that takes in data about each customer from every touch point—including web and mobile sites, email, social media and online ads—and immediately feeds back information about each shopper to systems driving website personalization, email marketing and online advertising.
The second piece is an “asset data platform” that includes 25 million pieces of content—including product images, marketing materials, offer overlays and videos.
Layered on top of both systems is machine learning, Knab says. Applied to the customer data that allows Overstock.com to segment customers based on preferences and purchase history, as well as where they are in their purchasing journey. A shopper researching a high-priced item, for example, might click on several items while deciding which one she wants, then later on click several times on the same item as she narrows her search.
“If the shopper is early on the purchasing journey, she may be looking for an inspirational piece of a buying guide, or looking for reviews,” Knab says. “The system attempts to understand what part of the journey the consumer is on to serve the appropriate content.”
The machine-learning technology applied to the asset database monitors the 40 million monthly visits to Overstock.com to learn how different types of customers respond to various pieces of content as they search among the 6 million products Overstock offers. The large quantity of data it is continually receiving enables the system to more effectively rank the impact of each asset for each group of shoppers, Knab says. “The system is learning. It doesn’t just ask whether the asset is relevant, but whether it is the right asset to serve to that consumer.”
Now able to assess a shopper’s intent based on what the consumer is doing currently, as well as her past behaviors, and with a better idea of what type of content will appeal to her, Overstock.com has upgraded its personalization so far in two areas: the landing pages consumers see after clicking on Overstock ads on Facebook and Pinterest, and on marketing emails triggered by customer behavior. The customer and asset management systems feed data into market automation technology from Braze to create personalized marketing emails.
“The content the system is choosing is leading to much higher engagement,” Knab says. A key indicator, he says, is how many consumers click beyond the initial landing page from social media ads. “Engagement click-through rates climb on pages where we’re implementing these systems.”
In one project, Overstock used machine learning to optimize emails based on the time of day and registered an increase of 9% in email open rates overall, and a 31% increase among “non-engaged” consumers who had not opened the e-retailer’s emails in the previous 30 days. A separate test that optimized email by serving relevant content and product suggestions based on prior shopping behavior raised the overall email open rate by 10-15%.
The online retailer will continue deploying the new marketing technology to other channels, which Knab notes is a big job. He says there are different data format requirements for pushing content to a Facebook ad, for example, versus a product recommendation panel on a web page, and Overstock must map out the requirements for each channel before the new systems can be implemented.
Better testing required
The “pageless” approach to creating individual web pages for Overstock.com also means the e-retailer must upgrade its testing capabilities before applying the new system to product recommendations on Overstock.com, Knab says. There are millions of possible variations of each page based on the individual customer and the products he is considering, and Overstock wants to be sure that the customized pages the new systems produce generate positive results.
“We need testing systems that can look at whether this is resulting in a better customer experience,” he says. To that end, Knab says, Overstock.com has just signed a deal with a provider of testing technology, but he declined to name the company.
He adds that testing must not just evaluate short-term results but the longer-term impact on how much each customer will buy from Overstock. For example, he says, a bigger discount will almost certainly lead to more sales than a smaller discount, but bigger discounts cut into a retailer’s profit margin.
Knab says the machine-learning system will use Overstock’s historical customer data to evaluate promotions, looking at “not just what’s right for tomorrow, but also for a week, month or a year from now.” The aim, he says, is to strike a balance between what will engage the customer today with what is best in terms of lifetime customer value.
However, he emphasizes that Overstock only personalizes content for shoppers who have either accepted cookies or signed into their account. “We let our customers decide the level of personalization or anonymity they receive when shopping on our site,” he says.
Look for Internet Retailer’s upcoming research: The 2018 State of E-Commerce Personalization Report