eBay Incorporates Machine Learning to Overhaul Email Marketing Platform - eMarketer

eBay is no stranger to sending batch-and-blast emails—for years, that was the company’s primary email strategy. But as customer expectations grew, the strategy—and the technology that supports it—had to mature. Alex Weinstein, director of marketing technology and CRM (customer relationship management) at eBay, spoke with eMarketer’s Maria Minsker about how the company built its own personalization platform, and explained how the platform now supports a number of marketing and advertising channels.

eMarketer: For a while, eBay’s approach to email marketing was largely batch-and-blast. Why did that eventually stop working for you?

Alex Weinstein: In the past, we had our marketing operations organization handle email. If our fashion marketing team wanted to let customers know about a shoe sale, they would ask our marketing operations team to create an ad hoc email that would go out to several million customers and advertise that shoe sale. That approach was OK, but the problem was that it was addressing our customers in large groups. There wasn’t as much one-to-one personalization as we wanted.

eMarketer: How did you change your approach to email?

Weinstein: We understood that we needed to reinvent our strategy to be more customer-focused and treat each customer as an individual, rather than a member of a group [of millions]. That’s why we decided to invest quite meaningfully into a one-to-one personalization platform. We created an approach where humans make the creative and the raw materials, but machines do the personalization.

eMarketer: Can you talk more about this personalization platform? How does it work?

Weinstein: Imagine a customer is browsing eBay for shoes. Marketers constantly create deals—there are some deals on shoes, some on electronics and some on accessories for cars. As those deals are created, they’re placed into virtual “buckets.”

For every piece of marketing content, such as our newsletter, a machine learning model looks through all these buckets and decides that based on a customer’s browsing history, a shoe deal would be most relevant. It’s the best of both worlds—content is filled by a machine learning model selecting from deals that humans have created.

eMarketer: Why did you decide to build this platform in-house instead of buying a prebuilt one?

Weinstein: We evaluated a bunch of third-party offerings, but there were two reasons for doing this in-house. The first reason was eBay’s sheer scale. We are one of the largest marketplaces on the planet, with a billion items for sale and 167 million active buyers. Third-party solutions struggled with the scale.

The second reason was our internal decision to prioritize this work and be one of the best in the world at it. That’s why we decided we have to develop in-house expertise that will enable us to deeply understand every element of the stack, develop machine learning models and create real-time data processing pipelines to differentiate ourselves.

eMarketer: What role does real-time data play in your personalization platform?

Weinstein: With the batch-and-blast approach, we struggled with data delays. When someone made a purchase, that data had to be processed and eventually placed in a data warehouse. Our marketing system would operate on top of that data warehouse, but by the time a purchase was processed, a campaign that would have benefited from that data may have already run.

Now, we have a real-time data pipeline that powers our downstream marketing campaigns. Whenever an action takes place on the site—a customer buying an item, browsing or just seeing an ad—the activity is tracked by our real-time engine, which updates the profile of the customer and sets off triggers that we have embedded in the system. The triggers apply to both customers and items: For example, the moment the price changes on an item a customer has viewed, we can automatically send that customer an email. No batches involved.

eMarketer: What other challenges has real-time data helped you solve?

Weinstein: Another problem we faced is we would send an email to our customers with a great deal, but by the time they opened that email a few hours later, the item advertised was already sold.

As a result, we decided to use the real-time data pipeline to provide customers with the most relevant offer for them at the moment they open the email, not the moment we send it. This was a huge technological challenge because it meant that the email would have to render dynamically the instant it’s opened, but we’ve made sure that every ad slot loads in 20 milliseconds or less.

eMarketer: How does the personalized email platform fit into your CRM system?

Weinstein: Our journey with CRM started with building this real-time data pipeline for email. Our next step was adding machine learning to that pipeline. And the third step was expanding this system to all outbound channels, including the customer’s experience on-site.

It’s not just about being personalized in email, but also in our display ads, which is why we connected our display stack to the same pipeline. As we continue to add channels, we’re developing a cross-channel CRM system.

eMarketer: You mentioned that machine learning is a component of your platform. What’s your outlook on where machine learning is heading?

Weinstein: Machine learning has become foundational to how we see personalization. My invitation to my colleagues in the industry is to not try to boil the ocean, but rather dip their toes in a little bit. For example, use a light machine learning model to improve your newsletter. Try it, and you’ll see results.