At Zulily, a small approach to big data

October 19, 2014, 7:00 PM UTC
Paul Bradbury—Getty Images/OJO Images RF

Racks of clothes, piles of toys, and stacks of boxes clutter the offices and rooms inside the Seattle headquarters of online retailer Zulily. Each day, 9,000 new products are added to the company’s website to be sold in flash sales. Each day, Zulily staff takes thousands of photographs of those products—everything from canvas shoes and leather boots to jewelry and plastic doll houses. Upstairs, engineering director Mike Errecart has a big job: predict which pictures will most likely make each of Zulily’s 4.1 million customers click the ‘buy’ button, then write software to serve up personalized pages.

The approach to such a big data project? The same as it has always been at the five-year-old startup: Think small. Really small. Only a couple of engineers will tackle this project on their own and they’ll manage to do it in just a few months. They’ll take enormous amounts of data about customers and break it down into small bites, perform small experiments using small customer samples, and check in with senior executives for incredibly short meetings each week.

That culture of small and fast offers a lesson to other companies trying to manage giant pools of data. Amy Webb, founder of the consulting and strategy firm Webbmedia Group, says she consistently hears complaints from big companies that everyone is siloed and they don’t talk to each other. “You cannot work that way in the big data space,” Webb says.

Just as Zappos and Amazon (AMZN) aggressively applied predictive modeling to ensure customer purchases were not a one-time deal, Zulily is following the same growth pattern, if not more accelerated. Sales of the five-year-old startup hit $695 million last year; that number is expected to hit $1.2 billion this year. Its stock price (ZU) soared to $72.75 a share last spring, although it has since hovered around $37 a share.

Zulily’s business is in retail, but its DNA is in technology, using data and machine learning to make decisions. In any moment, the company serves up millions of variations of its home pages, marketing emails, or mobile apps with personalized products. Each screen is tailored to an individual customer based on data like what items and brands each customer likes, where they live, what time of day they shop and what they tend to click on or buy.

In 2013, Zulily set out to take that customer data one step further, building software that could predict which photos resonated with which customer and how. Did the photo of workout pants shown on the model work best for the mother of three in Texas? Or would she click on an image of clothes on a hanger—or maybe an action shot? Does the photo make her more apt to hit buy in the morning or evening? Zulily wanted software that could automatically predict that behavior of millions of customers who visit the site.

The photo project is one of 60 to 80 different big data projects going on at once at the company, and despite more than 1,700 employees, teams rarely surpass four to six people. Standup meetings may last less than 10 minutes, operating on what’s known as “Zulily time.”

“If we get more than 10 people in the room, we’re asking ourselves why,” says Luke Friang, the company’s chief information officer. “It really slows us down with too many ideas and too much input.”

Friang and CEO Darrell Cavens take a relatively hands off approach to those kinds of projects, turning over control to small teams with periodic checkins.

Each morning, Errecart arrives to plan new tests for the photo project. He holds 10-minute standup meetings with two team managers and three software engineers. Then they’re off to write code, sharing work online using online collaboration software. To avoid getting caught up in the minutia of data, Errecart’s team of data scientists will pop in to sit in on a customer focus group or visit with merchandisers and marketers to better understand what customers think. “The engineers we hire spend as much time with customers as they do writing software,” Friang says.

To dissect how photos impact buying decisions, Errecart’s team looked at 15 different shots of a product and tested them on the homepage, watching the activity of a few thousand customers. Did one photo lead to more people putting it in their cart?

They looked at the wealth of data about a product, including color, fit, style, size, brand, and texture. One test told them women who shop in the morning tend to be more tactile when buying and they snap up things more quickly.

Each day and hour, the team tweaks the predictive software based on that kind of information, giving frequent ad hoc updates to the leadership team at Zulily. These miniature experiments let engineers make mistakes, learn from them, and move on. The process is a necessity, due to the massive amount of data generated by the site, Friang says. “It can be completely overwhelming,” he says. “It’s about breaking it down, testing, and trial and error.”

That process may be more challenging as time goes on. Zulily faces stiff competition from the likes of Amazon, Target (TGT), and Walmart (WMT) and flash-sale sites like Fab and Gilt Group. Culling big data is a big advantage. As Zulily grows, so does its data. The company’s greatest challenge ahead? Staying small, of course.