Personalizing Beyond the Point of No Return

Eric Colson
- San Francisco, CA

There is an old story of a commander who, upon landing on the beach of his adversary, ordered “burn the boats” so that his warriors would have no other choice but to triumph. The imposed constraint provided clarity. The lack of a fallback mechanism and the high cost of failure focused all efforts towards victory.

Image of several data scientists

The Stitch Fix business model [1] imposes similar constraints on us. The service was designed to remove many of the hassles of shopping: the trip to the mall, the online searching, the sifting, the reading of reviews, etc. We had to think very differently about personalization. We would have to go beyond what modern recommender systems provide. To be sure, recommender systems play a huge role in helping customers find the things that are most relevant to them:

  • Amazon says that 35% of its sales are driven from its recommendation engine. [2]
  • Linkedin boasts that as much as 50% of connections are made from its recommendation algorithms. [3]
  • Netflix states that 75% of what people watch on its service are discovered through its recommendations. [4]

Yet, each of the above services also has fallbacks. In the event that the recommendations are not relevant, the customer can navigate the service to find things on her own. At Stitch Fix, our customers are specifically seeking expert advice and suggestions. Fallbacks would erode much of our value proposition and therefore we don’t have them:

  • At Stitch Fix, 100% of merchandise is sold through recommendations.

That is, 0% of what our customers purchase is merchandise they found themselves. There is no shopping on; our customers expect us to do the finding. We leave ourselves no fallback - no search functionality, browsing options, or any other mechanism that shifts the burdensome tasks of shopping back to the customer. We are solely responsible for getting relevant things in front of each customer. In addition, we face substantial penalties if we fail: the cost of shipping (both ways), the carrying cost of inventory, and a very upset customer.

The lack of a fallback mechanism and the high cost of failure act as constraints, focusing our efforts. This results in profound consequences:

  • Crystal Clear Goals. If we were to allow a fallback—say, an online module that enables the customer to browse or search on their own in the event that she didn’t like what we picked out for her—we would find it all too convenient to rely on it. We would likely make subsequent investments to improve the fallback mechanism, adding new features, and optimizing others. Such investments would come at the expense of our ultimate goal: delivering an effortless, more personalized experience for the customer. Absent the distractions of fallbacks, we can focus our energy towards that goal.
  • A Culture of Innovation. With such high penalties for getting it wrong, we are motivated to try new ways of doing things. For example, the cost of synthesizing machine algorithms with expert-human stylists could not be justified in other business models. In our model, however, their additive contributions allow us to leverage more data and more processing in order to reduce the probability of failure. As another example, consider size and fit of apparel merchandise. There are no standards in the industry, resulting in disparity from brand to brand. The onus falls on the consumer to figure out what is right for her. In our model, we have every incentive to figure it out for her. Each garment is carefully studied for fit and size specifications - both through physical assessment and empirical findings. These processes provide big value to customers. Traditional retail models cannot justify their high costs and instead pass the burden on to the consumer. But in our model, the alternative is paying even more to fail and the motivation becomes apparent.
  • Long-term Symbiotic Relationships. The nuances of our business model provide natural, almost altruistic, incentives for both us and our customers. It only works if both parties are receiving value. There is no “selling” here—only relevancy. Given the high cost of failure it doesn’t make sense for us to send anything but extremely relevant merchandise that we believe the customer will love. Likewise, the customer has every incentive to help us help her. She does this by providing detailed preference data, which informs our human stylists and machine algorithms. She also provides thoughtful feedback on the selections we send her. This is different from the implicit feedback that powers most recommender systems (clicks, views, search strings, etc). By contrast, we get explicit and detailed feedback from customers that have actually experienced the products. This information is far richer than the fleeting reaction to an image in an app or on a website. This symbiosis also changes the way we manage and measure success. It becomes less about the transaction and more about long-term customer relationships.
  • An inspiring work environment. The constraints create an inspiring work environment where you are incentivized to be the best in the world at what you do. You don’t stop at ‘good enough’. You are supported by the entire company to continue to innovate. Rather than a niche, esoteric team, Algorithms is its own department reporting to the CEO (see “Advice for Data Scientist on Where to Work”) and is staffed by more than 40 (and growing) scientists from diverse quantitative fields. And, it's not just the Algorithms team - the support for innovation is pervasive throughout the company. Having no fallback mechanism means there is no ecommerce site to support. This enables our Engineering team to focus much of their talent toward building beautiful internal applications - an area so often under-supported by most companies. As another example, our Merchandising team is supported in finding new ways to buy and manage inventory that better leverages the personal preferences of our customers. This support comes not just from the company, but also the entire apparel industry that is so primed for a new way of doing things. The list goes on and on - every function in the company has the support it needs to make change. This makes it an inspiring place to work - you not only get to be creative but you get the resources you need to execute on your ideas.

Our business model was designed to be very different from the beginning. I doubt we could have made a transition from a traditional retail or ecommerce model to this one – we likely would not have had the courage to ‘burn our boats’. Nonetheless, the model inherently imposes constraints and a commitment that fosters innovation and creativity, leading to a better outcome for both our customers and ourselves. And, there is still so much more to figure out. For many of us, this is the most gratifying part. It is far more interesting to pioneer new ground than it is to optimize a known business model. Each day is another step into uncharted territory with no fallback plan and high cost of failure. Yet, there is no trepidation—only the exhilaration of the new.

[1] The Stitch Fix model provides shopping-as-a-service. The company delivers five highly curated pieces of clothing to its customers’ homes, either on-demand or as a subscription. The customer pays a $20 styling fee but if they decide to purchase any of the five items the fee is applied toward the purchase. There is no cost to the customer for shipping. If a customer decides to buy all five items in their shipment, they receive a 25% discount off the entire purchase.

[2] Marshall, Matt. “Aggregate Knowledge Raises $5M from Kleiner, on A roll.” VentureBeat. N.p., 10 Dec. 2006.

[3] Tunkelang, Daniel. “Content, Connections, and Context.” Content, Connections, and Context (2012): n. pag. Sept.-Oct. 2012. Web.

[4] Amatriain, Xavier and Basilico, Justin . “The Netflix Tech Blog: Netflix Recommendations: Beyond the 5 Stars (Part 1).” Netflix, 6 Apr. 2012.

Tweet this post! Post on LinkedIn

Come Work with Us!

We’re a diverse team dedicated to building great products, and we’d love your help. Do you want to build amazing products with amazing peers? Join us!