Stitch Fix, the online personalized styling service, has made personal styling more accessible to more than 3.5 million consumers. Because personal styling is just that – personal – the company’s premise is that customers can share their preferences about the products they like (and dislike). Stitch Fix stylists will then finetune recommendations over time by gaining a better understanding of each customer.
Stitch Fix employs thousands of stylists who create a “fix,” a curated set of five items for each customer. After customers receive a fix, they can decide to keep or return any number of the items.
Achieving Personalization With Data, AI, and Machine Learning
In addition to human stylists, the company’s styling processes use machine learning-based recommendation engines. “Data, AI, and machine learning have been at the heart of what we do for a very long time,” said Jeff Cooper, director of data science at Stitch Fix.
Cooper asserted that Stitch Fix has a competitive advantage as a styling service compared to other types of retailers, as clients recognize that the company would need to know them well to understand their preferences.
When customers get started with Stitch Fix, they receive a questionnaire about their preferences. Additionally, clients provide feedback after each fix, some of which are easily machine-readable. Both structured and unstructured feedback is processed by internal tools and continually continually refine their preferences. The data then feeds into automated recommendation systems and contributes to the training of AI models.
“We think of our algorithms and our human stylists as being in this close partnership, where our stylists bring great fashion expertise, an ability to make personal connections, and [understanding of] our clients,” Cooper noted. “Our algorithms bring the ability to process, summarize, and search through our vast amounts of data.”
The Stitch Fix algorithms have a dual purpose: They help identify overall trends (e.g., are jumpsuits increasing in popularity?) and offer personalized recommendations to individual customers.
Using AI-Generated Text To Create Product Pages
Stitch Fix has long aimed to craft descriptions for product pages. Product pages are important because they help both clients and stylists develop a detailed understanding of the clothes and outfits.
In addition, product pages are useful for ranking on search engines. Accurate product descriptions “can help clients find their way to our website,” Cooper said.
For a period, Stitch Fix potentially missed out on organic leads through online searches because it did not have detailed product pages. The increased need for the pages didn’t come about until the company launched Freestyle, a feature that lets clients shop for individual items within their personalized shopping feeds.
Understanding that product pages can be useful is one thing, yet the sheer volume of products within the company’s catalog is intimidating. “We have tens of thousands of products that we have offered over the years, with several thousand new ones coming in every month,” Cooper explained. The goal was to develop a scalable process for generating product descriptions that could be applied to both existing and new products. “We’re looking to speed up the time that it takes to draft, review, and approve the copy and improve the quality of the descriptions,” he added.
Stitch Fix didn’t want product pages that mechanically regurgitate information, but rather to offer guidance to clients on how to wear and pair outfits. The challenge was to use AI-generated text while still sticking to the brand’s voice.
Tailoring the AI Model
Development began with expert copywriters generating a collection of high-quality examples of product description pages. Stitch Fix then experimented with ways of feeding those examples into GPT-3, a language prediction model. “GPT-3 allows for finetuning where you can feed the inputs for the product and also the style of output that you prefer,” Cooper said.
The first proof of concept underwent review by copywriters and merchandise experts, who made necessary tweaks. There was significant back and forth involved in training the model with enough product description examples so it could generate persuasive product copy that checked off all the boxes. The process took a few months and several thousands of examples for GPT-3 to hit the sweet spot.
The newly implemented generative AI model now creates product descriptions for all products that come in. Stitch Fix can generate 10,000 product descriptions in 30 minutes, with each description requiring less than a minute of review time, according to the company.
A key technical advantage of using large language models (LLMs) like GPT-3 is the ability to finetune it to individual needs, Cooper said. His advice to other companies looking to work with LLMs: First, have a bank of historical data that you can pipe into GPT-3 or an equivalent as training examples. Second, ensure you have a pool of experts who can help craft examples and refine the process.
About the authorPoornima Apte is a trained engineer turned writer who specializes in the fields of robotics, AI, IoT, 5G, cybersecurity, and more. Winner of a reporting award from the South Asian Journalists’ Association, Poornima loves learning and writing about new technologies—and the people behind them. Her client list includes numerous B2B and B2C outlets, who commission features, profiles, white papers, case studies, infographics, video scripts, and industry reports. Poornima is also a card-carrying member of the Cloud Appreciation Society.