How do you make an e-commerce platform more accessible to shoppers who are searching for specific products when all products in the database are user-generated content?
Keep was built initially as a discovery shopping site. Users browsed for products by navigating an infinite feed of images, dynamically curated by a popularity algorithm (read more on this here). While many users liked this serendipitous experience, it omitted a large portion of potential users: shoppers who were looking for something specific, not browsing as entertainment. We had search functionality, of course, but it wasn’t enough.
Studying these “specific shoppers,” we found many users prefer to use category and filter functionality as their main method of shopping.
For example: user Samantha thinks, “Fall is almost here. I need a new pair of tall black boots.” She navigates to Nordstrom’s website, hovers over the Women tab and finds Shoes > Boots. Once on the Boots page, she then uses the filters to select “Shaft Height: Knee High and Over-the-knee” and “Color: Black.” Instantly, she receives all the options available in her preferred boot style.
We learned that when users are on the mission for a particular item, they don’t want to sift through unrelated items. They want to see everything available at once in order to make an informed purchase decision efficiently. This was not possible on Keep.
Traditional e-commerce platforms upload the products they sell into their system using standardized feeds or sometimes manually. Regardless the method, this results in uniform data. The system knows everything about the product and can store the information like we would a spreadsheet: item title, brand, price, size(s), color(s), description, SKU number, wash and wear instructions, etc.
Keep is unique. It's not a traditional e-commerce platform and neither are its base of products. Instead of products being imported using a standardized data structure, we allowed both our editors and our users to bring products into the platform. The Keep team developed a browser extension and bookmarklet that is used by staff and open to the public. Users can shop within the Keep ecosystem, but the power users who are serious about shopping can take it to the next level. When on another retailer's website, users can deploy the Keep It bookmarklet to save products into their personal collections on the Keep website and app. This bookmarklet scrapes the retailer's site and dynamically figures out where important information (such as title, price, brand and image) lives on the page. This information is sent back to the Keep platform, instantaneously generating the new product in the system. This scraper is also used to update pricing and availability information regularly, so users are able to see (and get alerts) if any of their "Keeps" go on sale.
What's the problem?
Keep's scrapers may be able to determine the correct product information most of the time, but it's not a guarantee. Retailer websites vary immensely. Secondly, the scraping process is not able to determine product category. For example, we may know the product title is “Tory Burch Marsden Over the Knee Boot” or “Nike Presto Fly Sneaker (Women).” However, we aren’t able to determine from scraping a product page that both of these products should be categorized under shoes. This is because different retailers categorize their products differently, e.g. footwear, boots. This is not to mention products that are more ambiguous. Is the “Ava Mirror” a mirror for the dining room wall or a portable cosmetic mirror?
Due to the volume of products entering the system each day, it was impossible to manually tag each product with a category.
Ideal solution: use visual search, AI and machine learning technology to read product titles and images and automatically determine a product category. Unsurprisingly, this technology is expensive, both to license or to build. The sheer volume of products in the system also makes this method a bit extravagant. A large volume of products that were brought into the system by users are never seen by anyone else; they aren’t interesting or nice and don’t become popular enough (as determined by the algorithm) to surface in highly-trafficked areas.
Until these technologies become more readily available and less expensive, it was more affordable to utilize our editorial staff. In order to accommodate the “specific shoppers,” we realized we didn’t need to categorize the millions of products in the system. We had learned that our special sauce is our curation (more about this here). Although it wouldn’t be as beneficial to our SEO as tagging products, we could still provide a robust range of shopping categories to mimic the flow users take on traditional e-commerce platforms.
Our editors were already manually curating dozens of products each day for our daily marketing emails. These hand-selected products were adored by users. 70% of traffic sent to retailers originated from our editorial team’s products. This blew the platform’s popularity algorithm out of the water. We attributed this success to our editors’ skills for styling; they know how to select the most original products with the most visually-interesting imagery for our female, millennial-skewing audience. That’s not something technology can do well quite yet.
We capitalized on the success of our curation and expanded the editorial team to build out an abundant set of product categories and subcategories. Using research about our users habits, I worked with our Marketing and Content Manager to develop the list of subcategories:
I designed a front-end page for the site and app that could be used for all the categories and subcategories, repurposing technology we had already built as much as possible. I also created an admin interface the team could use to manage and edit categories.