Developing a product taxonomy allowed us to support a more robust search experience, a proper top-level navigation and a more granular product creation process for our sellers.
Working with our Senior Product Manager, I led the UX research efforts for restucturing our product taxonomy and information architecture. The goal of the initial iteration of the taxonomy project was to increase the accuracy of search and improve discovery. Since this required products to be properly categorized and tagged, the project was dependent on seller adoption. Following this project, I was tasked to design other areas affected by taxonomy updates, including the seller-side product creation and buyer-side search UI for web and mobile.
Research Methods Used: Tree Sort, Card Sort, Competitive Analysis, Direct Interviews with Sellers, Stakeholder Interviews (Account Managers and Customer Happiness)
Through early user research, we understood that users had a hard time navigating the site (we didn't have site navigation!) and thus, relied on using the search tool. The problem with that was since we lacked a proper taxonomy, our search results weren't accurate. Establishing a solid taxonomy would help solve both of these issues.
Tree Test & Initial Research
Based on the initial Tree Test, the participants failed to locate categories outside of the traditional subject groupings. It was easy for participants to find a math product by going to math > addition, but less clear when locating products related to categories such as public speaking.
Our core marketplace users are homeschooling mothers who don't always structure their curriculum based on traditional subjects and progressions, and who might not search for products within the bounds of those subjects. Although we assumed that organizing our taxonomy according to traditional subjects would likely be a deterrent for them, what we found was almost the opposite - most of the card sort participants relied heavily on traditional groupings when structuring the categories.
OPEN CARD SORT
Using our existing product categories as a jumping off point for an open card sort, we had 134 users (70% homeschooling parents) sort ~113 cards and name the category groupings. From the open card sort, we were able to standardize 11 categories.
The standardization grid allowed us to see the raw number of times that a card had been placed into one of the 11 standardized categories. This helped gauge the confidence of any one card grouped within a standardized category.
The dendrogram shows the relationship of the groupings (with most frequent group labels) and at which point they relate or merge with another. In the example above, we can see where technology starts to connect with the more general science arm of the graph.
Using the similarity matrix, we looked at the number of times that a user agreed that two cards had been put into the same group. The matrix shows the strongest groupings along the diagonal edge.
The Participant-Centric Analysis (PCA) shows the top most preferred groups based on similar responses, structured in three vastly different patterns for comparrision.
Closed Card Sort
Following the open card sort and analysis, we ran a closed card sort with ~102 cards and the 11 standardized categories. The Results Matrix shows how many times a card has been grouped into one of the categories, where the popular placements matrix breaks down the percentage of participants that sorted each card into the corresponding category.
Our closing question asked the participants what (if any) categories were missing and the bulk of the responses had noted foreign language and technology/computer science. We had removed foreign language from the card sort because of the confidence level in that category and to reduce the number of cards for the study, but added technology as its' own category based on the feedback.
Internal Card Sort
With the open and closed card sorts, our users were able to lay the new foundation for our product taxonomy and help us carve out a structure for the information architecture of our web and mobile navigation.
After analyzing the results from our OptimalSort studies, we laid out the base structure and ran a 'mini card sort' to make some final tweaks with a few team members. With the bulk of the structure set, we focused on cards that had been split between two or more groups. These included cards that may have been placed in the dominant group by ~70% of participants but placed in a different group by ~30% or similar, and excluded cards that had a more definitive split such as 90/10.
Our users need to find what they need in ways that makes sense to them. Through user research, we were able to rename the bulk of our existing categories using language that is more relevant to our users and TreeTesting/Card Sorting allowed us to spot inefficiencies in our existing taxonomy and model the updated structure based on actual user behavior.
With the results from multiple card sorts, our Product Manager worked on mapping the initial to the updated taxonomy. As we worked on this project in parallel with the development of the web and mobile site navigation, this was used to immediately populate the JSON file for the new nav. The results not only helped establish site navigation, but improved search and discovery with updated search algos.
We cleared out defunct categories and updated category names were reflected on existing products. Additionally, categories can automatically be assigned to products based on their titles and descriptions.