Card Sorting at Scale
Four card sorts across four years and four very different products, from an internal dashboard's chart settings to YouTube's own content naming.
The same method, aimed at four different naming problems
Card sorting is one of the most reliable ways to learn how people actually group and label things in their own heads, instead of the way an org chart or engineering model happens to organize them. Over several years, that method got reused to answer very different information-architecture questions: an internal dashboard’s dense settings menu, a public help center’s topic list, naming for a new YouTube content tier, and a public design system’s own site navigation.
Open sorts to find the categories, closed sorts to test them
Each study paired an open card sort, where participants create their own category names, with a closed card sort, where participants sort into names already proposed, then relied on similarity-matrix and dendrogram analysis to see which items reliably clustered together versus which ones split opinion. Sample sizes ranged widely by context: 97 Plx Dashboards users for an internal chart-editing menu, 37 external AdWords advertisers for the public Help Center, 240 participants for YouTube’s content naming, and 37 designers and developers for Google’s Material.io design system site.
Naming split opinion far more often than grouping did
- 01Plx Dashboards chart editor (N=97).
Two clean groupings emerged around table aesthetics and table formatting, but nearly a quarter of items landed in a vague "miscellaneous" bucket. Grouping worked; several of the items themselves needed clearer labels before any category name could help.
- 02AdWords Help Center (N=37 external advertisers).
Eight clear topic clusters emerged, targeting, bidding, account management, campaign creation, keywords, views, reporting, and ad creation, giving the editorial team a validated structure instead of a guess about how advertisers actually think about the help content.
- 03YouTube Premiere naming (N=240).
Movies, TV Series, and YouTube Originals reliably grouped together (90% agreement on the strongest pair), but no single proposed name, Featured, Explore, or Discover, performed well across all three at once.
- 04Material.io design system (N=37).
Even among people already using Google's own design system, Bootstrap was mentioned as a reference point more than twice as often as Material itself. The category people reached for first wasn't the one being tested.
Across all four studies, the strongest predictor of a good category name was whether it matched language people already used out loud, more than how clever it sounded. Generic or borrowed-sounding names split opinion every time: "Featured" worked for YouTube Originals but split badly on Movies, and "Miscellaneous" absorbed anything the Plx study didn't have a clear word for.
A repeatable method, reused whenever a naming decision needed evidence
Each study fed a real decision for its own team: which chart-editing settings to consolidate, how to restructure a help center, which name to ship for a new YouTube content tier, and how to talk about a design system to people who don’t work at Google. Together they became a standing method other teams reached for whenever an information-architecture decision needed to be settled by evidence instead of an opinion in a room.