Skip to content

Card Sort or Example Sort

User-research method — instantiates Prototype-Centered Category Modeling

Has people sort real examples into piles so the category's natural dimensions, sub-groups, and fuzzy edges surface from behaviour rather than from a definition.

Before anyone writes a definition, a Card Sort hands people a stack of real examples and asks them to group the ones that belong together. What makes it this mechanism and not its siblings is that the category structure is an output of observed behaviour, not an input: nobody supplies the criteria, the piles, or the labels up front — they emerge from how a spread of people actually cluster the cards. Where a rubric or benchmark imposes a structure to be applied, a card sort elicits the structure already in people's heads, revealing the dimensions they sort on, the distinct sub-groups they form, and the cards that refuse to sit still. It is the discovery step that tells you what shape the category really has.

Example

A help centre's sixty-odd articles have grown into a menu nobody can navigate. Rather than redesign the taxonomy in a room, the team prints one card per article and runs an open sort with a dozen participants — support agents and actual customers — each free to make their own piles and name them. Aggregated, the sorts show more than a tidy tree. Two different mental models of "billing" appear: one group sorts by product ("the mobile app," "the web dashboard"), another by task ("change my plan," "get a refund"). A handful of cards — "why was I charged twice?" — land in a different pile for almost every sorter. The readout isn't a taxonomy; it is a map of how people think: the axes they use, the sub-groups that are genuinely distinct, and the articles that live on a boundary and will need a routing rule wherever they're filed.

How it works

The distinguishing move is to treat grouping behaviour as the data. Participants sort; their groupings are aggregated — typically into a co-occurrence or similarity matrix across sorters — so the piles reflect a consensus rather than one person's scheme. Piles that recur across sorters read as candidate prototypes; the labels people give them expose the dimensions they are sorting on; and cards that scatter across different piles mark the radial edges of the category. An open sort (participants invent the groups) is used to discover unknown structure; a closed sort (participants file into your groups) is used to test a structure you already suspect.

Tuning parameters

  • Open vs. closed — open sorts discover unknown structure; closed sorts validate a proposed one. Open early, closed to confirm.
  • Panel size and diversity — more and more varied sorters average out idiosyncrasy and surface competing mental models; a small homogeneous panel just launders its own assumptions.
  • Card granularity — fine-grained cards reveal subtle sub-groups but tire sorters and add noise; coarse cards are faster but hide radial detail.
  • Aggregation threshold — how tight a co-occurrence counts as "same pile" (where you cut the dendrogram) decides how many prototypes you end up seeing.
  • Think-aloud vs. silent — narrated sorts capture the why behind a grouping; silent sorts are cleaner to aggregate but mute the reasoning.

When it helps, and when it misleads

Its strength is catching the moment your intended category fights the way people actually think — it surfaces the users' own family-resemblance structure, complete with the sub-groups and boundary cases a top-down definition would flatten.[1] It is the cheapest way to discover that a category has two centres rather than one.

It misleads when the panel is small or homogeneous, because a card sort faithfully reproduces whatever bias its sorters bring; and participants will rationalise a grouping after the fact, so the labels can be tidier than the thinking. The classic misuse is the closed sort run to rubber-stamp a taxonomy already chosen — the exercise then confirms a structure instead of testing it. The discipline is to run an open sort first, recruit deliberately across perspectives, and treat every emergent pile as a hypothesis for a downstream instrument to confirm, not as a finished category.

How it implements the components

A card sort fills only the discovery-side components an elicitation method can produce:

  • similarity_dimension_map — the axes participants sort on (by product, by task, by tone) are read off the pile labels and the co-occurrence structure.
  • multi_prototype_structure — recurring, distinct piles are the evidence that the category has more than one centre, and which ones they are.
  • radial_extension_map — cards that scatter across piles trace the category's fuzzy outer band, the members held only by loose resemblance.

It does not curate those piles into a vetted anchor set — that's Golden Case Benchmark — nor turn the discovered axes into a scored instrument, which is Similarity Dimension Rubric; degree-of-membership scoring belongs to Graded Membership Table.

Notes

A card sort discovers structure but says nothing about how to handle a given case; it sits upstream of scoring and governance. Its output is a hypothesis about the category's shape, best confirmed on a larger sample before it drives a taxonomy people have to live with.

References

[1] Family resemblance — Wittgenstein's observation that the members of a category may share no single defining feature, only a web of overlapping similarities. Emergent piles are that web made visible, which is why a sort can find structure a necessary-and-sufficient definition misses. ↩