Card sorting: a practical guide

A practical guide to card sorting for UX research: open vs. closed sorts, card and participant counts, and how to analyze the results.

Card sorting asks participants to organize items, usually pages, features, or product content, into groups that make sense to them. It helps information architecture work by showing how participants expect content to be organized. You can compare their groupings with an existing navigation or category system. The method includes open, closed, and hybrid formats, each suited to a different stage of design.

Open card sorts vs. closed card sorts

There are two basic formats, and picking the right one depends on what you already know about your content.

An open card sort gives participants an empty canvas: they group the cards into as many or as few categories as they want, and they name each category themselves. Open sorts are exploratory. Use them before committing to labels or groupings, when you want to learn how people think about your content. The output provides raw material for building an information architecture; it does not evaluate one that already exists.

In a closed card sort, you provide a fixed set of categories and participants decide where each card belongs. Closed sorts are confirmatory. Use one to evaluate a proposed structure, perhaps from an earlier open sort, or to compare competing structures. Fixed categories also make results easier to compare across participants.

A hybrid sort sits between the two: participants sort into your predefined categories but can also create a new one if nothing fits. It is often the more realistic version of a closed sort in practice, since a truly fixed category list rarely covers every edge case a participant runs into.

Many teams run one of each on the same project: an open sort early to build the structure, then a closed or hybrid sort later to validate it before launch.

How many cards, and how many participants

Card sorting works best with a card count a participant can hold in their head over one sitting. As a practical starting point, keep a single sort under about 60 cards; well beyond that, fatigue sets in and groupings made toward the end tend to get sloppier than the ones made at the start. If your content library is bigger than that, split it by section and run several smaller sorts rather than one long one, or pilot a subset first to sanity-check your card wording before committing everyone to the full set.

Card wording matters as much as card count. Each card should describe one thing, in the participant's language rather than your internal naming, and without hints about which category it belongs in. A card that leaks its own answer, "Billing settings" sitting next to a category called "Settings," tells you less about how people actually think.

Closed sorts can work with a modest group per segment because you are testing a fixed structure and looking for convergence. Open sorts benefit from more participants per segment because recurring categories and groupings matter more than one-off opinions. How many participants you need for user research explains how this reasoning applies across methods.

Segment participants before analysis. If you are designing for two audiences, run and analyze the sort separately for each one. Pooling groups with different mental models can hide a structure that works well for one audience.

Analyzing the results

Do not stop at eyeballing a handful of individual results. Any one participant's groupings will look a little idiosyncratic on their own; the real pattern only shows up once you aggregate across the group.

Two views support most analyses. A similarity matrix shows how often participants placed each pair of cards in the same group. A high score indicates broad agreement that the items belong together and may belong near each other in the structure. A dendrogram applies hierarchical clustering to the pairwise data and shows how cards and groups nest inside larger clusters. Use those clusters as candidates rather than automatic categories.

For closed sorts, a category breakdown, how often each card landed in each category, does a similar job: a card that lands in one category the large majority of the time is a settled decision, while one that splits closer to evenly across two or three categories is telling you the structure itself is ambiguous, not that participants got confused.

Treat the output as evidence, not a verdict. A sort shows how people group and label items when that is their only task. It does not show whether they can find content several levels deep in a real navigation. A tree test answers that separate question. Run the card sort first to propose a structure, then a tree test to evaluate whether people can navigate it.

Common card sorting mistakes

• Writing cards in your own internal language. If a card only makes sense to someone who already knows your product's naming, you are testing vocabulary, not information architecture.
• Packing in too many cards. A long sort produces rushed, lower-quality groupings toward the end, which quietly drags down your whole dataset.
• Skipping the pilot. Card wording and instructions that seem obvious to you are not always obvious to someone outside the project; a quick pilot with one or two people catches this before it costs you a full round.
• Running only a closed sort. If you never ran an open sort first, a closed sort can end up validating a structure that never matched anyone's actual mental model to begin with.
• Pooling distinct audiences into one dataset. Averaging two different mental models produces a structure that fits neither group well.
• Treating the sort as the final word. A card sort tells you how people would organize content in the abstract; it does not confirm they can find that content inside your real navigation, which is what a follow-up tree test is for.

Running a card sort in Versive

Card Sort is one of Versive's choices-and-sorting question types, alongside multiple choice, ranking, and tree test. You add your cards and your categories in the question builder, and you can randomize the order of both so position does not bias the results. One setting determines which sort format you get: Custom categories ("Allow users add new categories") in the builder. Leave your category list empty and turn the setting on, and participants build their own categories from scratch, which is a true open sort. Provide a fixed category list with the setting off, and participants only place cards into the categories you gave them, which is a closed sort. Provide a fixed category list and turn the setting on, and participants sort into your categories but can add their own if nothing fits, which is a hybrid sort. A separate, unrelated Custom cards setting ("Allow users to add new cards") lets participants add cards of their own, useful if you suspect you have not captured everything that belongs in the set. See the full question types reference for every setting.

Participants sort using drag-and-drop on desktop or tap-to-place on mobile, with a progress indicator showing how many cards remain as they work, so a larger card set stays manageable during the task itself.

Results land directly in the study dashboard: a matrix view showing how often each card and category paired up, a similarity view for which cards were grouped together most often, a dendrogram for the hierarchical clustering, and a category breakdown for how each category was used, without exporting to a separate analysis tool.

A card sort addresses structure, while a tree test addresses findability. Plan the card sort first, then evaluate the candidate structure with a tree test in the same study or a follow-up. If you need participants for either study, how to recruit research participants covers four recruitment methods, including Versive's panel.

Add a Card Sort question to a study and pilot a small open sort before recruiting a full round to validate one structure.