In short
Card sorting is a research method in which participants sort labelled cards representing pages or content items into groups that make sense to them, and often name the groups. Open sorts let participants create their own categories and reveal their mental model; closed sorts ask them to place cards into categories you provide and test a proposed structure. Fifteen to twenty participants give reliable patterns. It is the standard method for designing or fixing a site's information architecture and navigation, and it is fast and cheap to run online.
What it is
You write each piece of content, page, or feature on a card. You ask a participant to sort the cards into piles that belong together, and, in most versions, to label each pile. You do this with fifteen or twenty people. Then you look at which cards were grouped together most often and what names people gave the groups.
The result is a picture of how your users think your content is organised, which is the information a navigation menu and a site structure should be built on. Teams that skip it build the structure around their own departments and products, and users cannot find things because the categories reflect the company, not the customer.
The three variants
Open card sort. Participants create their own groups and name them. Best for discovering the mental model when you are designing a structure from scratch or suspect the current one is wrong. The group names are as valuable as the groupings, because they are the labels users would understand.
Closed card sort. You provide the categories; participants place cards into them. Best for testing a proposed structure: if most participants put a card where you expected, the structure works for that item; if they scatter it, the item does not belong where you put it or the category name is unclear.
Hybrid. Participants place cards into your categories but can create new ones if nothing fits. A pragmatic middle for refining an existing structure.
A related method, tree testing, does the reverse: it gives participants a proposed navigation tree and asks them to find where a given item lives. Card sorting builds the structure; tree testing checks it.
When to use it
Designing the navigation for a new site or product. Restructuring one whose analytics and support tickets show people cannot find things. Merging two sites. Deciding how to categorise a large catalogue, a documentation set, or a settings area. Settling a disagreement about structure with evidence rather than opinion.
It is less useful for sites with a handful of pages, where the structure is obvious, and for content whose organisation is dictated by an external standard.
Preparing the cards
Thirty to sixty cards is the workable range; fewer than thirty gives thin results, more than sixty exhausts participants. Choose items that represent the content, at a consistent level of granularity: pages or major content types, not individual articles. Write each card in plain, user-facing language, not internal jargon, and avoid labels that give away the grouping. Include a few items you expect to be ambiguous; where they land is often the most useful finding.
Recruiting and running it
Fifteen participants from the actual audience typically reveal the main patterns; twenty to thirty improve confidence for closed sorts. Recruit from customers, a research panel, or, for a public site, anyone in the target group. Expertise matters: sorting a medical site with non-patients produces the wrong model.
Run it online with a tool such as Optimal Workshop, Maze or UXtweak, which present the cards, record the sorts and produce the analysis; each session takes participants ten to twenty minutes. In person, physical cards on a table work well for small groups and allow you to hear people think aloud, which adds reasons to the groupings.
Give one instruction: sort these into groups that make sense to you, and name each group. Do not explain your intended structure. Allow participants to leave cards out if they cannot place them; that is data too.
Reading the results
Similarity matrix. For each pair of cards, the percentage of participants who grouped them together. High-similarity clusters are your candidate categories; cards with low similarity to everything are the ones that do not fit anywhere and may need a home of their own or a different label.
Dendrogram. A tree showing how cards cluster at different levels of agreement. Cut it at a level that gives four to eight top-level groups and see whether they make sense.
Category names. From an open sort, the labels participants used. Where many people chose similar words, that is your navigation label. Where the names diverge, the category is unclear and may need splitting.
Standardised categories. Tools let you merge synonymous group names to see the consensus structure.
The output is a proposed information architecture with user-derived labels, ready for a tree test to confirm it.
What it does not tell you
Where in the structure the most important items should be promoted for business reasons. How the navigation should look or behave. Whether a label works in context on a page. What users are trying to accomplish, as opposed to how they group things. And anything about content you did not put on a card. It answers one question, how people group and name these things, and it answers it well.
The practical sequence
Open sort with fifteen people to discover the model; draft a structure from the clusters and labels; closed sort or tree test to confirm; adjust; build. The whole cycle fits inside two weeks and costs a few hundred dollars in tools and incentives, which is cheap insurance against a redesign that ships a navigation users cannot use. If you would like one run for a site whose structure has grown out of control, book a call.
Common questions
What is the difference between open and closed card sorting?
In an open sort, participants create and name their own groups, revealing how they mentally organise the content; use it to design a structure from scratch. In a closed sort, you provide the categories and participants place cards into them; use it to test whether a proposed structure and its labels work. A hybrid lets participants add categories when yours do not fit.
How many participants do you need for a card sort?
About fifteen for an open sort reveals the main patterns reliably; twenty to thirty improves confidence, particularly for closed sorts where you are measuring agreement with a proposed structure. Beyond thirty the patterns rarely change. Participants should come from the real audience, since expertise shapes how people group things.
How many cards should a card sort have?
Between thirty and sixty. Fewer gives thin results; more tires participants and lowers the quality of the later sorts. Cards should represent pages or content types at a consistent level of detail, written in plain user-facing language without labels that hint at the intended grouping.
What is the difference between card sorting and tree testing?
Card sorting asks participants to build groups from content items, which generates a structure. Tree testing gives participants a proposed navigation tree and asks them to find specific items in it, which validates a structure. Run a card sort to design the architecture, then a tree test to confirm people can navigate it.
What tools can you use for online card sorting?
Optimal Workshop (OptimalSort), Maze, UXtweak, UserZoom and Lyssna all offer remote card sorting with participant recruitment and automatic analysis including similarity matrices and dendrograms. Sessions take participants ten to twenty minutes and can be run unmoderated, which makes fifteen to twenty participants achievable within a few days.
