How to run a card sorting study

Card sorting is a user research method where participants group labeled cards into categories that make sense to them and then name those categories. You run it to learn how real users mentally organize your content, so you can design navigation and information architecture that match their expectations instead of your internal org chart. The fastest way to get reliable results is to sort the right content with the right users, then look for grouping patterns that repeat across people.

This guide walks through what card sorting is, when to use open, closed, or hybrid sorts, how many cards and participants you need, moderated versus unmoderated formats, and how to analyze the results at a conceptual level. It also covers how card sorting feeds your information architecture and pairs with tree testing.

What card sorting is and when to use it

A card sort presents participants with a set of items, one per card, such as page titles, product features, article topics, or menu entries. Participants sort those cards into groups based on how they think the items belong together. Depending on the method, they either invent their own group names or place cards into groups you provide.

The output tells you two things. First, which items users consider related, which shapes how you cluster content. Second, what language users apply to those clusters, which shapes your labels and menu names. Both are hard to get from opinion or internal debate, because teams tend to organize content around how the business is structured rather than how customers think about the problem.

Card sorting sits in the discovery and structure-building part of a project. It is most useful when you are designing a new site or app section, restructuring navigation that people struggle with, consolidating overlapping content, or naming categories that keep causing confusion. It is one of several methods worth knowing, and our complete walkthrough to product research methods shows where it fits alongside interviews, surveys, and usability testing.

Card sorting is not the right tool when you need to test whether a specific design works, measure task success, or understand why users behave a certain way in a live product. For those questions you want usability testing or interviews. For a broader view of the discipline, see our complete guide to user research.

Open, closed, and hybrid card sorts

The first decision is which type of sort to run. The choice depends on whether you are building a structure from scratch or refining one that already exists.

Method	How it works	Best for	Trade-off
Open	Participants create and name their own categories	Designing a brand new structure, discovering user mental models and vocabulary	Messier data, more analysis time, harder to compare across people
Closed	Participants place cards into categories you define	Validating or refining an existing structure, testing proposed navigation	Will not surface categories you did not think of
Hybrid	Participants use your categories but can also create new ones	Refining a draft structure while leaving room for discovery	Some participants over-rely on your categories, some ignore them

An open card sort is generative. Because participants name their own groups, you learn both how they cluster items and what words they use for those clusters. This is the right starting point when you have no structure yet or when the current one is clearly broken. The cost is analysis effort, since everyone produces slightly different category names that you later have to reconcile.

A closed card sort is evaluative. You already have candidate categories, and you want to know whether users can place content into them confidently. Closed sorts produce cleaner, more comparable data and are faster to analyze, but they cannot reveal a better structure you have not imagined.

A hybrid sort is a middle path. Participants start with your categories and can add their own when nothing fits. Hybrid sorts are useful when you have a reasonable draft structure but suspect gaps. In practice, many teams run an open sort early to build a structure, then a closed or hybrid sort later to validate it.

How many cards to include

The size of your card set has a direct effect on data quality. Too few cards and you will not see meaningful grouping patterns. Too many and participants get tired, rush, and produce noisy sorts.

Most card sorts use between 30 and 60 cards. Below roughly 20 cards, there is not enough content for distinct groups to emerge. Above 60, completion rates drop and the quality of grouping decisions declines, especially in unmoderated sessions where no one is there to keep participants engaged.

If your content set is much larger, do not include everything. Sample a representative subset that spans the range of topics and item types. Write each card as a short, self-explanatory label, avoid internal jargon, and make sure no two cards are near duplicates, since ambiguous or overlapping cards create artificial confusion that has nothing to do with your structure.

How many participants you need

Participant count depends on whether you are running the sort for numbers or for reasoning.

For quantitative sorts, where you rely on how often items are grouped together, agreement patterns tend to stabilize somewhere around 15 to 30 participants per distinct user group. Research from Nielsen Norman Group on card sorting suggests that correlation between results climbs quickly and then flattens, so pushing far beyond 30 adds cost without changing the picture much. You can read their guidance on card sorting fundamentals.

For qualitative, moderated sorts, where the value is in hearing why people group things the way they do, 8 to 12 participants per group is usually enough. You are collecting reasoning, not statistical agreement, so a handful of thoughtful sessions goes a long way.

The bigger issue is not headcount but fit. A card sort is only as good as the people doing it. If you recruit general consumers when your product serves procurement leaders, the groupings you get will reflect the wrong mental model. This matters even more in B2B, where the right participant might be one specific role at a specific company size. Our guides on how to recruit participants for user research and how to recruit B2B research participants go deep on getting the sample right.

This is where recruiting infrastructure matters. CleverX is a B2B research platform with more than 8 million verified professionals, plus B2C reach, across 150 or more countries. Because participants are identity and employment verified, you can target a card sort to the exact roles whose mental model you are trying to capture, and you get enough of the right users for the groupings to hold together rather than shift with every new sort.

Moderated versus unmoderated card sorting

The second format decision is whether to observe the sort live or let participants complete it on their own.

Moderated card sorting is run in a session, in person or over video, with a researcher present. You watch decisions happen, ask why a card went into one group instead of another, and probe hesitation. This surfaces reasoning that raw data never shows, such as a participant who almost put an item in two places or who misread a label. The cost is that moderated sessions are slower and harder to scale, so you run fewer of them.

Unmoderated card sorting is completed by participants on their own through a tool, with no researcher present. It scales easily, reaches people across time zones, and is well suited to quantitative sorts where you want 20 or more participants. The trade-off is that you see the result but not the thinking, so a confusing card can quietly distort the data without you knowing why.

A common and effective sequence is to run a small moderated round first to catch confusing cards and understand reasoning, then launch a larger unmoderated round to confirm the patterns at scale. This combines depth and volume without overinvesting in either.

How to analyze card sort results

Analysis looks intimidating but rests on a few conceptual ideas. You do not need to compute anything by hand, since card sorting tools generate these views for you. The goal is to understand what each view is telling you.

The similarity matrix

The similarity matrix is the foundation of quantitative analysis. It is a grid of every card against every other card, where each cell shows the percentage of participants who placed that pair in the same group. A pair that 90 percent of participants grouped together is a strong, stable relationship. A pair that only 20 percent grouped together is weak. Reading the matrix, you look for blocks of high agreement, which represent natural clusters, and for cards that do not agree strongly with anything, which are the items people find hard to place.

Dendrograms

A dendrogram is a tree diagram that shows how cards cluster together based on the similarity data. Cards that were frequently grouped together join at a low level, and clusters merge into bigger clusters as you move up the tree. Conceptually, it answers the question, if I wanted to divide this content into 3 groups, or 5, or 8, where would the natural cut lines fall? You read a dendrogram by choosing a level to cut across it, and the branches below that cut become your candidate categories. It is a fast way to see both tight groupings and loose ones that could go either way.

Most card sorting tools produce two flavors of dendrogram using slightly different clustering assumptions, so treat them as suggestions to interpret rather than answers to copy. Combine them with the labels participants used in open sorts, and pay attention to the cards that never settle, since those often signal content that needs rewording, splitting, or removing.

The point of analysis is not a perfect algorithm output but a defensible structure you can explain and test. If you already run interviews and other qualitative work, our piece on analyzing user interview data covers a similar mindset of moving from raw input to a decision.

How card sorting feeds information architecture

Card sorting does not hand you a finished navigation. It hands you evidence about how users group and name content, which you translate into an information architecture through judgment.

In practice that means taking your clusters, resolving the cards that would not settle, and reconciling user labels with terms your product and brand can actually use. A category that 70 percent of users grouped together and named with similar words is a strong candidate for a top-level section. A cluster that split evenly might become two smaller groups, or a signal that the underlying content needs rethinking.

The insight only matters if it changes what you build. Turning a similarity matrix into shipped navigation is a decision-making step, and our guide on how to turn product research into better product decisions and our post on turning product research into better product decisions walk through making that leap without losing the evidence along the way.

Why card sorting pairs with tree testing

Card sorting and tree testing are two halves of the same information architecture workflow, and neither is complete without the other.

Card sorting is generative. It helps you build a structure by learning how users would organize content. Tree testing is evaluative. It takes a structure you have already built, strips away visual design, and asks users to find specific items by clicking through the labels alone. Tree testing answers a question card sorting cannot: given this navigation, can people actually locate what they came for?

The natural sequence is card sort first, tree test second. You use card sorting to design candidate categories and labels, draft the navigation, then run tree testing to confirm that real users can find things in it. If tree testing shows people getting lost, you revise the structure and test again. Running both means your final architecture is not just built the way users think, but proven to work when they try to use it. For B2B products specifically, where audiences are narrow and expert, our B2B user research playbook explains how to sequence methods like these for a specialized audience.

Running a card sort with the right participants

The method is only as strong as the sample. To recruit enough of the right professionals for stable groupings, quickly and without a panel contract, you can recruit verified participants on CleverX. The platform gives you access to more than 8 million verified professionals plus B2C respondents across 150 or more countries, with typical delivery in about 2 to 5 days and pay-as-you-go pricing, so you can target a card sort to the exact roles whose mental model you need to capture.

Frequently asked questions

What is a card sorting study?

A card sorting study is a research method where participants group labeled cards into categories that make sense to them, then name those categories. It reveals how users mentally organize content, which helps you design navigation and information architecture that match their expectations rather than your internal structure.

What is the difference between open and closed card sorting?

In an open card sort, participants create and name their own categories, which is ideal for discovery and building a new structure. In a closed card sort, participants place cards into categories you predefine, which is ideal for validating or refining an existing structure. A hybrid sort lets participants use your categories but also add their own.

How many participants do you need for a card sort?

For quantitative card sorts that rely on similarity data, 15 to 30 participants per user group is a common range because agreement patterns stabilize around that point. Moderated or exploratory sorts can run with 8 to 12 participants since you are gathering reasoning rather than statistical agreement. What matters most is recruiting the right users, not just enough of them.

How many cards should a card sort include?

Most card sorts use between 30 and 60 cards. Fewer than 20 rarely produces meaningful grouping patterns, and more than 60 tends to fatigue participants and lowers data quality. If you have a large content set, sample a representative subset rather than including every item.

Should I use moderated or unmoderated card sorting?

Use moderated card sorting when you want to hear the reasoning behind each decision and can invest in fewer, deeper sessions. Use unmoderated card sorting when you want to reach a larger, geographically spread sample quickly and care most about grouping patterns. Many teams run a small moderated round first, then a larger unmoderated round to confirm.

How does card sorting relate to tree testing?

Card sorting is generative and helps you build a structure by learning how users group content. Tree testing is evaluative and checks whether users can find items in a structure you have already built. Run card sorting first to design the information architecture, then run tree testing to validate that the navigation actually works.