Research methods
Card sorting in UX: sort with 15 real users before you draw the menu
Card sorting in UX means users group labeled cards so menus match how they think. Run an open sort with at least 15 people from your real audience.
Card sorting in UX is a research method where people who use your product group labeled cards (one per page, feature or piece of content) into piles that make sense to them, so you can build menus and categories around how they think. Run a UX card sort before you design navigation, with at least 15 people from your actual audience, and treat the team's own instincts as the hypothesis being tested.
The last part matters most. Your team already knows where everything lives, which is exactly why its sense of what belongs together is a poor guide. (Nobody on the team has ever looked for "Settings" under "Account" and given up.)
Pick the type of card sort by the question you need answered
Nielsen Norman Group describes three versions. Each answers a different question.
| Type | What participants do | Use it when |
|---|---|---|
| Open | Create and name their own groups | You are designing a structure, or want to see how users think about the content |
| Closed | Place cards into categories you defined | You want to check whether existing categories hold up |
| Hybrid | Use your categories and may add their own | You are confident about one or two categories and unsure about the rest |
Start with an open sort. NN/g calls it the most common type and says it gives the fullest picture of how users think about the content. The group names participants invent are useful evidence about their vocabulary, though they are rarely good enough to ship as labels.
Skip the closed sort for checking a finished menu. NN/g recommends tree testing for evaluating navigation categories, because a tree test asks people to find specific things in your proposed structure, which is closer to real use than dropping cards into buckets. NN/g also advises against hybrid sorts in most cases, since the categories you supply nudge people toward your structure.
Recruit 15 people for an open sort, and more if you want percentages
Jakob Nielsen's 2004 analysis compared smaller groups against results from 168 participants. Five participants produced results that correlated at 0.75 with the full set, 15 reached 0.90 and 30 reached 0.95. His recommendation was at least 15, three times the usual number for a usability test, with 30 reserved for large, well-funded projects.
NN/g's current guidance splits the number by goal: at least 15 people for a qualitative study where you hear their reasoning, and 30 to 50 for a quantitative study where you count how often cards land together. The same page recommends 30 to 50 cards, because past 50 people get tired and a large "Miscellaneous" pile appears.
So: 15 people if you will watch and listen, 30 or more if you will run it unmoderated and lean on the numbers. The tradeoff is that the extra people add confidence in the pattern while adding little new understanding of why it exists.
Moderate a few sessions even when the main study runs unmoderated
In a moderated sort, a researcher watches live (in person or over a video call) and asks follow-up questions. In an unmoderated sort, a tool gives the instructions and collects the result with nobody present. NN/g matches moderated sessions to qualitative goals and unmoderated ones to quantitative goals, and recommends running a few moderated sessions alongside a large unmoderated study.
Do that. The numbers tell you two cards ended up together; only a person thinking aloud tells you it was because both mentioned invoices, and the cards were otherwise unrelated. Ask participants to talk while they sort, and save your questions for when they hesitate over a card.
Short version: run three to five moderated sessions first, fix any confusing card labels, then launch the unmoderated study.
Write cards that never share a key word
NN/g warns that identical words in different card labels make participants group those cards automatically. Its example is car models: put the brand name on every card and people sort by brand, missing differences such as size and seating.
Rewrite cards so each describes the thing in plain terms, and vary the wording across cards that share a word in your product. If five cards start with "Report", people will make a pile called Reports and you will learn nothing about how they think about the five reports. Writing the cards takes longer than you expect. That is fine; it is where most of the study's quality is decided.
Recruit people whose mental models differ from the team's
NN/g calls recruiting participants who represent your target audience crucial. For consumer products that is mostly a sampling question. For specialist B2B tools it decides whether the study is worth running.
Here is a made-up example. A software company sells a laboratory information system to hospital labs. The product team organized the menu by system component: Instruments, Rules Engine, Integrations, Reports. The people who use it daily are lab managers and medical technologists, and in this example they think in terms of where a sample is in its journey: before testing (collection, labeling, accessioning), during testing (instrument runs, quality control) and after testing (result review, reporting). A card sort with the company's own support staff would mostly reproduce the engineering menu. A card sort with 15 lab managers might split "Instrument interfaces" across two piles, because to them connecting an analyzer and reviewing its results are different jobs at different times of day.
That gap only shows up when the participants are the specialists themselves. Write a screener that checks they do the work, not that they hold the title. Research screener questions covers how. Lab managers, procurement leads and clinicians have little spare time, so plan an incentive they will take seriously and more lead time than for a consumer study. Research participant recruitment covers channels and incentives for business participants.
Sort with two groups when your users split by role. If technologists and lab directors both use the system, run 15 of each and compare. A structure that works for one group and baffles the other is a finding, and it might mean the product needs two entry points.
Read the similarity matrix for strong pairs, then for orphans
Most card sorting tools produce a similarity matrix: a grid with every card along both axes, where each cell shows the percentage of participants who put that pair in the same group. Maze, which sells a research platform with card sorting, gives the calculation as the number of participants who paired the two cards, times 100, divided by the total number of participants. Darker cells mean stronger pairs, and the most closely related pairs cluster along the edges of the grid.
Read it in three passes:
- Strong clusters. Blocks of dark cells are groups most people agree on. In the lab example, if 13 of 15 participants put "Quality control rules" with "Levey-Jennings charts", which is 87%, and those belong together.
- Orphans. A card with no dark cells anywhere is one people could not place. Usually the label is unclear, or the card describes something users do not think of as a separate thing. Reword it and retest.
- Split cards. A card that pairs at around half with two different clusters belongs in two places, or needs a cross-link from the other section. Check your moderated notes to see which reasoning drove each choice.
Treat the matrix as evidence for a draft structure. Then draw the menu, and test it with a tree test and later a usability test once screens exist. If you also ran moderated sessions with lots of notes, affinity mapping is a quick way to group what people said about why they sorted the way they did.
Your next step
List the 30 to 50 things your navigation needs to hold, rewrite any labels that share a word, and book three moderated sessions with people who do the job your product supports. Instant Expert can find people matching a description such as "lab managers at hospital clinical laboratories who use a laboratory information system daily" and invite them to a short paid call; the offer you set is what they receive, Instant Expert's fee is added on top, and you pay only for calls that get booked. To get advice on study design itself, start with the directory pages for user research experts and UX consultants.