The sort closed on Friday. Thirty participants grouped the same 45 cards, and no two of them built quite the same menu. Card sorting analysis is how you find the structure those sorts share, measure how firmly participants agree on it, and explain the cards where they split.
The numbers show you where agreement is strong. The sessions behind them show you why a card was hard to place. Below, we cover both, in the order a research team would work through them, and finish with a draft structure you can defend in front of stakeholders and then validate.
TL;DR
- Check your sample first: NN/g recommends at least 15 participants for a qualitative sort and 30 to 50 for a quantitative one.
- Read agreement with the similarity matrix and dendrogram, then read disagreement in the sessions.
- Treat a card split between two groups as a labeling or cross-linking decision, rather than a failure.
- Validate the draft structure with a tree test before it ships.
Know What Your Card Sort Can Answer
Before opening the results, be clear about the two kinds of evidence a card sort produces. Each one answers a different question, and a strong analysis uses both.
- The quantitative layer counts how often participants placed cards together or into the same category, answering “what belongs together, and how sure are we?”
- The qualitative layer comes from what participants said and did while sorting, answering “why did this card land here, and why did that one stall people?”
Card sorting also has limits worth naming before anyone asks. According to Nielsen Norman Group’s guide to card sorting, participants create a single level of grouping, so the method will not hand you subcategories or full navigation paths. The cards also appear without the page context, imagery, and links that guide people on a live site.
Those limits shape how you report the results. A card sort tells you how people group your content in their heads; a tree test later tells you whether they can find it in the structure you build.
Check Whether Your Sample Can Carry the Analysis
Card sort data stabilizes more slowly than usability test findings. Confirm the sample is large enough before you read meaning into small differences in the matrix.
The best-known evidence comes from Thomas Tullis and Larry Wood, who ran a card sort with 168 participants and then measured how closely smaller random subsets matched the full result. Jakob Nielsen summarized their findings in an NN/g article on how many participants a card sort needs:
NN/g recommends at least 15 participants for a qualitative sort, where you ask people to explain their groupings, and at least 30 to 50 for a quantitative sort where you want the groupings to generalize.
If your study fell short, you can still learn from it. Read the strongest clusters as directional, avoid ranking cards by small percentage differences, and plan to confirm the structure with a tree test. When you need more sorts quickly, you can recruit participants who match the audience for that part of the site.
Read the Five Core Outputs
Most card sorting tools present the same handful of views. Each one answers a narrower question, so read them in sequence rather than jumping straight to the most colorful chart.
Card and Category Tables
Start with the plain counts. The card view lists every card with the categories participants chose for it and how often, while the category view shows which cards ended up inside each group.
A card that lands in one or two categories is easy to place. A card scattered across six categories is your first signal of a vague label or a concept your audience does not share.
Similarity Matrix
The similarity matrix compares every pair of cards. Each cell holds the share of participants who put both cards in the same group, so 100% means everyone paired them and 50% means half did.
Look for sets of cards that all score high with each other, since those sets mark candidate categories. Then look for a card with moderate scores to two separate sets, since that card is pulling in two directions.
Dendrogram
A dendrogram turns the pair data into a tree. Cards that participants grouped together often join near the leaves, and cards that were rarely grouped join only near the root.
The useful decision is where you “cut” the tree. Cutting low gives you many small, tight groups; cutting high gives you a few broad ones. Try two or three cut points and compare each set of groups against the number of top-level menu items your design can realistically hold.
Result Matrix
The result matrix crosses cards with categories and shades each cell by how many participants made that placement. In a closed or hybrid sort, it is the fastest way to see whether your predefined categories are pulling the cards you expected.
Individual Sorts and Recordings
Aggregate views flatten everyone into one answer. Individual sorts show the reasoning that got averaged away, such as a participant who sorted by audience while everyone else sorted by task.
In Userlytics, the Help Center guide to card sorting metrics shows how to open each participant’s full set of choices, the time they spent, and the video of their session. The same Metrics area includes a Recommendation tab that tags each card as Recommended, Needs Review, or Uncategorized based on how consistently participants placed it.
How to Analyze Card Sorting Results in Six Steps
The steps below follow the order that keeps you from building conclusions on bad data. Plan on going back and forth between the numbers and the sessions; the order is a starting point.
1. Check the Data Before You Analyze It
Review sorts that stand out before deciding whether to exclude them. Fast completion, very few groups, or a large “Other” category can signal low effort, but they can also reflect a participant’s understanding of the content.
Check the recording and comments, where available, to understand those choices. Exclude a sort only when there is clear evidence that the participant did not engage with the task or follow the instructions, and document the reason.
Pay particular attention to large catch-all groups. These may reflect fatigue, unclear labels, or unfamiliar content. Use the session evidence to investigate before drawing conclusions.
2. Standardize Labels in Open and Hybrid Sorts
In an open sort, 30 participants can produce 30 different names for the same idea: “Billing,” “Payments,” “My Invoices,” and “Money stuff” may all describe one group. Merge labels that clearly refer to the same concept into a standard category so the counts reflect meaning rather than spelling.
Keep a log of every merge and the original labels behind it. The raw vocabulary is evidence in its own right, because the words participants chose are strong candidates for your final navigation labels.
3. Lock In the Strong Clusters First
Use the similarity matrix and dendrogram to identify the groups most participants agree on. Write these down as your provisional categories and move on, since spending an hour on cards everyone agrees about delays the work that matters.
Before you accept a strong cluster, check for keyword matching. NN/g’s research on terminology matches shows that participants often group cards that share a word, such as every card starting with “Employee,” instead of grouping by meaning. A cluster held together by a repeated word may say more about your card labels than about your audience’s mental model.
4. Triage the Cards That Split
Now turn to the cards without a clear home. Sort them into three piles:
- Two-home cards: placed in two categories at similar rates.
- Scattered cards: spread thinly across many categories. The label is usually the problem, or the concept is unfamiliar to the people you recruited.
- Orphans: cards that sit mostly in a catch-all group. These need a closer look at the sessions before you decide anything.
The two-home card is the one teams most often get wrong. Forcing it into one category goes against how about half your participants grouped it, so treat it as a cross-linking or labeling decision and let a tree test show which location people can find.
For example, if a “Parental leave” card on a benefits portal splits almost evenly between “Time Off” and “Benefits,” participants are telling you it genuinely belongs in both places.
5. Watch the Sessions Behind the Hard Cards
The numbers tell you which cards are hard. Only the participants can tell you why. NN/g advises running a few moderated sessions alongside a quantitative unmoderated sort for exactly this reason.
If participants used the think-aloud protocol while sorting, go straight to the moments where they picked up a split card. Listen for three things: a participant who did not understand the label, one who understood it but saw two valid homes, and one who sorted by a different principle altogether, such as by department or by audience.
On the Userlytics platform, the AI Insights feature can help you find those moments across many recordings. You ask a question about the study in plain language, and each answer links back to the sessions, transcripts, and annotations it drew on, so you can check the evidence before you act on it.
6. Draft the Structure in Participants’ Words
Combine the strong clusters, your triage decisions, and the session evidence into a draft structure. For each top-level category, record which cards it holds, the agreement behind it, and the label most participants used.
Resist renaming categories into internal language at this stage. If participants consistently called a group “Help,” a label like “Customer Success Resources” undoes part of the work you just did.
Need a second analyst on a messy sort? Userlytics’ UX consultants can take on the analysis for you, from study design through a report and presentation with recommendations. Talk to a UX consultant
Match Each Pattern to an IA Decision
Analysis earns its place when it changes the structure. Use the table below to move from what you see in the results to what you do next.
Each row is a hypothesis. Treat the table as a list of decisions to validate, rather than a final answer.
Adjust the Analysis for Open, Closed and Hybrid Sorts
The sort type changes what you read first. The card sorting tool page describes all three formats; here is how the analysis shifts for each.
Open sorts are generative, so their value lies in the groupings and vocabulary you did not anticipate. Closed sorts are evaluative, so their value lies in where your proposed categories fail to hold their cards.
Validate the Draft Structure With a Tree Test
A card sort suggests a structure, but it does not prove people can navigate it. NN/g’s guide to tree testing describes card sorting as a generative method and tree testing as the way to evaluate the resulting hierarchy.
The handoff from one study to the next is where your analysis pays off twice. NN/g recommends including categories where card sort participants disagreed as tree test tasks, so every two-home card and every relabeled scattered card from step 4 becomes a task.
When you run a tree test, participants look for items in a text-only version of your menu, so you learn whether the new labels and locations work before design starts. Our post on website information architecture covers how the two methods fit into a full IA project.
Present the Findings So Stakeholders Approve the Structure
The readout decides whether the structure gets built. Most stakeholders will not read a similarity matrix, so lead with decisions and keep the charts as backup.
- Open with the proposed structure. Show the top-level categories and their cards on one slide.
- Show the evidence behind each category. How often did participants group two cards together? For example, “24 of 30 participants grouped ‘Invoices’ and ‘Payment history’ together.” Avoid presenting a pair’s agreement score as agreement on the entire category.
- Play two or three clips. A clip of a participant explaining why a split card belongs under two headings persuades faster than a heatmap.
- List the open decisions. Name the two-home and scattered cards, and what you propose for each.
- State the limits. Note the sample size, the single level of grouping, and the tree test that will confirm the structure.
A readout built this way connects the research to the decision in front of the room: which structure the team commits to and what it still needs to verify.
Your Next Step After the Sort Closes
Before you open a single chart, check your sample size and remove the sorts that show no effort. Then list every card that split between two or more groups, as those cards are where your structure will succeed or fail.
Watch the sessions for those cards first, draft the structure in your participants’ own words, and turn every open question into a tree test task. If you are planning your next sort, our card sort template gives you a ready-made starting point that already asks participants to think aloud.
See how Userlytics turns card sorts into a structure your team can defend. Book a free demo
Frequently Asked Questions
Short answers to the questions research teams ask most often once their card sort results arrive.
