How to Interpret Dendrograms in Card Sorting Analysis
Dendrograms look intimidating but reveal powerful patterns. Learn how to read card sorting dendrograms and turn cluster data into clear IA decisions.
How to Interpret Dendrograms in Card Sorting Analysis
You ran your card sort, collected the data, and now you're staring at something that looks like an upside-down tree diagram. That's a dendrogram, and it's one of the most useful outputs from a card sorting study — once you know how to read it.
A dendrogram shows you how participants grouped your cards, visualized as a hierarchy. Cards that people consistently placed together sit close to each other at the bottom. Cards that rarely ended up in the same group only connect way up near the top. The height where two branches merge tells you how strong (or weak) that relationship is.
This guide walks you through reading a dendrogram step by step, so you can turn that tree diagram into real information architecture decisions.
What You'll Need
- A completed card sorting study with at least 15-20 participants
- Access to dendrogram analysis tools (built into most card sorting platforms)
- A CardSort account (free at freecardsort.com)
- A basic sense of your content categories and what your users are trying to do
Step 1: Examine the Overall Tree Structure
Start by zooming out. Look at the full dendrogram and find the main branches — the thick trunks that split off from each other at the top of the tree. These are your primary categories, the big buckets that participants naturally created when they sorted your cards.
The way a dendrogram works: individual cards sit at the bottom as their own tiny branches. Moving upward, cards that were frequently grouped together merge first (at lower heights). Cards with weaker connections merge later, higher up the tree. So the lower a merge point, the stronger the association between those items.
Pro tip: Take a screenshot of the full dendrogram before you start drilling in. It's easy to lose the forest for the trees (literally) when you're analyzing individual clusters.
Step 2: Identify Cluster Separation Points
Now look for the big gaps. You want to find places where branches stay separate for a long stretch before finally merging — those tall vertical lines between merge points. These gaps are your natural category boundaries.
Think of it this way: if a group of cards all merge together at low heights, they belong together. Then there's a big jump before that group merges with anything else. That jump is telling you "these are different categories in your users' minds."
Example: Say your cards merge steadily at heights around 0.2, 0.3, and 0.35. Then nothing merges again until 0.6. That big gap between 0.35 and 0.6 is your signal — draw a horizontal line there, and everything below it falls into distinct groups.
Step 3: Count and Validate Distinct Groups
Pick a cut-off height and count how many separate branches exist below it. Those are your proposed categories.
A few things to watch for:
- Aim for 3-7 main groups. Fewer than 3 means your cut-off is too high and you're lumping unrelated things together. More than 7-8 probably means you're slicing too fine.
- Check that groups are roughly balanced. If one cluster has 15 cards and another has 2, your cut-off point might need adjusting.
- Try a few different cut-off heights. Move the line up and down and see how the groupings change. There's rarely one perfect answer — you're looking for the height that produces the most sensible, usable set of categories.
Pro tip: If you end up with one giant group and a bunch of tiny ones, that's a sign the giant group actually contains subcategories. Try a lower cut-off to split it apart.
Step 4: Analyze Within-Group Relationships
Once you've identified your main clusters, look inside each one. The internal structure tells you about subcategories and content relationships that can inform your navigation hierarchy.
Cards that merge at very low heights within a cluster are tightly connected — people almost always put them together. These are your "definitely belongs here" items. Cards that merge late within a cluster (close to the cut-off line) are more loosely associated. They fit in this group, but just barely.
Example: Inside a "Customer Support" cluster, you might see "FAQ" and "Help Articles" merging first at a very low height — people see those as basically the same thing. Then "Contact Us" joins the cluster much later. That tells you FAQ and Help Articles should be neighbors in your nav, while Contact Us can sit a bit further away (maybe as a separate link or a different section within Support).
Step 5: Cross-Reference with Participant Data
A dendrogram is an aggregate view — it shows you the average pattern across all participants. But averages can hide important differences.
Go back to the individual sorting data and check:
- Do most participants agree with the clusters you're seeing? If a cluster only reflects what half your participants did while the other half split those cards differently, that's a weak grouping.
- Are there distinct user segments with different mental models? Sometimes a dendrogram looks muddy because two different user types organized things in fundamentally different ways.
- Which items are "floaters"? Some cards end up in different groups depending on who's sorting. These borderline items are prime candidates for cross-linking in your final design.
Pro tip: Borderline cards that could fit multiple categories are actually some of the most valuable findings. They tell you where users might look in the wrong place — so you know where to add shortcuts, cross-links, or redundant navigation.
Step 6: Document Insights and Recommendations
Now translate your analysis into something your team can act on. For each recommended category:
- List the cards that belong in it and note how tightly they cluster
- Call out any subcategory structure you found within the group
- Flag items that borderline between categories
- Suggest where cross-links might help
It's also worth documenting what the dendrogram tells you versus what it doesn't. A dendrogram shows you groupings, but it doesn't tell you what to name those groups (that's what your category labels from the sort are for) or how to prioritize them in navigation.
Example: "The dendrogram suggests 5 primary categories. 'Products' and 'Services' each have strong internal groupings — cards within each cluster merged early and stayed together. But the two clusters barely connect to each other, which supports putting them as separate top-level sections rather than combining them."
Pro Tips
-
Try multiple cut-off heights. Analyzing the dendrogram at 2-3 different levels gives you both a broad view (fewer, larger categories) and a detailed view (more, smaller subcategories). Both are useful for different parts of your IA.
-
Sanity-check against the actual card labels. If a cluster is statistically tight but the card names in it don't make logical sense together, dig into the individual data. Confusing card labels can cause misleading clusters.
-
Factor in real-world constraints. The dendrogram might suggest 7 top-level categories, but your nav bar only fits 5. That's okay — use the dendrogram to understand the ideal structure, then make pragmatic compromises where needed.
-
Flag weak signals for follow-up. If some groupings are ambiguous, note them as open questions. They're good candidates for tree testing or A/B testing later on.
Common Mistakes to Avoid
-
Forcing your existing categories onto the data. If the dendrogram tells you something different from your current site structure, listen to it. The whole point of card sorting is to discover how users actually think, not to confirm what you already built.
-
Reading too much into tiny differences. Small variations in merge height — especially with fewer than 30 participants — might just be noise. Focus on the big, obvious separations first.
-
Ignoring how card labels affect results. Participants sort based on what they read on the cards. If a card was poorly worded or ambiguous, the dendrogram will reflect that confusion. Bad input, bad output.
-
Slicing into too many groups. If you end up with 10+ categories, you've probably gone too granular. Most people can only hold about 5-9 categories in working memory, and your navigation should reflect that.
Further Reading
- What is Card Sorting? Complete Guide
- Card Sorting (UX Glossary)
- Information Architecture (UX Glossary)
- How To Run Your First Card Sort Study
Frequently Asked Questions
How long does dendrogram interpretation take for card sorting analysis?
For a straightforward study, expect to spend around 15-30 minutes on your first pass — just getting a feel for the main clusters and obvious patterns. Writing up your findings and recommendations takes another 30-60 minutes on top of that. If you're working with a larger study (50+ cards or unusual patterns that need investigation), budget a couple of hours to do it properly.
What tools are required for dendrogram interpretation in card sorting?
You probably don't need anything beyond what your card sorting platform already provides. Tools like OptimalSort, UsabilityHub, and CardSort all generate dendrograms and similarity matrices for you. If you want to do more advanced statistical analysis, you can export raw data into R or Python, but for most UX research projects the built-in tools are plenty.
How do I validate the accuracy of my dendrogram interpretation?
Look for a few things: clear visual separation between your main clusters (not everything blending together at similar heights), groups that are reasonably balanced in size, and labels within each cluster that actually make sense together. Then go back to the individual participant data and check whether the majority of people sorted things in a way that matches the clusters you identified. If only a slim majority agree, treat that grouping with some skepticism.
What merge height differences indicate meaningful category separations?
There's no universal magic number, but as a rough guide: a gap of 0.2 or more between merge heights usually points to a genuine category boundary in your users' minds. Very small differences (under 0.1) are more likely noise, especially with smaller sample sizes. The most reliable approach is to look for the largest jumps in the merge height progression — those are your strongest natural dividing lines.
How many participants do I need for reliable dendrogram patterns?
Aim for at least 15-20 participants. Below that, your dendrogram might look clean but the patterns won't be stable — run the study again with different people and you'd likely get different clusters. With 30 or more participants, you can be much more confident that the patterns you're seeing reflect genuine shared mental models rather than the quirks of a small sample.