Card Sorting
Write features on cards and let users group them—more accurate than guessing yourself.
Write each feature on a card, let users group them and name each group. You'll walk away with a categorization logic completely different from your own. Card sorting is one of the cheapest research methods in information architecture, used for decades to uncover how users mentally group things. card-sorting-wiki
What you'll run into:
- No matter how you change the categories, someone says they can't find things, and you can't say where the problem is
- Your team argues endlessly about whether a feature belongs in one category or another
- The navigation labels follow your internal vocabulary, and users don't understand them
How to run it
| Step | What to do |
|---|---|
| Prepare | Write each of twenty to forty features on a card. Use language users understand, not internal jargon |
| Open sort | Let them group freely, any number of groups, then name each group. This round captures their mental model |
| Closed sort | Give them your group names and have them place the cards. This round tests whether your scheme can be understood |
| Probe | For any card they hesitate on for more than a few seconds, ask what they were thinking. Those cards mark the blurry category boundaries |
The open-then-closed order matters. If you do the closed sort first, your group names pull them along and you never see how they'd have sorted it on their own.
Five to eight people is enough. By the fifth person, almost no new groupings appear.
You can also have the AI play the user for a first pass—but know the boundary: it gives you the common splits of similar products, not your users' thoughts. It can eliminate bad schemes; it can't pick the best one.
How to read the results
- Look at which cards always stay together. If most people group them together, these features are one thing in users' minds, and they should sit together in the interface.
- Look at which cards wander everywhere. If everyone places them differently, the feature's positioning itself is blurry—you may need to rename it or split it.
- Look at the group names they come up with. Use these words directly as navigation labels; they'll be more accurate than what you'd invent. "Stuff about money" should probably be called "Checkout," not "Financial Management."
- Look at where they hesitate. Going back and forth on a single card means you've cut your categories along the wrong dimension—this card belongs to two at once.
How to do it solo
If you can't run a formal study, there are a few low-cost alternatives.
| Approach | How to execute |
|---|---|
| Run it online with three target users | Use any whiteboard tool, lay out the cards, let them drag. Twenty minutes per person |
| Use a reverse test instead | Give them a task and watch where they click first. Faster than card sorting, but you get no group names |
| See how competitors divide it | The navigation structures of three or four similar products. Users' mental models were already shaped by them |
| Let AI play the user for a first pass | You get no real mental model, but it quickly exposes obviously bad groupings, good for filtering out the worst schemes |
Know the boundary of the last one: AI offers the common splits of similar products, not your users' thoughts. It can rule out bad schemes; it can't choose the best one.
How to verify the split works. After launch, watch two numbers: search usage inside the navigation, and support feedback about "can't find things." If search stays persistently high, users' mental groupings still don't match your navigation's—run another card-sorting round and swap in the new group names, which beats fine-tuning the current categories.