Run a card sort

Outcome

A draft set of top-level categories, each with a label taken from participants' own names for their groups, and a list of the cards participants could not agree on. The draft is ready to be checked with a tree test.

When to use it

• You are building or reorganising navigation, a help centre, a settings area or any place where people browse to find things.
• The current categories were decided in a meeting and nobody outside it has seen them.
• You need labels as well as groups, and want them in the words your users use.

When not to use it

• You already have a structure and want to know whether people can find things in it. That is a tree test.
• The content is small enough to show on one screen, such as seven items. There is nothing to organise.
• People find things by searching for a known name, such as a product code. Invest in search instead.

Before you start

• The content inventory. Everything the structure has to hold, from a sitemap, a spreadsheet or the current navigation.
• Participants. At least fifteen people who would use the structure. Recruit across the user groups who will browse it.
• The tool. A remote card sort tool, such as the Card Sort in Pulse, or paper cards if the sessions are in person.
• Time. A day to choose and write the cards, a few days to collect sorts, half a day to analyse.

Steps

Step 1: Decide between an open and a closed sort

Choose open if you need categories and labels: participants create and name their own groups. Choose closed if you already have categories and want to see whether content lands in them. A first sort for a new structure is almost always open.

You should now have: a decision of open or closed, written next to the reason.

Step 2: Choose and write the cards

Pick 30 to 60 items from the inventory that represent the range of content, including the items you expect to be hard to place. Write each card in plain words a participant would understand without context. Avoid starting several cards with the same word, which makes people group them by the word.

You should now have: 30 to 60 cards, written in plain words, without repeated leading words.

Step 3: Set up the sort and write the instructions

Enter the cards in the tool. Write one short instruction: group the cards in a way that makes sense to you, and name each group. Say there are no right answers and that a card can go in a group called "not sure". Add one question after the sort about which cards were hardest to place.

You should now have: a working sort with one instruction and one follow-up question.

Step 4: Run a pilot with two people

Send the sort to two people who fit the profile. Check how long it takes, which cards they did not understand and whether the instruction was clear. Rewrite unclear cards and remove any that everyone misread.

You should now have: a piloted sort, with unclear cards rewritten and a known completion time.

Step 5: Collect at least fifteen sorts

Send the sort to the participants. Watch the first few results to make sure nothing is broken. Close the sort once you have fifteen or more completed sorts from people in the profile, and drop any that took under a minute.

You should now have: fifteen or more completed sorts from people in the profile.

Step 6: Read the similarity matrix for groups that recur

Open the similarity matrix, which shows how often each pair of cards was placed together. Look for clusters of cards that were paired by more than half the participants. Each strong cluster is a candidate category. Cards with no strong pairing are the hard-to-place list.

You should now have: a set of candidate categories from strong clusters and a list of hard-to-place cards.

Step 7: Name the categories from participants' labels

For each candidate category, list the names participants gave the groups that held most of its cards. Pick the name used most often, or the clearest of the common ones. Write the draft structure with these labels and the hard-to-place cards noted.

You should now have: a draft structure with labels taken from participants, and a list of cards to resolve.

Worked example

A software company is reorganising its help centre, which has 140 articles under eight categories named after internal teams.

1. Open sort, because the team wants new categories and new labels.
2. 48 cards, including Change the email address on your account, Download an invoice and Connect to Slack. Three cards beginning "Manage…" are rewritten so they start differently.
3. The instruction asks people to group the cards as they would expect to find them in help.
4. The pilot takes eleven minutes. Both pilot participants misread SSO setup, which becomes Let your team sign in with your company login.
5. 22 customers complete the sort.
6. Five strong clusters appear: account, billing, team members, integrations and data export. Audit log and API keys pair with nothing.
7. Participants called the billing groups "Billing" eleven times, "Payments" six times and "Money stuff" once. The draft uses Billing, Your account, Team, Connect other tools and Export your data. The two hard-to-place cards go into the tree test as open questions.

Checklist

• The choice of open or closed sort is written down with a reason.
• There are 30 to 60 cards, written in plain words.
• No group of cards shares a leading word.
• The sort was piloted with two people and fixed.
• At least fifteen people in the profile completed it.
• Categories come from clusters most participants agreed on.
• Labels come from the names participants gave their groups.
• Hard-to-place cards are listed for the next step.

Common mistakes

• Using the current labels on the cards. If a card says "Workspace settings", people will group it under anything named workspace. Describe the content, not the menu.
• Treating every cluster as a category. Two or three participants agreeing is noise. Look for pairings most people made.
• Stopping at the sort. A card sort suggests a structure; it does not prove people can find things in it. Test the draft.
• Ignoring the group names. The labels participants wrote are half the result, and the half teams most often throw away.
• Too many cards. Above about 60, participants tire and start grouping carelessly.

Related

• Card Sort: run an open or closed sort with the panel and read the similarity matrix.
• Tree Test: check the draft structure once the sort is done.
• Run a tree test: the step that follows a card sort.
• Match between system and the real world: why labels should use the users' words.
• Information architecture and card sorting course: the method in lessons with knowledge checks.
• What Is a Taxonomy in Information Architecture?: why categories built from card sorts outlast categories from a whiteboard.