How to run a card sort (step by step, with a worked example)

A card sort reveals how your users naturally group and label content — the raw material for an information architecture people can actually navigate. Participants organize a set of items ("cards") into groups that make sense to them, not to your org chart. This guide takes you from a blank study to a defensible IA, with a worked example the whole way through.

When to use card sorting

You're designing or restructuring navigation, a menu, or a content hierarchy.
You want categories and labels in users' language, not internal jargon (users don't search for "Solutions" or "Resources" — they search for what they're trying to do).
You're early enough that the structure isn't locked. Card sorting generates options; tree testing validates them.

Open vs closed card sorting

Open sort: participants create and name their own groups. Best for discovery — you learn the groupings and the labels users would choose.
Closed sort: participants sort cards into your predefined categories. Best for validation — does your proposed structure hold up?
Hybrid: predefined categories, but participants can add their own — validates your structure while surfacing gaps.

Start open when you're exploring; go closed when you're confirming. Full breakdown: open vs closed card sorting.

Step 1: Choose your cards

List the content items or features to organize. Keep each card a single, concrete item with a clear name. 30–60 cards is the comfortable range — too few and you learn little, too many and participants fatigue and sort carelessly.

Worked example — a SaaS help center. You're restructuring the support site and want to know how users would organize these 20 topics:

Reset password · Update billing card · Cancel subscription · Invite a teammate ·
Export data · API keys · Set up SSO · Change plan · Download invoices ·
Two-factor authentication · Delete account · Notification settings ·
Connect integrations · Usage limits · Refund policy · Keyboard shortcuts ·
Mobile app setup · Data privacy · Roles & permissions · Contact support

Add a one-line description to any card that could be ambiguous ("API keys — generate and revoke developer access tokens"), so people sort on meaning rather than guessing.

Step 2: Pick open or closed

Discovery → open. Validation → closed. If you're unsure, run an open sort first; its results often become the categories for a later closed sort. In our help-center example we'll run an open sort, because we don't yet know the right categories — that's exactly what we're trying to learn.

Step 3: Set your sample size

Card sorting stabilizes faster than most methods. Guidance:

15–20 participants per audience surfaces the major patterns.
Go to 30 for diverse audiences or high-stakes IA.
More participants mainly sharpen the agreement percentages; they don't usually change the big clusters after ~15.

Step 4: Recruit and distribute

Share a link with your panel, email list, or a QR code. Card sorts run unmoderated and asynchronously, so participants sort on their own time — desktop or mobile. A good card sort takes participants 5–10 minutes.

Step 5: Analyze the results

Worked example — what came back. After 18 participants sorted the 20 help-center cards, clear clusters emerged:

"Account & Login" — most people grouped Reset password, Two-factor authentication, SSO, and Delete account together. 15 of 18 put reset password + 2FA in the same group.
"Billing" — Update billing card, Download invoices, Change plan, Cancel subscription, Refund policy, Usage limits clustered tightly (16 of 18 grouped billing card + invoices).
"Team & Access" — Invite a teammate, Roles & permissions, Contact support? No — Contact support scattered everywhere.
Ambiguous cards — Data privacy landed sometimes in Account, sometimes in a standalone "Legal" group. Contact support went everywhere (people don't think of it as a topic — it's a fallback).

Read the analysis in this order:

Similarity matrix — the percentage of participants who put each pair of cards in the same group. Tight clusters of high percentages = natural groups. (16/18 = 89% for billing card + invoices is a strong signal.)
Suggested clusters / dendrogram — the tool proposes groups by linking cards above an agreement threshold (say 50%). Use it as a first draft.
Group labels (open sorts) — mine the names participants gave. If ten people independently write "Billing" or "Payments," you've found both a category and its label.
Ambiguous cards — items that scattered (Data privacy, Contact support) need a deliberate placement decision or clearer labeling.

Deeper walkthrough: card sort analysis guide.

Step 6: Turn the analysis into an IA

From the example, a defensible draft help-center IA becomes:

Account & Login  (Reset password, 2FA, SSO, Delete account)
Billing          (Billing card, Invoices, Change plan, Cancel, Refunds, Usage limits)
Team & Access    (Invite teammate, Roles & permissions)
Developers       (API keys, Integrations, Export data)
Getting Started  (Mobile app setup, Keyboard shortcuts, Notifications)

…with Contact support promoted to a persistent element (not a topic), and Data privacy placed under Account with a cross-link to a Legal page. Every one of those decisions traces back to participant data — that's the point.

Step 7: Validate with a tree test

A card sort proposes a structure; a tree test proves people can navigate it. Always follow up with a tree test on the draft IA before you build. If a branch tests poorly, revisit the card-sort data or the labels and re-test.

Common mistakes

Too many cards (60+) → fatigue and sloppy sorting. Split into two studies if needed.
Vague card labels → people sort on guesses. Add short descriptions.
Only running a closed sort when you actually needed discovery — you'll only ever confirm the categories you already thought of.
Eyeballing raw groupings instead of using the similarity matrix — you'll over-index on a few vocal participants.
Forcing every card into a clean group. Ambiguous cards are data, not noise — they're telling you something is genuinely unclear.
Skipping validation. A card sort alone is a hypothesis; the tree test is the experiment.

Run a card sort free

ResearchRocket includes open and closed card sorting with a built-in similarity matrix and agreement-based suggested clusters, plus tree testing to validate the result — all in one platform, no exporting to spreadsheets. Try the free card sorting tool.

FAQ

How do you run a card sort? Choose 30–60 cards, pick open (discovery) or closed (validation), recruit ~15–20 participants per audience, share a link, then analyze with a similarity matrix and suggested clusters — and validate the resulting IA with a tree test.

How many cards should a card sort have? 30–60 is the sweet spot. Fewer yields thin insight; more risks participant fatigue and careless sorting.

How many participants do I need for a card sort? Around 15–20 per audience surfaces the major patterns; 30 for diverse or high-stakes audiences. Beyond that, extra participants mainly sharpen the agreement percentages.

Open or closed card sort — which should I use? Open to discover categories and labels; closed to validate a structure you already have; hybrid to validate while surfacing gaps.

How long does a card sort take for a participant? Typically 5–10 minutes for 30–60 cards. Keep it under that to protect data quality.

What do I do with ambiguous cards that scatter everywhere? Treat them as a finding. Either the label is unclear (fix it and re-test) or the item isn't a real "topic" (like "Contact support") and belongs as a persistent element rather than a category.

Is there a free card sorting tool? Yes — ResearchRocket offers a free card sort with unlimited participants and built-in analysis, no account required.

Run a card sort free on ResearchRocket →