Card sort analysis: how to read the results (with a worked example)

Running a card sort is the easy part. The real value is in the analysis — turning dozens of participants' groupings into a defensible information architecture instead of a gut-feel guess. This guide explains every output that matters, shows you a worked example with real numbers, and gives you a repeatable process for going from data to IA.

What a card sort gives you

In a card sort, participants group your content items ("cards") into categories — either open (they name their own groups) or closed (they sort into your predefined categories). Across many participants, two kinds of pattern emerge: what belongs together (grouping) and what people call it (labels). Good analysis extracts both.

The three analyses that matter

1. The similarity matrix (start here)

The similarity matrix shows, for every pair of cards, the percentage of participants who put them in the same group. It's the single most useful card-sort output because it's built directly from agreement — no interpretation required.

Worked example. Suppose 20 participants sorted a set of e-commerce account cards. A slice of the matrix looks like this (each cell = % who grouped that pair together):

	Update card	Invoices	Reset pwd	2FA
Update card	—	90%	15%	10%
Invoices	90%	—	20%	10%
Reset pwd	15%	20%	—	85%
2FA	10%	10%	85%	—

Read it as a heatmap. Two obvious clusters jump out: Update card + Invoices (90% — a "Billing" group) and Reset password + 2FA (85% — a "Security/Login" group). The cross-pairs (Update card × Reset pwd = 15%) are low, confirming the two clusters are genuinely distinct. That's your IA skeleton, straight from the data.

How to read thresholds:

70%+ → a strong pairing; these cards belong together.
40–69% → a soft pairing; context-dependent, worth a look.
Under 40% → these cards are seen as unrelated.

2. Suggested clusters / dendrogram

A dendrogram (or agreement-based cluster list) turns the matrix into candidate groups by linking cards that exceed an agreement threshold. Picture a tree: cards that were grouped together most often join at the bottom (high agreement); weaker links join higher up. Where you "cut" the dendrogram determines how many categories you get.

Cut low (e.g., only merge pairs above 70%) → many small, tight groups.
Cut high (merge above 40%) → fewer, broader groups.

The dendrogram answers "if I had to draw category boundaries, where would they naturally fall?" Use it as a proposal, then apply judgment — it's a starting point, not a verdict.

3. Category labels (open sorts)

In an open sort, the names participants give their groups are gold. Cluster the label language: if twelve of twenty people call the billing cluster "Billing" or "Payments," you've found both the category and its label — in users' own words. This is the part spreadsheets miss and that most teams under-use.

The standardization step (don't skip)

In open sorts, people name the same group differently — "Payments," "Billing," "Money stuff," "$$." Before you count labels, standardize them into canonical buckets so your label analysis isn't fragmented across synonyms. Then the frequency count becomes meaningful ("14/20 named this group some variant of Billing").

Turning analysis into an IA (repeatable process)

Start from the similarity matrix — identify the tight clusters (high mutual agreement).
Cross-check with the dendrogram / suggested clusters to see where boundaries naturally fall and how many categories the data supports.
Name each group using participants' own standardized label language from open sorts.
List the ambiguous cards — items that landed in many different groups. For each, decide: relabel it, place it deliberately, or make it a cross-link / persistent element.
Draft the IA and validate it with a tree test before building.

Reading ambiguous cards (this is where the insight is)

Cards that scatter across many groups aren't noise — they're your clearest signal that something is unclear. Two common patterns:

A card split evenly between two groups (e.g., "Data privacy" — 45% Account, 40% Legal) → the item genuinely belongs to two mental models. Solution: place it in the primary and cross-link from the other.
A card that lands everywhere (e.g., "Contact support" — no group above 25%) → it isn't a topic at all. It's a fallback action. Solution: pull it out of the taxonomy and make it a persistent element (a header link, a floating button).

How many participants do you need for reliable analysis?

Card sorting stabilizes faster than you'd expect. 15–20 participants per audience typically surfaces the major clusters; the similarity percentages keep sharpening as you add more, but the big groups rarely change after ~15. Go to 30 for diverse or high-stakes audiences. If your matrix is full of 40–60% values with no clear clusters, that's usually a content problem (genuinely ambiguous items), not a sample-size problem.

Common analysis mistakes

Eyeballing raw groupings instead of using the similarity matrix — you'll over-weight a few vocal participants.
Treating the dendrogram as the answer — it's a proposal; where you cut it is a judgment call, and ambiguous cards still need decisions.
Ignoring label language — the names users choose are half the value of an open sort, and the half most teams throw away.
Skipping standardization — un-merged synonyms fragment your label counts and hide the real consensus.
Skipping validation — always tree-test the resulting IA. Analysis produces a hypothesis; the tree test is the experiment.

Analyze a card sort free

ResearchRocket's card sorting includes a cross-participant similarity matrix (shaded by agreement) and agreement-based suggested clusters automatically — no exporting to a spreadsheet, no manual matrix-building. It's part of an all-in-one platform, so you can card sort, tree test, and analyze in one place. Try the free card sorting tool.

FAQ

How do you analyze a card sort? Start with the similarity matrix (pairwise agreement between cards), confirm boundaries with the dendrogram / suggested clusters, standardize and count participants' group labels from open sorts, flag ambiguous cards for deliberate decisions, then validate the resulting IA with a tree test.

What is a similarity matrix in card sorting? A grid showing, for each pair of cards, the percentage of participants who grouped them together — the clearest, least subjective view of what belongs with what. Values above ~70% indicate strong pairings.

What is a dendrogram in card sorting? A tree diagram that links cards by how often they were grouped together; cards with high agreement join low on the tree. Where you "cut" it determines how many categories your IA has.

How many participants do I need for a card sort? Around 15–20 per audience surfaces the major patterns; 30 for diverse audiences. More participants mainly sharpen the agreement percentages rather than changing the big clusters.

What do I do with cards that don't fit any group? If a card splits evenly between two groups, place it in the primary and cross-link. If it scatters everywhere, it's probably an action (like "Contact support"), not a topic — pull it out as a persistent element.

Do I have to build the similarity matrix by hand? No — a good card sorting tool generates it automatically. ResearchRocket includes the similarity matrix and suggested clusters built in.

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

Analyze a card sort free on ResearchRocket →