Quantitative Card Sorting Sample Calculator

Use it when the card sort is meant to validate an architecture, not to discover one:

You will analyze with a similarity matrix, a dendrogram or agreement percentages, not by reading groupings one by one

The resulting structure has to be defended to someone who asks for evidence

You are comparing two candidate taxonomies and need to know which one groups better

When not to use it: for an exploratory card sort, where the goal is to understand people's vocabulary and mental model, Jakob Nielsen recommends 15 participants (Nielsen, 2004) and that figure still holds. A well-moderated qualitative sort with 15 people will tell you more about why they group that way than 30 unmoderated ones.

Tom Tullis and Larry Wood (Tullis & Wood, 2004) compared the structures obtained from different subsamples against a full study of 168 participants. Correlation with the final structure:

15 participants: 0.90 correlation

20 participants: around 0.93

30 participants: 0.95 correlation

Nielsen read the same data and concluded that 15 are enough for most projects, because returns drop off quickly past that point. It is not a contradiction: it answers a different question.

This calculator starts at 30 because the use case differs. 15 are enough to guide an architecture decision; 30 support the statistical analysis when the structure will be justified with figures.

Worth saying plainly: the gap between 0.90 and 0.95 is small. If your budget does not reach 30, a card sort with 15 participants is still a useful study. You lose precision, not validity.

The calculator solves n = min(30 * card_factor * group_factor, 100) and then adds 10% for no-shows.

Card factor: the number of cards divided by 50. Below 50 cards the factor is 1 and changes nothing.

Group factor: expected groups divided by 5. Below 5 groups the factor is 1.

A ceiling of 100 participants: deliberate. Past that point recruiting cost grows faster than the stability you gain. If the calculation goes over 80, the calculator warns that simplifying the study beats enlarging the sample.

These two factors do not come from the Tullis and Wood study, which worked with a fixed card set. They are a practical adjustment with simple reasoning: more cards and more groups mean more possible pairs, and a sparser similarity matrix needs more observations to stabilize.

Open or closed changes the effort, not the sample. In a closed sort you supply the categories and the task is faster; in an open sort participants name them and the real vocabulary surfaces, at the cost of longer sessions and slower analysis.

Long sets cause fatigue. Past 60 cards the quality of the last groupings tends to drop: people start sorting by elimination. If your inventory is large, splitting it into two studies yields better data than stretching one.

It runs unmoderated. At this scale quantitative card sorting is executed with remote tools. That makes recruiting cheaper, but you lose the why: if you can, moderate a handful of sessions separately to interpret the odd groupings.

Watch the panel bias. Recruiting only among frequent users produces an architecture that works for people who already know the product. For a public taxonomy, the profile that matters is usually the first-time visitor.