How many participants do you need for a card sort?

It is the first question everybody asks, and the honest answer is smaller than most people expect.

The short answer

About 15 for a directional read. 20 to 30 for a firm one. Rarely more than 30.

That is not a rule of thumb somebody invented. It comes from two pieces of work that have held up for two decades.

Nielsen Norman Group's guidance puts 15 participants at roughly a 0.90 correlation with the result you would get from a far larger group. Tullis and Wood tested this directly in 2004, running a card sort with 168 people and then sampling smaller groups out of it. At 20 to 30 participants the correlation rises to somewhere around 0.93 to 0.98, and then it flattens. Adding people past that point buys you decimal places, not decisions.

Why so few?

Because a card sort is not a survey. You are not measuring how many people hold an opinion, you are looking for structure, and structure emerges fast.

If eleven of your first fifteen people put "free returns" and "easy size exchange" in the same pile, the twelfth, twentieth and fiftieth are very unlikely to change that. The pairs that people agree on stabilise early. The pairs they disagree on tend to keep disagreeing, and more participants just makes the disagreement more precisely measured rather than resolving it.

This is the opposite of the intuition most people bring from opinion polling, where the margin of error is the whole game.

When to go to the top of the range

Push toward 30, or run separate sorts, when:

You have distinct audiences you expect to think differently. New customers and long-standing ones often group things differently. If you plan to compare them, each group needs its own defensible sample, not 15 split into two halves of seven.
The list is long. A 60-card deck asks more of each person and produces more variation in how far they get. More participants smooths that out.
The decision is expensive or hard to reverse. If a navigation rebuild hangs on the result, the extra ten people are cheap insurance.

When 15 is plenty

You are exploring. You want to know what themes exist at all, before committing to anything.
You are sanity-checking a structure your team already drafted. A closed sort with 15 people will tell you very quickly if a category name is not landing.
You are iterating. Two rounds of 15 with a change in between teaches you more than one round of 30.

How to tell when you have enough

Do not guess. Watch the agreement.

In SortedResearch the overview shows how strongly participants agreed, and the dendrogram reports an average within-group agreement percentage that updates as you change the number of groups. When you add five more people and that figure barely moves, and the clusters keep the same shape, you have your answer. That is a better stopping signal than any target number, because it responds to your actual list rather than to an average of everybody else's.

What about people who rush it?

Sample size is a quantity question. It does not protect you from a bad response, and a rushed sort can distort a small sample more than a large one.

Every response in SortedResearch is checked for two behaviours that give away low effort: placing cards faster than they could physically be read, and "grab and dump", where somebody drops cards into piles in exactly the order they were presented without ever reorganising. Flagged responses are marked but never dropped for you. You decide whether to include them, and you can see the result both ways with a checkbox.

If you bought your audience through us, a rejected response is sent back to the panel and replaced, so it does not count toward your total or your bill.

The practical answer

Start with 20. Look at the agreement figures when they land. If the clusters are clear and stable, you are finished. If two themes are fighting over the same cards, another ten people will usually settle it, and if they do not, that disagreement is itself the finding: your customers genuinely do not agree on where those things belong, and no amount of extra sampling will invent a consensus that is not there.