For one round of qualitative testing on a single group of users, five participants is a genuinely useful number. Research popularised by Jakob Nielsen found that testing with five people uncovers about 85 percent of a product’s usability problems, with each additional person adding less new information than the last. That number is a rule of thumb built for one specific job. It stops applying the moment your test changes shape, and a lot of teams reach for it without knowing which job it was built for.
Where does the five-user number come from?
In a widely cited 2000 piece, Nielsen wrote up findings based on earlier work with Thomas Landauer: across a set of studies, the probability that any one user would stumble on a given usability problem averaged around 31 percent. Run the math on that detection rate and a curve falls out. Your first user finds roughly a third of the problems. Your second finds most of what’s left, plus a few new ones. By the time you reach a fifth person, you’re mostly re-hearing complaints you already logged, and the return on a sixth or seventh tester keeps shrinking.
That’s the whole mechanism. It isn’t a claim about human behaviour in general. It’s an observation about how fast a specific kind of signal saturates when everyone in the room shares roughly the same context.
What “usability problem” means in that number
The 85 percent figure is about surfacing friction fast, in one round, in one homogeneous group. It answers “what’s broken and how badly,” not “how many of our users hit this” or “did the redesign actually move the number.” Nielsen’s own follow-up work is explicit about the difference: five participants are enough for a formative, qualitative round meant to catch obvious problems before you ship a fix. They are not enough to measure anything you’d want to put in a report with a confidence interval attached.
Five users isn’t a magic number. It’s a shortcut for one job: catching the loudest problems fast, in one round, with one kind of user.
When five users isn’t enough
Three situations break the assumption underneath the rule.
Multiple distinct user segments. If new signups and power users hit your product differently, testing five people drawn from a blend of both groups gives you an average of two half-formed pictures. Nielsen’s own guidance here is to treat each segment as its own five-user round, not to stretch one round across both.
Quantitative or benchmarking work. The moment you want a number you can defend, “we improved task completion from 62 percent to 81 percent,” you’ve left the five-user rule’s territory entirely. Statistical confidence needs a much larger sample, often several dozen participants depending on the effect size you’re trying to detect, because you’re now measuring a rate rather than spotting a pattern.
Decisions with real cost attached to being wrong. A checkout redesign that fails quietly costs more than a five-person round saved you. The size of your sample should scale with what a wrong call actually costs, not with habit.
A practical way to decide your own number
Start with the decision the round needs to move, not with a target headcount. If you’re trying to catch obvious friction before a launch, five people who genuinely look like your real users is a defensible starting point. Watch for the saturation signal directly: if your fourth and fifth participants are repeating complaints your second and third already raised, that round is done, whatever number you originally had in mind.
Then plan for another round rather than a bigger one. A second small test after you ship a fix tells you whether the fix actually worked, which a single oversized round never will, because it can’t observe the “after.”
Whichever number you land on, the round only pays off if you run it. Pick people who look like your real users, decide what this specific round needs to move, and put the question in front of them this week. That’s the whole job: Fiuto exists to make that round quick to set up, drafting a study from what you want to learn and sending it straight to your own users rather than a recruited panel.
