By Eric, BlinkBench Founder · Last reviewed: July 2026
The reaction time test, the aim trainer, the typing test and the chimp test all say the same sentence somewhere on their page: reported as a median, not a mean. It looks like a small statistical preference. It is actually load-bearing, and the reason is visible in five numbers.
A worked example
Five reaction-time trials, in milliseconds, in the order they happened:
241, 253, 902, 238, 262
Four of those are an ordinary fast reaction. One — the phone buzzed, the eyes drifted, whatever it was — took 902 ms. That is not a slow reaction. It is a lapse: a different event from the other four, mixed into the same set of five numbers.
The mean of all five is 379 ms. Take the same five numbers, sort them, and take the middle one — the median — and it is 253 ms. Drop the lapse entirely and the remaining four average 249 ms. The median of all five, lapse included, sits within a couple of milliseconds of the four-trial average with the lapse thrown out. That is the whole argument. The median arrived at roughly the same place as manually discarding the bad trial — without anyone having to decide which trial was bad.
| Sorted trials (ms) | Statistic | Value |
|---|---|---|
| 238, 241, 253, 262, 902 | Mean of all 5 | 379 ms |
| Median of all 5 | 253 ms | |
| Mean with the lapse dropped | 249 ms |
Why this isn't special pleading
A mean treats every trial as equally informative and adds them up. That is the right tool when the thing you're measuring is symmetric — errors above the true value and below it are equally likely and roughly equal in size. A quick reaction can only be so much quicker than average; the floor is your nervous system. A slow one can be arbitrarily slower, because there is no ceiling on how distracted a person can be for one trial out of five. That asymmetry is what “right-skewed” means, and it is why one lapse can move a five-trial mean by tens of milliseconds while the other four trials, which are the actual measurement, sit still.
The median is not immune to a bad run — enough lapses and it moves too. What it resists is exactly one outlier dominating a small set, which is the specific failure mode a five-to-twelve-trial test is most exposed to.
The part that actually matters: showing the trials
A median has a cost the mean doesn't: it throws information away. Two runs with very different spreads can report the identical median. That is exactly why every timed test on this site prints the individual trials next to the result rather than only the aggregate — the number tells you the middle, the list tells you whether that middle was a tight cluster or one lucky trial surrounded by noise. A site that reports only the median without the trials is making the same move it criticizes elsewhere: showing you a summary and asking you to trust the part you can't see.
Where else this shows up
Aim trainer reports median throughput across scored moves for the same reason, and separately shows the raw median movement time labelled as not comparable across window sizes — two different numbers, kept apart on purpose. Typing speed takes the median of five sentences rather than the mean, because one stumbled sentence is exactly the same shape of problem as one lapsed reaction-time trial, and shows no score at all if fewer than three sentences come back valid — a median of two is not a measurement. Chimp test reports the median study time per numeral across rounds, for the same reason it reports a median anywhere: study time is right-skewed too.
The pattern generalizes past this site. Any time you see a browser test boast a mean reaction time under 200 ms, ask what happened to the trial where the tab lost focus for a second. A mean has no way to tell you. A median mostly shrugs it off, and showing the trials tells you whether it had to.