A percentile is a position in a queue, not a score. If a result puts you at the 71st percentile, roughly 71 percent of the measurements it was compared against fall below yours in value. The number says almost nothing about the quality of your result on its own terms, and almost everything about the company it is keeping.
Which end of that queue is the good end depends on what is being measured, and this trips people up constantly. On a quiz score, higher is better, so a high percentile is a good result. On a reaction time, lower is better, so the good result is a low percentile. Speed tests commonly flip it round and report how many people you beat instead, which is why the same measurement can be described as the 17th percentile or as faster than 83 percent without either being wrong. Read the units before you read the rank.
The US National Institute of Standards and Technology, in its e-Handbook of Statistical Methods, defines it from the sorted data: the pth percentile is a value such that at most 100p percent of the measurements are less than it, and at most 100 times (1 minus p) percent are greater. "Percentiles split a set of ordered data into hundredths", the handbook says, and "The 50th percentile is called the median."
That is the whole idea. Everything difficult about percentiles comes from what happens next. There is no single agreed way to compute one from a sample. Reaction times are not shaped the way everyday statistics assume. A best-of-several score is not a measurement of ability at all. And the group you are being ranked against was almost certainly not a random slice of anybody.
Key Takeaways
- A percentile is a rank among other measurements. A percentage is a score against a fixed maximum. The same percentage can be a top or a bottom percentile depending on who else is in the group, and on whether high or low is the good end.
- There is no single definition of a sample percentile. Hyndman and Fan catalogued nine in common use in 1996, and they still disagree, most sharply on small samples and near the tails.
- Reaction-time distributions have a long right tail, so the mean sits above the median and roughly six attempts in ten come in faster than the average.
- The best of several attempts is biased by construction. The fastest of n tries sits on average at the 100/(n + 1) percentile of your own distribution, so best-of-twenty flatters you far more than best-of-five.
- A percentile from a self-selected online sample is a rank among the people who chose to show up, which is not the same group as everyone.
A percentile is a rank, and a percentage is a score
The two words look alike and measure completely different things, which is why the confusion is so persistent.
A percentage is absolute. Getting 88 out of 100 is 88 percent whether nobody else took the test or a million people did. A percentile is relative. It exists only once there is a group to rank against, and it moves when the group moves, even though your own performance has not changed by a hair.
Two ten-person groups make the gap obvious. Score 88 while the other nine score 91, 92, 93, 94, 95, 96, 97, 98 and 99, and you have beaten none of them: a high percentage, the bottom of the ranking. Now score 60 while the other nine score 30, 33, 35, 38, 40, 41, 44, 47 and 51, and you have beaten all nine: a much lower percentage, the top of the ranking.
Nothing about your work improved between those two scenarios. Only the comparison group changed. So when a test tells you that you are faster than 63 percent of everyone, the sentence is a statement about them at least as much as about you, and a percentile can never be better than the set it was drawn from.
Why nine different percentile definitions give nine different answers
Ask a computer for the 90th percentile of a set of numbers and there is no single right answer to hand back. This surprises people who work with data for a living.
The problem is that a sample almost never has a data point sitting exactly on the boundary. With ten measurements the 90th percentile falls somewhere between the ninth and tenth values, and something has to decide where. Different traditions decided differently, and the differences were baked into different software.
Rob Hyndman and Yanan Fan wrote the standard survey of this in The American Statistician in 1996. Its abstract opens bluntly: "There are a large number of different definitions used for sample quantiles in statistical computer packages." They rewrote all of them in one common notation, found nine, tested each against six desirable properties, and argued "that there is a need to adopt a standard definition for sample quantiles so that the same answers are produced by different packages and within each package." Thirty years on, that has not happened. The nine survive as the nine types of R's quantile function, and NIST's own handbook notes drily that "Although this has not in fact happened, the article does provide a useful summary and evaluation of various methods for computing percentiles."
Take ten plausible reaction times, in milliseconds, as a worked example: 211, 224, 231, 238, 246, 255, 269, 288, 317 and 402. Their 90th percentile, under each of the nine definitions in turn:
| Type |
How it places the kth of the sorted values |
90th percentile |
| 1 |
Inverse of the empirical distribution function |
317.00 ms |
| 2 |
As type 1, but averaging at the jumps |
359.50 ms |
| 3 |
Nearest even order statistic |
317.00 ms |
| 4 |
Linear interpolation of the empirical distribution, at k/n |
317.00 ms |
| 5 |
Midway through each step, at (k - 1/2)/n |
359.50 ms |
| 6 |
At k/(n + 1), the Weibull and Gumbel position, used by Minitab and SPSS |
393.50 ms |
| 7 |
At (k - 1)/(n - 1), the default in R, and the method Excel uses |
325.50 ms |
| 8 |
At (k - 1/3)/(n + 1/3), median-unbiased |
370.83 ms |
| 9 |
At (k - 3/8)/(n + 1/4), tuned for normal data |
368.00 ms |
The answers run from 317.00 ms to 393.50 ms: a spread of 76.50 ms across the same ten numbers, produced entirely by a choice of convention that no result screen anywhere mentions. That is about 30 percent of the middle of the sample, which the same nine definitions place somewhere between 246.00 ms and 250.50 ms.
This is not a fringe worry. NIST works the same problem on twelve silicon-wafer resistivity measurements and gets 95.1981, 95.1972 and 95.1957 for the 90th percentile under three of the nine, noting that "Most statistical and spreadsheet software use one of the methods described in Hyndman and Fan." Recomputing all three from the handbook's own twelve values reproduces those figures exactly.
Why the middle of a distribution is steadier than its edges
The nine do not disagree equally everywhere, and knowing where they diverge tells you which numbers to trust.
On those same ten reaction times, the nine span 4.50 ms at the median, 8.75 ms at the 25th percentile, 16.75 ms at the 75th, and 76.50 ms at the 90th. The middle is nearly unanimous; the tail is a free-for-all. That ordering is not a coincidence. Near the middle, data points crowd both sides of the boundary and every interpolation rule lands in much the same place. Out at the 90th percentile of ten values there is one enormous gap, and each rule picks a different point inside it. Push further still and the spread narrows again, to 42.50 ms at the 95th, but only because ten values leave the nine rules nowhere to land except between the top two.
Sample size fixes it, quickly. Drawing 2,000 independent samples from one fixed right-skewed shape and asking all nine for the 90th percentile each time, the average disagreement falls from 43.96 ms at ten values to 4.39 ms at a hundred, 0.44 ms at a thousand and 0.05 ms at ten thousand. Ten values is not just wide but wildly unstable, with the middle four-fifths of those samples landing anywhere between about 5 ms and about 97 ms. By a thousand observations it has settled to under a fifth of a percent of the distribution's median.
So a percentile quoted from a large reference set is stable, and the arithmetic behind it barely matters. One quoted from a handful of measurements, especially an extreme one, is soft in a way the decimal places conceal. That applies to your own five goes at a reaction time test exactly as it applies to a laboratory sample.
Why an average reaction time is not the typical reaction time
Most statistical intuition is trained on the bell curve, where the mean, the median and the peak sit in the same place. Reaction times are not like that, and neither is any other measurement with a hard floor and no ceiling.
NIST puts the general case simply: "Some measurements have a lower bound and are skewed right", giving reliability testing as its example, since "failure times cannot be negative." Reaction time has exactly that shape. Robert Whelan's 2008 review in The Psychological Record, the standard reference here, reports Luce's finding that a genuine reaction time has a floor of at least 100 ms, the time the body needs for perception and for the motor response. There is no ceiling at all, because attention can lapse for any length of time.
Whelan describes the resulting shape directly. Response-time distributions "are not Gaussian (normal) distributions but rather rise rapidly on the left and have a long positive tail on the right", and the consequence follows: "the mean is not reflective of the typical response if the distribution is skewed, because the mean is distorted in the direction of the skew."
How far? Whelan reports that reaction times are well described by the ex-Gaussian shape, and quotes fitted parameters from Hervey and colleagues (2006), a study of children with attention deficit hyperactivity disorder: for the control group, a normal component averaging 319 ms and an exponential tail averaging 144 ms. The mean of such a distribution is exactly the sum of the two, 463 ms. Simulating four million draws from that shape puts the median between 420 ms and 434 ms. That band comes from sweeping the width of the normal component from 20 ms to 80 ms, because the source does not report it. Either way the mean lands 28 ms to 43 ms above the middle, and roughly six attempts in ten come in faster than the average, between 58 and 63 percent of them depending on the width assumed.
That is why a percentile beats an average here. Faster than average is a weaker claim than it sounds when most attempts already beat it. The median is beaten by exactly half of them, by definition, which is why Whelan notes that "many researchers report the median RT as a central tendency parameter, because it is less susceptible to departures from normality".
The median has its own catch. Whelan points out that "unlike the sample mean, it is a biased estimator of the population median when the population is skewed", and that "the bias becomes more extreme as the sample size becomes smaller". Five attempts is a small sample; read your median as an estimate rather than a verdict. The step before it is unsettled too. Ivan Vankov's 2023 paper in Frontiers in Psychology notes that there is no overall agreement on how to deal with outliers in response-time data, and shows by simulation that the choice between the available methods can inflate p-values and distort effect sizes.
Why your best-ever score is not a measurement of your ability
Every leaderboard on the internet reports a best, and a best is the one summary statistic guaranteed to improve the longer you sit there.
The maths is exact and does not depend on the shape of the distribution, as long as no two attempts land on exactly the same number. Hyndman and Fan's sixth definition rests on the fact that for n independent draws the kth smallest sits, on average, at the k/(n + 1) point of the underlying distribution, and they state the consequence for the extremes directly: the chance that a fresh draw beats the smallest of n, or is beaten by the largest, is 1/(n + 1) either way. Read that for a personal best. The fastest of n attempts sits, on average, at the 100/(n + 1) percentile of your own distribution, meaning the whole population of attempts you would make if you kept going.
One attempt lands at the 50th percentile of that distribution, which is what typical means. Best of three lands at the 25th. Best of five lands at the 16.67th, faster than 83.33 percent of the attempts you would make. Best of twenty lands at the 4.76th, and best of a hundred at the 0.99th. Simulating 400,000 trials from three unrelated shapes, a lognormal, an ex-Gaussian and a flat uniform, gives mean percentiles of 16.64, 16.67 and 16.68 for the best of five against the predicted 16.67, so the result really is indifferent to the underlying shape.
In raw units the drift is large. Taking the same children's control-group shape used earlier, whose single-attempt average is 463 ms, the best of five averages 327 ms, the best of twenty 276 ms and the best of a hundred 236 ms. Nothing about the simulated player changed. Only the number of attempts did, which is worth remembering before you compare your best on a click speed test with a stranger's.
A percentile from an online test is a rank among the people who showed up
The last thing a percentile inherits is the way its reference group was assembled, and no amount of arithmetic repairs that.
Statistics has a precise term for what an open web test collects. Jelke Bethlehem, writing in the International Statistical Review in 2010, defines it: "Self-selection means that it is completely left to individuals to select themselves for the survey." The questionnaire, or the game, is simply put on the web, and the respondents are whoever happens to find it and decides to take part. "The survey researcher is not in control over the selection process."
Bethlehem derives what that costs. The bias in an estimate from a self-selected sample works out to the covariance between the thing being measured and each person's probability of taking part, divided by the average probability of taking part. It vanishes in only three situations: everyone is equally likely to participate, everyone has the same value of the measured thing, or participation is unrelated to that value. His conclusion for web surveys generally is that "self-selection leads to unreliable survey outcomes."
Apply that to any test that ranks you against its own takers. The people who seek out a speed test and finish it are likely to be those who expect to do well, or who are curious enough to try repeatedly. Participation is then correlated with the very quantity being measured, and that correlation closes the third of Bethlehem's escape routes, the one that requires participation to be unrelated to the value. The resulting distribution is real and it describes real people. It is not the general population, so a percentile drawn from it should be read as a rank among the people who took that test, never as a rank among humans.
A ranking can also be produced without ranking you against anybody present, by fitting a curve to published norms instead. That moves the problem rather than solving it, because those norms were themselves collected from somebody, and the same question applies to whoever that was.
That is not a reason to ignore the number. It is a rank in a named crowd rather than a fact about the species, which is a smaller claim and a true one. It is also why tracking your own results over time, on your own equipment, tells you more than any cross-site comparison, a point the reflex test hub makes about hardware for the same reason.
Frequently asked questions
What does a percentile mean in a test result?
It means your rank within the group you were compared against. Being at the 71st percentile means roughly 71 percent of the measurements in that reference group fall below yours in value, which is a good result where higher is better and a poor one where lower is better, as on a reaction time. NIST's e-Handbook defines the pth percentile as a value such that at most 100p percent of the measurements are less than it, and notes that percentiles split ordered data into hundredths, with the 50th percentile being the median. Speed tests commonly invert the figure and report how many people you beat. Either way it describes your position, not the quality of your score in absolute terms.
What is the difference between a percentile and a percentage?
A percentage is an absolute score against a fixed maximum. A percentile is a relative rank against other people's results. Score 88 in a group whose other nine members scored 91 through 99 and you have beaten nobody, despite the high percentage. Score 60 in a group whose others scored 30 through 51 and you have beaten all nine. The percentage is a property of your answer sheet; the percentile is a property of the room.
Why do two websites give different percentiles for the same score?
Two reasons, and both are real. First, they are ranking you against different groups of people, and a percentile is meaningless without its reference set. Second, there is no single agreed formula: Hyndman and Fan catalogued nine sample quantile definitions in 1996 and no standard was ever adopted, so identical data can yield different percentiles. On ten reaction times the nine definitions put the 90th percentile anywhere from 317.00 ms to 393.50 ms.
Should I use my average or my best score?
Neither on its own. The average is pulled upward by the long right tail of slow attempts, so on a realistic reaction-time shape roughly six attempts in ten beat it. The best of several is biased the other way and improves the more you try: the fastest of n attempts sits on average at the 100/(n + 1) percentile of your own distribution, so best-of-twenty looks far better than best-of-five with no change in ability. The median of a run of attempts is the most honest single number.
Is an online test percentile the same as a percentile for the general population?
No. Open web tests recruit by self-selection, which Bethlehem describes as leaving it entirely to individuals to select themselves, with the researcher not in control of the selection process. People who seek out a speed test are unlikely to be a random slice of humanity, and participation is likely to correlate with the thing being measured, which is precisely the condition under which the bias fails to vanish. Read the result as a rank among the people who took that test.
Sources
- NIST/SEMATECH, e-Handbook of Statistical Methods, "Percentiles" - the definition of percentiles and order statistics, the worked twelve-measurement example, and the note that no single interpolation method is standard. Retrieved 13 August 2026, https://www.itl.nist.gov/div898/handbook/prc/section2/prc262.htm
- NIST/SEMATECH, e-Handbook of Statistical Methods, "Measures of Skewness and Kurtosis" - right skew, and measurements with a lower bound. Retrieved 13 August 2026, https://www.itl.nist.gov/div898/handbook/eda/section3/eda35b.htm
- Hyndman, R. J. and Fan, Y., Sample Quantiles in Statistical Packages, The American Statistician, volume 50, number 4, pages 361 to 365 (1996) - the nine definitions, the six properties, and the case for standardising. Retrieved 13 August 2026, https://robjhyndman.com/papers/sample_quantiles.pdf
- R Core Team, R documentation for
quantile - the nine Hyndman and Fan types as implemented, and which packages use which. Retrieved 13 August 2026, https://stat.ethz.ch/R-manual/R-devel/library/stats/html/quantile.html
- Whelan, R., Effective Analysis of Reaction Time Data, The Psychological Record, volume 58, pages 475 to 482 (2008) - the shape of response-time distributions, the mean under skew, and the median's own small-sample bias. Retrieved 13 August 2026, https://link.springer.com/article/10.1007/BF03395630
- Bethlehem, J., Selection Bias in Web Surveys, International Statistical Review, volume 78, number 2, pages 161 to 188 (2010) - self-selection, the bias of the sample mean, and the three conditions under which it vanishes. Publisher of record https://doi.org/10.1111/j.1751-5823.2010.00112.x ; the full text read here was the University of Amsterdam repository copy, retrieved 13 August 2026.
- Vankov, I. I., The hazards of dealing with response time outliers, Frontiers in Psychology, volume 14 (2023) - on the absence of agreement about handling response-time outliers. Retrieved 13 August 2026, https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2023.1220281/full
