Your average support reply went out in two hours last month, a number you are quietly proud of. But hidden inside that tidy average is a group of customers who waited a full day, wrote three angry follow ups, and then left without a word. The average did not just miss them, it hid them, folding their misery into a comfortable middle number that made everything look fine.
I run my own small businesses, and I have been reassured by a gentle average right up until the moment it fell apart. Percentiles are the tool that pulls those buried customers back into the light. An average tells you about a middle that may not even exist. A percentile tells you about the people your worst experiences actually happen to, and the best part is that it assumes nothing at all about the shape of your data.
What a percentile actually is
A percentile is the value below which a given share of your data falls. The ninetieth percentile is simply the number that ninety percent of your values sit below, with the top ten percent stretching out above it. If the ninetieth percentile of your delivery time is six days, then nine in ten parcels arrive within six days and one in ten takes longer.
It is a plain statement about position, nothing more. You line every value up from smallest to largest, walk ninety percent of the way along that line, and read off the number standing there. That number is your ninetieth percentile, and it took no clever maths to find.
The median is a percentile you already know
You have already met the most famous percentile without knowing its other name. The median, the true middle of your data with half the values above it and half below, is nothing but the fiftieth percentile.
That is worth holding onto, because it means the median, the ninetieth percentile, and the ninety fifth are all members of one family, all read the same way, all just positions in a sorted line. Once you see the median as the fiftieth percentile, the rest stop feeling like intimidating statistics and start feeling like a ruler you can lay across any set of numbers you own.
Why they beat the average on messy data
Here is the quiet superpower. Because a percentile only cares about position in the line, a single monstrous value cannot drag it around. One customer who spends ten thousand dollars will yank your average order value upward and make the whole business look richer than it really is. But your median order, your fiftieth percentile, barely moves, because that one whale is still just a single position at the far end of the line.
On skewed data, where a long tail hauls the average away from where most people actually live, the percentiles keep quietly telling you the honest truth about the typical case.
No bell shape required
The two thirds rule, the one where most values sit within one standard deviation of the average, only works when your data piles up in a rough bell shape. An enormous amount of business data does not. Income, order sizes, page load times, and days to pay an invoice all clump low and then trail off into a long high tail.
Percentiles do not care in the slightest. They make no assumption about the shape at all. They simply read whatever your real data does, which is exactly why they are the safer tool for the messy numbers a small business actually produces.
The tail is where the pain lives
Customers do not churn over their average experience, they churn over their worst one. The person who usually gets a one hour reply but once waited two days remembers the two days. The shopper whose page usually loads fast but once hung for ten seconds remembers the hang.
Your average smooths all of that away into a number nobody actually experienced. The ninety fifth percentile walks you straight to the unlucky one in twenty whose bad day is quietly deciding whether they ever come back.
A founder sized example
Treat these as illustration, not benchmarks to copy. Say your average page load is one and a half seconds, which sounds perfectly healthy, so you move on. But you pull the ninety fifth percentile and it is eight seconds.
| Measure | Value | What it tells you |
|---|---|---|
| Average load | 1.5s | Looks perfectly healthy; you move on |
| Median (p50) | 1.2s | The typical visitor waits about a second |
| p95 | 8s | One in twenty visitors stares at a nearly dead page for eight seconds |
The average told you everything was fine. The percentile told you where a real chunk of your visitors were actually suffering the whole time.
Percentiles read your money too
This is not only for speed and support, it reads your revenue just as well. The fiftieth percentile of your order values tells you what a genuinely typical order looks like, unbothered by the handful of giant ones. The ninetieth tells you roughly where your big spenders begin.
The gap between your median order and your average order is itself a signal. When the average sits well above the median, you know a small number of large orders are stretching the top, and that your typical customer is a good deal more modest than the headline average ever admitted.
Quartiles and the middle fifty
There is a tidy way to describe spread using nothing but percentiles, and it is the honest cousin of the standard deviation. Split your data at the twenty fifth, the fiftieth, and the seventy fifth percentiles, the quartiles, and the distance between the twenty fifth and the seventy fifth is called the interquartile range.
It captures the middle half of your customers while completely ignoring the wild values at both ends. So when your data is skewed and a standard deviation would be inflated by a couple of freak orders, the interquartile range gives you a measure of spread that keeps its head, because the extremes simply are not invited.
If you want the fastest snapshot of any column, ask for five percentiles at once: the smallest value, the twenty fifth, the fiftieth, the seventy fifth, and the largest. That set is often called the five number summary, and some tools draw it as a simple box with whiskers so you can read the shape at a glance.
Percentiles make honest promises
Once you start thinking in percentiles, the way you set targets changes. An average is a weak promise, because you can hit an average reply time of one hour while a fifth of your customers wait far longer, and you would never know it from the number.
A percentile target is a real promise to nearly everyone. Saying that ninety five percent of replies go out within two hours is a commitment about almost the whole crowd, not a blurry claim about a middle that some invisible tail is dragging around. The good teams I learn from set their targets on a high percentile for exactly this reason.
You need enough data for a tail
A high percentile needs a decent pile of data underneath it to mean anything at all. If you have only fifteen orders, a ninety fifth percentile is basically one lucky or unlucky value wearing a fancy label, and it will jump around wildly from week to week.
The median holds up on thin data far better than the extreme percentiles do. So on small samples, lean on the fiftieth, treat the ninetieth as a rough hint, and be honestly suspicious of a ninety ninth percentile until you have thousands of events feeding it.
The mistakes to sidestep
A few traps catch people again and again. The first is reporting only the average and letting the tail stay hidden, which is the entire problem we started with. The second is computing a dramatic high percentile on a tiny sample and then treating that one jumpy value as gospel. The third is forgetting which direction good points in, because a high percentile of order value is wonderful, while a high percentile of wait time is exactly the pain you are trying to kill.
Getting these numbers is genuinely easy. Every spreadsheet has a percentile function and a quartile function that turn a column into any percentile you ask for in a single line, and most analytics tools will show you the fiftieth, ninetieth, and ninety fifth beside the average once you go looking for the setting. The skill is not the calculation. It is the habit of never again reading an average without reading the median and at least one tail percentile next to it.
The honest limit
Hold the whole idea in proportion. A percentile describes where the pain is, it does not explain why the tail exists, and it will not lift a finger to fix it. On thin data the extreme ones wobble far too much to trust. All a percentile really does is refuse to let a few values, high or low, lie to you about the typical experience, and put the unlucky tail somewhere you can finally see it.
A number that shows you the customer having the worst day is worth far more than an average that pretends that customer was never there. Every metric and method like this one is explained in plain English at dataresearchanalysiscollection.com, so you can read it again slowly with your own numbers open. No hype, no promises about your results, just the idea explained until you can see the customers your comfortable averages have been hiding all along.
Get new guides and videos first — join the Telegram channel.