I spent about two months building a product for a customer who does not exist.
He spent $62 a month. That was the average, and it looked like a sensible thing to design around, because nothing I already sold sat anywhere near it. So I built a middle option, priced it in the fifties, and shipped it. Eleven people bought it in the first quarter and four of those were existing customers moving down.
The whole problem was sitting in a chart I had never made. My customers came in two piles. Around a hundred of them spent roughly $15 a month, casually, whenever they happened to need something. About twenty spent north of $200 every month without thinking about it. Between $40 and $150 there was almost nobody, and $62 lands in the middle of that empty stretch.
The average had not made an error. Add it up, divide by the number of people, and $62 is what falls out. It is a correct answer to a question nobody asked.
What an average reports when there are two groups
An average finds the balance point. Put a hundred small weights on one end of a plank and twenty heavy ones on the other, and there is a spot where it balances. Nothing is sitting on that spot.
The mean does not know or care whether your customers are one crowd or two. It hands you the balance point either way, and only in the one crowd case does the balance point land anywhere near a real person.
What makes this expensive rather than merely wrong is what comes next. Nobody reads an average and then stops. You price against it. You write the headline for it. You pick which feature to build for the person it describes. A number describing a hypothetical middle customer becomes, quietly and without anyone deciding, the customer your business is designed for.
This is a different problem from outliers
Worth being blunt about, because the symptom looks similar and the treatment is the opposite.
An outlier is a single row far from the rest. A duplicate charge, a test order nobody deleted, one enterprise deal ten times your usual size. The work is to check whether it is genuine, then decide whether it belongs in the calculation.
Two populations is a different animal. Nothing in the data is anomalous. Every one of my $200 customers was real and completely typical of the twenty people like them. There is no bad row to go and find.
So outlier tools actively hurt you here. Trimming the top of the range throws away a sixth of the customer base and most of the revenue, and it feels like tidying up while you do it.
An outlier method answers whether a value is real. The question in front of you is how many kinds of customer you have. No amount of outlier work gets you to the second one.
The median does not rescue you either
The standard fix for a misleading average is to reach for the median, and most of the time that is good advice.
It fails in this particular case. With a hundred small customers and twenty large ones, the median is $15, because the middle of a sorted list falls inside whichever group is bigger. You have traded a number that describes nobody for a number that describes your small customers while presenting itself as a number about everybody.
The distance between the mean and the median is still worth watching as an alarm, but it is a poor diagnosis. A wide gap says the data is lopsided, and lopsided has two quite different causes: one hump with a long tail of large values, or two separate humps with a valley between them. Those want opposite responses, and no pair of summary statistics can tell them apart.
The two minute check almost nobody runs
Stop reading summaries. Go and look at the values.
Put one row per customer in a column and build a histogram. Buckets across the bottom, count of customers going up. Every spreadsheet has done this for decades. Then ask one question about the shape: one hill, or two humps?
That is the technique. It takes two minutes. The reason it never happens is that a histogram is not a metric, so it gets no tile on a dashboard and never gets built.
One warning about bucket width, because that is where this goes wrong. Wide buckets smear two humps into a single comfortable mound. Narrow buckets turn a hundred customers into fifteen little spikes of nothing. Run it at three or four widths. A genuine split shows up at all of them, and something that appears at one width and vanishes at the others was never there. That test costs nothing and will save you more embarrassment than anything else on this page.
If a histogram feels like too much work, sort the column and scroll it. A jump from $15 to $210 with nothing in between is visible to the naked eye.
And if the bell curve is in the back of your mind, set it aside here. A bell has a single peak by definition. Two peaks means every shortcut built on the bell is already off the table.
Signs you should go and plot it
Before any chart, a few things should raise an eyebrow.
The mean and the median sit a long way apart. As above, that is a prompt to go and look rather than a verdict.
A small fraction of customers accounts for most of the money. If a sixth of them bring three quarters of the revenue, an average across all of them is blending two things with nothing in common.
Your inbox holds two distinct kinds of question. The people asking how to get started are not the people asking whether you can invoice quarterly.
And the one I lean on most: describe your typical customer out loud, to another person, using only your own numbers. If what comes out is “somebody who spends a moderate amount moderately often”, you are describing an average. A real typical customer is somebody you could picture and give a name to.
Two honest numbers beat one tidy one
Once the average is discredited, the instinct is to go hunting for a cleverer average. A weighted one, a trimmed one, something geometric. I think that is the wrong direction and I am happy to be argued with about it.
Every one of those is still a single number standing in for two populations. Trimming deletes your best customers on purpose. Weighting decides in advance which group the answer will be about, without ever saying so where anyone can see it. There is no honest single number, because the thing being described does not have one middle.
So split it and report both lines. 104 customers at an average of $16. 19 customers at an average of $240. Both lines, in every report, with the counts next to the averages. An average with no count behind it is half a fact.
Name the groups in language a person would use, not “segment A” and “segment B”. The ones who order once a year around Christmas and the ones who reorder monthly. Individuals and companies. If you cannot put a plain name on a group, you probably have not understood it yet.
Your report gets uglier. One number was tidy, two are not, and every dashboard tool ever made will fight you on it, because they are built to display one big number in one big font. Take the ugly report you believe.
Choosing the groups well, and working out how many earn their keep, is a separate job. This is only about noticing that there are two.
Telling a real split from noise
Now the caution, because I have fooled myself in this direction as well.
With 40 customers you can find a pattern in anything. Drop 40 numbers into buckets and there will be lumps. Random data has lumps. So here is the bar a split has to clear before I act on it.
I have to be able to name the mechanism. Not the dip in the chart, the reason for the dip. These are companies and those are individuals. These arrived from search and those came from a partner sending me his overflow. If the best I can manage is “there is a gap around $80”, I have found a bucket width rather than a business.
It has to survive next month, on customers who were not in the first chart. One month of anything is a story.
And each group has to be large enough to act on. If the second hump is five people, then it is five people. Go and talk to all five and skip the segment.
I once found what I was certain was a third group in the middle, about 30 accounts, and spent an evening writing them their own email. By the following month they had scattered into the other two and the pattern never returned. 30 people is not enough people to have a shape.
It happens to numbers that are not customers
Average response time, in a business that answers some messages in ten minutes and lets the rest sit until the following morning. Average order value in a shop selling a $5 accessory alongside a $300 main product. Average session length when half the traffic leaves in eight seconds.
Any time a number is built from two different kinds of event, its average is a made up event. There is a nastier relative of this where splitting a group reverses a trend entirely, which deserves its own piece.
What the chart will not tell you
Seeing two humps does not tell you which one to chase. It says you are running two businesses. It says nothing about whether to concentrate on the twenty large customers or go and find four hundred more small ones, and I have watched both work.
That decision is about what you want to run and what you are good at. All the histogram does is stop you making it on behalf of somebody who does not exist.
More plain English walkthroughs on the numbers a one person business runs on are at Data Research Analysis Collection.
Get new guides and videos first — join the Telegram channel.