The average outcome is often one that nobody actually gets

One number standing in for a whole shape

Reporting an average is a compression, and like any compression it discards what did not fit. For a symmetric, tightly clustered distribution, almost nothing is lost and the mean is a fair description of a typical case. For anything skewed, heavy-tailed or split into groups, the mean can sit in a region where few or no actual outcomes occur.

The clearest everyday example is any quantity where the bottom is bounded and the top isn’t — durations, incomes, waiting times, repair costs, project overruns. These are almost always right-skewed, with a long tail of large values pulling the mean above the middle of the pack. The median, the value with half the cases either side, stays where most of the experience is.

Why the gap matters for a decision

If you are planning your own experience of something — how long a task will take, what a repair is likely to cost — the median is usually the better guide, because it answers the question of what a typical instance looks like. If you are planning in aggregate, across many instances, the mean is what you need, because totals are sums and sums track means.

Using the wrong one produces predictable errors in opposite directions. Budget every instance at the median and the total will come in short, because the tail cases aren’t included anywhere. Plan a single instance at the mean and you will usually over-provide, and will still be caught out when the genuine tail case arrives, since the mean is not the maximum either.

Variance is a cost even when the mean is fixed

Two arrangements with identical average outcomes are not equally good if one of them swings and the other does not. Volatility consumes attention, forces buffers to be held, and creates the possibility of hitting a floor from which recovery is slow or impossible. A steady income and a lumpy one with the same annual total are different products.

This is not merely a preference. When outcomes compound, or when hitting zero ends the process, the spread affects the result directly rather than only the experience of it. That is why arrangements which look inefficient in a table of averages — reserves, smoothing, insurance — can be correct once the distribution rather than its centre is in view.

Multiple populations wearing one number

A second way an average misleads is by describing a mixture. When a group actually contains two or three distinct subgroups with different behaviour, the overall mean lands somewhere between them, describing nothing. Average satisfaction across users who love a thing and users who can’t make it work is a number that no individual holds.

The tell is usually visible in a histogram and invisible in a summary. When a distribution has two humps, or a large mass at one extreme, any single-number description will misrepresent it, and the useful question becomes what distinguishes the groups rather than what the average is.

What to ask for instead

A more informative summary states the middle and the spread together: the median, some sense of the range within which most cases fall, and an explicit note about the tail. Where the data is available, a few percentiles convey more than any amount of discussion about the mean.

When only an average is offered, the useful follow-up is whether the underlying quantity is bounded below and unbounded above, which predicts skew reliably, and whether the population is likely to be a mixture. Both questions can usually be answered from knowledge of the subject rather than from the data, and either one is enough to know how much to trust the number.

The limits of this advice

None of this makes averages bad. They are the right tool for totals, for comparing large groups, and for any quantity whose distribution is roughly symmetric, which covers a great deal of measurement. The mean also has properties that make it the correct input for many calculations, and replacing it with the median everywhere would introduce errors of its own.

The discipline is narrower: know which question you are asking, know the rough shape of the distribution, and treat a lone summary statistic as a claim about a shape rather than as a fact about a case. Most of the disagreement about whether things are getting better or worse turns out, on inspection, to be a disagreement about which part of a distribution the speakers are looking at.