Average order value and average customer value are reported almost everywhere and describe very few actual customers. The distribution behind them is skewed enough to make the mean misleading.
Spending is not distributed evenly
In most categories a small proportion of customers accounts for a large share of revenue, and a long tail buys once and never returns.
An average taken across that shape sits between the two groups, describing a customer type that barely exists in the data.
Decisions calibrated to that midpoint therefore serve neither group well, since the two behave differently in almost every respect.
The mean moves for the wrong reasons
A single very large order can shift a monthly average noticeably, particularly in businesses with a wide range of transaction sizes.
That movement is then interpreted as a change in customer behaviour, and explanations are constructed for something that was one order.
Medians are far more stable and are almost never the number that appears in a summary report, largely out of habit.
The instability also makes month-to-month comparison of averages unreliable, since the difference between two periods can be smaller than the effect of a handful of unusual orders.
Acquisition targets built on averages misprice
Setting a maximum acquisition cost from average customer value implies that every customer is worth the same at the point of acquisition.
In practice the value depends heavily on which segment and which channel the customer arrived through, and those differ by multiples rather than by margins.
A single blended target therefore overpays for sources that deliver one-time buyers and underpays for sources that deliver repeat customers.
What to report instead
Distributions rather than points. A simple breakdown by value band shows where revenue actually sits and how many customers occupy each band.
Cohort views add the other missing dimension, showing how value accumulates over time for customers acquired in the same period.
Both are straightforward to produce from data every business already holds, and they change decisions in ways a mean cannot.
Where averages remain appropriate
For symmetric, tightly clustered measures such as page load time or delivery duration, an average summarises the data honestly.
They also work for very high volume, low variance events where individual outliers cannot move the total.
The check before using one is simply whether a small number of cases could dominate the result, and for customer value the answer is almost always yes.