Why averages mislead, and what to use instead

Take ten orders: nine of them between ₹400 and ₹900, and one at ₹40,000 from a bulk buyer. The average order value is a little over ₹4,300 — a figure higher than nine of the ten orders actually placed. It is arithmetically correct and descriptively useless, because no order resembles it and no decision based on it will be right. Anything you plan around a ₹4,300 order — packaging, delivery, staffing, stock, the free-shipping threshold — is designed for a customer who appears once in ten.

This is not an unusual case. It is the normal shape of business data. Order values, customer lifetimes, response times, days to payment and job durations are all bounded below and unbounded above: nothing can be less than zero, and a single instance can be enormous. Distributions with that shape have most of their mass at the low end and a long tail to the right, and for them the average sits well above the typical case. The mistake is not using an average; it is assuming that a single number can describe a distribution, which is only true when the data is symmetrical and clustered — and business data rarely is.

The median is the middle value: line up every order from smallest to largest and take the one in the middle. In the example above it is around ₹650, which describes a typical order rather than an arithmetic artefact. Its usefulness comes from the fact that it counts positions rather than amounts, so the size of the largest order is irrelevant — the bulk buyer could have spent ₹4,00,000 and the median would not move. That resistance to a single extreme value is exactly the property you want when the extremes are real but rare.

The most efficient habit is to report both and treat the gap between them as information in itself. When average and median are close, the data is well-behaved and either figure describes it. When the average sits far above the median, you have a right-skewed distribution and a small number of large cases driving the total — which is worth knowing explicitly, because those large cases usually deserve their own handling rather than being blended into a figure meant to describe everyone. The gap is a diagnostic, and it costs one extra calculation.

Two numbers are better than one, and a few percentiles are better still. The 25th percentile is the value below which a quarter of your orders fall, the 75th is where three-quarters fall, and the 90th tells you what a genuinely large order looks like. Reading the 25th, 50th and 75th together gives you the middle half of your business — a range rather than a point — and that range is usually what an operational decision actually needs.

Percentiles matter most where the tail is the thing you care about. Delivery times are the clearest case: an average delivery time of two days is consistent with a tenth of customers waiting a week, and it is the tenth waiting a week who complain, leave and tell other people. The average describes an experience nobody is upset about while concealing the one that is costing you customers. Any measure of how long something takes is better read at the 90th percentile than at the mean, because the worst experiences are what people remember and repeat. The same logic applies to days taken to pay an invoice, where the tail is where the cash problem lives.

Before choosing a summary statistic, look at the data's shape, and the cheapest way is a histogram: buckets across the bottom, count of orders in each. It takes minutes in a spreadsheet and it answers a question no summary can, which is whether the data has one cluster or several. A single hump means a summary figure is meaningful. Two humps mean you have two different populations mixed together, and no single statistic — average, median or otherwise — describes either of them.

Two clusters are common and consequential: retail and wholesale buyers, one-off and repeat customers, two products with different price points, or a business and consumer segment sharing one sales record. In every case, the honest move is to split the data and summarise each group separately, because a figure computed across both describes a population that does not exist. This is the general form of the lesson: segment before averaging. Once the groups are separated, an average within each is often perfectly informative — the problem was never the average, it was averaging across things that should not have been combined.

Customer lifetime value is the worst offender, because it compounds the problem. It is usually computed as an average order value multiplied by an average purchase frequency over an average lifespan — three averages, each drawn from a skewed distribution, multiplied together. Errors do not cancel in a product; they amplify. The resulting figure is typically dominated by a handful of exceptional customers and describes almost nobody, and it is then used to justify what you can afford to spend acquiring an ordinary customer, which is precisely the wrong comparison.

The better version is median-based and segmented: the median value of a customer from a given channel or product, over a fixed window rather than a projected lifetime. It is a smaller, duller number and it supports the decision you actually make. The same trap appears wherever averages are multiplied — average deal size times average close rate, average visitors times average conversion — and the tell is a forecast whose output is confidently precise while every input was a mean of something lumpy. Averaging also hides the extremes at both ends, which is a business problem as much as a statistical one: your best customers and your worst are the two groups most worth understanding, and an average is the operation that removes both.

The median is not simply a better average, and using it reflexively causes its own errors. It tells you nothing about totals: if you want to know how much revenue you took, you need the sum, and the average multiplied by the count is the honest route to it. The large orders the median ignores are real money, and a business that plans capacity around the median will be unable to serve the tail it depends on for revenue. Both figures answer real questions; neither answers the other's.

The median also moves erratically when there are few data points, and with an even count it is a convention rather than an observation. Neither statistic protects you from the more basic problems: data that was recorded inconsistently, a period chosen to flatter, comparing groups that are not comparable, or a sample of customers that is not representative of the market. And no summary statistic — mean, median or percentile — tells you why the distribution has the shape it has. That requires looking at the cases themselves, which is why the most productive thing to do with a histogram is often to pick out the orders in the extreme buckets and read them individually.