Is the “Average Customer” Hiding Who You’re Overpricing and Undercharging?

Average deal size, average order value and price elasticity show up in every management report, and they are usually the numbers quoted when a pricing decision gets made.

The problem is that an average describes a whole population, while a price is paid by one customer at a time. If your base contains groups with genuinely different economics, the average describes a customer who does not exist, and the price you build around that customer fails in two directions at once. It prices one group out completely, and it undercharges another. Both failures arrive in your accounts as ordinary revenue, which is why neither one ever shows up in the report.

The arithmetic of an average price

Imagine a market of 100 customers, kept simple so the mechanics are visible.

Willingness to pay is the most a customer will hand over in a given buying situation. 50 of these customers are price sensitive and will pay at most $50. The other fifty will pay up to $150. Average willingness to pay across the base is exactly $100.

Price at the average, RM100, and here is what happens:

• the RM50 group buys nothing

• the RM150 group buys 50 units

• total revenue is RM5,000

• every premium buyer keeps RM50 you could have charged, RM2,500 in total.

Run two prices instead, RM50 and RM150:

• the price sensitive group buys 50 units at RM50

• the premium group buys 50 units at RM150

• total revenue is RM10,000

• nothing is left on the table from either group.

Same customers, same product economics, half the revenue.

The example is deliberately clean. Real markets have cost-to-serve differences, tiers that cannibalise each other, and plenty of customers sitting in the middle. The structural point still holds: an average price is only the right price when your customers are genuinely similar.

The same assumption shows up in the analysis

Pricing at the average is one trap. On the other hand, the methods used to estimate willingness to pay often carry the same assumption, so the number they hand you is wrong in a predictable direction.

A 2026 discrete-choice study of the wine market found substantial unobserved preference heterogeneity, meaning buyers differed in ways the observable variables did not pick up. Models that ignored those differences systematically overstated average willingness to pay. An overstated average leads to prices set too high, which pushes away exactly the buyers the model never saw in the first place.

Someone has put a number on the cost. Work by Abhishek Hosanagar and Fader on a large sponsored-search dataset estimated revenue losses of up to 11% from bidding and pricing decisions made on aggregate data that ignored these differences. As a line item, 11% would trigger an investigation. Spread across an average, it disappears.

The customers who are invisibly unprofitable

Price sensitivity is only half the story and cost to serve is the other half. It's distributed even less evenly.

A 2025 case study of an Indonesian printing and packaging company looked at 147 of its B2B customers. Average profitability made the base look stable. The distribution said otherwise:

• 20.41% of customers produced 58.39% of total contribution margin, mostly specialised accounts that wanted particular configurations and fast turnaround

• 59.86% sat around break-even while consuming support at standard rates

• 19.73% were actively losing the company money

Set your service levels, lead times and pricing around the average customer and you under-serve the fifth of the base generating close to 60% of your profit, while subsidising the fifth that destroys value. Both groups are invisible in an average and obvious in a distribution. Worth noting that this is a printing and manufacturing business, the kind usually assumed to have fairly uniform customer economics.

Where more segmentation stops helping

All of that argues in one direction, so the counterargument deserves proper space rather than a caveat at the end. More segmentation is not automatically better.

Micro-segments built on small statistical wobbles give you weak sample sizes and unreliable models. You end up fitting noise, then pricing against non-repeatable patterns. The operational cost compounds too, because fifty price points and fifty configurations all have to be sold, supported, billed and explained, and that overhead can eat more value than the segmentation created. Customers notice this as well. When pricing looks fragmented or opaque, buyers start to assume the structure exists to extract from them rather than to serve them, and that kind of suspicion affects trust.

There is also the case where the average is simply correct. If elasticity, value drivers, and cost to serve barely vary across your base, an average is both accurate and easier to run. It becomes dangerous only when economically distinct groups get compressed into it, not as a general property of arithmetic.

So the principle I would hold onto is this: don't segment because customers are different; segment when the difference changes a commercial decision.

That gives you an order of operations. Difference, then decision, then economics. If two groups behave differently but you would still price, package, acquire and serve them the same way, the segmentation is not commercially useful, however real the difference is.

A test you can run this month

Rank your customers by contribution margin rather than revenue and look at the shape of the list

If it is reasonably flat, your averages are doing their job; If the top fifth carries most of the profit and the bottom fifth is negative, your average customer is moderately profitable and moderately demanding, and nobody in your base is either of those things. You are still pricing for that person:

• your best accounts are underserved (They pay more and get less)

• your worst accounts are overserved (They pay less and get more)

This means Margin moves from the first group to the second every month, and none of it shows up in your reporting.

Then apply the second test before acting: for each group you are tempted to separate, name the decision that would change, and if you cannot name one, leave it alone.

Sources & Further Reading

• 2026 discrete-choice study of consumer preferences in the wine market, finding substantial unobserved preference heterogeneity and showing that models ignoring it produce materially overstated estimates of mean willingness to pay.

• Abhishek, Hosanagar and Fader, sponsored-search study estimating revenue loss of up to 11% from ignoring preference heterogeneity in aggregate-level data.

• 2025 case study of an Indonesian printing and packaging company, customer profitability across 147 B2B customers (20.41% generating 58.39% of contribution margin; 59.86% near break-even; 19.73% unprofitable).

• Peer-reviewed retail price-response research (2021) on aggregation bias in elasticity modelling, showing that aggregate-level models mask store-level and segment-level differences. Further reading on the same mechanism.

• The 100-customer model is a hypothetical construct used to show the arithmetic, not a documented case.

The over-segmentation limits, the segmentation principle and the contribution-margin test are drawn from practice rather than from published research.

The Morning Owl SDN. BHD.
(202501015530 / 1616945-K)