ARPU and the Averages That Hide Your Business

A mean over a distribution with no middle

Take a book that looks like most subscription businesses at some point: 10,000 registered accounts, $100,000 billed in the month. Seven thousand of those accounts pay nothing. Two thousand pay $10, eight hundred pay $25, a hundred and fifty pay $100, forty pay $500, and ten pay $2,500. That is the whole dataset, and it adds to exactly $100,000. Divide by the 10,000 accounts and you get an ARPU of $10.00. Now ask what the median account pays. It pays zero, because the middle of a list of 10,000 accounts of which 7,000 pay nothing is a non-payer. The average sits $10.00 above the median, and it describes no account in the book.

Throw the non-payers out and it barely improves. Three thousand accounts pay something; their mean is $33.33 and their median is $10.00, so the mean is 3.33 times the middle of its own distribution. Ask how many payers actually bill at or above the average payer and the answer is 200 — 6.67% of payers, 2% of registered accounts. Put the other way, 93.3% of your paying customers bill less than what you tell the board a paying customer is worth.

The decile view says it in one line. Split the 3,000 payers into ten equal groups by what they bill. The first six deciles all average $10.00. The seventh averages $15.00, the eighth and ninth $25.00, and the tenth averages $208.33 — 20.8 times the bottom decile. That top decile holds 62.5% of all revenue, and the top percentile, thirty accounts, holds 35.0%. An average taken over that shape is not wrong, it is simply uninformative: it is a single point standing in for a curve that spans two orders of magnitude.

The denominator decides the answer

Keep the revenue fixed at $100,000 for the month and change nothing about the business. Divide by the 10,000 accounts that ever registered and ARPU is $10.00. Divide by the 4,000 that logged in during the month and it is $25.00. Divide by the 3,000 that were billed and it is $33.33 — that figure has its own name, ARPPU, and mixing the two is the single most common error in a deck. Divide by the average of 1,250 daily actives and it is $80.00. One month, one revenue line, an eightfold spread depending on which population you put underneath it.

The period does the same work more quietly. Run the same business for a quarter and it bills $300,000. Count the users who were active at any point in those three months and you find 6,000, not 4,000, because a wider window catches the people who show up once every few weeks. Quarterly ARPU is $50.00, which per month is $16.67 — against the $25.00 the monthly view reported. Nobody changed a plan, a price or a habit. The window moved.

There is a useful identity buried in all this: ARPU equals ARPPU multiplied by the paying rate. Here, $33.33 × 30% = $10.00. It matters because it tells you which of the two levers moved. A rising ARPU with a flat paying rate is real price realisation. A rising ARPU with a flat ARPPU is conversion. A rising ARPU with both flat is arithmetically impossible, which is a useful check that someone changed the denominator between two slides.

Two opposite businesses, one ARPU

Run the book forward a year in two different ways. In the first, everything grows broadly: 20% more accounts in every price band and 20% more price in each of them. That gives 12,000 registered accounts, 3,600 payers and $144,000 of revenue. ARPU is $144,000 ÷ 12,000 = $12.00, up 20% on the $10.00 baseline. Revenue is up 44%, accounts are up 20%, and the shape of the distribution has not changed at all.

In the second, the base erodes. Every band loses 20% of its accounts, so 2,400 payers survive and bill $80,000 between them, and free registrations shrink to 5,592. Then eight enterprise contracts arrive at $2,000 a month, adding $16,000. Total: 8,000 registered accounts, 2,408 payers, $96,000 of revenue. ARPU is $96,000 ÷ 8,000 = $12.00. The same $12.00. Revenue is down 4%, accounts are down 20%, and the headline metric reports the identical 20% improvement.

This is the failure mode that makes ARPU dangerous rather than merely imprecise. It is a ratio, and a ratio improves whenever the denominator falls faster than the numerator. Losing the least valuable half of your users is indistinguishable, in that one number, from earning more from the users you kept. Any metric of the form X per user has the same defect — cost per user, sessions per user, support tickets per user. The base shrinking always flatters them.

ARPPU does not rescue you either

The obvious fix is to quote ARPPU instead — revenue over paying customers, immune to the free base. Run it on the two scenarios. The growing book reports $144,000 ÷ 3,600 = $40.00. The shrinking one reports $96,000 ÷ 2,408 = $39.87. A gap of thirteen cents between a business that grew 44% and a business that shrank 4%. ARPPU is a better number than ARPU, and on this pair it is just as blind.

What does separate them is concentration. In the growing book the top decile of payers still holds 62.5% of revenue and the top percentile 35.0% — unchanged, because everything scaled together. In the shrinking one the top decile holds 68.5% and the top percentile 41.7%. Six points of concentration moved into the top tenth, which is exactly what happens when eight contracts replace six hundred small accounts. That drift, not the mean, is the early warning.

The second diagnostic is even simpler: count the customers. Payers went from 3,000 to 3,600 in one scenario and to 2,408 in the other. A count cannot be flattered by a shrinking denominator because it has none. Whenever a ratio and a count disagree, believe the count.

Gross, net, and the money you never see

Before any argument about the denominator, settle the numerator. Suppose 30% of that $100,000 is billed through an app store taking a 30% commission. You never receive $9,000 of it. Net revenue is $91,000 and ARPU is $9.10, not $10.00 — a 9% haircut applied before anybody has said the word "user". Refunds, chargebacks, sales tax collected on behalf of a government and payment processing all bite the same way, and none of them appear in a bank-agnostic "revenue" field pulled from a billing table.

Currency is the other quiet edit. A book billed in several currencies and reported in one moves with the exchange rate whether or not a single customer changed plan. Companies that report ARPU seriously publish a constant-currency figure alongside it for exactly that reason, and the gap between the two is often larger than the growth they are claiming.

What to publish instead of a single mean

Four lines replace the one. First, ARPU with its denominator and period written in the same sentence — "$25.00 per monthly active user, August" is a number; "ARPU $25" is a rumour. Second, ARPPU beside it, so the paying rate is recoverable by division. Third, the payer count and the paying rate as raw counts. Fourth, three points of the payer distribution: the median, the ninetieth percentile and the ninety-ninth. On our book those read $10.00, $25.00 and $500, and anyone who sees them immediately knows not to trust the mean.

Then cohort it. A blended ARPU mixes customers who signed last week with customers who signed three years ago, and those two groups behave nothing alike. Compute ARPU per signup cohort at the same age — month three for the January cohort against month three for the February cohort — and expansion, contraction and mix separate cleanly. That is also the only version of the metric that can be fed into a lifetime-value model without circularity, since the LTV calculator wants a per-cohort revenue curve, not a company-wide average, and the CAC payback article in this series shows what happens downstream when the input is off by a quarter.

Frequently asked questions

Is ARPU the same thing as ARPPU?

No, and the two are linked by an exact identity: ARPU = ARPPU × paying rate. On the book used throughout this article, $33.33 × 30% = $10.00. ARPU divides revenue by everyone in the chosen population; ARPPU divides it only by the accounts that were billed in the period. A freemium product will always show a large gap between them, and the gap is not noise — it is the conversion rate, stated backwards.

Our ARPU is rising while revenue falls. How?

The denominator is falling faster than the numerator. In the second scenario above, revenue drops 4% and accounts drop 20%, so ARPU rises 20% to $12.00 — exactly the figure produced by the healthy scenario, where revenue rose 44%. Check three things before celebrating: the paying customer count, the free-to-paid conversion rate, and the share of revenue held by your top decile of accounts. If the count is down and the concentration is up, the ARPU move is churn wearing a disguise.

What is a good ARPU?

There is no answer to that question without the denominator, the period and the pricing model attached, which is why a cross-industry benchmark for ARPU is close to meaningless. A consumer app monetising a large free base and a business tool selling seats can differ by three orders of magnitude and both be healthy. The comparisons that do carry information are internal and directional: your ARPU against your own previous cohorts at the same age, and your ARPU against your customer acquisition cost, since the ratio between them is what decides whether growth pays for itself.

Should ARPU use gross or net revenue?

Net — after refunds, chargebacks, platform commission and any tax you collect on behalf of a government, because none of that money is yours. The effect is not cosmetic: if 30% of a $100,000 month is billed through a store taking 30%, net revenue is $91,000 and ARPU falls from $10.00 to $9.10 before any other adjustment. Whatever you choose, state it next to the number, and keep it identical across periods — a metric that silently switches basis between two quarters is worse than no metric.

Why not just report the median revenue per user?

Because in any freemium book the median is zero and stays zero for years, which is true but useless for steering. The usable version is the median of the paying accounts — $10.00 here, against an ARPPU of $33.33 — reported together with the ninetieth and ninety-ninth percentiles. Those three numbers cost nothing extra to compute, cannot be gamed by moving the denominator, and immediately show whether your revenue rests on a broad base or on a handful of accounts.

How does ARPU feed MRR and lifetime value?

MRR is ARPPU multiplied by the number of paying accounts, so the two are the same statement seen from either end; lifetime value is roughly ARPPU × gross margin ÷ churn rate. That chain is why a denominator error compounds. Inflate ARPU by 20% through a narrower denominator and you inflate the lifetime value by 20%, which inflates the acceptable acquisition cost by 20%, which is spent in cash that the customers never return. Model the pieces separately with an MRR and a lifetime-value calculator rather than deriving everything from one average.