GASP

Research

The definitive guide to churn

By Gareth O'Shea ·

Steven Noble once suggested that Shopify give up on months.

February has 28 days and March has 31, so a business that loses customers at the same steady pace all year will lose more of them in March than in February. Its monthly churn rate goes up and its customers have done nothing new. Noble had a remedy:

(As an aside I once suggested that we stop reporting figures for actual months and instead chop up the year into 30 day periods and only report figures for these 30 day Shopify months. I didn’t really get an answer as much as a I got a look that said did we really just hire this guy? I needed to remind myself that analysis needs to work within the world that we are in if it is to be useful; not an idealized version of the world. The globe isn’t a sphere. The market isn’t efficient. And months aren’t all 30 days.)

(Steven H. Noble, Shopify Engineering, 2011)

The months survived. Noble’s essay is about a larger puzzle than the length of March, one no calendar could repair.

Take a business that opens July with 10,000 customers. During the month 500 of them cancel. It also signs 5,000 new customers and 125 of those cancel before the month is out. August is July again. Five per cent of the customers it opens with cancel. Another 5,000 sign up and 125 of them cancel.

Now work out the churn rate the way most people would, as everyone who cancelled during the month divided by the customers the month began with. July comes to 6.25%. August comes to 5.87%.

Noble called it “the exact same behaviour in both months”, yet the second month looks better than the first. A manager shown those two figures would assume that something had started to work. Noble’s comment on that manager was “You would have told yourself that your churn rate is improving when nothing has changed.”

The fraction counts two different groups

The improvement comes from the 125 newcomers who cancelled.

They cancelled in July, so they are counted above the line of the fraction. They were not customers on the first of July, so they are absent below it. In July those 125 extra cancellations are set against 10,000 opening customers. By August the business has grown to 14,375 and the same 125 are set against that, where they weigh less. The customers are exactly as loyal in August as they were in July. The business is bigger, which is all the falling churn rate has picked up.

A fraction means something only when its top and its bottom describe the same group of people. In Noble’s July they describe two groups. There are two ways to bring them back together, since the newcomers can be added to the bottom or taken off the top.

The first fix counts new customers below the line

In 2004 Netflix rented DVDs by post and reported its churn to investors every quarter.

Churn is a monthly measure defined as customer cancellations in the quarter divided by the sum of beginning subscribers and gross subscriber additions, then divided by three months. Customer cancellations in the quarter include cancellations from gross subscriber additions, which is why we include gross subscriber additions in the denominator.

(Netflix, Form 10-K for fiscal 2004)

Netflix had seen the 125. Its cancellations included people who had joined during the quarter, so it put everyone who had joined during the quarter below the line. On that basis it reported churn of 4.4% a month for the last quarter of 2004.

By then the formula had been taken to court. In July 2004 shareholders sued Netflix with the denominator at the centre of their case. They had a fraction of their own, cancellations over the average of the subscribers at the start and the end of the quarter, which they called “actual churn”. By their arithmetic the 5.2% a month that Netflix had reported for the third quarter of 2003 should have been 6.5%. The court was being asked which of two fractions was the true one.

There is, however, no official formula for calculating churn.

(Judge Fern M. Smith, In re Netflix, Inc. Securities Litigation, 2005)

Netflix had printed its formula in a footnote to each financial report, so the rates the formula produced, the judge wrote, “cannot be considered to be false”. The case was dismissed in June 2005 and again, for good, that November. No accounting standard has defined churn since. A company may still use the fraction it prefers as long as it prints the footnote.

The repair works on its own terms, since every cancellation now belongs to someone in the denominator. But a subscriber who signed up on the last day of the quarter counts below the line as fully as one who was there for all of it, so a quarter with many sign-ups reports lower churn. Run Noble’s two months through the Netflix denominator and they come to 4.17% and 4.36%. Two identical months still disagree and now it is the second that looks worse.

The second fix counts only the customers of day one

The other repair is to count a cancellation only when the customer was there on the first day. The 125 leave the top of the fraction while the 10,000 stay on the bottom, which brings July to 5.00%. August comes to 5.00% as well.

How the rate is worked outJulyAugust
Everyone who cancelled, over the customers on day one6.25%5.87%
Everyone who cancelled, over the day-one customers plus the month’s sign-ups4.17%4.36%
Day-one customers who cancelled, over the customers on day one5.00%5.00%

The first row is Noble’s own arithmetic. The second and third are worked here from his inputs.

The GASP Standard uses the third row. In the standard’s formula grammar the last clause is the instruction to follow the customers of day one.

metric.churned_customers / metric.total_customers at period_start over cohort(field.customer_id, period_start)

The test quarter the standard ships with opens on 1 April 2026 with 15 customers. Two of them have cancelled by 30 June. A sixteenth customer signs on 13 May and cancels on 20 June. Count everyone who cancelled and Logo Churn Rate is three over 15, which is 20%. Follow the customers of day one and it is two over 15, which is 13.3%.

The same quarter reads 20% by the first row of the table and 13.3% by the third.

A rule on a page can still be ignored by whoever writes the query, so the standard does not leave the query to be written. Each of its formulas compiles to SQL in the dialect of the company’s warehouse and the cohort clause compiles with it. A company that takes its churn query from the standard gets the third row of the table whether or not anyone on its team has heard of Noble’s puzzle.

The customer who came and went in under six weeks is still on the books. The standard’s rule reads “A customer who joins and churns inside the period appears in Churned Customers and sits outside this rate.” Churned Customers is a count and the short-lived customer is in it. The percentage is kept for one question, which is what became of the customers the period opened with.

Revenue churn follows the same customers

Counting customers treats them all alike. A company that loses two of its 200 customers has lost 1% of them. If those two paid a fifth of the revenue it has lost 20% of its income, so the standard keeps a second rate in monthly recurring revenue (MRR) and says of it “Losing a $50K customer hits harder than losing a $500 customer.”

Revenue Churn Rate follows the day-one rule as well.

(metric.churned_mrr + metric.contraction_mrr) / metric.mrr at period_start over cohort(field.customer_id, period_start)

The billing tools that many subscription businesses read their churn from work the same way. Paddle, Chargebee and ChartMogul each leave out the customer who signs up and cancels inside the same month. Paddle’s help pages give it as a reason a churn figure might look lower than expected, saying “Metrics excludes users who sign up and churn within the same month from MRR and churn numbers.”

On the standard’s example dataset the rule moved revenue churn for the first half of 2026 from 21.9% to 19.9%. Those two points belonged to one customer, who signed in February and cut its plan by $300 in June. Under the day-one rule the rate and Gross Revenue Retention (GRR) follow the same customers, so the two add up to 100%.

A customer who pays less moves revenue churn and leaves logo churn alone

A customer who cancels has churned. A customer who moves from a $1,000 plan to a $600 plan has not, in the standard’s vocabulary, churned at all. They are still a customer and Logo Churn Rate does not move. The missing $400 has a name of its own, Contraction MRR, which the standard defines as “The reduction in MRR from existing customers who remain customers.”

A customer on $1,000 a month who switches to annual billing for a 10% discount keeps every seat and every feature and pays $10,800 a year, which is $900 a month. That $100 is Contraction MRR too. The definition continues “Every decrease in committed recurring revenue counts, whether or not the service changed.” The ledger labels the reason, so a discount can be told apart from a customer buying less.

Either way the revenue is gone, so Revenue Churn Rate counts it beside Churned MRR, the MRR of the customers who left. The two rates can then tell opposite stories about one quarter. Take a company with 200 customers on that $1,000 plan. All 200 stay and 40 of them drop to the $600 plan. Its Logo Churn Rate for the quarter is 0%. Its Revenue Churn Rate is 8%, because $16,000 of the $200,000 it opened with is gone.

A reader who hears “churn” and pictures departures would take the 0% for the whole story. In the standard, Churned Customers and Churned MRR belong to customers whose MRR reached zero. Revenue Churn Rate is the wider figure because it adds the reductions of the customers who stayed, as GRR and NRR do.

A planned ending still counts

Some losses are written into the contract.

A customer paying $1,000 a month takes a $600 analytics add-on for a three-month campaign. On 1 July the add-on expires as agreed and the customer’s MRR returns to $1,000. Another customer signs up for a six-month client project and closes its account in the sixth month, as it said it would on the day it signed. Neither ending surprised anyone, which makes it tempting to keep both out of the churn figures.

The standard counts both. The add-on is Expansion MRR in April and Contraction MRR in July, so it raises Revenue Churn Rate in the quarter it lapses. The project customer is a churned customer in Logo Churn Rate and its fees are Churned MRR.

While they run, both sit in MRR and therefore in Annual Recurring Revenue (ARR), which the standard defines as MRR multiplied by twelve. Each leaves those figures on the day its contract ends and no sooner. ARR is a run rate. It states today’s recurring billing at the scale of a year and makes no forecast, so a six-month project appears in it at twelve times its monthly fee until the month it stops.

Each MRR movement in the standard’s ledger can be marked as permanent, as temporary or as the scheduled reversal of a temporary one. The mark is read from the contract: “A temporary upgrade is temporary because the contract says so, not because someone thinks it might reverse.” The mark sits on the movement itself, so the add-on and the six-month project stay in ARR while anyone who reads the ledger can filter on the mark and see how much of that ARR has an end date. A company that fills the label in can report Adjusted NRR and Adjusted GRR, which leave the temporary movements out, beside the standard figures that keep them in. The standard figures are the ones for the board and for investors. The adjusted pair shows a management team how much of a quarter’s movement was ever meant to last.

A customer has left when their MRR reaches zero

The day-one rule settles who is in the fraction. It leaves open what a customer has to do to count as gone.

Peloton sells exercise bikes and treadmills with a monthly subscription for the classes. A member can pause that subscription instead of cancelling it, which raises the question of whether a paused member has left. The annual report for fiscal 2023 says of the churn figure “This metric does not treat a pause of a Connected Fitness Subscription as a churn event.”

Starting in fiscal year 2024, we are making changes to our reported operating metrics. We will no longer include paused Connected Fitness subscriptions in our new Ending Paid Connected Fitness Subscriptions metric, and will now treat a pause action as a churn event in our Average Net Monthly Paid Connected Fitness Subscription Churn as described below.

(Peloton, Form 10-K for fiscal 2023)

Later reports give the reason in a clause, “Because we do not receive payment for paused Connected Fitness Subscriptions”. Peloton reported churn of 1.2% for fiscal 2023 under the old rule and 1.4% for fiscal 2024 under the new one. How much of that rise was members leaving and how much was the definition, the two figures alone cannot say.

The GASP Standard starts where Peloton ended up. It builds churn from MRR and a subscription has MRR only while it is active. In the standard’s words, “A paused, cancelled or expired subscription contributes nothing.” A customer whose MRR reaches zero has churned that day, whatever the reason.

The two differ over the member who comes back. Peloton counts an unpause “as a reduction in our churn count in that period”, so a return cancels out a departure. The standard records the return as Reactivation MRR, a movement of its own, “counted apart from New MRR, so win-back and acquisition can be judged on their own”. Churn stays whole and the win-back shows up beside it.

Revenue churn is reported gross

Every subscription business would like to report the figure David Skok named.

It is possible to run a SaaS, or any other kind of recurring revenue, business in such a way as to get what I call Negative Churn. This happens when the expansions/up-sells/cross-sells to your current customer base exceed the revenue that you are losing because of Churn.

(David Skok, “Unlocking the Path to Negative Churn”, For Entrepreneurs)

Negative churn is real and worth having. It can also cover a leak. Take a company that loses 2% of its MRR in a month while the customers who stay buy 2% more. Its net churn is zero. A reader given only that zero cannot tell the company from one that lost nothing at all.

So the standard’s revenue churn is gross. It counts the MRR lost to cancellations and reductions and leaves expansion out. Of that choice the metric’s page says “It is the canonical formula because it isolates the leakage.” The net view has a metric of its own, Net Revenue Retention (NRR), where expansion is allowed to offset the losses under a name that says so.

State the period and fix the segment

Beside Logo Churn Rate the standard’s glossary carries a short warning, “monthly × 12 ≠ annual”. A company that loses 2% of its customers every month keeps 98% of them twelve times over, which leaves 78.5%. Its annual churn is 21.5% and multiplying by twelve would have said 24%. At 5% a month the gap is wider, 46% against 60%.

Small customers leave far more readily than large ones. The standard’s benchmark notes, citing SaaS Capital, put it at “SMB churn is ~8x higher than enterprise.” A blended rate therefore reports the mix of customers as much as their loyalty, so the standard breaks the rate out by segment, which needs a rule for the customer who changes segment in the middle of a period. The standard’s rule is “Each customer counts in the segment they held on the first day of the period.”

In the test quarter, one of the two customers who cancelled was mid-market on 1 April and was relabelled enterprise on 1 May, six weeks before leaving. The customer who signed in May was enterprise as well. Go by the label on the day of the cancellation and enterprise lost two of its four customers, a churn rate of 50%, though one of the two was mid-market when the quarter opened and the other had yet to sign. Go by day one and enterprise lost none of its four while mid-market lost two of its five.

An AI model without the definition makes one up

If you go to three different analysts looking for a definition of “churn rate,” they will all agree that it’s an important metric and that the definition is self evident. Then they will go ahead and give you three different definitions.

(Steven H. Noble, Shopify Engineering, 2011)

Every fraction in this article looked self evident to whoever wrote it down.

Software makes the same mistake. In the GASP benchmark an AI model, Claude Sonnet 5.5, was asked for Net Revenue Retention for March 2026 with only the database tables to go on. In four runs of five it summed the start and end MRR columns of each subscription’s latest row and returned exactly 100%. The true figure was 98.2%. Given the standard’s written definition it returned 98.2% in all five runs.

How the GASP Standard solves it

Noble let the 30-day month go. Analysis, he reminded himself, “needs to work within the world that we are in”, where months are uneven and a customer can sign on 13 May and be gone by 20 June. A churn rate cannot be made immune to any of that. The choices behind it can be written down beside it, which is what kept Netflix’s figure standing in court.

The GASP Standard solves the problem by writing those choices down once, for everyone. Its churn definitions are canonical. Each rate has one formula, published in the open and compiled into the SQL a company can run, so the three analysts in Noble’s essay would hand back the same number. They are also realistic in the sense Noble meant, written for the customer who cancels six weeks after signing, for the downgrade, for the discount, for the pause and for the six-month project that ends on schedule.

Metrics must specify their level. “Churn rate” is ambiguous. Specify logo churn (customer), MRR churn (subscription) or seat churn (license).

(GASP Standard, Definitions)

The level is the first of eight things to state.

DecisionThe GASP Standard’s answer
Customers or moneyBoth, under two names: Logo Churn Rate and Revenue Churn Rate
Who is countedThe customers active on the first day of the period
What leaving meansMRR reaching zero. A pause counts and a return is Reactivation MRR
A customer who pays lessContraction MRR, with or without a change in service. It counts in Revenue Churn Rate and never in Logo Churn Rate
A planned endingCounted like any other and labelled, so the adjusted figures can set it aside
Gross or netGross. NRR carries the net view
The periodStated with the figure. A monthly rate reaches a year by compounding
The segmentThe one held on the first day of the period

A churn figure that arrives with those eight answers can be set beside another company’s or beside last year’s. One that arrives without them may be Noble’s August.

Sources

What the words mean

Plain wordsIn the GASP Standard
Customers lost, as a shareLogo Churn Rate, the market form of churn
Revenue lost, as a shareRevenue Churn Rate, the operating form of churn
Customers lost, as a countChurned Customers
Revenue from customers who leftChurned MRR
Revenue lost from customers who stayedContraction MRR
Revenue from customers who came backReactivation MRR
The net viewNet Revenue Retention (NRR)
Losses alone, as retentionGross Revenue Retention (GRR)
A change the contract will reverseChange nature on the ledger: temporary, then a scheduled reversal
Retention without the temporary movementsAdjusted NRR and Adjusted GRR
The customers of day oneA cohort: the customers active at period start
The record of each MRR movementThe Commercial Event Ledger

This article is also available as plain Markdown. More from Research.

Try searching for:

↑↓navigate↵select