How is customer retention rate calculated?

There is no single calculation behind the phrase. Seven are published right now — in platform documentation, in vendor glossaries, and in the third-party guides that answer the question.

They are not variants. Each counts a different population, so one store can report a different figure under each. Four benchmark numbers circulate for the name, and they attach to some of these calculations and not to others, on bases that cannot be compared. One of the seven is published in two incompatible readings, so there are eight distinct quantities in play.

Two companies each publish two of them. One formula can return more than 100%.

Customer retention rate, as published on 2026-08-19. Klaviyo’s glossary divides repeat buyers by one-time buyers. Shopify’s returning customer rate divides returning customers by every customer who ordered in the period. Zendesk, Triple Whale, LoyaltyLion and Gorgias publish the same arithmetic in four different renderings — end customers minus new, over start — where “End” means a customer base to Zendesk and the period’s buyers to LoyaltyLion. Recharge divides churned subscriptions by average daily active subscriptions.

What does “retention rate” name?

Seven calculations get offered as the answer. A share of a period’s buyers. A cohort’s survival. A two-segment repeat rate on fixed windows. A ratio of repeat buyers to one-time buyers. The (End − New) ÷ Start formula, which has two incompatible readings. Subscription churn, counted on subscriptions. A customer-level churn complement.

Five of them are published under the name. Two are not: one is called a returning customer rate, the other a churn rate. Both get read as retention anyway.

Each is a defensible thing to measure. The problem is not any one of them.

The problem is that they share a label. A figure and the benchmark it is held against can rest on different calculations, and the name of the metric does not say which.

What does a store dashboard’s returning customer rate measure?

Shopify’s field reference defines its Returning customer rate as “Percentage of returning customers relative to all customers who placed orders”, and prints the arithmetic beside it: “Formula: returning customers / customers”.

That is a composition figure. It answers what fraction of the people who bought this month had bought before.

The same reference states the range: “Higher rates reduce acquisition costs. Most stores display 20-40%.”

What else moves a returning customer rate?

The denominator contains the numerator. Customers is defined as “Number of customers who placed an order”, and returning customers are a subset of them.

So the ratio falls when acquisition rises, with retention unchanged. A strong new-customer month pushes the figure down.

Read as a retention signal, that is backwards. Read as a composition signal, it is exactly right.

What does a cohort retention rate measure?

The same document carries a second metric under a second name, Customer retention rate: “Percentage of customers in the given cohort who placed an order in the given period”.

Three sibling variants repeat that definition string. This is a survival rate, not a share.

Two differently defined metrics sit in one vendor’s field reference — one of them named a retention rate, one of them not — and neither name warns you about the other.

Does a second order in the same month count?

It does, in period 0. Shopify groups customers “into cohorts based on the date that they placed their first order”, and the first column counts activity inside that same period.

The documentation gives its own example:

“John made their first purchase in February 2022. John then made repeat purchases in February 2022 again, and in June 2022, and in September 2022. In a monthly cohort analysis for 2022, John would be in the February cohort and would be counted as a repeat customer for Month 0, Month 4, and Month 7.” — Shopify Help Center, Customers reports, 2026-08-19

So a cohort’s period-0 figure reports fast second orders, not long-run loyalty. The later columns are where survival actually shows.

How do you measure retention with two segments?

Klaviyo’s help centre builds it from two segments. The first is customers “who purchased from your store last year (between 365 to 730 days ago) who have also purchased recently (e.g., in the last 6 months)”.

The second drops the recency condition: “customers who purchased last year (between 365-730 days ago)”.

Divide one by the other. The published example puts 2,500 profiles over 10,000, which is 25%.

The six-month condition is deliberate. Without it, the guide says, “you risk including customers who made a second purchase immediately after the first, and who then never returned to your store.”

Why does one vendor publish two formulas?

The same company’s glossary defines the metric differently: “[customers who have purchased more than once] / [customers who have purchased only once] x 100 = customer retention rate”.

That is not the help centre’s calculation. It has no time window, no cohort, and a denominator built from a different population.

The glossary links to that help-centre guide as further reading on the metric it has just defined another way.

Why can one retention formula go above 100%?

“Purchased more than once” and “purchased only once” cannot overlap. The denominator is not the total; it is the other group.

So the quotient passes 1 whenever repeat buyers outnumber one-time buyers. A healthy store can publish a retention rate above 100%.

A rate whose ceiling depends on which group is larger is not measuring survival. It is a composition ratio wearing a percentage sign.

What do the same customers look like under a second formula?

The glossary supplies both counts in its own worked example: 36,000 customers who purchased more than once, and 56,000 who “have only made one purchase within the last year”. Its formula returns 64%.

Now change one thing. Divide the same 36,000 by every customer in the example rather than by the one-time buyers alone. That is 36,000 ÷ 92,000, or 39%.

Same two counts, 25 percentage points apart. Nothing about the store moved — only the denominator did.

That is the whole mechanism. Klaviyo’s glossary divides by a population that excludes its repeat buyers. Klaviyo’s help-centre segments and Shopify’s cohort report divide by populations that include them. The two shapes answer different questions, and both are printed as retention rates.

Why does retention jump when nothing changed?

Changing the denominator puts a step in the trend line, and the step looks like a result.

The gap between two denominators on identical counts was 25 points in the example above. That gap comes from the formulas alone. No customer behaviour is involved in producing it.

So a store that switches reports, adds a tool, or inherits a new dashboard sees its retention “improve” or “collapse” on a date when nothing about its customers changed.

That gives one cheap diagnostic. When the number jumps, check whether the report definition moved before looking for a cause in behaviour.

And it gives one rule worth holding. Record the numerator and denominator next to the figure, with the date the basis was last changed, so a chart cannot silently splice two measurements together.

Is the standard formula actually standard?

Outside the platforms, the advice layer converges. Four unrelated guides publish the same arithmetic.

Zendesk writes it ”[(E-N)÷S] x 100”. Triple Whale writes “CRR = ((End-New)/Start) x 100”. LoyaltyLion writes ”( E-N/S)*100”. Gorgias spells the same arithmetic out in words.

That looks like consensus. It is agreement on the arithmetic, not on what the letters count.

Do those four guides count the same customers?

They do not. Zendesk defines E as “The number of customers at the end of a given period” and S as “The number of customers at the start of a given period” — a base, held at two moments.

LoyaltyLion defines E as “the total number of customers who made a purchase in a time period” and S as “the number of customers who had made a purchase during a comparative period” — buyers, counted inside two windows.

Same letters. Same arithmetic. Different measured objects.

A customer base is a headcount held at a moment. A period’s buyers is a count of who ordered between two dates. Which one a store can produce decides which reading it is using.

Which reading do a store’s period reports populate?

Shopify’s period fields are counts of buyers: “Number of customers who placed an order”, and “Number of new customers who placed an order”, noted as “First-time purchasers only.”

Those are buyers-in-a-window quantities. Shopify’s period reports publish who ordered, not a customer base standing at a date.

So an operator plugging dashboard numbers into the standard formula is using one guide’s definition of the letters, whichever guide they read.

Are the worked examples wrong?

No. Each published example is arithmetically correct on its own inputs.

Triple Whale’s “((225 - 45) ÷ 285) x 100 = 63%” is right. Zendesk’s 90% is right. LoyaltyLion’s 33.33% is right. Gorgias’s ”[(80 - 45) / 100] x 100”, stated as 35%, is right.

They differ because they describe different situations, not because anyone slipped. That is what makes the divergence hard to notice.

How do subscription tools count retention?

They change the unit. Recharge’s dashboard metric is measured on subscriptions: “The churn rate is defined as the percent of average daily active subscriptions churned within the period.”

Two things move there at once. The object counted is a subscription, not a person. The denominator is an average across days, not a headcount at either end.

A store averaging more than one subscription per customer is therefore counting a different population than any customer-level rate counts.

Is a cancelled subscription a lost customer?

Recharge treats that as a separate question, and reports the answer as its own metric: “The percentage of subscription cancellations that resulted in customer churn.”

It draws the line explicitly across two dashboards. On the customer view, “The customer is only considered churned after all subscriptions are cancelled.” On the subscription view, “Cancellations are tracked per subscription and attributed to their original cohort.”

A seventh calculation sits one layer up. Polar Analytics computes “Retention Rate is 1 - (Churned Customers / Total Customers)” from that same subscription data, at customer level.

What are the published benchmarks attached to?

Four figures circulate. Only one is published by the page that also defines the calculation behind it.

Three of the four are quotations. A quoted figure does not carry the formula of the page quoting it, so the calculation behind each of those three is whatever the original used.

That is the practical problem. A store cannot check whether a benchmark was built the way its own number was.

Where does the industry table come from?

Both publishers print the same fourteen percentages: 84%, 84%, 83%, 83%, 81%, 80%, 78%, 78%, 77%, 77%, 75%, 67%, 67%, 55%, under industry labels whose wording differs between them.

Then each adds a fifteenth row of its own. Exploding Topics adds retail at “63%”. Triple Whale adds “Ecommerce: 30%” and carries no retail row.

So the only two rows that name a store’s own sector — retail and ecommerce — are the two that are not in the shared list.

Where that list came from is itself disputed. Triple Whale credits it to Statista, whose page is a paid statistic titled “Customer retention rate of businesses worldwide in 2018, by industry”, drawn from 468 respondents. Its public summary matches the list at both ends: retention is “highest in the media and professional services industries”, lowest in “hospitality, travel and restaurants”. Exploding Topics’ section credits Customer Gauge and Aspect instead.

Neither appended row traces to a figure a reader can read. Exploding Topics’ two reports cannot be reached. Statista’s values sit behind a subscription, so whether its 2018 table carries an ecommerce row at all is not checkable from outside.

Can you check how a benchmark was calculated?

For three of the four, no. Omniconvert’s benchmark page, the source Gorgias credits for 31%, now redirects to its homepage. Statista’s figures sit behind a paid subscription. Of the two reports Exploding Topics names, one refuses the connection and the other lands on a resources index.

None of that makes the numbers false. It means the calculation behind each one cannot be read, and that is the single thing a benchmark has to disclose to be usable.

A benchmark whose method you cannot see is a number, not a standard.

Which number answers which question?

Pick the basis from the decision, not from the dashboard.

The second row reads like the first and is not. It moves when acquisition moves, because acquisition sits in its denominator. A cohort’s survival is computed inside one cohort.

What should you check before comparing?

Three checks.

Write down the numerator and denominator your figure actually uses. Not the metric’s name — the two counts.

Do the same for the benchmark. If its basis is not published, the comparison has no meaning and the benchmark should be dropped.

Then hold one definition still. A retention figure is worth far more compared against the same store last quarter than against any published average.

When does the definition stop mattering?

Some catalogues make the whole question quieter. Klaviyo’s own glossary puts it directly: for a business built on “expensive one-time purchases”, it says, “a lower retention rate is less concerning because most consumers don’t need to buy a mattress at the same frequency as buying a new shirt or face cream.”

And if one calculation is used everywhere internally, the label matters less. What breaks comparison is a change of denominator, or a change in what the letters count — which is why the name on the metric settles nothing.