How to measure retention in a recovery studio

Ask three operators for their retention rate and you will get three numbers computed three different ways, none of them written down. That is not carelessness. It is the predictable result of dashboards that print a percentage next to the word retention without stating a denominator, a window or an event definition, and of a category where the underlying behaviour is genuinely harder to measure than in a gym or a salon.

The difficulty is real. A recovery member does not have a fixed weekly slot, does not necessarily attend on a schedule, and can be absent for three weeks without anything having gone wrong. So the question of whether someone has left is a judgement call rather than an observation, and every metric in this guide is an attempt to make that judgement call explicit and consistent instead of implicit and drifting.

None of this needs a data team. All of it needs you to decide what the words mean before you compute anything, which is why the definitions come first and the arithmetic comes second.

Churn, attrition and lapse are three different words

Churn, in the sense borrowed from subscription businesses, is the loss of a paying relationship: a membership that stopped billing, whether by cancellation, expiry or failed payment. It is a financial event, it has a date, and your payment system knows about it.

Attrition is net shrinkage over a period. It folds losses and gains into one figure, which makes it useful for planning and useless for diagnosis. A studio losing members fast and replacing them just as fast reports the same attrition as a studio nobody leaves. It is a far more expensive business, and the number cannot tell you which one you are.

Lapse is a behavioural event: a member who has stopped visiting, whether or not they are still paying. It has no automatic date attached, because nothing happens in any system at the moment a member stops turning up. Lapse is the leading indicator and churn is the trailing one, and in a recovery studio the interval between them can be months. If you measure only churn, you are measuring decisions that were made a quarter ago and are no longer available to influence.

Why the rate on your dashboard is a snapshot

The standard monthly churn figure divides the members lost during a month by the member base at the start or the middle of that month. That is a snapshot rate, and its weakness is the denominator: it includes members who joined a week ago and members who have been coming for three years, groups with completely different probabilities of leaving. Since new members are far more likely to leave than established ones, the composition of the denominator drives the result.

That misleads in both directions, and specifically. A studio in a growth phase, adding many new members, dilutes its denominator with the very cohort most likely to leave, and its snapshot rate can improve while its actual member survival is getting worse. A studio that has stopped growing sees the same rate deteriorate for the opposite reason, with no change whatsoever in how well it retains anybody. Two studios with identical retention behaviour can post very different snapshot rates purely because one is growing.

The same problem defeats month-to-month comparison, which is the main thing operators use the number for. A marketing push, a seasonal signup wave or a corporate account landing in one month changes the denominator's composition and therefore changes the rate, and you will attribute the movement to whatever retention work you happened to be doing at the time.

Metric one: cohort survival by join month

The fix is to stop mixing tenures. Group members by the month they joined, then track each group separately over time: of the members who joined in a given month, what share were still visiting one month later, three months later, six months later. Each cohort is a fixed population that only shrinks, so the denominator cannot drift and the comparison between cohorts is meaningful.

A cohort table is the single most informative retention artefact a studio can build, because it separates two things a snapshot rate fuses together. Reading down a column tells you whether members are surviving longer than they used to. Reading across a row tells you where in a member's life the losses concentrate, which is almost always earlier than operators expect. If most of the fall happens between the first and second month, no amount of work on your long-tenured members will change your overall numbers.

Build it on visiting rather than on paying, or build both. A cohort measured on active billing will show members as retained while they are quietly not attending, which is exactly the population you want to see. Measured on visits, the same cohort tells you when the behaviour stopped, which is the event you can still respond to.

Metric two: cadence as a distribution

Average visits per member per month is one of the least useful numbers a studio can compute, because the average sits in a valley that almost nobody occupies. A recovery studio's member base typically contains a group visiting frequently and a group barely visiting at all, and the mean of those two lands somewhere neither group lives. Watching that mean move tells you nothing about which group changed.

Report the distribution instead: how many members visited zero times last month, how many visited once or twice, how many visited more. The zero bucket is your unreported churn, sitting in plain sight and still being billed. The one-to-two bucket is the population whose value calculation is about to fail, because they are dividing a monthly rate by a visit count that no longer justifies it. Both are actionable this week; a mean is not actionable at all.

Track the shape over time rather than any single bucket. A studio doing effective cadence work sees members move up out of the low buckets, and a studio with a capacity problem at peak hours sees them slide down while nothing about member satisfaction has changed. The distribution distinguishes those two stories, which is the diagnostic work an average cannot do.

Metric three: lapse, defined against a cadence

Since lapse is a judgement call, the definition has to be written down and applied consistently, or the metric will drift every time someone new builds the report. The failure mode to avoid is a fixed threshold across the whole member base. Thirty days without a visit means something entirely different for a member on a twice-weekly plan than for a member who has always come once a fortnight, and a fixed rule catches the second group constantly while missing the first for weeks.

Define lapse relative to the member's own established pattern: a multiple of their typical interval between visits, with a floor so that a member with only two visits of history does not trigger on noise. Members without an established pattern, which means most members inside their first weeks, need a separate rule tied to the cadence you recommended at intake, not to a pattern that does not exist yet.

Then hold the definition still. The temptation, when the lapsed count looks bad, is to widen the threshold, and a metric whose definition moves with the mood of the person reporting it is not a metric. Write the rule into the query, date it, and treat a change to it the way you would treat a change to your accounting policy: rare, deliberate, and noted alongside the numbers it affects.

Metric four: next-step coverage

The three metrics above all report on what has already happened. The fourth is a leading indicator and it is the only one that tells you, today, whether next quarter's cohort will hold: the share of completed visits that ended with a next visit scheduled. Call it next-step coverage. It is computable from booking data, it responds within days to a change in front-desk behaviour, and it precedes every lagging retention metric you have.

Its value is that it is a measure of your operation rather than of your members. Cohort survival tells you that something went wrong two months ago; next-step coverage tells you whether the thing that causes it is happening right now, on this shift, at this counter. That makes it the number to put in front of whoever runs the floor, because it is the only one they can move today.

Praxium computes the same quantity inside its paid protocol layer and calls it sequence completion, which is why the term appears here at all. The measurement is yours regardless: any booking export carrying a member id, a visit timestamp and a future-booking flag will produce it in a spreadsheet.

Assembling the report when your platform will not

Most studio platforms will not compute a cohort table, and exporting your way to one is more tractable than it sounds. You need very little data, and the discipline is in defining the fields rather than in obtaining them.

• Export a visit table: member id, visit timestamp, service or modality. This alone supports cohort survival, the cadence distribution and lapse.
• Export a membership table: member id, start date, plan, status, and the date of any status change. This gives you tenure and the paying-versus-visiting split.
• Export failed and retried payments separately, because a member lost to a declined card is a different problem from a member who chose to leave.
• Write each definition into the query itself, with a date, and never into a spreadsheet cell someone will overwrite next quarter.
• Recompute the whole thing from raw exports each period instead of appending to a running sheet, so a definition change reprices history instead of creating a seam in it.
• Keep unaddressable exits, such as members who moved away, in a separate column rather than removing them, so the total still reconciles to your billing.

What not to spend the effort measuring

A few numbers absorb a lot of studio attention and repay very little of it. Satisfaction scores collected from members who are still attending mostly measure the politeness of people who like you enough to answer, and they move too slowly to guide anything. Total visits per month is a demand figure dressed as a retention figure: it rises with signups and falls with seasonality, and it can hold perfectly steady while your longest-tenured members quietly leave and are replaced.

Public review ratings belong in the same category for this purpose. A rating tells you something about how you are perceived by people choosing a studio, which is a real and separate question about discovery and reputation. It tells you almost nothing about whether the members you already have are still visiting, because the people writing reviews and the people lapsing are largely different populations at different points in the relationship. Watching a rating as a retention metric substitutes a number you can see for one you would have to compute.

The general rule is worth stating plainly, since it disqualifies most of what a dashboard offers by default: a retention metric has to be computed over a defined population of members across a defined window, with a defined event. Anything that lacks one of those three is a number, not a measurement, and it will move for reasons that have nothing to do with the work you are doing.

Reading small numbers without fooling yourself

A single-location studio deals in small monthly counts, and small counts move a great deal for no reason. A cohort of a few dozen members losing two more than usual in a month produces a swing that looks dramatic in a percentage and means nothing. The discipline is to read direction over several periods, to look at the underlying counts alongside every rate, and to resist redesigning an operation on one month of movement.

Be equally careful with outside comparison. Retention benchmarks circulate widely in this category and most of them arrive with no sample size, no date, no definition of the event being counted and no description of the studios behind them. A number missing any of those cannot be compared with yours, because you do not know whether it counts cancellations, lapses or net attrition. The standard to hold is the one applied to the single first-party figure on this page, which names the records it counts and the date of the freshest one; anything that will not state its own population and clock belongs in a conversation, not in a report.

The most useful comparison available to a studio is against itself. Cohort against cohort, quarter against quarter, with definitions that have not moved. That comparison is internally valid by construction, it needs nobody else's data, and it answers the only question the work is really for: is what we changed making members stay longer than they used to.