Skip to content

Churn diagnostics · live-service games

The churn that matters is the one your metrics can't see.

Dividual is a diagnostic for live-service games. We find where your players actually leave, test which of your assumed causes is real, and tell you the one thing worth instrumenting — before you spend a quarter chasing the wrong fix.

The problem

Three things one retention number hides

Your retention number is one number. It hides at least three problems, and they need different fixes.

  1. Players don't all leave the same way.

    Some hit a cliff — full activity, then gone. Some fade over weeks. Your funnel averages them into one curve, and the average is nobody. The fix for a cliff and the fix for a fade are not the same.

  2. Your satisfaction metric may be grading its own homework.

    When the matchmaker optimizes for engagement and the satisfaction score is built from engagement, the score measures your own objective. A high number can mean the loop is working. Whether the player is actually happy is a different question, and that metric can't answer it.

  3. Your reporting cadence is biased.

    A quarterly read catches the churn that leaves loudly — friction, rage-quits, a broken update. It misses the churn that leaves quietly — the player who was satisfied into boredom and drifted off. Those two are opposite problems.

Saying these is free. Which one is true of your game, and what to do about it, is what the diagnostic answers.

The proof

We show our work where the truth is known

Any analyst can fit a churn model that scores well. None can show, on your data, that it found the right answer — because your data has no answer key.

So we prove the method on synthetic data where we planted the truth, then run the method blind to it. Three things it recovered:

  • The real shape of exit — the cliff-versus-fade split, read straight from the event stream.

  • The real cause, and only the real cause — it accepted the true driver (a large, unambiguous effect) and rejected a plausible decoy that had none.

  • It refused the confound — the sampling trap that hands a careless model a perfect, useless score, barred by construction.

The real shape of exit
Illustrative synthetic data — demonstration only
0 50 100 150 200 0.0 0.5 1.0 1.5 2.0 2.5 3.0 exit abruptness (A) players churned abrupt A ≥ 0.8 · recovered 40% (planted 53%)
Exits split by abruptness — a sharp-cliff tail sitting apart from the gradual fades. Read straight from the event stream, before any model is fit.
Two exit processes on different clocks
Illustrative synthetic data — demonstration only
0.00 0.05 0.10 0.15 0.20 0.25 0 5 10 15 weeks since acquisition cumulative incidence CIF₁ — exit type 1 CIF₂ — exit type 2
Competing-risks incidence: two kinds of leaving accrue on different clocks. Averaging them into a single retention curve hides both.
The real cause — and only the real cause
Illustrative synthetic data — demonstration only
Gates accept the true cause, reject the fake 0.0 0.5 1.0 1.5 2.0 2.5 social decline (true cause) performance (false cause) accept threshold |d| = 0.3 |Cohen d| It refuses the confound 0.4 0.5 0.6 0.7 0.8 0.9 1.0 recency (the trap) equal-exposure (honest read) chance = 0.5 AUC (churn)
Left: the validation gate accepts the true driver and rejects a plausible decoy that has no real effect. Right: a careless model scores a perfect AUC off a sampling artefact; barring it by construction returns the honest read at chance.

The same result holds under the survival-analysis toolkit a quantitative team expects by name: time-varying Cox regression, Aalen–Johansen competing risks, a random survival forest scored by Harrell's C and time-dependent AUC, and King–Zeng correction for case-control sampling. Run against a known truth, every one recovers the real cause and refuses the confound.

What you get

The diagnostic

A fixed-scope read on a slice of your telemetry, under NDA. Three to four weeks, one mid-point check-in, about an hour of your team's time to hand over data.

What you get
a short, plain-language report and a working session.
What it tells you
where your churn actually sits and in what shape; which of your assumed causes survive a validation gate and which don't; the confounds your data won't support; and the one thing worth instrumenting next.
What you do Monday
instrument that one thing, and run the experiment we pre-register with you to confirm the cause.

Fixed scope, fixed price, no black-box weeks.

Book a diagnostic call

How I work

How I work, and what it costs

Transparent by default — the pricing is on the page because the honesty is the point.

  1. A free second opinion.

    The method demonstration above, plus a 30-minute call reading your churn problem through it. No obligation.

    Free
  2. A pilot read

    a one-week look at a small slice, if you want proof on your own world first.

    ~$3k, or a founding-client arrangement
  3. The diagnostic Start here

    the fixed-scope engagement above.

    ~$12–25k, value-priced
  4. The build

    the fitted early-warning instrument and the causal experiments, once the diagnostic has scoped them.

    Project or monthly
  5. Ongoing advisory

    a churn-science retainer once we're working together.

    Retainer

Value-priced by the outcome. Every engagement is fixed scope, fixed price, fixed timeline, NDA-first.

Book a diagnostic call

Who

Who

AS

Adnan Selimovićcritical theorist and analyst. I wrote the framework this practice runs on: a formal model of how engagement, satiation, and churn actually work in interactive media, turned into a measurement program a quantitative team can check line by line.

The stance is rigorous and player-respecting at the same time, because in a market where players can always leave, those are the same thing. A captured player — satiated, quietly on the way out — is a player the game is losing. The diagnostic is built to catch exactly that, early enough to matter.

Why "Dividual."

To the loop, a player stops being an individual and becomes a dividual — a bundle of signals it can predict and keep feeding. We measure that prediction, find where it breaks, and give you an honest read of who's genuinely satisfied and who's just still clicking.

The churn that matters is the one you can't see. Let's find it.

Trouble with the scheduler? Email to book instead .