SaaS retention analysis
Fix what counts as churn, then test what saves it
A list of lost accounts is not yet a diagnosis. This page walks through the definitions, cohort checks and risk-rule backtests that sit between that list and a decision, and the holdout test that tells you whether a save play worked. ZeroTwo drafts the analysis from the files you attach. People approve every offer.
Everything below is method and placeholders. No customer data or results are shown.
Churn definition sheet
- Churn type
- Logo, or revenue (gross or net)
- Eligible population
- Paying accounts active on [start date]
- Observation window
- [start date] to [end date]
- Loss type
- Voluntary, or involuntary (failed payment)
Out-of-time check
History used for scoring
Only data that existed before the score date
Score date
Rule frozen
Outcome window
Who actually churned
Illustrative layout, not a result.
Six choices that change the churn number
Churn rates are only comparable when the definition stays fixed. None of these is a standard. Write yours down once, attach it to every export, and keep it unchanged between periods.
| Choice | What you decide | Why it changes the answer |
|---|---|---|
| Logo vs revenue churn | Count accounts lost, or count recurring revenue lost (gross, or net of expansion). | A few large accounts can move revenue churn while logo churn stays flat, and the reverse. |
| Eligible population | Which accounts could have churned in the window, and where free and trial accounts sit. | Accounts that never paid are conversion misses, not paid churn. Mixing them changes both the numerator and the denominator. |
| Observation window | Start and end dates, and whether a cancellation counts on the request date or when the paid period ends. | A 90-day window and a calendar quarter give different rates for the same accounts. |
| Voluntary vs involuntary | Cancelled on purpose, versus lost to a failed payment or an expired card. | A failed card needs a payment-recovery flow, not a save email. Blending the two hides which fix applies. |
| Reactivations and pauses | Whether an account that returns inside the window still counts as churned. | Settle it before counting, or the rate depends on who is counting. |
| Missing data | How accounts with no usage export, no ticket history or no exit reason are reported. | Give them their own row so that unknown does not quietly become not at risk. |
From an association to a claim you can act on
The riskiest sentence in a churn memo is a cause stated as fact, such as churn is an activation problem, not a pricing problem. A count of accounts shows an association. Three labels keep each conclusion honest about how far it has been checked.
- Observed
Accounts that never finished setup churned more often than accounts that did.
It needs
Both counts and both denominators, the window, and the definition of finished setup.
Write it as
[n] of [N] without setup churned, against [m] of [M] with it, over [window].
- Hypothesis
An incomplete setup is one reason accounts leave.
It needs
A second source such as exit notes, ticket text or interviews, plus a check on other explanations like plan, company size and acquisition source.
Write it as
Likely driver: [x]. Alternatives checked: [y]. Not yet checked: [z].
- Tested
A guided setup session reduces churn for similar accounts.
It needs
A holdout: a comparable group that did not get the session, chosen at random before anyone was contacted.
Write it as
Offered group [x] of [N] churned; holdout [y] of [M] churned, over [window].
Backtest a risk rule before you route people to it
Declining logins, quiet accounts and frustrated tickets sound like warning signs. Whether they predict cancellation in your product is a measurable question. Run the rule against the past, out of time, and read the result before it ranks anyone. The AI agent for data analysis covers the calculation workflow itself.
Freeze the rule
Write the rule down with its thresholds, for example no login for a set number of days or a failed payment, before looking at outcomes.
Score as of a past date
Apply the rule using only the data that existed on that date. Anything recorded afterwards leaks the answer into the score.
Read the outcome window
List which accounts actually churned in the period after the score date, using the same churn definition as everywhere else.
Report precision, recall and coverage
Show the counts behind each measure below, not only the percentages.
Compare with a plain baseline
A rule that cannot beat a simple one, such as any account with no activity lately, is not worth routing people to.
| Measure | The question it answers | How to calculate it |
|---|---|---|
| Precision | Of the accounts the rule flagged, how many churned? | Flagged and churned, divided by flagged. |
| Recall | Of the accounts that churned, how many did the rule flag? | Flagged and churned, divided by all churned. |
| Coverage | How many accounts could not be scored because inputs were missing? | Count and share of accounts with a missing input. |
| Baseline gap | Does the rule beat a simpler comparison on the same accounts? | Precision and recall of the rule next to the baseline. |
A rule tuned on the same quarter it is judged on will look better than it is. That is why the score date comes before the outcome window.
Match the save play to the loss, and name who approves it
A save email is the right tool for some losses and the wrong one for others. Discounts and release promises are commitments, so each row names the person who has to sign off before anything is sent.
| Segment | Check before acting | Play worth testing | Approval needed from |
|---|---|---|---|
| Setup or activation gap | Compare with accounts on the same plan and cohort that did finish setup. | Offer a guided setup session and measure it against a holdout. | Customer success lead |
| Switched to another product | Read the exit notes and tickets. Check every claim about the other product on its own site. | A factual comparison in which each claim about the other product has a source. | Product marketing owner |
| Price objection | Ask whether the account is on the right plan for its usage, and whether the real issue is value not seen. | A plan-fit review first. A discount only after finance has approved the terms. | Finance or pricing owner |
| Missing feature | Count how many accounts cite the same gap, and whether the feature exists but was not found. | A reply that states what exists today. No release date unless product has committed to one in writing. | Product owner |
| Payment failure | Confirm the failure, and check which retry and card-update messages went out. | A payment-recovery flow with retries and a card-update reminder, not a retention offer. | Billing owner |
| No reason recorded | Count them. Do not assign a reason by guessing. | A short exit question, with this group kept as its own row. | Lifecycle email owner |
Test the play with a holdout
Accounts you save might have stayed anyway. Only a comparison with a similar group you left alone shows what the play added.
Fix the eligible list
Use the churn definition from the first section, so the list and the later count mean the same thing.
Split at random
Divide the list into an offered group and a held-out group before anyone is contacted. Do not choose who gets the offer.
Write the outcome first
Name the outcome and its window, for example still paying [N] days after the message, before the first send.
Send to the offered group only
Leave the held-out group alone. Compare both groups when the window closes and report counts as well as rates.
Small lists
A short list proves little
If only a handful of accounts qualify, a holdout cannot tell you much. Report the counts, call the result a lead, and confirm it in conversations before you scale the play.
What ZeroTwo does here, and what stays with you
ZeroTwo can
- Read the exports you attach: billing history, usage, ticket text and exit notes.
- Calculate churn, cohort retention and rule precision from those files, and show the formulas so you can check them.
- List the assumptions it made and the rows it could not use.
- Draft the memo and the save emails for a person to review and approve.
Stays with your team
- Decide who receives an offer, or approve a discount or a release promise.
- Prove that one factor caused an account to leave. Only a test can do that.
- Send emails, or change anything in your billing system.
- Make up data your files do not contain. Gaps should come back as gaps.
Cohort work usually needs several exports at once. On the Free plan you can attach one file per message and start with 100 free credits. Plus ($14.99/mo) and above allow multiple file uploads per task. See pricing for the full comparison.
Prompts that carry the definitions with them
Each prompt asks for counts, formulas and exclusions, so the answer can be checked against your source data. Replace the bracketed parts and attach your files.
Define and count churn
Using the attached billing export, compute logo churn and gross revenue churn for [window]. State the eligible population, how you treated free and trial accounts, and split voluntary from involuntary loss. Show the formulas and list every row you excluded.
Open this prompt in ZeroTwoGroup by reason, with caveats
Group these churned accounts by the reason in the exit notes: setup gap, switched product, price, missing feature, payment failure, none recorded. Give counts, not only percentages, quote one line per group, and mark any group smaller than [n] as too small to rank.
Open this prompt in ZeroTwoBuild the cohort table
Build a monthly cohort table from signup date to churn date. Report retention by month since signup for each cohort, flag cohorts with fewer than [n] accounts, and describe the pattern without proposing causes.
Open this prompt in ZeroTwoBacktest a risk rule
Here is a rule for at-risk accounts: [rule]. Apply it using only data available on [date]. Then compare it with what happened in the following [window]: report precision, recall, accounts that could not be scored, and a comparison with a simple baseline.
Open this prompt in ZeroTwoDesign a save-play test
Design a save-play test for [segment]: the eligible list, a random split, the offer, the outcome and its window, and the group size below which the result would be too small to conclude anything. Mark any discount or release promise as needing approval from [owner].
Open this prompt in ZeroTwoDraft a save email for review
Draft a save email for an account that left because of [reason]. Use only the facts I give you: [facts]. Do not offer a discount or promise a release date. Leave a clearly marked placeholder where an approver must add terms.
Open this prompt in ZeroTwoWrite the churn memo
Write the churn memo: definitions used, results with counts, what is only associated and what has been tested, data gaps, and the next experiment. Label every conclusion Observed, Hypothesis or Tested.
Open this prompt in ZeroTwo
Questions that come up in churn reviews
What is the difference between logo churn and revenue churn?
Logo churn counts accounts lost. Revenue churn counts recurring revenue lost, either gross or net of expansion from other accounts. They can disagree: one large cancellation moves revenue churn a lot and logo churn very little. Pick both, define each, and report which one a conclusion is about.
Does a drop in logins mean an account will churn?
Not by itself. A drop can reflect a finished project, a holiday or a change of owner. Treat it as a candidate signal, then test it as a rule against past outcomes before using it to rank accounts.
Can ZeroTwo predict which accounts will churn?
ZeroTwo can help you write a risk rule, apply it as of a past date and report precision, recall and coverage from your own files. It is not a validated churn-prediction model, and this page does not claim it forecasts churn for your business.
How many churned accounts do I need before the segments mean anything?
There is no fixed number. With small groups, report counts instead of percentages, do not rank segments against each other, and treat each finding as a hypothesis to check in interviews or a test.
Should every cancelling account get a discount?
No. A discount changes revenue and sets expectations, so have the terms approved first and test the offer against a holdout. Some losses, such as a failed payment or a poor fit, should not get one at all.
Where does recurring metric reporting belong?
This page covers diagnosing churn and testing save plays. Weekly metric delivery sits with the SaaS KPI review use case, and competitor evidence with the competitive research use case.
Neighbouring workflows
- Onboarding and activation
If setup is where accounts stall, work on the onboarding path itself.
- Customer discovery and ICP
Exit interviews are discovery data. Turn them into themes you can test.
- Weekly KPI review
Recurring metric delivery lives here, once your churn definition is fixed.
- Competitive research
For evidence about competitors, source it separately from your churn data.
- AI agent for data analysis
The calculation workflow behind cohort tables and rule backtests.
- Support triage
Ticket text is one of the inputs to a risk rule. Sort it before you use it.