# Customer health scores for marketers > What belongs in a health score, building one without a CS platform, syncing bands into campaign triggers, and backtesting the score against churned cohorts. Source: https://saas-marketing.net/guides/customer-health-scores-for-marketers/ Topic: SaaS Customer Marketing Type: guide Published: 2026-09-11 Last updated: 2026-09-11 Publisher: SaaS Marketing (saas-marketing.net) License: CC BY 4.0. Quote or republish with attribution and a link to https://saas-marketing.net/guides/customer-health-scores-for-marketers/ ## Short answer A customer health score combines usage breadth and depth, champion and admin presence, support sentiment, invoice health and advocacy activity into a single number used to band accounts by renewal risk. Marketers usually inherit one they did not build. Before campaigning against it, backtest it against last year's churned accounts: if the score did not separate churned from retained accounts 90 days out, it has no predictive power and should be rebuilt around two variables instead of twelve. ## Key takeaways - Usage breadth, meaning how many distinct features and users are active, predicts renewal better than total usage volume. - Champion departure is the single strongest churn signal in most B2B SaaS, and most health scores do not track it. - You can build a usable score with three warehouse queries and four CRM fields. No Gainsight required. - Sync bands, not raw scores, into your marketing automation platform, so campaign segments update without manual list pulls. - Backtest against last year's churned cohort. If the AUC is near coin flip, your score is decorative. - Most health scores are unvalidated composites of whatever data was available when someone built them. --- Somebody in customer success built a health score two years ago using the fields that happened to be available, weighted them by intuition, and shipped it. Nobody has checked since whether it predicts anything. Now marketing is being asked to run retention campaigns against its bands. That is the normal situation, and it is fixable in about two weeks. Start by finding out whether the score you inherited works at all, because the answer determines everything else. ## What actually belongs in a health score? Five input families, and they are not equally useful. Most scores fail by including twelve inputs where three carry the signal, because weakly correlated variables dilute the strong ones. Champion tracking deserves special attention because it is the highest value input most teams do not have. When the person who bought your product leaves the company, renewal risk jumps sharply, and the signal is available cheaply: an email bounce, a LinkedIn title change, a login that stops. A large share of B2B SaaS churn has a champion departure in the preceding quarter, and it is often the only signal that appeared before the notice. Add a rule that flags an account when a contact tagged as champion or admin has a bounced email or no login for 30 days. It costs almost nothing to implement and in most accounts it surfaces more genuine risk than the rest of the composite combined. ## How do you build one without Gainsight or Totango? Three queries, four CRM fields, one scheduled job. Roughly two weeks of work for one analyst and one marketing ops person. The definitional groundwork is in [what is a customer health score](/glossary/customer-health-score/), and what follows is the build. **A two week health score build** Why bands rather than raw scores in the marketing platform? Because a raw number drifts constantly, which means every dynamic list recalculates daily and campaign entry becomes noisy. An account bouncing between 61 and 64 does not change treatment. An account dropping from Healthy to Watch does. **2 to 4** Inputs in a health score that typically outperforms a twelve variable composite ## How do you backtest the score against churned accounts? This is the step that separates a useful score from a decorative one, and almost nobody does it. The method takes an afternoon. Take every account that reached a renewal decision in the last twelve months. Recompute the score as it would have looked 90 days before that renewal date, using only data available at that time. This is the important constraint: if you score them using data from after the churn event, you will prove the score works and be wrong. Then compare distributions. Put churned accounts and retained accounts side by side. | Backtest result | Median score, churned | Median score, retained | Verdict | | --- | --- | --- | --- | | Strong separation | 34 | 71 | Trust it, build campaigns | | Weak separation | 52 | 61 | Rebuild around the two strongest components | | No separation | 58 | 60 | The score is decorative, stop using it | | Inverted | 66 | 55 | Something is wrong in the data pipeline, investigate | If separation is weak, run the same comparison on each component individually. You will usually find that one or two components separate cleanly and the rest are noise that was cancelling them out. Rebuild the score with just those, accept that it looks crude, and note that a crude score that predicts beats an elaborate one that does not. Scoring churned accounts using their final month of data will show beautiful separation, because usage collapses right before churn. That tells you nothing useful. You need a 90 day warning to act, so score at 90 days out or the whole exercise is theatre. Pair the backtest with a qualitative pass. Read the churn notes on twenty lost accounts and check whether the reason recorded there was something your score could ever have seen. If half the churn was caused by an acquisition or a budget cut, no health score will predict it, and you should size your retention programme accordingly. The diagnostic sequence in [churn risk audit checklist](/checklists/churn-risk-audit/) covers that review properly. ## What campaign belongs in each band? Different bands need different treatment, and the biggest error is sending the at risk band an email sequence. Healthy, roughly 70 and above. These accounts get expansion offers, seat growth nudges, advanced feature education and advocacy asks. This is where your reference and review requests should come from, which is the natural feed into a [customer advocacy program](/playbooks/customer-advocacy-program/). Expansion plays for this band are covered in [expansion marketing plays](/playbooks/expansion-marketing-plays/). Watch, roughly 45 to 70. The most valuable band and the most neglected. These accounts are using the product but narrowly. The right campaign is specific: name the feature they are not using, explain the outcome it produces for accounts like theirs, and invite them to a 30 minute workshop. Generic newsletters do nothing here. At risk, below 45. Do not send a campaign. Send a person. An automated re engagement email to an account whose champion left and whose tickets are unresolved reads as tone deaf and can accelerate the decision. Marketing's job for this band is to arm the CSM with the reason and the relevant asset, not to email the account. The sequencing for this sits in [churn prevention marketing](/playbooks/churn-prevention-marketing/). That last row is worth building as a separate trigger rather than folding into the score. It is actionable in a specific way, it has a clean play attached, and the window to act is short. ## How do you keep the score from rotting? Ownership and a review cadence. Health scores decay because the product changes, new features launch, a feature in the breadth count gets deprecated, and nobody updates the query. Four maintenance commitments, and they are small. - One named owner, reviewed in the same meeting as pipeline. If nobody's name is on it, it is nobody's. - Quarterly backtest on the trailing twelve months of renewals. Fifteen minutes once the query exists. - Feature list review whenever the product ships something significant, because the breadth count needs updating. - An annual decision about whether to simplify. Scores accumulate inputs over time and almost never lose them, which is how a three variable model becomes a fourteen variable one that predicts worse. The broader programme context sits in [customer lifecycle marketing](/glossary/customer-lifecycle-marketing/), because a health score is only useful inside a lifecycle where somebody acts on each band. ## The honest limits A health score is a probability estimate built on the subset of reality your systems happen to record. It cannot see a new CFO running a software consolidation review, a competitor's aggressive displacement offer, or the acquisition that will consolidate your buyer onto another platform. In many portfolios those causes account for a meaningful share of churn, and no amount of modelling will surface them in advance. It also creates a real risk of self fulfilling neglect. If at risk accounts get de prioritised because they look likely to leave, the score becomes a prophecy rather than a prediction. Watch for that by comparing renewal rates of accounts that entered the at risk band and received intervention against those that did not. And the score will produce false positives. Seasonal businesses drop in usage every year without any renewal risk. Build a seasonality exception before your first campaign rather than after an awkward email lands in a customer's inbox during their quiet quarter. ## What to do this week Run the backtest on your existing score before building anything new. It takes an afternoon and it tells you whether the last two years of health score reporting meant anything. If separation is weak, rebuild around the two strongest components and add champion departure tracking, which is usually the highest value addition available. Then sync bands rather than raw scores into your marketing platform, assign a campaign per band, and put the quarterly backtest in someone's calendar. Wider programme design and worked examples are in [SaaS customer marketing](/saas-customer-marketing/), [customer marketing campaign ideas](/guides/customer-marketing-campaign-ideas/) and [SaaS customer marketing examples](/examples/saas-customer-marketing-examples/). ## Frequently asked questions ### What is a customer health score? A customer health score is a composite number, usually 0 to 100, that estimates how likely an account is to renew and expand. Typical inputs are product usage breadth and depth, presence of an active admin or champion, support ticket volume and sentiment, payment history and advocacy behaviour. It is used to band accounts into groups like healthy, watch and at risk, and to trigger different treatment for each. ### What inputs actually predict renewal? Usage breadth across distinct users and features, plus admin and champion activity, carry most of the predictive weight in B2B SaaS. Support ticket sentiment and invoice health add a little. Login counts and total session time add almost nothing once breadth is accounted for, because a single heavy user does not protect a renewal decision made by their manager. ### Can marketers build a health score without a customer success platform? Yes. Three warehouse queries covering active users per account, distinct features used and days since last admin action, joined to CRM renewal dates and support ticket counts, will get you a workable score. Write four fields back to the CRM and sync bands to your email platform. That is a two week build, not a six figure purchase. ### How do you validate a customer health score? Backtest it. Recompute the score as it would have looked 90 days before renewal for every account that came up for renewal last year, then compare the distributions for churned and retained accounts. If the median score of churned accounts is not clearly lower, the score is not predictive and no campaign built on it will work. ### What campaigns should each health band get? Healthy accounts get expansion and advocacy asks. Watch accounts get education on the feature they are not using and an invitation to a workshop. At risk accounts get a human, not an email sequence, because a nurture campaign sent to an account that hates you accelerates the exit rather than preventing it. ### How often should health scores refresh? Daily for usage inputs, weekly for the composite score, and never more than monthly for any manual inputs like a CSM sentiment rating. Faster refresh does not mean better decisions, but a score more than a month stale will route campaigns based on a situation that has already changed. ### Why do most health scores fail? They are built from whatever data was available rather than from what predicts churn, they include too many weakly correlated inputs that cancel each other out, they are never validated against actual churn outcomes, and nobody owns them after the person who built the model leaves. The fix is fewer variables and one honest backtest.