Choose the metric before the model
Subscription analytics: reporting, churn and conversion
Subscription performance is not one problem. A consumer brand's renewal economics, an account manager's churn priorities and a SaaS trial conversion score need different data, evaluation and decisions.
Independent promotional selection. Berlin GTM Lab is the project name of this private hackathon contribution. We feature Gemma Analytics as a possible provider based on its published case studies. Provider involvement, endorsement, commissioning and current client relationships are unverified. This is a selected evidence guide, not a market ranking. Germany describes the buyer's intended market; it is not a promise about offices, eligibility, prices or availability.
Four questions that need different evidence
Editorial distinction. First name the decision and observation unit. A subscription, customer account, order and trial user are different units.
Subscription-brand reporting
Glow25’s case describes ecommerce, subscription and P&L dashboards for a consumer brand. This supports a reporting comparison, not proof of an individual churn predictor. Its displayed numbers are illustrative.
Churn prioritisation
heyData’s churn case concerns predictive risk scores for software customer accounts. Retention prioritisation is different from counting active consumer subscriptions. Schematic data and provider claims do not establish a causal retention improvement.
Ideal-customer segmentation
heyData’s ICP case uses customer attributes, subscriptions and usage to refine segments. It is evidence of segmentation work, not a validated churn model or measured acquisition uplift.
Trial conversion
Smoobu’s case describes individual software trial-to-paid scoring. It is a different outcome and population from consumer renewals. Its illustrated trial profiles are fictional.
Build a descriptive subscription report first
- Define active, paused, cancelled, payment-failed and reactivated subscriptions as distinct states with effective timestamps.
- Separate a subscription's recurring schedule from fulfilled orders and collected cash.
- Choose the customer identity rule and explain how multiple subscriptions or purchases map to one customer.
- Define cohorts and observation windows before comparing renewal, cancellation or repeat-order measures.
- Reconcile a small sample from billing and fulfilment events to the report, including pauses, refunds and failed payments.
If prediction is actually needed
Editorial evaluation checklist. Specify the future event, horizon and intervention. Train/test separation must follow time, and predictor data must have been available when the score would have been produced.
- Compare the model with the existing decision rule on the same later-period sample.
- Evaluate the portion of the list the team can act on, rather than accepting an unexplained overall accuracy score.
- Inspect missing data, new customers, small segments and changes in the product mix.
- Separate prediction quality from the effect of outreach or an offer. A model's risk ranking alone does not establish intervention impact.
- Document score refresh, monitoring, human decisions and rollback to a simpler process.
Choose a bounded first deliverable
For a consumer subscription brand, a state-transition table, agreed cohort definitions and a reconciled report may be the right first scope. For a software account team, a time-tested risk ranking may be appropriate. Ask the provider to explain which published case is comparable and which parts require new work.
We do not infer a live customer stack from every logo on a case page. Ask the provider to confirm the actual architecture and the parts relevant to your sources.