Predicting and Preventing Subscription Churn on Shopify: The Exact System to Build

The churn signal architecture — what data you actually need

Before building any model, you need to identify which data signals actually predict churn. The inputs that matter most, roughly in order of predictive power:

1. Order frequency deviation: if a customer's subscription is set to 30-day delivery but their previous 3 subscription gaps averaged 38 days (suggesting they're skipping or pausing more), that's a churn signal. Customers who pause are meaningfully more likely to cancel than those who don't.

2. Email engagement decline: a subscriber who opened every email for 3 months and then stopped opening is sending a signal. Klaviyo's engagement data feeds directly into the model — a sharp drop in email engagement over a 30-day period is one of the strongest individual churn predictors.

3. Subscription portal activity without action: a subscriber who visits the cancel page but doesn't cancel is in decision mode. Connecting ReCharge's customer activity data to the model is technically more complex but worth the effort — it's the clearest signal of imminent churn.

4. Customer service ticket history: customers who submitted a complaint in the last 60 days that wasn't resolved to their satisfaction churn at a meaningfully higher rate than baseline. Syncing support ticket data from Freshdesk or Zendesk into the churn model adds a genuinely useful signal.

5. Product usage frequency signals: for consumables, if a customer's declared usage pattern (captured in the onboarding quiz) drops off, the expected product usage rate changes — and so does the reorder frequency that should trigger a check-in.

Building the churn score without an ML engineer

For most D2C brands, a full custom ML model is overkill. A weighted scoring system using Klaviyo's predictive analytics combined with custom properties gets you most of the model quality in a fraction of the time.

Here's the scoring framework we use:

Step 1 — Create custom Klaviyo properties: for each churn signal, create a custom property that scores it 0–10 (0 = no signal, 10 = strong churn indicator). These properties update automatically via Klaviyo flows triggered by each signal event.

Step 2 — Build a composite churn score: use Klaviyo's "Update profile property" action to calculate a composite score (weighted sum of individual signals). A typical weighting: email engagement decline (30%), order frequency deviation (25%), portal activity (20%), support history (15%), usage signals (10%).

Step 3 — Set intervention thresholds: subscribers with a composite score above 60 enter a "watch list" (monitoring + gentle onboarding content). Above 70 → light intervention. Above 80 → full intervention sequence. Above 90 → escalate to a human team member for a personal outreach.

Step 4 — Validate and calibrate: after 60 days, compare predicted churn scores against actual churn events. Adjust the weightings to improve precision. The model improves continuously as you accumulate more labelled data.

The intervention sequence that recovers at-risk subscribers

The intervention sequence is not a generic "we miss you" email. It's personalised to the specific subscriber's history, subscription tenure, and the signal that triggered the intervention.

Message 1 — Personalised check-in (triggered at 70+ churn score): subject line references their specific product and asks a genuine check-in question, with a tip related to their stated goal. No discount. No urgency. Just a human-sounding touchpoint — this is consistently the single highest-retaining message in the sequence, precisely because it doesn't read as promotional.

Message 2 — Frictionless pause offer (if no open/click in 3 days): introduces the pause feature as a benefit, not a consolation prize. The CTA goes directly to a one-click pause, not the full portal. A subscriber who pauses instead of cancelling has a meaningfully better chance of resuming within 60 days than one who cancels outright.

Message 3 — WhatsApp message (if opted in, at day 7 with no engagement): a short, personalised text referencing their specific product and asking if they have questions about how they're using it. Conversational, not promotional.

Message 4 — Dynamic loyalty offer (at day 10): triggered only if Messages 1–3 produced no engagement. A personalised offer based on subscription tenure — shorter-tenure subscribers get a modest discount, longer-tenure ones get a free product or loyalty gift. The offer is tied to tenure because it feels earned, not desperate.

What a well-run system typically achieves

Across subscription brands running this kind of system end to end, the pattern is consistent even though the exact numbers vary by category and starting point:

Churn reduction: well-calibrated systems commonly cut monthly churn by 40–55% relative to baseline — a brand starting around 15–20% monthly churn typically lands in the high single digits once the scoring model and intervention sequence are both live and calibrated.

Intervention effectiveness breakdown: the personalised check-in (Message 1) alone typically retains the largest single share of the at-risk cohort, precisely because it's the least promotional. The pause offer, WhatsApp touchpoint, and loyalty offer each recover smaller, meaningful additional slices — combined, a mature sequence commonly retains roughly a third of subscribers who would otherwise have churned.

Model accuracy: after 60–90 days of calibration, a properly weighted scoring model typically identifies most subscribers who go on to churn while keeping false positives low enough that intervention budget isn't wasted on subscribers who were never actually at risk.

For any subscription brand at meaningful scale, building this system is almost always high-ROI work — every subscriber retained generates months of additional revenue at near-zero marginal cost.

FAQs

What's the single strongest individual churn predictor? Subscription portal activity without action — a subscriber visiting the cancel page but not completing the cancellation is in decision mode, and it's the clearest signal of imminent churn to track.

Do I need a data scientist to build this? No — a weighted scoring system using Klaviyo's predictive analytics combined with custom properties gets most of the model quality of a full custom ML build, without dedicated ML engineering.

Which single intervention message recovers the most at-risk subscribers? The personalised check-in is consistently the strongest performer across brands running this sequence — more than any message that follows it, because it's genuinely non-promotional and reads as a real touchpoint rather than a retention tactic.

How long before a churn model becomes reliable? Around 60–90 days of calibration against actual churn events — the model improves continuously as labelled data accumulates, so early weeks will be noisier than the pattern that emerges after a full quarter.

Our take

Meaningful churn reduction doesn't come from a single clever email — it comes from a genuine system: real signals, a calibrated scoring model, and an intervention sequence that escalates in personalisation as risk increases. That's replicable for any subscription brand with the data infrastructure to support it. If you're running subscriptions and don't have a systematic churn intervention in place yet, talk to us — this is consistently one of the highest-ROI builds we do for subscription brands.