How should we use customer data to predict who will churn?
Our answer
Predict churn from your own order history. Score recency, frequency and spend, flag customers as at risk at 1.5 times their usual gap between orders and lapsed at about 2 times, and use platform predictions to time flows.
- Confidence
- Medium
- Strongest proof
- 7 official rules
- 2 practitioners, 4 opinions or claims
- Sources
- 7 sources
- 13 statements
- K
PAOL
New here? Take the 1-minute tour
Why this is the answer
3 reasons and our judgmentWithout a cancel button, churn is a multiple of the usual gap between orders. Polar Analytics and Omniconvert put the line at about 2 times that gap and Darkroom at 2 to 3 times. Omniconvert starts win-back at 1.5 times.
Platform predictions suit groups better than single customers. Klaviyo says its predictions work best averaged over many customers rather than for one person. Its 180-day recency default can flag monthly buyers too late, so set thresholds by buying cycle.
You can start without a data scientist. Balistro (media and measurement agency) builds a three-tier score from order history alone, and Polar Analytics says no model is needed. Balistro reports one beauty brand's email revenue rose 34% in two months after splitting messages by group.
Our judgment
Flagging at risk at 1.5 times the usual gap between orders and lapsed at about twice it is our pick. They follow the churn lines from Polar Analytics and Omniconvert, sit at the low end of Darkroom's, take the point where Omniconvert starts win-back and Darkroom makes its first offer, and match the lapse point in Churn for one-time buyers. Start with a simple three-tier score on recency, frequency and spend, refreshed monthly, rather than a bought churn model. Balistro builds one in 1 to 2 weeks, Klaviyo's own recency, frequency and monetary (RFM) report uses three tiers, Darkroom suggests a 1 to 3 scale below 200,000 customers, and Polar Analytics says no model is needed. Use platform predictions to time flows and rank groups rather than to judge one customer, and replace the 180-day recency default for products bought monthly. Klaviyo itself prefers expected date of next order to the churn percentage for targeting.
Our judgment, built from Darkroom, Polar Analytics, Omniconvert, Balistro and Klaviyo: our reading of these sources, not a finding from any one of them.
What the sources say
13 statements from 7 sourcesKlaviyo says its expected date of next order prediction is more useful for targeting than churn risk, for example a flow triggered a few days before that date. Custom CLV needs its paid analytics plans.
“Our standard CLV metrics are calculated based on a set window of 365 days.”
Klaviyo's predicted CLV needs at least 500 customers with orders, 180 days of order history, recent orders and some repeat buyers. It predicts revenue over a year and works best averaged over many customers.
“predictions work best when averaged over many customers and are not expected to be exact for any single individual.”
Shopify's predicted spend tier sorts past buyers into High, Medium and Low by purchase frequency, order value, order count and recency, for use in segments. A store needs over 100 sales.
“To use predicted spend tier, your store must have made over 100 sales.”
Klaviyo's expected next order date ignores which products were bought, and Klaviyo advises against counting down to it in a flow or using it in place of replenishment flows.
“Expected date of next order doesn’t consider which products the customer ordered.”
Klaviyo tells senders to mail its engaged segments (opened or clicked in the last 30, 60 or 90 days, or recently joined) and exclude unengaged ones from broad sends to protect deliverability.
“Your engaged segments are the primary groups of people that you should send to.”
Darkroom, a retention agency, says a store with no cancel event can call a customer churned once the time since their last order is two to three times the median gap between first and second orders.
“Look at the median number of days between a customer’s first and second order.”
Klaviyo allows up to 4 RFM models per account, each needing 500 customers with orders, and suggests separate models for subscriptions and distinct product lines with recency set to each purchase cycle.
“Set recency to that population's purchase cycle.”
Balistro, an agency, says tools like Klaviyo can build RFM segments from 12 to 18 months of orders, three tiers each. One client's email-attributed revenue rose 34% in two months, with no holdout.
“split each dimension into three tiers (high/medium/low) rather than a full five-point scale”
Darkroom, an agency, says to give top RFM segments early access and loyalty perks and keep them out of blanket discounts, since they buy anyway. Guidance from client work, no data.
“excluding champions from blanket promotions is often the fastest margin win in the whole system.”
Lifetimely, which sells profit analytics, says give top RFM customers early access and referral asks, not discounts, and win back lapsed top customers personally. One apparel store had 15,108 of them.
“should get early access and referral asks, not discounts. They would have paid full price anyway”
Polar Analytics, which sells analytics software, says a store with no cancel button can call a customer churned after about twice their median gap between orders, tuned per category. It gives no data behind the 2x line.
“If your median time between orders is 45 days, then 90 days of silence is your churn line.”
Omniconvert, which sells customer intelligence software, says to set the churn line at about two purchase cycles without activity, and to trigger a 3-email win-back at 1.5 times a customer's usual cycle.
“Trigger a 3-email sequence when a previously-active customer hasn't purchased in 1.5x their typical purchase cycle.”
Klaviyo's RFM report needs 500 customers with orders and 180 days of history, scores customers 1 to 3 per dimension, and labels groups such as needs attention and at risk, warning against over-messaging the latter.
“Try not to over-message as they are a churn-risk.”
In the full answer
For readers who sign in- What to do this week3 steps
- The numbers behind it3 figures from cited sources
- Where each source standsA chart placing each source on the question
- The strongest disagreementThe best case against the answer, from a cited source
- How sure we areWhat its confidence rests on, and what would change the answer