Using Klaviyo's predictive analytics
Klaviyo computes four forward-looking metrics on every qualifying profile: churn risk, predicted customer lifetime value, expected date of next order, and average time between orders, with models that retrain weekly. Most accounts that qualify never build a single segment on them. This guide covers what the metrics measure, the eligibility thresholds, and the three segments that earn their keep first.
By Dana Okafor, who builds these segments for Sendnexa clients. Everything below assumes Klaviyo, but Omnisend and Customer.io expose comparable signals and the segment logic transfers.
What do the four predictive metrics actually measure?
Churn risk estimates the probability a customer will not purchase again, from their order pattern against the account's overall behavior. Predicted CLV projects future spend. Expected date of next order estimates when the next purchase should land if the customer's cadence holds. Average time between orders is the cadence itself. None of this is generic AI: the models train on your account's own orders, which is why Klaviyo gates them behind real data thresholds.
Who qualifies, and what if you don't?
The thresholds: 500+ customers with at least one order, 180 days of order history, orders in the last 30 days, and three or more orders per qualifying customer. Below the bar, approximate with recency-frequency-monetary segments; a 90-day non-engager exclusion behaves much like a churn-risk suppression while history accumulates. If you are migrating platforms, import historical orders so the clock starts from your real history rather than from zero, a point covered in ourESP-migration service.
Segment one: the churn-risk suppression
Definition: churn risk above the threshold you calibrate, excluded from routine campaigns. This is the highest-impact predictive segment because it touches every send. The coffee-subscription engagement in ourcase-study library cut send volume 22% with revenue flat and complaints down from 0.19% to 0.06%, comfortably under Google's 0.1% guidance. Start conservative, watch revenue hold, tighten.
Segment two: CLV tiers
Split the customer base into three tiers by predicted CLV and give each a cadence and offer policy. The failure this fixes: discounting people who buy at full price. A 12-location retailer we profile measured $3.10 revenue per recipient in its top tier against a $0.95 list average; the tiers moved its win-back program to service-led mail for the top tier and reserved incentives for at-risk profiles. The full story is in the retail win-back case study.
Segment three: next-order-date cohorts
Trigger replenishment and reminder flows a category-sized window ahead of each customer's expected date: about 7 days for consumables, 14 to 30 for considered purchases. Swapping a fixed 30-day timer for these triggers took a pet-supplies replenishment flow from 1.1% to 2.4% placed-order rate and $1.60 to $2.75 revenue per recipient in 90 days, with identical creative. Timing did all of it.
How do you know it's working?
Report in placed-order rate and revenue per recipient, not opens; Apple Mail Privacy Protection has inflated open counts since September 2021. Context numbers worth keeping on the dashboard: Klaviyo's platform averages of 1.69% campaign click rate, 5.58% flow click rate, and flows producing roughly 41% of email revenue from 5.3% of sends. If your flow share sits far below that after these segments ship, the flows are the next project.
What are the requirements for Klaviyo predictive analytics?
Four thresholds, all mandatory: at least 500 customers who have placed an order, at least 180 days of order history, orders within the last 30 days, and at least three orders per qualifying customer. Meet them and predictions appear on profiles automatically. Below them, build recency-frequency proxies and import historical orders to shorten the wait.
How accurate is the expected date of next order?
Accurate enough to time email around, provided your category has a real purchase cycle. The models retrain weekly on your account's own order patterns, so a coffee subscriber trends toward their true cadence rather than a category average. Categories without repeat cycles, furniture being the classic case, produce noise; skip the metric there.
Should low churn-risk customers get fewer emails?
Usually yes, and the discount logic should change too. A referenced retail engagement measured its top-CLV tier at $3.10 revenue per recipient against a $0.95 list average, while the old program sent everyone identical discount blasts. High-value, low-risk customers buy at full price; save incentives for profiles the model flags as genuinely at risk.
Can I use predictive segments outside of flows?
Yes, and campaign suppression is the highest-value use of the scores. Excluding high-churn-risk profiles from routine campaigns cut one client's send volume 22% with revenue flat, while spam complaints fell from 0.19% to 0.06%. Flows get most of the attention, but the suppression list quietly protects every single send the program makes.
Qualify, but nothing built on it?
That is the most common account state we see. A working session sizes what the three segments would change in yours.
Book a working session