Behavioral cohorts past RFM that actually move revenue

Where RFM runs out of road

Recency, frequency, monetary is a fine starting grid. It sorts a list into buckets fast, and for a brand with no segmentation at all it will lift revenue in the first month. We use it. We just do not stop there, because RFM describes what someone did, never why, and it cannot see what they are about to do.

The limit shows up in practice. RFM puts a one-time buyer who spent $400 and a four-time buyer who spent $400 in the same monetary band, even though their next-90-day behavior could not be more different. It treats a subscriber who browses a category three times a week the same as one who has not opened the site since checkout. That blindness is where revenue leaks. Better ecommerce segmentation has to model behavior and intent, not just transaction history.

Event-based cohorts

The first layer past RFM is cohorts built on behavioral events, not order facts. We instrument the storefront and Klaviyo to capture the signals that precede a purchase, then group people by what they are doing right now.

The events we lean on:

• Category affinity. Three or more product views in one category within 14 days. This person is shopping that category whether or not they have ever bought from it.
• Cart and checkout intent. Started checkout in the last 7 days with no order. Different from a passive cart-abandon because they reached the payment step.
• Repeat-browse without buy. Four or more site sessions in 30 days, zero orders. High intent, some friction we have not solved yet.
• Post-purchase silence. Bought, then no site session for a period longer than their median repurchase interval. The early-warning signal for churn, weeks before RFM would flag them as lapsed.

Each cohort gets its own message, because the right email for a category-affinity browser is a product email, and the right email for a checkout-intent profile is a friction-removal email about shipping or returns. We have seen checkout-intent cohorts convert at 6 to 9 times the rate of a blast to the full list, because we are talking to people at the exact moment they are deciding.

The operational trick is that these cohorts are time-windowed, so membership expires. A checkout-intent profile drops out of the cohort 7 days after their last started checkout, because intent decays. RFM segments are sticky and slow to update. Event cohorts are deliberately short-lived, which is what lets them carry messages that are timely instead of generic. We refresh membership nightly, and for the highest-intent cohorts we refresh in near real time so the friction-removal email goes out within hours, not days.

Predicted-LTV bands

The second layer is forward-looking. Instead of banding people by what they have spent, we band them by what they are likely to spend.

We build a predicted lifetime value score per profile from a small set of features that hold up across most catalogs: first-order value, days between order one and order two, product category of first purchase, discount depth on first order, and email engagement in the first 30 days. Klaviyo exposes a native predicted-CLV value, and for smaller accounts we use it directly. For accounts with enough order volume, we fit our own model because the native one underweights the second-order timing signal, which in our experience is the single strongest predictor of whether someone becomes a repeat buyer.

We then cut the list into three bands. The top band, usually the top 15% to 20% by predicted value, gets early access, no discounts, and a higher service tier, because discounting people who will pay full price is pure margin loss. The middle band gets the standard lifecycle program. The bottom band, the predicted one-and-done buyers, gets a lighter touch and almost no acquisition-style spend, because the model is telling us the reactivation math will not clear.

The move that pays is matching offer depth to predicted value. We stop sending 20% codes to high-LTV profiles who would have bought anyway, and we stop spending recovery budget on low-LTV profiles who will not return. That single reallocation has lifted blended margin by 3 to 5 points on accounts where the previous program discounted everyone equally.

How we build them

The stack is unglamorous and that is the point. Events flow from the storefront and Klaviyo into a warehouse table, one row per profile per event. We compute cohort membership and the LTV score on a nightly job, then sync the results back as profile properties Klaviyo can segment on. The flows read those properties. Nothing about the day-to-day campaign work changes for the operator. The intelligence lives in the data layer, and the marketing surface stays simple.

We keep the cohort definitions in version control as plain SQL, so a definition change is a reviewed commit, not a setting someone clicked and forgot. When a brand asks why a cohort grew or shrank, we can point at the exact query.

The predicted-LTV model gets retrained on a schedule, usually monthly, against actual realized value so we can measure how well last quarter's predictions held up. We track the model's calibration: of the profiles we put in the top band 90 days ago, what fraction actually became high-value buyers. When that number drifts, it tells us a feature has gone stale, often because the brand changed its acquisition mix or its discount strategy and the old signal no longer means what it did. A segmentation system that is never checked against reality slowly turns into superstition, so we hold ourselves to the same calibration standard we would hold any forecast.

A concrete example

A homewares brand came to us running pure RFM with five segments and a single discount ladder. We added four event-based cohorts and three predicted-LTV bands on top.

The checkout-intent cohort alone, with a friction-removal email about their 60-day return window, recovered orders at an 11% conversion rate against a list-blast baseline of under 1.5%. The post-purchase-silence cohort caught churn-risk buyers 5 weeks earlier than the old lapsed segment, and a no-discount product email to them held a 4% reorder rate. On the LTV side, pulling the top band out of the discount ladder recovered roughly 4 margin points on full-price-eligible buyers who had been getting 15% off for no reason.

Net, over the first quarter, email-attributed revenue rose 19% while total discount spend fell, because we stopped paying people to do what they would have done anyway.

The mistakes we watch for

Two failures show up over and over when teams build past RFM, and we design around both.

The first is cohort sprawl. It is tempting to keep adding cohorts until there are forty of them, at which point nobody knows which message goes to whom and several cohorts overlap so a profile lands in three at once and gets three emails. We cap the active set, usually at six to eight cohorts plus the LTV bands, and we make membership mutually exclusive where the messages would collide. A profile in checkout-intent does not also get the category-affinity email, because checkout intent is the stronger signal and it wins.

The second is treating the predicted-LTV score as fact. It is a forecast with error bars, and acting on it as if it were certain leads to writing off customers the model got wrong. We use the bands to shift effort and offer depth, never to fully cut a profile off, because a meaningful share of the predicted one-and-done buyers do come back, and a brand that stops mailing them entirely never finds out. The model guides where we lean. It does not get a veto.

The payoff for getting both right is a program where every email a customer receives maps to something they actually did or are likely to do, which is what makes the list feel curated instead of spammed, and a curated list is one people stay subscribed to.

How we think about it

RFM tells you who someone has been. Event cohorts tell you what they are doing this week. Predicted-LTV tells you what they are worth going forward. We run all three together, because the revenue that RFM leaves on the table lives in the gap between past behavior and likely future behavior, and the only way to capture it is to model both.

The full cohort definitions and the LTV feature set are written up at https://www.arthea.ai/article/beyond-rfm-segmentation.