Beyond RFM: Why Customer Trajectories Matter More Than Segments

RFM tells you where customers have been, not where they are going

Traditional RFM segmentation tells you where customers are. Trajectory-based intelligence tells you where they're going. The difference is the gap between reactive campaigns and proactive customer management.

A segmentation method built for the catalogue era is still the default in most customer bases. It produces an accurate picture of yesterday's buyer and a weak brief for tomorrow's marketing.

RFM, recency, frequency, monetary, has had a remarkable run. It survived the move from catalogue to retail, from retail to e-commerce, from e-commerce to omnichannel, and from manual segmentation to cloud data warehousing. It still sits at the centre of how most brands think about their customer base. The question worth asking inside any CRM review is not whether RFM is useful, because it remains useful. The question is whether RFM is still where the next commercial decision should be made, and the answer is no more often than the industry has yet admitted.

A segmentation method built for a different marketing problem

RFM was not built for a modern marketing function. It was built for the operational constraints of direct mail.

The technique has a long lineage. Catalogue retailers in the 1930s, working with index cards and hand-written ledgers, noticed that three attributes reliably predicted whether a customer would respond to the next catalogue. How recently had they bought. How frequently did they buy. How much had they spent. The attributes were chosen because they could be recorded on a 3x5 card and retrieved quickly when a mailing list was being built. Decades later, direct marketers like Sears and L.L. Bean turned the same attributes into a formal scoring model, using it to decide which households got the expensive winter catalogue and which got a cheaper version or nothing at all.

Bult and Wansbeek, writing in 1995, gave the method its first serious academic treatment. They confirmed what catalogue operators had been doing in practice, that recency, frequency, and monetary value were meaningful predictors of direct mail response. The paper formalised RFM as a segmentation approach and, in the way these things tend to go, set it as the default for the next thirty years of customer marketing.

The operational fit with direct mail was clean. A mailing list had a fixed cost per piece. A piece sent to a low-RFM-score customer lost money. A piece sent to a high-RFM-score customer made money. The goal of segmentation was to decide, under a tight postage budget, which households made the cut. RFM answered that question well, because the decision was binary, the data was scarce, and the intervention was identical for every customer who received it.

The customer marketing problem today is not that one. The decisions are not binary. The data is not scarce. The intervention is not identical. The method that fit the constraints of the 1930s is being asked to solve problems it was never shaped to answer.

What RFM still does well

Fair dealing first. RFM is not obsolete, and dismissing it on principle would be a mistake.

The method remains a genuinely useful lens for three reasons. It is simple enough that a marketing team can reason about the outputs without a data science function translating between them. It is interpretable, which means a segment described as "recent, frequent, high-value" is comprehensible to a CMO, a merchant, and a trading director in the same meeting. It is cheap, computationally and operationally, which matters in a customer base where the segmentation is (hopefully) being recalculated every day.

RFM is also reasonably predictive of next-period response for customers whose behaviour is stable. A customer who shops every four weeks at an average basket of $120 will (hopefully), most of the time, continue to do so, and RFM will place them in a segment whose next-period likelihood of response is high. For that customer, the method does the job it was designed to do and does it well.

The method is also a reasonable starting point for brands that have historically had no segmentation at all. Moving from "one send to everyone" to "four sends to four quartile-based RFM segments" is an improvement. The gains are real. They are also the easy ones.

Where RFM stops contributing is at the line between a customer whose behaviour is stable and a customer whose behaviour is changing. That line is where most of the commercial opportunity in a customer base lives, and where RFM goes quiet.

Where RFM stops earning its keep

Three structural problems show up in every customer base where RFM is being asked to do more than it was designed to do.

The first is that RFM is backwards-looking by construction. Every input, recency, frequency, monetary, is a statement about what the customer has already done. The method has no way of weighting a customer differently because their behaviour is accelerating, decelerating, or flat. A customer who spent $600 over eighteen months and has been quiet for six may not look very different to a customer who spent $600 over six months and bought last week. The method cannot tell the difference, because the inputs do not contain the information that would.

The direction of travel is what most determines a customer's value over the next twelve months. A lapsed customer trending toward active is a different commercial proposition from an active customer trending toward lapsed. RFM treats the two as potentially identical, because their recency, frequency, and monetary scores can happen to land in the same bucket on the date the model is run. The method is accurate about history. It is silent on trajectory.

The second problem is that RFM is category-blind. The method weights every transaction equally, as long as the monetary values are equivalent. A fashion retailer's customer who bought three pairs of basics scores the same as one who bought one coat at three times the price, or one who bought once at the same price six months ago. The first is probably a regular low-consideration buyer. The second may be a high-consideration customer whose next interaction is likely to be a year away. The third is ambiguous. RFM collapses the three into a single bucket because the monetary column says they are the same.

The limitation is visible in the output. Most brands running RFM find that the top two or three segments are overpopulated with customers whose next-best action is very different, and the marketing team either sends them all the same message, which underperforms, or overlays a second segmentation layer on top of RFM by hand, which defeats the point of the model.

The third problem is that RFM does not model the customer's journey through the lifecycle. A new customer whose first purchase looks identical to a lapsed customer's last purchase will score identically on frequency and monetary. The lifecycle position of the two is completely different. One has just started a relationship. The other is close to ending one. The next-best action for each is almost the opposite. RFM, by its construction, cannot tell a marketing team which of the two situations it is looking at.

The combination of these three problems is why RFM tends to plateau in commercial contribution after the initial rollout. The first pass improves targeting visibly. The second pass produces diminishing returns. The third and fourth passes are adjustments to the cutoffs and look-back windows, not new commercial value. At some point, the marketing team realises the model has stopped adding, and the question is what replaces it.

Trajectory is the missing dimension

The alternative is a segmentation method that treats customer behaviour as a trajectory rather than a score. The difference is not philosophical. It is practical.

Trajectory-based segmentation adds the dimension RFM leaves out, which is the direction and rate at which the customer's behaviour is moving. Instead of asking "what has this customer done", the method asks "where is this customer going, and at what speed". The inputs are still behavioural, but they include the derivative, how behaviour has changed over successive windows, not only the current snapshot.

The segments that fall out of the method are different in kind from those produced by RFM. A trajectory-based model identifies customers whose behaviour is accelerating, who are worth a different intervention than customers whose behaviour is flat at the same score level. It identifies customers who are decelerating and separates them from customers who are genuinely active but temporarily quiet. It identifies customers whose relationship with the brand is still being defined, and distinguishes them from customers whose relationship is mature and stable.

The commercial consequence is that marketing activity can be matched to the trajectory, not only to the score. A customer whose scores are currently mid-tier but whose trajectory is accelerating warrants a different intervention, and often a different budget, from a customer at the same score level whose trajectory is flat or declining. RFM cannot distinguish between the two. Trajectory-based segmentation is built to.

Five lifecycle stages, five commercial questions

The trajectory view becomes most useful when it is organised around the stages of the customer lifecycle, because each stage maps to a distinct commercial question.

The Acquire stage covers customers who have not yet made a purchase but whose behaviour suggests they are close. The question is whether a targeted intervention will move them to a first purchase at a commercially sensible cost. SIVV's work in this stage has delivered a 9% lift in conversion alongside a 2% increase in average order value, by prioritising the customers whose behavioural trajectory indicates they are closest to converting and by matching the offer to their emerging preferences rather than to a generic welcome series.

The Early Life stage covers customers who have made a first purchase and whose second purchase is the commercial pivot. Most retail brands sit on a one-time purchaser rate in excess of 70%, which means the Early Life segment is usually one of the largest and most commercially valuable in the base. The question is which of these customers are likely to come back on their own, which need a specific intervention, and which were never going to return regardless. Trajectory-based segmentation answers the question in time for the intervention to land. SIVV's work in this stage has reduced the one-time-only rate by 8%, which is a meaningful shift in a metric that determines the payback on every acquisition dollar the brand has spent.

The Nurture and Grow stage covers the middle of the customer base, customers whose behaviour is established and whose trajectory is either stable or gently increasing. The commercial question for this stage is not reactivation. It is share of wallet. Which of these customers have room to spend more, in which category, at what cadence, and through which channel. Trajectory-based segmentation identifies the customers whose behaviour suggests they are expanding their relationship with the brand and flags the ones whose behaviour suggests a gentle plateau that can be re-engaged with. SIVV's work in this stage has lifted customer lifetime value by 10%, which compounds across a long horizon in a way a single-cycle gain from a price promotion does not.

The Prevent Churn stage covers customers whose trajectory is decelerating but who have not yet crossed into lapsed. This is where the commercial value of the trajectory view is most visible, because the intervention is cheaper than the alternative. A customer who is drifting can often be re-engaged with a margin-sensible intervention. The same customer, once lapsed, needs a more expensive win-back with a lower success rate. SIVV's work in this stage has delivered an 11% revival rate, which is a significant number given these are customers the brand would otherwise have written off as part of a quarterly attrition assumption.

The Win Back stage covers customers whose trajectory has crossed into lapsed. The question here is not whether to run a campaign, but which lapsed customers are worth pursuing, with what offer, and at what cost to margin. Trajectory-based segmentation sorts the lapsed file into recoverable and non-recoverable segments rather than treating it as a single list. SIVV's work in this stage has produced a 19% incremental uplift, funded by 2% sacrifice in average transaction value, which is a favourable trade for the brand because the incremental revenue meaningfully exceeds the margin cost. More importantly, the same intervention was not deployed against customers for whom it would have been a discount to a file that was never going to come back, which is where most undisciplined win-back spend ends up.

Five stages, five questions, five different commercial cases. The common thread is that the intervention is calibrated to the trajectory rather than to the historical score. The interventions that work in one stage would be wasted in another, and the value of the segmentation is that it places each customer in the stage where the right intervention will land.

The scores underneath the stages

The lifecycle stages are the layer a CMO interacts with. Underneath them, a set of predictive scores does the work of placing each customer, at each point in time, in the stage that reflects their current trajectory.

Four scores in particular carry the weight.

1. A churn probability score predicts, for each customer, the likelihood that they will lapse within a specified window. The score is what moves a customer from Nurture and Grow into Prevent Churn before the behaviour has crossed the lapsed threshold. The score is most valuable not when it is high, but when it is rising, because a rising churn probability is the earliest warning of a trajectory that is decelerating.

2.A lifetime value prediction score estimates the revenue a customer is likely to generate over a forward window, conditional on their current trajectory. The score is what separates a mid-spending customer whose future revenue is high from a high-spending customer whose future revenue is likely to decline. RFM cannot distinguish between the two. A lifetime value prediction can, and the marketing budget attached to each customer is better calibrated as a result.

3. A response propensity score predicts, for each customer, the likelihood that they will engage with a specific type of intervention. The score is what prevents the blanket send. It identifies customers for whom email is the right channel, customers for whom a different channel will carry more weight, and customers who are unlikely to respond to anything and whose inclusion in the send is a cost against the campaign.

4. A reactivation probability score predicts, for each lapsed customer, the likelihood that a targeted intervention will bring them back. The score is what sorts the Win Back file into customers worth pursuing and customers who are not. The 19 per cent incremental uplift described above depends on the score being accurate, because the uplift evaporates the moment the intervention is broadcast to the file rather than targeted to the customers the score identifies as recoverable.

The scores are calibrated on the brand's own customer base, not on a generic model that averages across thousands of brands. A fashion retailer's churn probability is not a telco's, and a travel brand's reactivation probability is not either. The scores are specific, and the lifecycle stages built on top of them are specific in turn.

The budget question

The commercial point of moving from RFM to trajectory-based segmentation is not the segmentation itself. It is the budget decision that follows.

A marketing function running RFM tends to allocate budget against historical score, which often means over-spending on stable high-value customers who would have bought anyway and under-spending on customers whose trajectory suggests an intervention would matter. A marketing function running trajectory-based segmentation allocates against the future value the intervention is likely to create, which tends to mean less spend on customers whose behaviour is self-sustaining and more on customers whose trajectory is at an inflection point.

The reallocation is not dramatic in aggregate. It does not require a bigger marketing budget. It requires the same budget, placed differently, against customers whose segmentation reflects where they are going rather than where they have been. Brands partnered with SIVV have made the move alongside their existing CRM infrastructure, which is to say the RFM work did not need to be thrown out. It needed to be lifted into a method that captured the dimension it was missing.

RFM will still be in the customer analytics toolkit for a long time. It deserves its place for what it is good at, which is describing a snapshot of a stable customer base. What it is not, and was never built to be, is the segmentation a brand runs campaigns against in a customer base where trajectory is where the commercial opportunity lives. The question worth asking is not whether the RFM report is accurate. It almost always is. The question is whether it is answering the commercial question the marketing function is actually trying to decide, and the answer, for most brands, is that it is answering a different one.

‍

About SIVV

SIVV is the customer intelligence and decisioning platform built for marketers who want to know what's actually working.

We sit above your marketing platforms, combining:

• Sophisticated customer intelligence (churn prediction, lifecycle segmentation, propensity modelling)
• Scientific campaign measurement (randomised control groups for every campaign)
• True incremental revenue reporting (revenue that reconciles to your actual business performance)

Clients across telecommunications, retail, gaming, entertainment, and travel use SIVV to:

• Measure true incremental revenue for every campaign
• Identify which audiences and offers drive real lift
• Make budget allocation decisions based on incremental ROI
• Optimise marketing performance based on what actually moves the needle

Stop optimising against attribution. Start measuring incrementality.

Learn more at sivv.net or contact us to discuss how incremental measurement can transform your marketing performance.

‍