When a campaign goes to an entire customer list, response rate is usually not uniform — a small portion of the list drives a disproportionate share of responses, and another large portion will essentially never respond no matter what's offered. Response modeling is the practice of predicting, before you spend a dollar, which customers fall into which group — so budget goes toward the people worth contacting and away from people who won't respond regardless.

RFM analysis: the simplest response model that still works

RFM scores every customer on three dimensions calculated purely from transaction history: Recency (how long since their last purchase), Frequency (how many purchases, or how often), and Monetary value (how much they've spent, typically total or average order value). Each dimension is usually scored 1-5 (quintiles), and the three scores combine into a segment.

RFM works because it requires no modeling expertise, no demographic data, and no third-party data — it's built entirely from data every business already has in its transaction history. That makes it the right starting point before investing in anything more sophisticated.

rfm_segment_examples.txt

text

RFM Score (R-F-M, each 1-5, 5 = best)    Segment            Typical response to campaigns
--------------------------------------------------------------------------------------------
5-5-5                                     Champions           Highest — reward, don't discount
5-1-3                                     New/Recent           Good — nurture toward frequency
1-5-5                                     At-Risk Loyalists    Moderate — needs a win-back trigger
1-1-1                                     Lost                 Lowest — rarely worth full-price campaign cost
3-3-3                                     Core/Average         Moderate — responds to standard offers

RFM is a segmentation tool, not a prediction — propensity models are the next step

RFM tells you a customer's past behavior pattern. It doesn't directly predict whether a specific future campaign will convert them — a Champion might still ignore an irrelevant offer. Propensity modeling goes further by training on historical response data to predict the probability of response to this specific type of campaign, using RFM-style features as inputs alongside others.

Propensity modeling: predicting who will respond

A propensity model is typically a logistic regression or gradient-boosted tree trained on historical campaign data: given a customer's attributes (recency, frequency, monetary value, tenure, category affinity, past response to similar campaigns, engagement signals like email opens), predict the probability they respond to this campaign. The model outputs a score, usually 0-1, for every customer in the eligible list, and that score is what determines targeting — not a static segment.

The model needs a labeled training set: past campaigns where you know who was contacted and who responded. This is the main practical blocker for teams trying to build their first propensity model — if contact and response history weren't captured cleanly, there's nothing to train on.

RFM vs. propensity modeling

Dimension	RFM analysis	Propensity modeling
Data required	Transaction history only	Transaction history + past campaign contact/response history
Output	A segment label (e.g. 'Champions')	A continuous probability score per customer per campaign type
Setup effort	Low — spreadsheet-feasible	Higher — requires a labeled dataset and some modeling skill
Adapts per campaign type	No — same segments regardless of offer	Yes — a model can be trained separately for win-back vs. upsell vs. acquisition
Best first step for	Teams with no modeling capability yet	Teams with enough campaign history and volume to justify the investment

How a response model changes targeting economics

The practical payoff of a propensity model isn't accuracy for its own sake — it's that it lets you set a response probability threshold and only contact customers above it, which directly improves ROI by removing spend against customers who were never going to convert.

Consider a list of 500,000 customers with a flat 2% expected response rate if contacted indiscriminately. If a propensity model can identify the top 30% of that list who account for 70% of total expected responses, contacting only that top 30% at the same budget per contact captures most of the value at roughly a third of the contact cost — and the saved budget can be redeployed to acquisition or to a different campaign entirely.

propensity_targeting_economics.txt

text

Full list contact (no model):
  500,000 contacts x $0.80 cost/contact = $400,000
  Expected responses: 500,000 x 2.0% = 10,000
  Cost per response: $40.00

Top-30%-by-propensity contact (model-driven):
  150,000 contacts x $0.80 cost/contact = $120,000
  Expected responses: 70% of 10,000 = 7,000
  Cost per response: $17.14

Same offer, same cost-per-contact — contacting the right 30%
captures 70% of responses at 30% of the spend.

Models decay — retrain on a schedule, not when results crash

Customer behavior, offers, and market conditions shift, and a propensity model trained on last year's response patterns slowly loses accuracy (model drift). Set a retraining cadence (commonly quarterly) rather than waiting until targeting results visibly degrade to notice the model is stale.

What's next

Response modeling improves who you target; the next lesson covers how to build the spreadsheet model that forecasts whether a campaign to that targeted audience will actually be profitable before you launch it.

Next: Building a Marketing Financials Worksheet →

I build these systems professionally.

Whether it's a RAG pipeline, analytics migration, or AI workflow — let's talk.

Work With Me →

Need production AI or MarTech expertise? I build custom LLM workflows, server-side GTM tagging, and analytics infrastructures professionally.

Need custom AI or MarTech setup? Let's build together.

Let's talk →