Loyalty Programs: Are You Paying for Repeat Business?

The report looks excellent. Loyalty members spend forty percent more than non-members and visit twice as often. The program is working, says the slide. It might be. The number on that slide cannot tell you, and it will look identical if the program is doing nothing at all.

The comparison that proves nothing

Think about who joins. A rewards card is worth more to somebody who shops with you weekly than to somebody who comes twice a year, so the frequent shoppers sign up and the occasional ones do not.

That means the member group was already higher-spending on the day it formed, before a single point was issued. Every comparison you run afterwards inherits that head start.

The uncomfortable version: a loyalty program that changes nobody behaviour at all will still produce a slide showing members spending forty percent more. The number is real. It is just answering a question about enrolment rather than about the program.

What each question actually requires

Redemption rate deserves its own warning. It measures whether people claimed a reward, which they will, because it is free money. A high redemption rate on an offer that changed nobody purchasing is an expensive success metric.

Hold a group out

The fix is unglamorous and it is the only thing that answers the question. Take a random slice of eligible customers and do not give them the program, or do not send them the offers, and then compare the two groups over a defined period.

Random is the word carrying the weight. A holdout picked by store, region or segment reintroduces exactly the bias you are trying to remove.

It feels wasteful, and people object that you are denying rewards to real customers. The cost is the rewards for that slice. What you get back is knowing whether the remaining spend on the program is buying behaviour or subsidising it, which is a baseline problem rather than a marketing one.

Pick one behaviour

Programs drift toward rewarding everything, because every stakeholder wants their metric included. The result is a structure that pulls in several directions and reliably produces a discount.

Decide what you are trying to change before designing the mechanics. Frequency, basket size, category breadth and survival through the period when customers usually drift away are four different goals, and the reward that drives one does not drive the others. A points-per-dollar scheme rewards spend you already had. A threshold that unlocks at a basket size slightly above your current average has a reason to change behaviour.

Watch the tiering especially. A permanent discount for your highest spenders converts reliable margin into a lower price for the customers least likely to leave, which shows up directly in where your margin actually goes.

Where AI earns its place

Grouping customers by how they actually behave rather than by how much they spend. Spend tiers are the default and they hide the useful distinctions: the weekly small-basket shopper and the monthly large-basket one may spend the same and respond to completely different offers.

Early drift detection is the second, and it is where the money usually is. A customer who normally buys every three weeks and has not been in for six is a signal you can act on, and it rarely gets checked until the customer is long gone. That is the same pattern as what your retention data already shows.

Matching reward shapes to groups is the third, and it depends entirely on running the comparisons properly. Without a holdout, a model will happily learn which customers were going to spend anyway and recommend rewarding them, which is the expensive failure dressed up as personalisation.

All of it rests on transaction records being clean enough to tie purchases to the same customer over time. Duplicate accounts and unlinked cards quietly split one customer into three, and the analysis then describes people who do not exist, which is the dependency in data cleansing.

Frequently Asked Questions