Every proposal for a loyalty programme, a retention campaign or a customer success hire arrives with the same statistic attached. Keeping customers is five times cheaper than finding them. A five per cent improvement in retention lifts profit by twenty-five to ninety-five per cent. The numbers are quoted with confidence, in slide decks, by people who have never traced them.
They are worth tracing. What you find at the end of the chain is not a law of business. It is one article, about banks, from 1990.
01 — What happened
The source is Frederick Reichheld and W. Earl Sasser’s Zero Defections: Quality Comes to Services, published in the September–October 1990 issue of Harvard Business Review. Its findings were specific and industry-bound: a five per cent reduction in the defection rate generated 85% more profit in one bank’s branch system, 50% more in an insurance brokerage and 30% more in an auto-service chain. The article also reports that when the credit card company MBNA America halved its 10% defection rate, profits rose 125%.
Three industries, plus a credit card company — a business with contractual, monthly, entirely measurable customers.
What happened next is ordinary citation drift. Bain’s own page for the article now summarises it as companies being able to “boost profits by almost 100% by retaining just 5% more of their customers” — the bank figure, generalised. The familiar “25% to 95%” range is a later compression across several studies rather than a finding from one. Try tracing the “five times cheaper” ratio and the attribution shifts with whoever is quoting it: Forrester, Invesp, Bain, Harvard Business Review. Nobody cites a study, because the chain of citation is the study.
Meanwhile, the research that did test the mechanism directly is less flattering and almost never quoted. Byron Sharp and Anne Sharp examined a large loyalty programme against Dirichlet benchmarks — the repeat-purchase pattern a brand’s market share already predicts — and published the results in the International Journal of Research in Marketing in 1997. They found only a weak trend towards excess loyalty, and not consistently across the participating brands. Members largely bought as their share predicted they would.
So the honest position is this. Retention is worth a great deal of money in some businesses, very little in others, and the circulated statistics cannot tell you which one you are running.
02 — Why it matters
Because those numbers get used to authorise spending. A retention app, a points scheme, a CRM module, a new role — all signed off on arithmetic borrowed from a 1990 study of credit cards and bank branches, by a business whose customers buy once every four years.
The structure of repeat purchase is not a detail. A dental practice has customers who should return every six months and will tell you nothing when they stop. An accountant on an annual retainer has contractual churn with a visible renewal date. A roofer has customers who, having had the roof done, correctly never come back. These three businesses need entirely different answers, and the quoted statistic gives all three the same one.
The question worth answering is narrow and local: in this business, is a unit of effort spent keeping a customer worth more than the same unit spent finding a new one? It is answerable, and it needs four numbers already sitting in the records.
03 — The economics
The four numbers:
- Repeat rate over a defined window. The share of customers who bought again inside the period that matches your purchase cycle — not a calendar year chosen for convenience.
- Gross margin per order, not revenue. Retention compounds margin, and the mistake most often made in this calculation is using turnover.
- All-in acquisition cost per customer. Advertising, plus the sales and quoting time that advertising consumes. Leaving out labour is what makes acquisition look cheap.
- When customers lapse. The normal gap between orders, and the point past which silence means they have gone.
Then the test. Take an illustrative service business, say 800 active customers, 30% of whom order again within a year, an average order carrying 40% gross margin, and an acquisition cost roughly equal to the margin on two orders. Those assumptions are made up to show the shape of the calculation; your figures will differ.
On those numbers, a new customer is break-even on their second order. All profit from that relationship sits in the third order and beyond — which only 30% of customers ever reach. Moving the repeat rate from 30% to 33% adds roughly 24 additional repeat orders a year. Getting those same 24 orders through advertising means acquiring about twelve new customers and waiting for each of them to order twice — and on these assumptions that route nets nothing, because acquisition has eaten both orders’ margin. The 24 retained orders keep theirs.
That is the whole argument for retention, and notice what it did not require: a statistic from 1990. It also produces a different answer the moment acquisition is cheap and repeat purchase is structurally rare. For a one-off purchase — a wedding, a house move, a new roof — the calculation says spend on referral and reputation instead, because there is no second order to protect. “Retention is cheaper” is simply false in those businesses, and it is sold to them anyway.
04 — What businesses should do
Set the window from your own purchase cycle. If the typical gap between orders is 90 days, treat 180 days of silence as a lapse. If it is annual, a missed renewal month is the signal. Pick the number from your transaction history, not from an article.
Calculate the four figures from the last 24 months. If that takes more than a day, the finding is not about retention. It is that the business cannot currently see its own repeat behaviour, which is a more urgent and cheaper problem to fix.
Compare the two routes honestly. Margin from one extra order per retained customer, against all-in acquisition cost for the equivalent margin in new ones. If acquisition wins, stop here and spend there. A calculation that tells you not to buy something has paid for itself. Judge the result by its effect on margin, never by activity — the trap in hours saved is the wrong way to value an automation.
If retention wins, act manually first. Pull a list of customers past their normal gap. Have someone contact them for 60 to 90 days. Record what happens and what they say. You will learn the save rate and the real reasons for leaving — both needed before anyone can specify a system sensibly. Businesses handling several hundred of these a month will see quickly whether the contact work is the constraint — the same test applied in the economics of automating customer follow-up.
Only then decide what to build, scoped to what the manual run proved.
05 — Where technology fits
You probably do not need an AI churn model for this yet.
The published work makes the point better than the vendors do. A 2026 study in Frontiers in Artificial Intelligence compared seven churn models on the widely used Telco dataset of 7,043 customers, the best of them reaching an AUC-ROC of 0.932. Strong results — and the features doing the predicting were contract type, tenure, support history and charges: fields any business with decent records already holds. The authors also note that their SHAP analysis explains model behaviour, not causation. And the dataset is a subscription telecom, where churn is contractual and dated. Most businesses have non-contractual churn, where “lapsed” is a judgement you have to define before any model can learn it.
What this actually needs is duller. One record per customer with every transaction attached, in one place — a CRM or a connected system rather than a billing tool and a mailbox that never speak, the familiar version of the real cost of disconnected business systems. An integration between whatever takes the money and whatever holds the relationship, so the lapse clock runs on real orders. A recurring report that lists who has gone quiet past their normal gap. Then a person who contacts them.
Automation belongs on the list and the reminder, not on the judgement. A monthly query and a task queue will outperform a model the business cannot explain to itself.
The legitimate nothing: with a few hundred customers and one person who knows them all, the report is a conversation. Buy no software. What a CRM gives you at that size is memory that survives staff changes, which is a real reason to have one but not a retention strategy, as your CRM is not a growth strategy argues.
06 — RevenueStack’s take
Our view: the 1990 numbers still circulate because that was the last time anyone did the arithmetic in public, and quoting it is easier than doing it privately.
The valuable claim was never the 85%. It was that defection can be measured, and that the reasons for it tell you where the business falls short. Thirty-six years on, most businesses have measured neither, and cite someone else’s measurement precisely so they do not have to produce their own.
A retention programme justified by a borrowed statistic is a marketing decision. A retention programme justified by your own four numbers is an operating decision — and in our experience it comes out smaller, duller and more profitable than the one originally proposed. A list, a cadence, and someone who calls. That is usually the whole system, and the reason it works is not loyalty. It is that nobody else noticed the customer had gone.
Sources
- Zero Defections: Quality Comes to Services — Bain & Company / Harvard Business Review, 1990-09-01
- Zero Defections: Quality Comes to Services (HBR OnPoint Enhanced Edition) — Harvard Business Review, 1990-09-01
- Loyalty programs and their impact on repeat-purchase loyalty patterns — International Journal of Research in Marketing, 1997-12-01
- Explainable AI-driven customer churn prediction: a multi-model ensemble approach with SHAP-based feature analysis — Frontiers in Artificial Intelligence, 2026-02-10
