Case study · Multi-country retail · 2024

Before, there were two segments: those who spend and everyone else

Premium retail group · customer segmentation, data and loyalty · three countries and an e-commerce channel · 2024

Illustration: an undifferentiated crowd crudely cut in two on one side, twelve distinct ordered segments on the other

When I arrived, the company had a base of several thousand customers and two segments: those who spent a lot, and everyone else. That is not a caricature, it is the rule that actually drove discounts. Assigning a customer to a discount group was done by hand, margin was eroding under stacked discounts, and most customers had no idea which category they were in.

A two-box segmentation is not a segmentation, it is a judgement. It does not tell you when a good customer is drifting away, which one is worth calling back, or how long normally passes between two purchases. So it enables no decision, except granting a discount at the till.

The work was to build a real typology from actual transactions rather than from shop-floor impressions, then translate it into a commercial system that store teams can apply.

01 · Context

A premium retail group across three countries, with a European e-commerce channel. The product was strong, customers were attached to the brand, and the data existed: every till transaction had been recorded for years. Nobody had ever used it for anything but accounting. The consequence fits in one sentence: the company knew how much it had sold, it did not know to whom, or whether that person would come back.

The constraint

A till database is not a customer database. Records were created on the fly, with no reliable unique identifier, with duplicates, empty fields and email addresses missing on the large majority of entries. Second constraint: whatever we built would be applied by store teams, in three countries and four languages, with no sophisticated CRM. A segmentation the till cannot read does not exist. And third: the mandate ran to development-finance standards, so everything had to be documented and transferable.

03 · The system

1 · Clean before segmenting, and say so.

The first deliverable is not a segmentation, it is the count of records set aside and why. Duplicates, zero-price transactions, records with no usable history. An analysis that does not publish its waste rate cannot be checked.

2 · Three methods compared, not one method imposed.

Recency, frequency, monetary value, handled three ways in parallel: a statistical segmentation by transformation and automatic grouping into four families; a quintile split with a mapping grid to ten named segments; and a weighted score, recency counting least and value most. Comparing three methods on the same base shows which ones agree and where they diverge. That disagreement is what teaches you something.

3 · Name the segments in the language of the shop floor.

Ten explicit segments rather than cluster numbers: champions, loyal, promising, new, needing attention, falling asleep, cannot lose, at risk, hibernating, lost. A sales assistant must be able to say which segment the person in front of her belongs to, without opening a spreadsheet.

4 · Measure the normal rhythm, not only the value.

The question that changes a retention strategy is not how much a client spends, it is after how many days her silence becomes abnormal. That delay was calculated by country, on the best clients, as a median and not an average, because the average is crushed by a few extreme values. That threshold is what triggers a call-back.

5 · Translate the typology into a commercial system.

A grid of cumulative spend tiers, with benefits and written rules for each, in both working languages. And a change of logic in the loyalty programme: instead of the more you buy, the bigger your discount, which eats the margin, the more you buy, the more credit you build up, which protects it and brings the client back.

6 · Fix the collection at the source.

A segmentation does not survive a database that degrades. A bilingual welcome script with a decision tree was written for the till, with the mobile number as the unique identifier, a list of mandatory fields, and a way to handle the objection on email. Collecting the data became a step in the sales journey, not a box to tick.

04 · Results

2024 mandate, three countries and an e-commerce channel.

From two discount categories assigned by hand to ten named, computed segments
Three segmentation methods built and compared on the same base, with the discard rate published
Abnormal-silence threshold computed per country, as a median, to trigger callbacks at the right moment
Customer tier grid written in both working languages, with benefits and rules per tier
Loyalty logic inverted, from the discount that erodes margin to the credit that brings the customer back
Bilingual welcome script at the till, unique identifier and mandatory fields, so the base stops degrading
Operational reactivation campaign built on the segments, with call script and appointment booking

05 · What it proves

Plenty of companies believe they have an acquisition problem when what they have is a customer knowledge problem. As long as the base only separates those who spend from everyone else, every media pound goes on re-buying people you already had. A useful segmentation is not an analysis, it is an instrument: it names groups the teams recognise, it gives a trigger threshold, and it fills itself because collection was fixed at the same time.