Segmentation Is Not Demographics

The segment that feels rigorous and tells you nothing

Most segmentation decks open the same way. Here is our customer split by age. Here they are by gender. Here is the postcode heat map, the household income bands, the neat pie chart of who our buyers are. It looks like homework. It looks like someone did the analysis. And nine times out of ten it tells you almost nothing about why anyone bought anything.

That is the uncomfortable bit, so let me say it plainly. Demographic segments feel rigorous and usually explain close to nothing about what people actually purchase. Knowing that a customer is a woman aged 35 to 44 living in a mid income postcode is knowing a lot about her census record and very little about her shopping basket. Two people who match on every one of those boxes can walk into your shop wanting opposite things, and you will have grouped them together and written one message for both.

Let me be fair to demographics before I spend the rest of this piece taking them apart. They are not useless. They are easy to collect, they are stable, and once in a while they really do line up with a need. Nappies and prams genuinely do sell to new parents, and nobody is ordering a stairlift for the fun of it. The trouble is that we treat a weak signal as if it were the whole story. Age is a proxy for a need at best, and a lazy one at that. The moment a stronger signal is sitting untouched in your order history, reaching for the census instead is a choice, and usually the wrong one.

Two identical people, opposite baskets

Picture two customers. Same age to the year, same gender, same postcode, near enough the same income. On every demographic chart you own they sit in the exact same cell. You would target them with the same ad and the same offer.

Customer one buys a pair of trail running shoes, then a fortnight later some gels, then a lightweight jacket, then more shoes three months on. She is training for something. She buys on performance, she reorders consumables, she cares about weight and grip and she will pay for the good stuff.

Customer two bought one pair of the same shoes in November, in a gift box, to somebody else's size, and has not been back since. She is a gift buyer. She bought on the recommendation and the returns policy, price mattered, and the running has nothing to do with her.

Same demographic cell. Opposite needs. If you send them the same email about a new midsole technology, you delight one and baffle the other. The demographic box told you they were the same. Their behaviour tells you they could not be more different. That gap is the whole point of this article, and it is the first thing worth sitting with.

Two very different people who want the same thing

Now flip it. A 24 year old student and a 61 year old company director share no demographic box at all. Different age bracket, different income, different everything on the census. Yet both of them come to your shop for exactly one reason: they want the item that arrives tomorrow, guaranteed, because they left it too late. They are both in the same segment. Call it the last minute segment. They want speed and certainty, they barely look at price, and the message that wins both of them is the delivery cut off clock, not the product story.

Demographics put them in different rooms. Behaviour puts them in the same one. Once you have seen a couple of examples like this you cannot unsee it: the thing that predicts what someone buys is what they do, not who they are on paper. Here is the a ha to hold on to. Demographics describe the person. Behaviour describes the need. You are not selling to a person, you are selling to a need, so measure the thing you are actually selling to.

What a real segment actually is

So here is the definition I work from. A real segment is a cluster of need and behaviour. It is a group of customers who behave alike, who want the same job done, who respond to the same trigger, regardless of whether they share a single line on a census form.

And behaviour is something you already measure, you have just been filing it under reporting instead of using it. How recently did they buy. How often. What is in the basket, and what goes with what. How do they respond to a discount, do they wait for one or ignore it. Do they browse for weeks or buy in ninety seconds. Do they return things. Do they reorder the same consumable on a rhythm. Every one of those is a behaviour signal, and stacked together they describe a person far better than their postcode ever will.

The job, not the person

There is an old idea in marketing that I keep coming back to because it refuses to age: people do not buy products, they hire them to do a job. Nobody wants a drill, they want a hole, and really they want a shelf, and really they want a tidy room. Your customer has a job to be done, and they hire whatever gets it done. Two people hire the same product for wildly different jobs, and that is exactly why the product category on your invoice is a poor guide to who your customer really is.

Segments are jobs wearing customers. When you cluster behaviour you are not really grouping people, you are grouping jobs, and the reason three or four groups keep falling out is that most shops serve three or four jobs. The gift job. The stock up job. The treat myself job. The emergency job. Once you name the job, the message writes itself, because you finally know what the customer came to get done. This is the second a ha, and it reframes the whole exercise: you are not sorting humans, you are sorting the reasons they showed up.

The behaviour signals worth measuring

You do not need a hundred fields to cluster well. You need a handful of honest ones. Here are the signals that carry the most weight in almost every shop I look at, and what each one quietly reveals about the job the customer is hiring you for.

Signal	What it measures	What it hints at
Recency	Days since the last order	Whether they are active, lapsing or gone
Frequency	Orders in a window	Habit versus one off intent
Monetary	Average and total basket	Value tier and headroom
Discount response	Share of orders on promotion	Whether price is the trigger
Basket shape	What sits together in a cart	The job being done
Time to buy	Minutes from landing to checkout	Considered versus impulsive
Return rate	Share of items sent back	Gifting, uncertainty or fit trouble
Reorder rhythm	Gap between repeat buys of a consumable	Subscription intent hiding in plain sight

You will not use all eight every time. Three or four strong signals beat eight weak ones, because every extra axis you add spreads your customers thinner and makes the clumps harder to see. Start with recency, frequency, basket and discount response, and only add more when a group refuses to separate.

Finding the clusters without guessing

Fine, you say, but I cannot eyeball ten thousand customers and sort them into groups by hand. Correct. This is exactly the job cluster analysis was built for, and it is the backbone of the segmentation chapters in Cutting Edge Marketing Analytics from 2014 and again in Business Intelligence Analytics and Data Science from 2018. The idea is lovely and simple. You put each customer in a space where every axis is a behaviour, one axis for frequency, one for average basket, one for discount sensitivity, and so on. Customers who behave alike sit near each other in that space. A clustering method then finds the natural clumps.

The workhorse method is k means. You tell it how many groups to look for, call that number k, and it does two things over and over until it settles. It assigns every customer to the nearest group centre, then it moves each centre to the middle of the customers it just collected. Repeat, and the centres drift until they sit in the heart of real clumps of behaviour.

Put every signal on the same scale first

Here is a trap that catches nearly everyone the first time, and it is worth an a ha of its own. Your signals live on wildly different scales. Frequency might run from 1 to 20 orders. Basket might run from 5 to 500 pounds. Distance in that raw space is almost entirely basket, because a swing of a few hundred pounds dwarfs a swing of a few orders. Cluster on the raw numbers and you are secretly clustering on money alone, and the machine will never see the pattern in frequency at all.

The fix is to standardise every signal before you cluster, so each one speaks with the same voice. The usual move is the z score: for each signal, subtract its average and divide by its spread.

$z = \frac{x - \mu}{\sigma}$

In plain words, take a value $x$, subtract the average $\mu$ of that signal across all customers, and divide by the standard deviation $\sigma$, the typical distance from the average. Now every signal is measured in the same unit, standard deviations from the middle, and a big move in frequency counts for exactly as much as a big move in basket. Skip this step and every clustering you ever run is quietly wrong. Do it and you have removed the single most common reason segments come out mushy.

The distance that decides who sits together

Clustering lives or dies on one question asked over and over: how far apart are these two customers. Once every signal is standardised, the honest answer is the straight line distance between them in the behaviour space, the same Pythagoras you met at school, just with more axes.

$d(a, b) = \sqrt{\sum_{j} (a_j - b_j)^2}$

For two customers $a$ and $b$, walk through every signal $j$, take the gap between them on that signal, square it, add the squares up across all signals, and take the square root. Small distance means they behave alike and belong together. Large distance means they belong apart. That is the entire notion of near that k means uses when it hands each customer to the closest centre, and it is why standardising first matters so much: the distance is only fair if every axis was put on the same scale before you measured it.

What k means is really chasing

So what is the method trying to make good? It is trying to make each group as tight as possible. We measure tightness with the within cluster sum of squares, the total squared distance from every customer to the centre of their own group.

$WCSS = \sum_{i} \| x_i - \mu_{c(i)} \|^2$

Read it in plain words. For each customer $i$, take their position $x_i$, find the centre of the cluster they belong to, written $\mu_{c(i)}$, measure the distance between the two, square it, and add all those up. When WCSS is small, every customer sits close to their group centre, which means the groups are tight and the members really do behave alike. k means is simply the machine hunting for the group centres that make this number as small as it can.

How many segments: the elbow method

There is one decision the machine cannot make for you: how many groups should there be. That is the k. Pick k too small and you crush distinct customers together. Pick it too large and you slice one real group into meaningless shards.

Here is the trick, and it is worth sitting with. Run the clustering for k equals one, two, three, and so on, and write down the WCSS each time. WCSS always falls as you add groups, because more centres can always sit closer to more people. But it does not fall evenly. It plunges while you are still splitting genuinely different customers apart, then it flattens out once the real groups are found and all you are doing is chopping up clumps that were already tight.

Watch it happen on a small example.

Number of clusters	WCSS	Fall from the row above
1	100
2	55	45
3	30	25
4	22	8
5	18	4
6	15	3

Look at the last column. Going from one group to two buys you 45. Two to three buys you another 25. Then three to four buys only 8, and every step after that is loose change. The big falls stop after three. That bend, where the steep drop gives way to the gentle slope, is the elbow, and the elbow is your answer. Here it says three, maybe four, real segments. Not because a rule told you to, but because the data stopped rewarding you for splitting further. Most businesses I look at land on three or four true groups, which is a very human number and no accident.

A second opinion when the elbow is blurry

Sometimes the elbow is obvious and sometimes it is a gentle curve with no clear bend, and then you want a second opinion. The one I trust is the silhouette. For a single customer it asks a simple question: are you closer to your own group than to the nearest rival group. Let $a$ be the average distance from a customer to the others in their own cluster, and $b$ the average distance to the members of the nearest other cluster. The silhouette for that customer is

$s = \frac{b - a}{\max(a, b)}$

It lands between minus one and one. Close to one means the customer sits comfortably inside their own group and far from the next, a clean fit. Near zero means they sit on the border, they could go either way. Below zero means they are closer to another group than their own, which is the machine telling you the label is wrong. Average the silhouette across every customer for each value of k, and the k with the highest average is the tidiest split. When the elbow and the silhouette agree, you can stop arguing with yourself. When they disagree, trust the one that gives you groups you can actually name and act on, because a segment you cannot describe to a copywriter is not a segment, it is a number.

From behaviour to segment to action

Here is the whole path on one picture, from the raw signals to what you actually do on Monday.

The mistake is to stop at the demographic chart, box A skipped entirely, and jump straight to writing one message for everyone. The value lives further along, where the behaviour becomes a segment and the segment becomes a different promise and a different price for each real group.

And here is the pipeline the analyst walks through to get there, so you can see where each formula from above earns its place.

What the clusters usually turn out to be

When you actually run this on a real shop, the groups that fall out have personalities, and they are never the ones on the demographic slide. You tend to meet some version of these.

The loyal regular. Buys often, medium basket, ignores discounts because they were going to buy anyway. Your profit lives here.

The deal seeker. Only ever buys on promotion, big basket when they do, vanishes at full price. Trainable, but only just.

The one off gifter. One purchase, often seasonal, high return rate, never a consumable. Easy to please, will not be loyal, and you should stop pretending they will.

The considered high spender. Long browse, few but large orders, reads everything, wants reassurance not a nudge.

You will not get exactly these four, that is the point, you find yours. But notice that a 30 year old and a 60 year old can both be loyal regulars, and two identical looking 40 year olds can split one into a deal seeker and the other into a high spender. The behaviour sorted them. The demographics never could.

A fifth face you sometimes meet: the lapsing loyalist

There is a fifth personality worth calling out because it is the most valuable one to catch and the easiest to miss. The lapsing loyalist looks like a loyal regular in every signal except recency. Same healthy basket, same habit of buying at full price, same history of frequent orders, and then a gap that keeps growing. On a demographic slide they are invisible, filed with everyone else their age. On a behaviour cluster they light up, because recency has drifted while everything else stayed put. This is the group where a single well timed we have missed you message earns more than any acquisition campaign you could run, and you only ever see them if you cluster on behaviour and watch recency as its own axis. Catch the lapsing loyalist early and you keep a customer you already paid to win. Miss them and you pay again to replace them.

A worked mini case: a coffee subscription shop

Let me make this concrete with a small worked example. Imagine a shop that sells coffee, beans and gear. We pull four signals for a sample of customers: orders in the last year, average basket in pounds, the share of their orders placed on discount, and days since their last order. Here is a handful of them, standing in for the thousands you would really have.

Customer	Orders per year	Avg basket	Discount share	Days since last
A	14	32	0.05	12
B	2	41	0.90	8
C	13	30	0.10	140
D	1	68	0.00	200
E	15	34	0.00	9
F	3	39	0.85	15
G	1	72	0.00	30
H	12	29	0.05	160

Do not read age or postcode, there is none here, that is the point. Read the behaviour. Standardise the four columns so basket does not drown the rest, then cluster and let the elbow tell you how many groups. On this shape the elbow lands at four, and the groups more or less pick themselves.

A and E cluster together: many orders, modest basket, almost never on discount, bought last week. That is the loyal regular, the beating heart of the subscription. B and F cluster together: rare orders, decent basket, almost always on discount, only buy when there is a deal. That is the deal seeker. D and G cluster together: a single large order, full price, never came back, a grinder or a gift set once and done. That is the one off high spender. And C and H cluster together: they used to look exactly like A and E, frequent and loyal, but their days since last has crept out past a hundred while everything else stayed identical. That is the lapsing loyalist, and they are the most urgent group on the page.

Notice what just happened. Four behaviour groups, four different jobs, four different next actions, and not one of them came from knowing anybody's age. The loyal regulars get the new single origin first. The deal seekers get the bank holiday offer and nothing at full price. The one off spenders get a gentle nudge toward a subscription they never considered. And the lapsing loyalists get a we saved your usual message before the gap becomes permanent.

Reading a cluster centre like a portrait

Once the clusters exist, the most useful thing you own is the centre of each one, the average customer of that group. Lined up side by side, the centres read like portraits, and the differences between the columns are your whole strategy on one page.

Signal at the centre	Loyal regular	Deal seeker	One off spender	Lapsing loyalist
Orders per year	High	Low	Very low	Was high
Average basket	Medium	Medium	High	Medium
Discount share	Low	Very high	None	Low
Days since last	Low	Medium	High	High and rising
The job	Keep me stocked	Get me a deal	Buy this once	I drifted away
The move	Reward and first look	Fenced offers	Nudge to subscribe	Win back now

When you can read a segment like this, off the centre rather than off a hunch, the argument about who to email and what to charge stops being a matter of opinion. The portrait tells you.

Messaging: one product, several different promises

Once you have your three or four groups, your marketing stops being one shouted message and becomes a handful of quiet accurate ones. Same product, different promise, because each group wants a different job done. The loyal regular wants to hear about the new thing first and be thanked. The deal seeker wants the offer and the deadline. The gifter wants easy, safe, returnable, wrapped. The considered spender wants the detail, the guarantee, the proof. Sending all four the same email is why your open rates are what they are. Sending each of them their own is the single highest return change most shops can make, and it costs nothing but the segmentation.

Pricing: the same shelf, the right price for each need

Segmentation should move your pricing too, not just your words. The deal seeker is telling you, with their behaviour, that they will only move on a discount, so your promotions should be aimed at them and fenced off from everyone else, because the loyal regular would have paid full price and you just gave the margin away. The considered high spender is telling you they buy on reassurance, so a premium bundle with a strong guarantee will out sell a price cut for that group every time. When you price to a behaviour segment you stop leaving money on the table at both ends, the discount you did not need to give and the premium you never thought to offer.

Common mistakes I see again and again

I have watched a lot of segmentation projects go sideways, and they nearly always trip on the same handful of things. Here are the ones to guard against.

Clustering on the raw numbers. If you do not standardise first, you are clustering on whichever signal happens to have the biggest range, usually money, and calling the result insight. Standardise, every time.

Chasing too many segments. The elbow says three or four and the temptation is to run twelve because twelve feels more precise. It is not more precise, it is more fragile, and nobody can write twelve different emails that actually differ. Fewer, truer groups beat a dozen shards.

Segmenting on things you cannot act on. A segment is only worth having if it changes what you do. If two groups get the same message and the same price, they are one group wearing two labels.

Building it once and never again. Behaviour drifts. The loyal regular lapses, the deal seeker reforms, new customers arrive. A segmentation is a photograph, not a monument, so re run it on a sensible rhythm, quarterly for most shops, and let the groups move.

Letting demographics sneak back in. The pull to add age back into the model is strong because it is familiar. Resist it unless it earns its place by making the groups genuinely tighter. Most of the time it just blurs them.

Naming the segment before you understand it. The name should come from the centre, from the actual behaviour, not from a persona somebody sketched in a workshop. Read the portrait first, then name it.

How to apply this without a data team

You can start this in an afternoon with a spreadsheet, long before you hire anyone or buy anything. Here is the sequence I would follow.

First, pull your order history into one table, one row per customer, and work out four columns: how many times they bought, their average basket, the share of their orders that used a discount, and how many days since their last order. That is your behaviour space, and it is already more honest than any demographic file you own.

Second, sort by each column in turn and just look. Patterns jump out even before any maths: the cluster of frequent full price buyers, the cluster that only ever bought on offer, the big baskets that came once and vanished. You will very likely see your three or four groups with the naked eye, because real segments are not subtle once you look at behaviour.

Third, if you want to do it properly, standardise the four columns and run k means for k from two to six in any tool you already have, a notebook, a spreadsheet plugin, whatever. Plot the WCSS, find the elbow, and take the number it gives you. Check it against the silhouette if the elbow is soft.

Fourth, name each group from its centre, write down the one job it is hiring you for, and decide one message and one price for it. Not a strategy document, four lines, one per segment.

Fifth, act, and then measure. Send each group its own email, fence your discounts to the group that needs them, and watch the open rates and the margin. The proof is in the next campaign, not in the deck. When the groups start to drift, and they will, run the whole thing again.

That is the entire method, and none of it requires permission or a big budget. The hard part is not the maths, it is the decision to describe your customers by what they do instead of who they are.

The one thing to do this week

You do not need a data science team to start, and you do not need to cluster ten thousand people on Monday morning. Pull your order history, the real behaviour, recency and frequency and basket and discount response, and split your customers into just three groups by hand using those signals. Ignore age and postcode completely for one afternoon. Then write three emails instead of one and watch what happens. That is the whole method in miniature, and it will beat the demographic deck before you have clustered a single thing properly.

If you want a hand turning your order history into real behaviour segments, and then into messaging and pricing that speak to each of them, that is squarely the kind of work I do. My data and ecommerce work is built around exactly this: finding the three or four groups you actually have, not the ones the demographic slide pretends you have. If it would help to talk it through, book a call and we will look at your order history together.

The one idea to leave with

Stop describing your customers by who they are and start describing them by what they do. Age and postcode feel rigorous and predict almost nothing. Need and behaviour look messy and predict almost everything. Two people in the same demographic box can want opposite things, two people in different boxes can want the same thing, and the only way to see the truth is to cluster the behaviour and let the elbow tell you how many groups you really have. Find those three or four, speak to each one on its own terms, price to each one on its own terms, and you will have replaced a slide that looked like analysis with a map that actually works.