Guides  /  customer analytics

Customer segmentation and conjoint analysis

Segmentation sorts customers into groups. Conjoint works out what they will trade off. Both are routinely run, routinely over-trusted, and routinely presented as though the output were a discovery rather than a choice the analyst made. This guide works both through with numbers, covers how many segments you actually have, and shows how to read part-worth utilities without overstating them.

Tara Sutaria Written and reviewed by Tara Sutaria, Senior Statistical Analyst
Updated 4 November 202613 min read
What is the difference between segmentation and conjoint analysis?

Segmentation divides existing customers into groups that resemble each other, using data you already hold. It is descriptive and the number of groups is a decision, not a finding. Conjoint analysis is experimental: you show people designed combinations of product features, observe their choices, and estimate how much each feature is worth to them. Segmentation tells you who your customers are. Conjoint tells you what they will give up to get something else.

Definitions

What segmentation actually does

Clustering algorithms always return clusters. Give k-means random noise and ask for four groups and it will produce four tidy groups with centroids and sizes. Nothing in the output tells you whether the structure was in the data or imposed on it.

That is the single most important thing to understand before presenting a segmentation to a board.

MethodHow it groupsWatch for
k-meansMinimises distance to cluster centresNeeds k chosen in advance. Assumes roughly spherical, similar-sized clusters. Sensitive to scaling.
HierarchicalBuilds a tree of nested groupingsNo need to fix k upfront, but it does not scale, and the linkage choice changes the answer.
Latent classModel-based, assigns probabilitiesGives fit statistics you can compare, which the others do not. Slower and needs more care.
RFMRecency, frequency, monetary value, usually in tiersCrude, transparent, and often more actionable than anything clever.
Scale your variables first

k-means works on distances, so a variable measured in pounds will overwhelm one measured in visits purely because its numbers are bigger. Standardise everything before clustering. Skipping this step is the commonest reason a segmentation turns out to be a map of whichever variable had the largest units.

Worked example one

How many segments?

Run k-means for k from 1 to 7 and record the within-cluster sum of squares, which measures how tightly points sit around their centre.

kWithin-cluster SSReduction on previous
11,000—
262038.0%
343030.6%
436016.3%
53308.3%
63125.5%
73003.8%

Reading it

Adding a second group explains 38% more. A third adds 30.6%. The fourth adds 16.3% and the fifth only 8.3%. The gains flatten after three or four, which is the elbow.

Note what the table does not do: it never stops improving. Within-cluster sum of squares falls at every k, all the way to one cluster per customer. There is no value of k at which the algorithm tells you to stop. The elbow is a judgement.

Three checks worth running alongside it

✓Silhouette width. Measures how well each point sits in its own cluster versus the next nearest. Above 0.5 is reasonable structure; below 0.25 means the clusters are largely arbitrary.
✓Gap statistic. Compares your within-cluster SS against what you would get from uniform random data with the same spread. If the gap is small, your structure is not much better than noise.
✓Business usability. Seven segments that nobody can act on is worse than three that marketing can address. This is a legitimate criterion, not a compromise.

The missing check

The question nobody asks: is it stable?

A segmentation that changes when you resample the data is not describing your customers. It is describing this particular extract.

Test it directly. Split the data in half at random, cluster each half separately, and see whether the same segments appear. Or bootstrap: resample with replacement fifty times, recluster, and measure how often pairs of customers land together.

The Jaccard index is the usual summary. Above 0.75 a cluster is stable and worth naming. Between 0.6 and 0.75 it is a pattern worth noting. Below 0.6 it will not reproduce and should not be built into a strategy.

Why this matters commercially

Segments get named, given personas, and written into campaign plans and sometimes into org charts. If the underlying grouping would not survive a resample, all of that is built on an artefact of one data pull. Running the stability check takes an afternoon. Discovering the problem after a rebrand does not.

The same caution applies to profiling. Once segments exist, people test them against every variable in the database to describe them, which is a large number of comparisons and will throw up differences by chance. Treat profiling as description, not as hypothesis testing, or correct for the multiplicity.

The simple alternative

RFM, and when crude beats clever

Before commissioning a clustering exercise, it is worth asking whether RFM would do the job. It often does, and it has advantages no algorithm offers.

Score every customer on three things: how recently they bought, how frequently, and how much they spent. Divide each into quintiles, 1 to 5, and you have a three-digit code per customer.

A worked cut

Take 10,000 customers and split each dimension into five equal groups of 2,000.

CodeMeaningRoughlyAction
555Bought recently, often, and a lot~80 customersProtect. Early access, named contact.
155Was valuable, has not returned~80Win back. This is the highest-value intervention available.
511New, one purchase, low value~80Onboard and develop.
111Long gone, infrequent, low value~80Suppress. Stop spending on them.

Cell sizes are roughly even by construction if the three dimensions were independent, which they are not, so in practice the corners fill up and the middle thins out. That is informative in itself.

Why it holds its own against clustering

✓It is reproducible by definition. No random starts, no seed, no stability question. Run it twice and get the same answer.
✓Anyone can explain it. A marketing manager can describe segment 155 to a colleague without a slide about Euclidean distance.
✓It is directly actionable. The segments map onto decisions, which is more than can be said for a cluster defined by a centroid in eleven dimensions.
✓It uses behaviour, not attitudes. What people did, not what they said in a survey.

Where it falls short

RFM says nothing about why. It cannot incorporate product mix, channel, demographics or stated preference, and it treats the three dimensions as equally important when they rarely are. For a subscription business where almost everyone buys on the same cadence, frequency carries no information at all and the method collapses to two dimensions.

A sensible order of work

Run RFM first. If it separates your customers in ways the business recognises and can act on, you may be finished. Move to clustering when you have variables RFM cannot use and a specific question it cannot answer. Starting with the complex method and working backwards is how segmentation projects run long and deliver personas nobody uses.

A different question

How conjoint works

Asking customers what matters to them produces a list on which everything matters and price matters most. Conjoint avoids this by never asking the question directly.

Instead respondents choose between whole product configurations, each a bundle of attribute levels. Because the bundles are designed so attributes vary independently, the choices reveal the trade-offs people will actually make.

✓Attributes are the features that vary: price, delivery speed, support level, brand.
✓Levels are the options within each: £20, £30, £40.
✓Part-worth utility is the value a respondent places on each level, estimated from their choices.
✓Importance is how much an attribute can swing total utility, derived from the range of its part-worths.

Choice-based conjoint, where people pick one bundle from several, is the standard today because it mirrors a purchase decision. Older ratings-based designs ask people to score configurations and are easier to run but further from real behaviour.

Worked example two

Part-worths and importance

A choice experiment on four attributes returns these part-worth utilities, averaged across respondents.

AttributeLevels and part-worthsRangeImportance
Price£20 = +0.85 · £30 = +0.10 · £40 = −0.951.8049.3%
DeliveryNext day = +0.42 · 3 days = +0.05 · 7 days = −0.470.8924.4%
Support24/7 = +0.30 · Office hours = −0.300.6016.4%
BrandKnown = +0.18 · Unknown = −0.180.369.9%

How importance is calculated

Take the range of each attribute's part-worths, then express each as a share of the total.

Price range = 0.85 − (−0.95) = 1.80
total of all ranges = 1.80 + 0.89 + 0.60 + 0.36 = 3.65
Price importance = 1.80 / 3.65 = 49.3%

The caveat that should always accompany it

Importance depends entirely on the levels you chose. Price looks dominant here partly because the range tested was £20 to £40, a doubling. Test £28 to £32 instead and price would look unimportant, not because customers stopped caring but because you stopped varying it.

Attribute importance is a statement about your experimental design as much as about your customers. Report the levels alongside the importances, every time.

Worked example three

Predicting preference share

The useful output is not the importance table. It is the ability to compare configurations that do not exist yet.

Two candidate products

A: £30, next day delivery, 24/7 support, known brand
utility = 0.10 + 0.42 + 0.30 + 0.18 = +1.00

B: £20, 7 day delivery, office hours support, known brand
utility = 0.85 + (−0.47) + (−0.30) + 0.18 = +0.26

Converting utility to share

Under the standard logit rule, share is the exponential of each utility divided by the total.

share(A) = e^1.00 / (e^1.00 + e^0.26) = 67.7%
share(B) = 32.3%

What this tells the business

The more expensive product wins by better than two to one. Charging £10 more buys back more than it costs, provided the service around it improves. That is a pricing decision supported by evidence rather than instinct, and it is the reason conjoint is worth its cost.

Share of preference is not market share

This number says what respondents would choose among the options you showed them. It takes no account of distribution, awareness, salesforce, switching costs, or competitors you did not include. Treating a 68% preference share as a 68% market share forecast is the most expensive mistake available in this method.

Honesty

What conjoint cannot tell you

✓It is stated preference, not revealed. People choose in a survey without spending money. Hypothetical bias tends to understate price sensitivity, so real-world demand at a given price is usually worse than the model suggests.
✓You cannot extrapolate beyond the levels tested. If £40 was the highest price shown, the model says nothing about £50. The relationship is estimated only across the range you presented.
✓Too many attributes break it. Beyond about six attributes respondents simplify, often by focusing on one or two and ignoring the rest. The model still returns tidy estimates for all of them.
✓Averages hide segments. A mean part-worth across a sample containing two groups with opposite preferences describes nobody. Hierarchical Bayes estimation gives individual-level utilities and is now standard for exactly this reason.
✓No competitor response. Your rival will react to your launch. The model assumes a frozen market.

Doing it

In R, SPSS and Stata

R

Segmentation, with scaling and the stability check that usually gets skipped.

d <- scale(customers[, vars])
km <- kmeans(d, centers = 4, nstart = 25)
factoextra::fviz_nbclust(d, kmeans, method = "silhouette")
fpc::clusterboot(d, B = 100, clustermethod = fpc::kmeansCBI, k = 4)

nstart = 25 matters: k-means starts from random centres and can land in a poor solution, so it should be restarted many times. clusterboot returns the Jaccard stability per cluster.

For choice-based conjoint, support.CEs builds the design and mlogit or bayesm estimates it, with ChoiceModelR for hierarchical Bayes.

SPSS

QUICK CLUSTER z_recency z_frequency z_value
  /MISSING = LISTWISE
  /CRITERIA = CLUSTER(4) MXITER(50) CONVERGE(0)
  /METHOD = KMEANS(NOUPDATE)
  /PRINT ID(custid) INITIAL ANOVA.

Standardise the inputs first with DESCRIPTIVES /SAVE, since QUICK CLUSTER will not do it for you. SPSS has a Conjoint module with the CONJOINT command, but it handles ratings-based designs rather than modern choice-based conjoint.

Stata

cluster kmeans z_*, k(4) start(random(12345)) name(seg4)
cluster stop seg4, rule(calinski)
tabstat spend visits, by(seg4) stat(mean sd n)

For conjoint, clogit choice attr1 attr2, group(taskid) fits the conditional logit, and mixlogit allows preferences to vary between respondents, which is the practical equivalent of hierarchical Bayes.

Set the random seed and record it

Both k-means and hierarchical Bayes depend on random starting values. Without a fixed seed the segmentation you present on Monday may not be the one you can reproduce on Friday, and you will not be able to explain why. Record the seed in the methods alongside the software version.

Avoid these

Six mistakes that cost you money

MistakeWhy it matters
Not scaling before clusteringk-means works on distance, so the variable with the largest units dominates. The segmentation becomes a map of that one variable.
Treating the number of segments as a findingWithin-cluster SS improves at every k. The algorithm never says stop. Choosing k is a judgement and should be presented as one.
Skipping the stability checkA segmentation below 0.6 Jaccard will not reproduce. Personas and campaigns then rest on one data pull.
Quoting attribute importance without the levelsPrice looks dominant at £20 to £40 and trivial at £28 to £32. Importance describes your design as much as your customers.
Reading preference share as market shareIt ignores distribution, awareness, switching costs and every competitor you left out of the design.
Averaging part-worths across opposed groupsA mean utility across two segments with opposite preferences describes nobody in the sample.
Would your segments survive a rerun?

Clusters always appear, including in random data, so the real question is whether the same ones come back on a different sample. Send the data and you get segments tested for stability, profiled, and sized so the commercial case is honest.

See survey and insight analysis
Tara SutariaTara SutariaSenior Statistical Analyst

Writing it up

How to report it

✓For segmentation: the variables used, whether they were standardised, the method, k and how k was chosen.
✓Cluster sizes and a stability measure. A segment holding 3% of customers with a Jaccard of 0.5 should carry a warning.
✓For conjoint: the full attribute and level list, because importance cannot be interpreted without it.
✓The design type, the number of tasks per respondent, and the estimation method.
✓Preference shares labelled as preference shares, with the limitation in the same sentence.
✓The random seed and software version, so the analysis can be reproduced.

Common questions

Frequently asked questions

How many customer segments should I have?

There is no statistical answer, because within-cluster variation falls at every additional segment all the way to one customer per group. Use the elbow in the within-cluster sum of squares, silhouette width above about 0.5, and whether the business can actually act on the number you pick. Three or four is common and usually more useful than seven.

Do I need to standardise variables before clustering?

Yes, for any distance-based method such as k-means. Spend measured in pounds and visits measured in counts are on completely different scales, so without standardising, the variable with larger numbers dominates the distances and the segmentation mostly reflects that one variable.

How do I know my segments are real?

Test whether they reproduce. Split the data in half and cluster each half, or bootstrap fifty resamples and measure how often the same customers group together. A Jaccard index above 0.75 indicates a stable cluster; below 0.6 it will not reproduce and should not be built into strategy.

What is a part-worth utility?

The value a respondent places on one level of one attribute, estimated from the choices they made between whole product configurations. Summing the part-worths of a configuration gives its total utility, and comparing total utilities predicts which option people prefer.

How is attribute importance calculated?

Take the range between the highest and lowest part-worth within an attribute, then express it as a share of the total of all such ranges. A price range of 1.80 against a total of 3.65 gives price an importance of 49.3%. The figure depends heavily on the levels tested, so always report them alongside.

Can conjoint predict my market share?

No. It predicts share of preference among the configurations shown to respondents in a survey. Real market share depends on distribution, awareness, salesforce, switching costs and competitor reaction, none of which the experiment contains. Treat it as a comparison between options, not a forecast.

How many attributes can a conjoint handle?

Around six before respondents start simplifying, typically by focusing on one or two attributes and ignoring the rest. The model will still return clean estimates for every attribute, which is what makes over-specification dangerous rather than obviously wrong.