---
name: audience-segmentation
description: >
  Builds and validates segmentations that describe real, stable, actionable groups
  rather than arbitrary partitions of a homogeneous market, starting from whether
  the category is segmentable at all. Use for "run a segmentation", "segment the
  market", "build customer segments", "cluster analysis", "who are our audiences",
  "how many segments should we have", "is this segmentation any good", "build a
  typing tool", "the segments do not feel real", "we need to refresh the
  segmentation".
category: 09 Segmentation and Audience Understanding
ref: 09.01
tier: 1
inherits: [K2, K3, K4, K5]
---

# Audience Segmentation

## 1. One-line description
Determines whether an audience divides into meaningful groups at all, and where it does, builds a segmentation on basis variables chosen from the decision it must serve, tests it for stability rather than accepting statistical separation as proof, profiles it on evidence not used to construct it, and reports honestly when the market turns out to be one thing rather than five.

## 2. What this skill is used for

**The research problem it solves.** Segmentation is the most frequently commissioned and least frequently validated deliverable in commercial research, and the reason is structural: a clustering algorithm always returns clusters. Feed it random numbers and it will partition them, report separation statistics, and produce groups that can be named, coloured and put on a wall. Nothing in the standard workflow asks whether the partition corresponds to anything in the world. So a study that should have concluded "this market is largely homogeneous on the dimensions that matter to your decision" instead delivers five segments with memorable names, and the business spends three years building products, campaigns and targeting rules for groups that do not exist. The second failure is quieter and more common: segments built on demographics, which are easy to collect, easy to reach and almost never the reason people behave differently. The third is a segmentation that is real and useless, because nobody can identify which segment a customer belongs to, or because the business would do the same thing for all five. This skill supplies the prior question, the basis-variable discipline, the stability tests that separate a finding from an artefact of one algorithm run, and the standard for calling a segmentation reportable.

**Where it sits.** Analysis and synthesis. It takes prepared quantitative data, behavioural evidence and needs evidence, and hands validated segments to persona development, to comparison work, to targeting and to activation. It can also run as an audit of a segmentation somebody else built.

**Typical use cases.**
- Building a needs-based or attitudinal segmentation from a purpose-designed survey.
- Deciding how many segments a dataset actually supports, when the client has asked for "four or five".
- Auditing an existing segmentation that stakeholders say does not feel real, or that nobody uses.
- Refreshing a segmentation that is several years old and may have decayed.
- Building a typing tool so that customers in a database or a new survey can be allocated to segments.
- Establishing that a market is not usefully segmentable, and saying so before money is committed.

**Who uses it.** Quantitative research directors and senior analysts running segmentation studies; client-side insight and strategy leads who commission them and have to defend them internally; brand and product strategists who inherit a segmentation and need to know how much weight it will carry; anyone reviewing a segmentation produced by an automated pipeline, where the absence of the prior question is the default rather than an oversight.

## 3. When to use it

- A decision requires the audience to be treated differently in groups, and the groups are not already known.
- You have or can field a dataset with candidate basis variables measured on the whole sample.
- An existing segmentation is in use and its validity, stability or currency is in question.
- Stakeholders disagree about who the customers are, and the disagreement is blocking a product, pricing or communications decision.
- A previous study showed strong bimodality or genuinely divergent behaviour on a key measure, suggesting more than one population is present.
- A typing tool is needed so that segments can be applied to a customer database, a media plan or a future wave.
- You are being asked to validate someone else's segmentation before the organisation builds on it.
- The honest answer might be that no segmentation is warranted, and someone needs to be able to say so with evidence.

## 4. When NOT to use it

- **The category may not be segmentable, and this has not been checked.** This is the precondition, not a caveat. Many categories are close to homogeneous on the dimensions that drive a given decision: people want largely the same things, for largely the same reasons, and differ mainly in how much they buy. Forcing a segmentation onto such a market does not produce weak segments, it produces arbitrary ones, and arbitrariness is invisible in the output because the statistics look identical either way. Run the heterogeneity assessment in step 2 before anything else. Where it fails, the deliverable is a description of the market, the evidence that it is not usefully divided on these dimensions, and the statement that a segmentation was not built. That is a professional result, and it is far cheaper than the alternative.
- **The groups already exist and the question is how they differ.** Customer tiers, screener-defined groups, markets, product holders, users and non-users are not segmentation output; they are given. Comparing them is **09.04 Segment Comparison**. Running a clustering algorithm on data that already has the relevant split in it is a way of losing the split.
- **The decision the segmentation must serve has not been established.** Basis variables are chosen from the decision. Without it there is no principled way to choose between a needs basis, an attitudinal basis and a behavioural basis, and the choice is then made by whatever was in the questionnaire. Go to **01.02 Business Problem to Research Question** first. A segmentation built without a decision is a taxonomy, and taxonomies do not get used.
- **The organisation cannot act differently towards different segments.** Actionability is a property of the business, not of the data. Where there is one product, one price, one channel and no ability to identify individuals, a statistically excellent segmentation changes nothing and consumes attention that a positioning or a prioritisation study would have repaid. Say this before fielding, not in the debrief.
- **The unit of variation is the occasion rather than the person.** In many categories the same person behaves in different ways on different occasions, and the variance between occasions within a person exceeds the variance between people. Segmenting people then averages away the thing that matters. Establish the unit first with **09.03 Behavioural Profiling**, and segment occasions if that is where the structure lives.
- **The sample cannot carry it.** Segmentation needs enough respondents that each candidate segment has a reportable base after the solution is chosen, and it needs a sample that covers the population rather than the reachable part of it. A convenience sample produces segments of the sample, not of the market, per K4 §3.3. Where the smallest plausible segment would fall below a reportable base, say so before the fieldwork rather than suppressing a segment afterwards.
- **What is wanted is a persona, not a segment.** Personas are a communication format built on evidence, and they are **09.02 Persona Development**. A persona built directly from a clustering run, without the validation steps below, inherits every weakness of the solution and hides it behind a photograph and a name.
- **The request is to reproduce a conclusion that has already been chosen.** Where a stakeholder has named the segments in advance and wants the analysis to find them, this becomes a search for confirming evidence, prohibited by K4 §4.2. Offer instead an explicit test of the proposed structure against the data, reported whether it passes or fails.
- **The data is not prepared, or the basis variables were not designed for this.** Segmentation is unusually sensitive to scale-use differences, missing data patterns and derived-variable construction. Run **04.02**, **04.03** and **04.04** first. Retrofitting a segmentation onto a questionnaire designed for something else is possible and should be labelled as opportunistic, with the limitation stated at the front.

## 5. Required inputs

**Required.** Without these the skill cannot run. If absent, ask. If the first is absent, stop.

- **The decision the segmentation must serve, and who owns it.** Named specifically: which product decisions, which communications decisions, which service decisions will be made differently by segment. This determines the basis variables and is the only defensible source of the actionability standard. **Without it, do not proceed to variable selection.**
- **A respondent-level dataset with candidate basis variables measured on the whole sample.** Basis variables asked of a routed subgroup cannot be used, because the solution would then exist only for that subgroup.
- **The questionnaire as fielded**, with wording, scales and their direction, and the routing. Attitudinal batteries are especially sensitive to wording, and a reverse-coded item entering a cluster solution unnoticed will distort every segment.
- **The data dictionary**, including how derived variables were constructed, per **04.04**.
- **Sample and fieldwork documentation.** What the sample represents governs what the segments represent.

**Optional, and what each one adds.**

- **Behavioural or transactional data linked at respondent level:** allows profiling on evidence that was not self-reported and was not used to build the solution, which is the strongest available validation. It also allows segment value to be estimated rather than claimed.
- **Qualitative evidence from the same population** (**07.01**, **09.05**): supplies candidate basis dimensions that a survey would not have thought to ask, and gives the eventual segments a mechanism rather than a label.
- **A prior segmentation and its typing tool:** allows continuity to be assessed, migration between segments to be estimated, and the organisational cost of switching to be made explicit.
- **A hold-out sample or a second wave:** allows out-of-sample replication, which is a stronger stability test than any within-sample split.
- **The customer database schema and the media or targeting capability:** determines reachability, which is half of actionability. A segment that cannot be found in the database or bought in a media plan is a segment the business cannot serve.
- **Category penetration and volume data:** allows segment sizing in value rather than in respondents, which is what a commercial decision needs.

## 6. Questions to ask before starting

1. **What will be done differently for different segments, specifically?** Determines the basis variables and sets the actionability bar. *Default if unanswered:* build on needs and behaviour rather than attitudes alone, and flag that actionability could not be assessed, per K5 §2.1.
2. **Is there prior evidence that this market is heterogeneous on the dimensions that matter?** Determines whether the prior question in step 2 is a formality or the main event. *Default:* treat it as an open question and report the heterogeneity assessment as a finding in its own right.
3. **How will segments be identified in the real world after the study?** Determines whether a typing tool is required, and therefore whether basis variables must be short, stable and askable outside a long survey. *Default:* assume a typing tool is needed, and build the solution so one is possible.
4. **What is the smallest segment the business could serve profitably?** Sets the minimum size threshold, which is a commercial number, not a statistical one. *Default:* flag any segment below 10% of the population as a size risk and any below 5% as probably unservable, and mark it for researcher review.
5. **Is there an existing segmentation, and what would it cost to replace it?** Determines whether the honest recommendation might be to refine rather than rebuild. *Default:* assume replacement has a real organisational cost and report continuity with any prior solution.
6. **Does behaviour in this category vary more between people or between occasions for the same person?** Determines whether the person is the right unit. *Default:* check it in the data if occasion-level measures exist, and state the assumption if they do not.
7. **Which variables are bases and which are descriptors, and who decided?** Prevents demographics entering the solution by default. *Default:* demographics, region and firmographics are descriptors unless there is a stated reason they cause the behaviour of interest.

## 7. Step-by-step methodology

**The position this method takes.** A segmentation is a claim that a population contains distinguishable groups whose differences matter for a decision. It is not a claim that an algorithm partitioned a dataset, which is guaranteed in advance and therefore evidence of nothing. Every step below exists either to test that claim or to make the resulting groups usable. Two standards govern the whole method: **statistical separation is not evidence that segments are real**, and **a segmentation that does not change what the business does is not actionable, however clean it looks.**

**1. Write the action table before touching the data.** For each candidate way of dividing the audience, write what the business would do differently: which product, which message, which channel, which price, which service level. If the column is the same for every group, that dividing principle is not actionable no matter how well it separates. Do this first, in writing, because doing it afterwards turns it into a justification exercise. *Correct result:* a one-page table naming the decisions in play, so that a later solution can be tested against it rather than defended.

**2. Establish whether the market is segmentable at all, and be willing to conclude that it is not.** This is the step most segmentations skip. Run four checks. **Distribution shape** on each candidate basis variable: unimodal and roughly symmetric distributions across the whole battery indicate a population that varies by degree rather than by kind, which is the signature of a market that is not usefully divided. Bimodality or clear multi-modality on several variables at once is the signature of one that is. **Correlation structure:** if the basis battery collapses to one or two dimensions on which everyone sits somewhere along a continuum, you have a spectrum, not segments, and reporting it as a spectrum is more useful and more honest. **Prior and qualitative evidence:** does the qualitative work describe recognisably different kinds of people with different logics, or the same person at different levels of engagement? **Behavioural dispersion:** do people differ in what they do, or only in how much? Where three of these four point to homogeneity, stop and report it. *Correct result:* an explicit, evidenced statement of whether the population is heterogeneous on decision-relevant dimensions, written before any solution is run so it cannot be reverse-engineered from one.

**3. Choose the basis variables from the decision, and separate them from descriptors.** Basis variables construct the solution. Descriptors describe the solution afterwards. The families available as bases are: **needs** (what the person is trying to achieve), **attitudes and beliefs** (how they think about the category), **behaviour** (what they do, buy, use, at what frequency and in what repertoire), **occasion** (the situations in which the category is used), and **value** (what they are worth, which is a business basis rather than a customer one). Choose the family that sits closest to the decision: a product decision usually wants needs or occasions, a communications decision usually wants attitudes and motivations, a resource-allocation decision usually wants value and behaviour. **Demographics are descriptors, not bases, in almost every case.** Age, gender, income, region and firmographics are cheap to collect, easy to target and rarely the reason anyone behaves as they do. They earn a place as a basis only where they are the mechanism itself: life stage in a category defined by life stage, or firm size where it determines the buying process. The convenience argument for demographic bases is real but it belongs at the reachability stage, where descriptors are used to find segments, not at the construction stage. *Correct result:* a written basis set with a one-line justification per variable tying it to the decision, and a separate descriptor list including every demographic.

**4. Prepare the basis set, and remove the artefacts that would otherwise become segments.** Three specific hazards. **Scale-use bias:** respondents differ systematically in how they use rating scales, and an unadjusted attitudinal battery will produce a "positive about everything" segment and a "negative about everything" segment that are measurement artefacts wearing the clothes of attitude. Standardise within respondent where the battery is agreement-scaled, and state that you did. **Redundancy:** near-duplicate items give one underlying idea several votes and let it dominate the solution. Check the correlation matrix and reduce, either by selection or by factor reduction, noting that factor reduction costs interpretability and can smooth away exactly the sharp differences a segmentation is looking for. **Missing data:** listwise deletion on a long battery can remove a quarter of the sample non-randomly, per **04.03**. *Correct result:* a documented basis matrix, with every transformation logged per K4 §4.4.

**5. Run many solutions, not one.** Vary the algorithm (at minimum a partitioning method and a model-based or hierarchical method), vary the number of segments across a range wider than the client's expectation, and vary the starting conditions. A single run with a single method at a single k produces a result with no way of knowing whether it is a property of the data or of that run. *Correct result:* a solution matrix, typically twenty or more runs, with fit statistics recorded but explicitly not treated as the decision criterion.

**6. Evaluate solutions on four gates, in this order, and let any one of them disqualify.** **Gate 1, separation:** are the segments distinguishable on the basis variables, and by how much? This is necessary and worth almost nothing on its own. **Gate 2, stability:** does the solution survive step 7. **Gate 3, size:** is every segment large enough to serve, against the threshold set in Section 6. **Gate 4, differentiability of action:** would the business genuinely do something different for each segment, checked against the step 1 table? A solution that passes 1 and fails any of 2 to 4 is not reportable as a segmentation, whatever its fit statistics say. *Correct result:* the solution matrix scored on all four gates, with the reason each rejected solution failed.

**7. Test stability three ways, and hold the standard.** **Split-half replication:** split the sample randomly, build the solution independently on each half, apply each half's rule to the other, and measure the agreement rate. **Alternative algorithm:** build the same number of segments with a different method and cross-tabulate the two allocations. **Alternative variable subset:** rebuild on a randomly chosen subset of the basis variables and compare. The standard that matters is not a specific agreement percentage, which varies by number of segments and by method, but consistency of structure: the same number of recognisable groups, with the same defining differences, and individuals mostly landing in the corresponding group. Where the segments change shape when the algorithm changes, the structure is an artefact of the method, and **an unstable solution is not reportable**. Report what the instability was, rather than repeating the analysis until a stable-looking run appears, which is the same error in a different costume. *Correct result:* three stability tests with their agreement measured and stated, and a decision on reportability that is defensible in either direction.

**8. Size the segments in the currency the decision uses.** Respondent share is the weakest sizing. Convert to population size where the sample supports it, and to category value where behavioural or spend data allows, because a segment that is 12% of people and 31% of volume is a different proposition from one that is 12% of both. Then apply the servability threshold: **a segment too small for the business to serve is not a segment, it is a finding about the tail.** Where a solution's only interesting group is unservably small, say so; it may still be a valuable observation about an emerging need, reported as such rather than as a segment.

**9. Profile on variables that were not used to build the solution.** This is the only honest test of whether the segments mean anything. Cross-tabulate the segments against everything held out: demographics, behaviour, category value, media use, life stage, and above all any behavioural or transactional data that exists independently of the survey. Two outcomes matter. If the segments differ meaningfully on held-out variables, they correspond to something outside the basis battery, and confidence rises substantially. If they differ only on the variables used to construct them, they are a restatement of those variables and nothing more, which is a result the report must state plainly. Test the profiling differences per **05.02** rather than reading the table, and watch the composition confound covered in **09.04**. *Correct result:* a profile table with tested differences, bases shown, and an explicit statement of how much of the segments' character comes from outside the basis set.

**10. Name and describe the segments without inflating them.** A segment name is a memory aid and it will outlive every caveat attached to it, so it must not assert more than the evidence carries. Names describing the defining difference ("Price-led switchers") are safer than names describing a personality ("The Sceptics"), because the second invites readers to fill in traits nobody measured. Write each description as: the defining difference, size in people and in value, what they do, what they need, what distinguishes them from the nearest other segment, and what is *not* different about them. That last element is routinely omitted and it is what keeps the reader honest. *Correct result:* descriptions in which every attribute is traceable to a variable and a base, per K2 §4.1.

**11. Build the typing tool, and measure how often it is wrong.** A typing tool is a short set of questions plus an allocation rule that assigns a new individual to a segment. Build it on a subset of basis variables, validate it by applying it to respondents whose true segment is known, and produce the confusion matrix. **Report the overall hit rate and the per-segment hit rate**, because a tool that is 78% accurate overall can be 40% accurate on the smallest and most interesting segment. Misclassification is not a defect to be hidden; it is a property of every typing tool, and downstream users who do not know the rate will treat allocations as facts. Where the tool cannot reach a defensible accuracy, say that the segmentation cannot currently be applied outside the study. *Correct result:* a typing tool with a published confusion matrix and a stated accuracy per segment.

**12. Write the limitations and mark the judgement points.** State what the segmentation is a segmentation of (this sample, this market, this period), how stable it was, what it does not distinguish, and when it should be re-validated. Mark for human decision, per K5: whether the segments are commercially material (§2.1), whether the naming and cultural reading hold in each market (§2.2), and whether the organisation can actually act on the structure (§2.3). *Correct result:* a segmentation a research director can defend in a room where somebody asks whether the groups are real.

## 8. Analytical framework

The chain for a defensible segmentation:

    Decision → Basis choice → Solution set → Stability → Size and reach → External profile → Action

**Decision.** What will be done differently. Everything upstream of the data.
**Basis choice.** The variables that construct the groups, justified individually against the decision.
**Solution set.** Many runs, many methods, many k. Never one.
**Stability.** Split-half, alternative algorithm, alternative variables. The gate that most solutions fail and few report failing.
**Size and reach.** Population and value size, and whether the group can be found in the real world.
**External profile.** Difference on variables not used to build it. The only honest test of reality.
**Action.** Whether the business would do something different, checked against the table written in step 1.

**The four gates, stated as pass or fail.**

| Gate | Question | Fails if |
|---|---|---|
| Separation | Are the groups distinguishable on the basis variables? | Overlap is so high that most individuals sit near a boundary |
| Stability | Does the same structure appear under different splits, algorithms and variable sets? | Segment shape changes with the method. Not reportable |
| Size | Is every segment big enough to serve, in people and in value? | A segment the business could not economically address |
| Differentiability | Would the business act differently towards each? | The action column is identical for two or more segments. Merge them |

**Against the K2 evidence chain**, a segmentation solution is Analysis, and the segment descriptions are Findings. "Segment C is motivated by control" is an Interpretation and must be signalled as one. The commonest level failure here is a segment name silently promoted to an explanation: a label is not a mechanism.

## 9. Output format

**1. Segmentability assessment.** The four checks from step 2, their results, and the conclusion, stated before any solution is presented. Where the conclusion is that the market is not usefully segmentable, this section is the deliverable and the rest is not produced.

**2. Basis and descriptor register.**

| Variable | Basis or descriptor | Justification against the decision | Transformation applied |
|---|---|---|---|

**3. Solution matrix.**

| Run | Method | k | Separation | Split-half agreement | Alt-algorithm agreement | Smallest segment | Gate outcome |
|---|---|---|---|---|---|---|---|

**4. Chosen solution and why**, including the solutions rejected and the gate each failed.

**5. Stability report.** The three tests, their results, and the reportability decision.

**6. Segment descriptions**, one per segment: defining difference, size in people and in value, needs, behaviour, what distinguishes them from the nearest segment, what is not different, and every claim carrying its variable and base.

**7. External profile table.** Segments against held-out variables, with bases and tested differences, and an explicit statement of how much character comes from outside the basis set.

**8. Typing tool and confusion matrix**, with overall and per-segment accuracy.

**9. Action table**, mapping each segment to what the business would do differently, carried forward from step 1.

**10. Limitations, currency and re-validation trigger.** What this is a segmentation of, and when it should be re-tested.

**When the evidence is thin.** The format is not filled to look complete, per K4 §1. A market that fails the segmentability assessment produces section 1 and a description of the population, not five segments with lower confidence. A solution that fails stability is reported as a failed solution with what was learned, not as a provisional segmentation. A segment whose base is too small to profile is shown with its base and a note, not with percentages. And where the segments differ only on the variables used to build them, that sentence appears in the summary, not in an appendix.

## 10. Quality checks

Run before anything is presented. These sit on top of K4 §8.

1. Was the segmentability of the market assessed and reported before any solution was built?
2. Is every basis variable justified against a named decision, and is every demographic in the descriptor list rather than the basis set?
3. Were basis variables measured on the whole sample rather than a routed subgroup?
4. Was scale-use bias addressed, and is the treatment stated?
5. Were multiple algorithms and multiple values of k run, and is the full solution matrix available?
6. Were all three stability tests run, with their agreement measured and reported?
7. Is any solution being reported that failed a stability test, and if so, is the failure stated in the summary rather than an appendix?
8. Does every segment meet the servability threshold, and is the threshold a commercial number rather than a statistical one?
9. Have the segments been profiled on variables not used to construct them, and is the result stated either way?
10. Are profiling differences tested per **05.02**, with bases shown for both groups?
11. Does any segment name assert a trait, motivation or explanation that was not measured?
12. Does the output say what is *not* different between segments, as well as what is?
13. Is the typing tool's misclassification rate reported overall and per segment?
14. Is segment size given in value as well as in respondents, where the data allowed it?
15. Does the action table show a genuinely different action per segment, and have any two segments with identical actions been merged or flagged?
16. Is the currency of the segmentation stated, with a re-validation trigger?

## 11. Common failure modes

| Failure | How to recognise it | How to prevent it |
|---|---|---|
| **Segmenting a homogeneous market** | Clean statistics, unimodal distributions on every basis variable, segments that differ only in degree. Nobody can describe a segment without using the word "more" | The step 2 assessment, run and reported before any solution exists |
| **Clustering treated as evidence** | The report presents fit statistics as proof that the segments are real | Fit is Gate 1 of four. Stability and external profile carry the reality claim |
| **Demographic bases** | Segments that are essentially age bands with attitude labels attached | Demographics are descriptors. Basis variables are justified individually against the decision |
| **One run, one method** | A single solution appears in the working files with no rejected alternatives | Solution matrix, minimum two algorithms and a range of k |
| **Instability laundered by repetition** | The analysis was re-run until a clean solution appeared, and only that one is documented | Log every run. Report the instability, do not iterate past it |
| **Circular profiling** | Segments differ dramatically on exactly the battery used to build them, and this is presented as validation | Profile on held-out variables. State how much character comes from outside the basis set |
| **The unservable segment** | A vivid, strategically exciting group at 3% of the market | Size threshold set from the business before the solution is chosen |
| **Name inflation** | A segment called "The Anxious Optimisers" whose measured difference is a four-point gap on two agreement items | Names describe the defining difference. Every descriptive claim carries a variable and a base |
| **The typing tool taken as truth** | Database allocations used operationally with no reference to accuracy | Publish the confusion matrix; state per-segment accuracy wherever allocations are used |
| **Silent decay** | A segmentation quoted five years on, with no re-validation and a category that has changed | Currency statement and re-validation trigger in the deliverable |
| **AI: filling the requested number of segments** | The client asked for five and five appeared, with the fifth thin and overlapping | The number of segments is a property of the data. Report the k the evidence supports and say what happens at the requested k |
| **AI: inventing segment character** | Rich descriptions of lifestyle, media habits and motivations that no variable in the study measured | Every attribute traces to a variable and a base, per K4 §2.1 and §2.2 |
| **Merging findings into a story** | Two segments combined in the narrative because the story is cleaner than the solution | The solution is the solution. Narrative simplification happens in 11.01 and is labelled |

## 12. AI guardrails

Skill-specific only. K4 applies in full and is not repeated here.

1. **Never present a clustering result as evidence that segments exist.** Separation is produced by the method regardless of the data. The reality claim rests on stability and on external profiling, and if those were not run, the output says the segments are unvalidated.
2. **Never build a segmentation without first assessing and reporting whether the market is segmentable.** Where the assessment says it is not, do not proceed to a solution because a solution was requested.
3. **Never place a demographic variable in the basis set** unless a stated mechanism makes it the cause of the behaviour of interest, and record that reason in the basis register.
4. **Never report a solution that failed stability testing as a segmentation.** Report it as a failed solution with what was learned.
5. **Never produce a specific number of segments because a number was requested.** State what the data supports and what is lost or invented at the requested number.
6. **Never describe a segment with an attribute that was not measured.** No inferred lifestyles, no inferred media habits, no inferred motivations, no illustrative individuals presented as observed, per K4 §2.2.
7. **Never let a segment name function as an explanation.** A label is not a mechanism, and a name repeated three times becomes a finding in the reader's memory.
8. **Never report profiling differences without a test and both bases**, per K4 §3.1 and **05.02**.
9. **Never publish a typing tool without its misclassification rate**, overall and per segment.
10. **Never omit what the segments have in common.** A comparison table showing only differences systematically overstates how different the groups are.
11. **Never carry a prior segmentation's names or story onto a new solution** because stakeholders are used to them. Report continuity and migration explicitly instead.

## 13. Best-practice principles

- **The prior question is the valuable one.** Determining that a market is not usefully segmentable is a real result, it is often the right one, and it is the single most expensive thing to get wrong. Nobody was ever fired for delivering five segments; the cost lands three years later.
- **Clustering is a description of a dataset, not a discovery about a population.** The algorithm partitions whatever it is given, including noise. Everything that makes a segmentation credible happens after the run.
- **Choose bases from the decision, and only from the decision.** A segmentation built on whatever happened to be in the questionnaire is a segmentation of the questionnaire.
- **Demographics find segments; they do not make them.** The right sequence is to build on needs or behaviour, then profile demographically so the segments can be reached. Reversing this is the most common structural error in the field.
- **Stability outranks separation.** A solution with modest separation that reappears under every method is worth more than a sharp one that dissolves when the algorithm changes.
- **The external profile is where a segmentation earns belief.** Differences on variables nobody used to build the groups are the only evidence that the groups correspond to something outside the analysis.
- **Segment size is a commercial threshold, not a statistical one.** Ask what the smallest servable group is before you see the solution, so the answer is not negotiated afterwards.
- **Say what is the same.** Segments in most categories share far more than they differ on, and a report that shows only differences produces a management team that believes it has five markets when it has one market with five accents.
- **A memorable segmentation and a true one are different objectives, and the pressure runs one way.** Names, colours and archetypes make a segmentation travel, and travelling is exactly what makes an unvalidated one dangerous.
- **A typing tool with an unstated error rate becomes a fact in the database.** Once allocations enter a CRM they are treated as attributes of the person, and the 30% who were misallocated are indistinguishable from the rest.
- **Segments decay.** They are built on a market at a moment. Attach a re-validation trigger, and re-test rather than re-quote, per K2 §7.
- **Two honest segments beat five decorated ones.** If the data supports a split between two kinds of buyer, report two. Filling out a five-box layout is a design decision impersonating an analytical one.

## 14. Worked example

*Fictional scenario, used to demonstrate method. The organisation, figures and findings below are invented.*

**INPUT.** A national home improvement retailer commissions a segmentation to inform range architecture and store format. Brief asks for "five or six segments we can build the category strategy around". Dataset: 2,400 respondents who bought in the category in the past year, with a 32-item attitudinal battery, a needs battery of 14 items, purchase behaviour, project type, and linked loyalty-card spend for 1,610 of them.

**PROCESS.**

*Step 1.* The action table names three decisions: which ranges to stock at which price tiers, which store format to invest in, and how to structure online guidance. Written before any analysis, it establishes that a segmentation that does not distinguish project type or advice need cannot serve any of them.

*Step 2, the prior question.* The attitudinal battery is unimodal and symmetric on 29 of 32 items, and collapses to two dominant dimensions on which respondents sit along a continuum. This is a spectrum, not a set of groups. But the needs battery and the project-type data behave differently: strongly multi-modal, with clear discontinuities between people undertaking structural work, cosmetic refreshes and maintenance repairs. **Judgement call:** the attitudinal segmentation the client expected is not supported, and saying so risks the brief. The resolution is to report both: the attitudinal dimensions as a spectrum with the evidence, and to build the solution on needs and project type, where the heterogeneity is real. Reporting the negative result on attitudes turned out to be the most-used page in the debrief, because it stopped a planned communications architecture built on attitudinal types.

*Steps 3 to 6.* Bases: 14 needs items plus three project-behaviour measures. Demographics, region, income and tenure all go to descriptors. Scale-use standardisation applied within respondent to the needs battery; three near-duplicate items reduced to one. Twenty-four runs across two algorithms and k from 2 to 8.

*Step 7, stability.* The six-segment solution the client wanted has split-half agreement of 61% and its two smallest segments swap members freely under a different algorithm. The four-segment solution reproduces at 84% across splits, holds under both algorithms, and survives rebuilding on two-thirds of the basis items. Six is not reportable; four is.

*Step 9, the honest test.* Profiled against the loyalty-card spend nobody used to build the solution (n=1,610), the four segments differ in annual category spend, in trip frequency and in the share of spend going to tools rather than materials, all tested at 95%. They differ far less on demographics than the client expected, which becomes an explicit finding: age and income do not identify these groups, so the store format decision cannot be made on catchment demographics alone.

*Step 11.* The typing tool reaches 81% overall accuracy on seven questions but only 58% on the smallest segment, which is reported prominently, because that segment is the one the range decision most depends on.

**OUTPUT.** A segmentability assessment concluding that the market is a spectrum attitudinally and genuinely grouped on needs; a four-segment solution with its stability record; the six-segment solution reported as failed with the reason; segment sizes in people and in category value; an external profile on linked spend; a typing tool with per-segment accuracy; an action table; and a **researcher decision required** marker on whether the 58% accuracy on the critical segment is fit for the range decision, per K5 §2.1.

## 15. Advanced usage

**Segmenting occasions rather than people.** Where occasion-level data exists, run the solution on occasions and then describe people by their occasion repertoire. This frequently outperforms person-level segmentation in categories where the same individual behaves in several ways, and it converts a stubbornly unstable person-level solution into a stable occasion-level one. It also changes the deliverable: the business targets moments, not people. Feed from **09.03**.

**Hybrid bases.** Combining a needs basis with a value basis produces a two-dimensional grid rather than a partition, which is often more usable than either alone: needs tell you what to build, value tells you where to spend. Report it as a grid, and resist the temptation to collapse the grid back into named cells, which reintroduces every problem of an unvalidated partition.

**Re-validating an inherited segmentation.** Apply the existing typing tool to a fresh sample, then independently build a new solution on the same basis variables and cross-tabulate. Three outcomes: the structure holds and only sizes have shifted (refresh the sizing, keep the segmentation), the structure holds but membership has migrated substantially (report migration, keep the framework, re-issue the typing tool), or the structure no longer appears (the segmentation has decayed, and the organisational cost of replacement is now a business decision, not a research one).

**Multi-market segmentation.** Building one solution across markets usually produces segments that are largely country markers, because cultural response style and category maturity dominate. Build within market first, then test whether the structures correspond. A genuinely global segmentation is one where the same needs structure appears independently in each market, with different sizes. Route the cultural reading to a human per K5 §2.2.

**When the standard approach does not fit.** Where the sample is too small for stable clustering, use a small number of a priori groups defined from qualitative and behavioural evidence, test whether they differ on held-out variables, and label the result an a priori typology rather than a segmentation. It is a weaker claim, honestly made, and it is often enough for the decision.

## 16. Skill chain

**Recommended previous skills:**
- **01.02 Business Problem to Research Question** and **01.07 Analysis Plan Development.** Hand over the decision the segmentation must serve and the pre-committed analysis approach, which is what makes the basis choice defensible rather than opportunistic.
- **04.04 Data Transformation and Dataset Preparation.** Hands over derived variables, scale treatments and standardisations with their construction rules logged, which segmentation is unusually sensitive to.
- **09.03 Behavioural Profiling.** Hands over what people actually do, the occasion structure, and whether the person is the right unit of analysis.
- **09.05 Needs, Motivation and Jobs-to-be-Done Analysis.** Hands over the needs dimensions that make the strongest basis variables, and the evidence that they are needs rather than stated preferences.
- **05.01 Descriptive Analysis.** Hands over distributions, including the bimodality that is the first signal a population contains more than one group.

**Recommended next skills:**
- **09.02 Persona Development.** Takes validated segments and turns them into a usable communication format, inheriting this skill's evidence annotations and confidence.
- **09.04 Segment Comparison.** Takes the segments and compares them rigorously, including on the variables used to build them, which must be labelled circular.
- **05.02 Statistical Testing** and **05.03 Cross-Tabulation.** Do the profiling work properly rather than by eye.
- **08.01 Finding to Insight Development.** Takes the segment differences and works out why they exist, which a segmentation describes but does not explain.

**Runs well alongside:**
- **13.03 AI Output Verification**, run against any segmentation produced by an automated pipeline, with the stability and external-profile checks as its focus.
- **13.04 Bias Detection**, particularly where a stakeholder named the segments in advance.
- **K3 §3.6**, because a segment description is a finding and a segment explanation is two steps further, and the two are routinely reported at the same confidence.

---
A Yazi Supplied Skill and resource.