How to Run: Market Segmentation Study

Introduction

Run a market segmentation study in five moves. Decide whether to segment by rule, by survey data, or not at all. Build a short needs battery kept separate from demographic descriptors. Field at least 70 completed responses per segmentation variable, rerun the analysis to see whether the segments reproduce, and test each segment with a holdout group before rollout. A rule on attributes you hold needs no clustering.

Last updated October 2026.

What a Segmentation Study Has to Prove

A segmentation study earns its cost only if the groups it produces change what a company does. That’s a harder test than most studies face. Clustering software returns groups whether or not real groups exist, and a deck of named segments can look finished long before anyone has checked whether the segments are stable, reachable, or worth treating differently.

This guide follows the study in the order you’d run it. It starts with the decision most teams skip, which is whether to segment at all. It then covers the instrument, the sample, the clustering and stability check, and the tests a segmentation should pass before it reaches a roadmap or a budget.

Every figure comes from a published source or from arithmetic shown on the page, and where no source exists, the guide says so.

What a Market Segmentation Study Is

A market segmentation study is research that divides a market into groups of buyers who respond differently enough to justify different treatment. Wendell Smith framed segmentation as a marketing strategy in his 1956 Journal of Marketing article, and the study that produces the groups is the first step of the sequence marketers call STP: segmentation, targeting, and positioning.

Basis Variables and Descriptor Variables

Every segmentation uses two kinds of variable, and keeping them apart is the most important design rule in this guide.

Basis variables are the ones the segments are built from, such as needs, benefits sought, usage behavior, and purchase behavior. Descriptor variables are the ones you use afterward to describe and reach each segment, such as age, region, company size, plan tier, and acquisition channel. A basis variable decides who belongs together. A descriptor variable tells you how to find them.

The Six Types of Segmentation

Textbooks name six types of segmentation after the variable they group on. They aren’t equal choices.

Behavioral variables are the basis with direct support in the evidence cited here, and needs-based segmentation is the closest survey equivalent, though no source cited here tests it directly. Firmographic variables are the exception in business markets, because company size and industry often decide which sales motion a customer gets. That makes them a sound basis for a rule-based segmentation even though they’re descriptors in a survey.

Segments, Personas, and Cross-Tabs

A segment is a sized group with a membership rule. A persona is a narrative built on top of a segment for a design team, and the persona research guide covers that step. A cross-tab compares groups you already have. Comparing results across existing segments is a different and cheaper operation than deciding what the segments should be.

When to Run a Market Segmentation Study

Run one when a team has to treat customers differently and can’t yet say, with evidence, which differences matter.

The clearest case is a portfolio decision. A company choosing which audience to build a new offer for, which market to enter first, or which customers to prioritize needs groups it can size and compare.

A second case is a product whose customers want contradictory things. When one group asks for more control and another asks for less setup, a single roadmap tends to satisfy neither, and a segmentation shows how large each group is.

A third case is messaging. Segment-specific messages can only help if the segments differ on something a message can address, and a segmentation tells you whether they do. If you’re choosing between specific claims for an audience you already know, a message testing survey is the lighter tool.

Segmentation also fits after a usage and attitudes study or a customer needs assessment has already collected the variables. A segmentation can start from a battery you’ve fielded instead of a blank questionnaire.

A segmentation also pays off when sales, product, and marketing disagree about who the customer is. A shared, evidence-based set of segments gives the teams one vocabulary, and it turns an argument about opinions into a question about which segment a decision serves. It’s also worth running when a rule-based segmentation has been in use for some time and people suspect the rules miss a distinction that matters.

Whichever case applies, write the decision down before the study starts. A useful test is one sentence: "If segment X turns out to be large, we’ll do Y." If nobody can finish that sentence, the study isn’t ready.

When Not to Run a Market Segmentation Study

A segmentation study can’t answer seven questions, and each has a better instrument. The first and the fifth matter most, so hold them hardest and state both on the first page of the readout.

Whether the Segments Are Real

A clustering algorithm returns groups whether or not groups exist. The better instrument is the stability protocol in the analysis section, which reruns the solution on resampled data and reports whether the same segments come back. The critique section gives the published evidence for why this check can’t be skipped.

The algorithm optimizes a distance measure, and any dataset has a best partition under any measure, including data with no structure at all. A tidy cluster plot therefore isn’t evidence that the groups exist. Reproducibility across reruns is the minimum evidence you should ask for before anyone treats the groups as real.

Which Brand a Person Will Choose

Fennell, Allenby, Yang, and Edwards (Quantitative Marketing and Economics, 2003) found that demographic and general psychographic variables predict product use but aren’t useful for predicting relative brand preference. For brand choice, run choice-based conjoint or MaxDiff.

What a Segment Will Pay

A segmentation scores people, not price points. A group that looks price-insensitive on a battery of needs may still balk at a specific price, so measure price inside each segment with a price instrument and don’t infer it from needs. Use Gabor-Granger, run within each segment.

Which Feature to Build for a Segment

Basis variables describe needs at whatever level of abstraction you wrote them. To rank features within a segment, run MaxDiff inside it, or run a monadic concept test per segment.

Whether Acting on the Segmentation Will Pay

A segmentation study measures how people differ. It doesn’t measure whether treating them differently raises revenue. Dibb and Simkin (Journal of Marketing Management, 2009) followed one European utilities company through a segmentation project and classified the barriers it met when it tried to implement the result. The better instrument is a holdout experiment, set out in the section on acting on the result.

Once every customer receives the segment-specific treatment, no comparison group is left, and the question of whether segmentation helped becomes permanently unanswerable.

How Large Each Segment Is in the Real Market

A survey of your users or a panel describes that frame, not the market. Segment shares from a customer-list study are shares of your respondents, and quoting them as market sizes overstates the segments your product already attracts. The better instrument is a weighted, incidence-calibrated panel study, with the segmentation carried across as a typing tool.

Why a Segment Behaves as It Does

Closed need statements show that a group differs, not the reasoning behind the difference. The better instrument is depth interviewing within each target segment. It’s a qualitative method outside the survey category, and you run it after the segments exist, not instead of them.

None of these limits is a reason to skip segmentation. Each is a reason to pair it with a second instrument. The stability protocol answers the first, and the holdout test answers the fifth. The others get their own follow-on study once the segments exist. If your decision hinges on one of them, such as price or feature choice, start with the better instrument and segment afterward.

Study Design and the Fork That Decides Everything Else

The first design decision is whether to segment by rule, by data, or not at all. The distinction underneath it’s old. In an a priori design you choose the segments in advance from variables you already hold. In a post hoc design, analysis finds the segments in survey data.

The Committed Default

Segment on needs or observed behavior. Use rules when you already know the variables that change how you treat someone. Use data only when you don’t, and only when you can field enough respondents. Profile with demographics afterward, and never let them into the basis.

The rest of this section gives the reasons.

Before you go further, finish this sentence in writing: "We’ll segment on [basis] because it changes [decision]." If you can’t fill in both blanks, you haven’t chosen a design yet.

Rule-Based Segmentation

Rule-based segmentation assigns customers to groups using variables you already hold, such as plan tier, company size, lifecycle stage, usage frequency, or acquisition channel. It’s reproducible by construction, and it’s actionable because the rule is the targeting criterion. Each reported segment still needs enough respondents for its own estimates.

It’s often the cheaper correct answer. Its weakness is that it can only find distinctions you already suspected.

Rules can be simple. A business-to-business team might segment on company size and lifecycle stage, and a consumer team on purchase frequency and channel. Either way, the test of a good rule is whether the two groups it produces would get different treatment, so write down that treatment before you adopt the rule.

Data-Driven Segmentation

Data-driven segmentation fields a survey battery and uses a clustering method to group respondents who answered alike. Use it when you don’t know what distinguishes your customers, when you can field the sample the sample-size section derives, and when the organization can act differently toward different groups.

This is the branch that needs the stability check. Treat the first solution as a hypothesis, not a finding.

No Segmentation

No segmentation is a legitimate outcome. It fits when you sell one product at one price with no way to treat groups differently, or when the honest answer to "what would we do differently for segment three" is nothing. It also fits categories where competing brands' buyers look alike, a position the critique section covers.

Which Basis to Segment on

The evidence on bases points one way. Fennell and colleagues (2003) examined a large set of demographic and general psychographic variables and found them predictive of product use but not useful for predicting relative brand preference. Their abstract states that the evidence is overwhelming that these covariates relate to brand use in an identical way for all brands.

Behavior has direct support too. Rossi, McCulloch, and Allenby (Marketing Science, 1996) built a target-couponing model on purchase histories and found that even a single observed purchase occasion produced about 50 percent more net couponing revenue than blanket couponing. That’s a coupon-targeting result in one setting, and it didn’t test demographics against behavior, so it supports targeting on behavior without proving that behavior beats demographics.

So segment on what people need or do. Keep demographics and firmographics in the survey as descriptors, because they’re often how you reach a segment once you’ve found it.

Factor Analysis Before Clustering

A frequently taught recipe runs factor analysis on the battery and then clusters the factor scores. Dolnicar and Grün (Journal of Travel Research, 2008) argue that this factor-cluster approach isn’t generally the best procedure for identifying homogeneous groups of individuals. Cut redundant items at the design stage instead. That also lowers the sample you need.

The Decision Tree

Work down these four questions in order. A no at question 1, 3, or 4 ends the tree, and a yes at question 2 gives you a design.

Would you treat customers differently if you knew which group they were in? If no, don’t segment.
Do you already hold the variables that would drive that treatment? If yes, a rule-based segmentation covers that treatment, and you add a survey only for distinctions the rules can’t see.
Can you field the sample the sample-size table calls for, at 70 respondents per segmentation variable? If no, cut variables, or run a rule-based segmentation now and a data-driven one later.
Will you rerun the analysis and test the result before naming segments? If no, the study isn’t ready to field.

Question three is where many studies should stop. A data-driven segmentation run on too few respondents often produces confident-looking segments that don’t reproduce, so cut the battery or start with rules instead of fielding anyway.

The two branches combine. The worked example later in this guide starts with rules and uses survey data only where the rules run out, which is often the cheapest design that still finds what attributes can’t. Data earns its cost when the needs you care about aren’t recorded anywhere you hold, when teams disagree about what customers want, and when the organization can run at least two distinct treatments.

In Sprig: the rule-based branch runs on user attributes, which hold strings, numerics, and booleans and can refine targeting and filter response data. The data-driven branch leaves the platform at the analysis step.

The Analyze Agent page lists crosstabs, t-tests, ANOVA, correlation, and Van Westendorp as coming soon and in preview. It names no clustering, latent class, or segmentation. Attributes can’t trigger a study, and the documentation says one can’t be deleted unless the visitor’s data is deleted too, so name a new attribute for each segmentation version.

Writing the Instrument

A segmentation instrument has five blocks, and the order protects the basis from contamination by everything else. The needs battery comes first among the questions that matter, and the descriptors come last.

The Instrument Spec

This is the instrument in fielding order. Replace the example need statements with your own and keep the structure.

The last column is the point. If basis and descriptor items mix, the segments are partly defined by the variables you later use to describe them. Print the basis or descriptor label next to every item in the draft instrument, so nobody has to remember which is which.

The item counts in the table are examples, not thresholds. No source supports a fixed battery length, so let the sample you can afford set the ceiling.

Writing Need Statements That Can Split

A need statement has to be something people genuinely disagree about. "I value quality" is agreed by almost everyone, so it carries no information for clustering. Three statements that do split:

I need to get set up without talking to anyone
I need to see exactly what changed before I approve it
I need this to work the same way on my phone and my laptop

Write every statement at the same level of abstraction. A battery that mixes broad values with specific feature requests clusters on the mix itself, not on any real difference. If you don’t yet know which needs people have, customer needs assessment research is the place to generate the items.

Write need statements in customers' own words, drawn from interviews, support conversations, or open-text answers, not in product language. Randomize the order of the statements within the battery so that position doesn’t become a pattern the clustering can pick up.

The Holdout Outcome Question

Add one question that the segmentation should predict but that doesn’t enter it, such as stated likelihood to upgrade, satisfaction with the current solution, or intent to switch. Validate against this question, not against the answers used to build the segments. A segmentation whose groups don’t differ on the holdout outcome has found response patterns, not markets.

Choose an outcome the business already cares about and can observe later. If you can link it to a behavior you log after the survey, such as an upgrade, the holdout becomes a real test and not just another stated intention.

Keep the Battery Short

Every basis item raises the sample you need and adds another dimension where structure can dissolve. Some guides advise a maximum number of items without a source. The arithmetic makes the case on its own. At 70 respondents per variable, 10 items already require 700 completes.

Add one open-text question after the battery asking what the respondent is trying to get done. The answers never enter the clustering, but they let you describe a segment in its members' own words instead of a label the team invented. Keep the whole instrument short enough to finish. A long battery plus descriptors, a holdout question, and open text is too much for a typical in-product intercept.

In Sprig: the Design Agent can draft a needs battery and screener from a written brief. Mark each item in your brief as basis or descriptor, so the draft carries the label.

Matrix questions hold a battery on one scale, but matrix rows and columns can’t be deleted or reordered after launch, so pilot the battery before you launch it.

Scale and Scoring Choices

Put every basis item on one scale, with the same length, labels, and direction. A battery on mixed scales has to be rescaled before clustering, and rescaling is a weighting decision that quietly drives the answer. The examples in this guide use a five-point agreement scale. No source cited here shows that five points beats another length, so the choice matters less than keeping it identical across items.

Keep reverse-coded items to a minimum in the basis. A reversed item that isn’t recoded before analysis flips its contribution, and the resulting segment can look real.

Agreement scales invite acquiescence, where respondents agree with everything. A battery in which most people agree with most statements carries little information for clustering. If you expect that, consider a forced-choice format such as MaxDiff, which gives each person a relative priority score and removes the everyone-agrees problem at the cost of a different analysis. The MaxDiff analysis guide covers that path.

Behavior anchors such as frequency or spend sit on different scales from the battery. Decide in advance whether each is basis or descriptor, and treat it as a descriptor when in doubt. Mixing a count of purchases into a battery of agreement ratings forces a rescaling choice that you then have to defend.

Decide how "don’t know" and "not applicable" are coded before fielding. Coding either one as the scale midpoint inserts a neutral opinion the respondent never expressed, and the clustering treats it as real. Treat them as missing, report how many there are per item, and let the missing-data rule from the quality section handle them.

Scale matters most for k-means, which measures distance and so lets the widest-ranging variable dominate unless you standardize. The analysis section explains why latent class analysis reduces that problem for some batteries, and why it doesn’t remove the choice everywhere.

Record two decisions before you field. First, write down the coded direction of every item, so higher always means more agreement. Second, write down whether you’ll standardize and why, because the segments you get will depend on it, and a later reader can’t tell from the output which way you went.

Sample Size and Precision

Plan for at least 70 completed responses per segmentation variable. That rule comes from Dolnicar, Grün, Leisch, and Schmidt (Journal of Travel Research, 2014), who simulated data of known structure and tested how well segments were recovered. Their abstract states that a sample of 70 times the number of variables proves adequate under all simulated data circumstances.

Where 60, 70, and 100 Come From

The three numbers in circulation have different origins. In the aggregated-data setting the 2014 paper found 60 times the variables adequate. It recommends at least 70 because that allows for real data being harder than the simulations, and it found no significant improvement above 60 in the aggregated setting. The simulations kept the best of 30 random initializations of k-means across 50 datasets per setting.

The 100 comes from Dolnicar, Grün, and Leisch’s 2018 textbook, Market Segmentation Analysis, which recommends 100 times the variables and cites a 2016 study by the same authors. That study’s abstract states no number, so cite the figure through the book.

Treat 70 as the planning figure, not a universal floor. It comes from one simulation study of k-means on tourism-style data, and no source cited here tests it for latent class models or other batteries. It still beats the unsourced totals that circulate in segmentation guides.

Choose between the numbers on purpose. Use 70 as the default. Consider the 60 figure only when your data are clean and the structure is strong, and move toward 100 when you expect noisy or biased responses, since the textbook figure comes from work on how biased survey data degrades segment recovery.

Cut Variables, Not Sample

The rule makes one design choice obvious. An 8-item battery needs about 560 respondents at 70 per variable, and a 20-item battery needs 1,400. Practitioners often do the opposite: they field a long battery to see what comes out, then run it on a few hundred people.

A 2002 review of 47 published tourism segmentation studies, reported in the 2014 paper, found a median of 22.4 respondents per variable. That’s about a third of the recommended 70.

The Derived Sample-Size Table

The table multiplies the number of segmentation variables by each rule. The last column shows how many people a segment holding 8 percent of the sample would contain at 70 per variable. It’s derived arithmetic, not a published norm.

| Segmentation variables | At 60 per variable | At 70 per variable | At 100 per variable | People in an 8% segment at 70 | |:----------------------:|:------------------:|:------------------:|:-------------------:|:-----------------------------:| | 5 | 300 | 350 | 500 | 28 | | 8 | 480 | 560 | 800 | 45 | | 10 | 600 | 700 | 1,000 | 56 | | 12 | 720 | 840 | 1,200 | 67 | | 15 | 900 | 1,050 | 1,500 | 84 | | 20 | 1,200 | 1,400 | 2,000 | 112 | | 25 | 1,500 | 1,750 | 2,500 | 140 | | 30 | 1,800 | 2,100 | 3,000 | 168 |

What Counts as a Segmentation Variable

Count every column the clustering will run on, not every question on the page. A matrix of 10 need statements is 10 variables. A multi-select question with six options becomes six variables once it’s coded for analysis.

Descriptors and the holdout outcome don’t count, because they never enter the clustering. Count before fielding and write the number on the study plan. Adding variables after you’ve seen the data changes the rule’s input and the honesty of the result.

The rule counts completed, clean responses. If you expect to lose some respondents to the exclusion rules in the quality section, add them to the invitation target before you start, so the final analysis file still meets the number.

What the Rule Doesn't Cover

The rule sizes the study for recovering segment structure. It doesn’t size the study for reporting on the smallest segment. A segment holding 8 percent of an 800-person sample rests on 64 people, and every percentage inside it carries the precision of 64, not 800. At 50 percent, a 95 percent interval on 64 people spans about 12 points either side. On 800 people it spans about 3.5.

Decide the precision you need for your smallest target segment, and size the total sample so that segment reaches it. If your own list can’t supply that many people, plan the invitations through a research panel or email before you fix the field date.

Audience, Frames, and Screening

The segments describe whoever you field to, so define that population before writing the battery. Current customers, all users including dormant ones, and the whole category are three different studies. Fielding in-product to active users and presenting the result as a market segmentation is a common frame error.

Match the frame to the decision. A decision about which new market to enter needs the category, typically reached through a research panel. A decision about which existing customers to prioritize can use your own list, by email or in the product.

Write down who’s missing from the frame. Churned customers, dormant accounts, and people outside your reach won’t appear in the segments, and the readout should say so. Report the response rate and who was invited, because people who respond to surveys differ from people who don’t.

Screening

Screen on category behavior, not on interest. A screener that previews the battery’s themes primes the answers you’re about to cluster on.

Ask about behavior in a bounded recent period that you name on the instrument, instead of asking whether someone has ever done something. A bounded period keeps the frame comparable across respondents and makes a later rerun possible.

Business markets need one more decision: whether the respondent is the user, the buyer, or both. A segmentation of daily users and a segmentation of budget holders will often produce different segments from the same companies. Screen for the role the decision cares about and record the other as a descriptor.

Quotas, Carefully

Set quotas on descriptor variables, never on basis variables. A quota on a basis variable fixes part of the segment sizes you’ll recover, and those sizes then get reported as a finding. Quotas balance cells you’ve already defined. They can’t balance segments you haven’t discovered, so they serve the rule-based branch better than the data-driven one.

In Sprig: response quotas are available on the Enterprise plan only. They count completed responses, and their screener questions are limited to single-select, rating scale, and NPS, so a quota can’t sit on a multi-select or a matrix row. A quota can’t be deleted from a live survey, so set the cells before launch.

Fielding and Delivery

Field the whole battery in one window. A segmentation assembled from responses collected months apart mixes people answering in different contexts, and the clustering can’t tell them apart.

A full needs battery plus descriptors is a survey, not an intercept. Email or panel delivery generally suits it better than an in-product prompt, and both give you a defined frame.

In-product delivery, through web apps and websites, works when the battery is short and the population really is your active users. Check how the delivery tool samples, because not everyone who qualifies is necessarily shown the survey.

Pilot the instrument on a small group before launch and read the item distributions. An item everyone answers the same way is dead weight in the clustering, and it’s far cheaper to replace before launch than after.

Check how the battery renders on mobile during the pilot, because a matrix often displays differently on a small screen. State the field window’s dates in the report so a later rerun can match the season.

Don’t pool responses across delivery modes without checking. The same battery can draw different answers by email, in the product, and from a panel, so field one mode or compare modes before you combine them. Keep the invitation wording neutral. Describing the survey as a segmentation primes respondents, so describe the topic instead.

During fielding, check completion and straight-lining as responses arrive, and fix a broken item early. Once the study is live, change as little as possible, because every change splits your sample into respondents who saw different instruments.

In Sprig: panels offer more than 300 demographic, professional, behavioral, and firmographic targeting attributes across dozens of countries, for B2B and B2C audiences. The page publishes no incidence rate, sample limits, or fielding turnaround, so get a written feasibility estimate before you commit to a variable count and a date.

Timing and Cadence

Segmentation is episodic. Run it when a targeting or portfolio decision is open, not on a schedule. Rerun it when the market or the product has changed enough that the basis no longer describes how people decide. If what you need is continuous measurement of how shares move, a brand tracking survey is the better instrument.

Plan the calendar backward from the decision. The sequence is instrument and pilot, fielding, cleaning and exclusions, analysis with its reruns, segment evaluation, and the design of the holdout test. Teams squeeze the analysis, but the reruns and the evaluation are the steps that make the result defensible. Build the holdout into the schedule from the start, not after the readout.

When you do rerun, field the identical battery. A battery someone improved between waves produces a change you can’t attribute to the market. Expect the segments to move even so, often in both count and size. A rerun on a frozen battery can return a different number of segments, which is itself evidence about how stable the original structure was.

Quality Control and Data Hygiene

Segmentation is unusually sensitive to careless responses, because a straight-liner produces a consistent profile that can cluster into its own group. Compute each respondent’s standard deviation across the basis items and flag anyone at or near zero. Review those records before you delete them.

Attention checks catch some inattentive respondents, but they cost goodwill and don’t catch everyone, so use them sparingly. Compare each completion time against your pilot, and review unusually fast records by hand. Check for duplicate respondents too, especially in a panel or an open link.

Then check the output for a response-style segment. A group defined mainly by using the top or bottom of the scale is usually an artifact of how people answer surveys, not a market.

Write the exclusion rules down before you look at the segments. A rule applied after you’ve seen which respondents produced an inconvenient cluster isn’t quality control.

Keep an exclusion funnel in the report: how many respondents started, how many completed, and how many each rule removed. A reader can then see how much the analysis file differs from the field file, and you can see whether any rule removed a suspicious share of one kind of respondent.

Decide the missing-data rule in advance too. Clustering tools either drop incomplete cases or require imputation, and both choices change the result. Keep every basis item answerable by every respondent, and report how many cases you dropped.

Bot detection covers the automated end of this problem. It doesn’t catch a person answering carelessly, so the variance check still applies.

Analysis: The Core Calculation

A rule-based segmentation needs no model. Each respondent is assigned by the rule, and the analysis is a set of comparisons between known groups, which the simple cross-tab guide covers. A data-driven segmentation needs a clustering method, repeated runs, and a stability statistic. This section owns all three.

Latent Class Analysis or K-Means

Both methods partition respondents. They differ in one way that matters for survey batteries. Vermunt and Magidson (2002), in their chapter "Latent Class Cluster Analysis" in Applied Latent Class Analysis (Cambridge University Press), argue that latent class clustering spares you decisions about the scaling of the observed variables, which standard non-hierarchical methods can’t avoid.

Read the condition. The property they describe holds for normal distributions with unknown variances, so it reduces the weighting problem for a battery of that kind and doesn’t remove it everywhere. Latent class models also offer formal model selection through information criteria such as AIC and BIC, and they handle mixed measurement levels.

K-means remains reasonable for a clean single-scale battery. It needs the discipline in the rest of this section to be trustworthy.

Choosing the Number of Segments

Choose the number with the variance ratio criterion (also called Calinski-Harabasz), BIC, or the gap statistic. With latent class analysis, BIC does this job as part of the model fit. Don’t use the elbow method. Schubert (2023) argues that the elbow criterion "severely lacks theoretic support" and shows that changing the scaling of the axes may well change the human interpretation of an elbow.

Then check stability before you commit to a value. The count should also reflect how many distinct treatments your organization can actually run.

The Stability Protocol

A single clustering run isn’t a result. K-means starts from randomly selected points, so run many random restarts and record the number. The 2014 sample-size simulations used 30. Then bootstrap the solution:

Draw two bootstrap resamples and cluster each one.
Use each solution to assign every case in the full original dataset to a segment.
Compare those two sets of assignments with the Adjusted Rand Index, which measures agreement between two partitions corrected for chance.
Repeat across many pairs and read the distribution.

Step two is the one people leave out, and without it the rest is meaningless. Two bootstrap samples contain different observations, so their partitions can’t be compared directly. An index computed without projecting both solutions back onto the same cases measures nothing.

No universal threshold licenses reporting, and anyone offering one has invented it. Set your own stability bar before you see the results. Global stability is also easiest at small numbers of segments, so a rule of taking the most stable solution will pick two segments almost every time. Narrow the candidates with stability, then choose among them on whether the segments are substantively distinguishable.

The Validation Checklist

Run these in order before anyone names a segment.

Mark every variable as segmentation or profiling, and cluster only on the first set
Record the scaling decision, including a decision not to rescale
Run many random restarts and record the number
Compute the solution across a range of segment counts, not one value
Choose the count by variance ratio criterion, BIC, or the gap statistic, never by eyeballing an elbow
Bootstrap the solution and report the Adjusted Rand Index across replications
Classify the result as natural, reproducible, or constructive, and say which in the report
Check the smallest segment’s count before reporting any percentage inside it
Check whether any segment is defined mainly by response style
Profile the segments on held-out variables, not on the clustering variables

Benchmarks and What a Good Result Looks Like

There’s no published cross-industry benchmark for a good market segmentation: no norm for the number of segments, no threshold for variance explained, and no minimum segment size any source supports. The respondent ranges some guides quote are unsourced.

What the literature offers instead is a reproducibility standard. Ernst and Dolnicar found naturally occurring segments in only 6 percent of 32 tourism datasets, so the useful question is whether the same segments reappear when you run the analysis again.

Two Rules of Thumb Without a Source

Two rules circulate in segmentation guides. One says more than five segments is rarely useful, and the other says a useful segment holds at least 10 percent of the population. No source cited here tests either one. They read as professional experience, which is worth something, but they shouldn’t appear in a deck as a standard.

What a Good Result Looks Like

A good segmentation has segments that reappear across reruns, at least one variable you control that identifies each target segment, clear differences on the holdout outcome, and a written statement of whether the structure was natural, reproducible, or constructed. It also changes something. The strongest evidence that a segmentation was good is a holdout test in which the treated segment outperformed its control.

Your best benchmark is your own previous segmentation on the same battery and frame. Compare stability and holdout differences against your last study, not against another company’s segment count.

If someone quotes a benchmark for your study, ask three questions. What was the sample, how many segmentation variables did it use, and did anyone rerun the analysis? A number with no answer to those questions is an opinion. A segment count is an output of design and of the number of treatments you can run, not a quality score.

Interpreting and Acting on the Result

Judge each segment on its own before choosing targets. A solution can hold one strong segment and several weak ones, and the strong one is worth acting on even when the overall solution is mediocre. Dolnicar and Leisch (Marketing Letters, 2017) make this point formally. Their abstract says criteria that assess the global quality of a solution can miss individual attractive segments. Segment-level stability, they say, can identify them.

The Segment Evaluation Scorecard

Score every segment on five criteria adapted from the conditions for an effective segment in Kotler’s marketing textbook: measurable, substantial, accessible, differentiable, and actionable.

A segment that fails the reachable test isn’t actionable, however clean it looks. If you can’t find its members in your data or your media, you can’t treat them differently.

A Worked Example

A project-management software company wants to decide which customers to build its next offer for. It has 40,000 active accounts and holds plan tier, company size, and seats as attributes. The numbers here are illustrative, and the sequence is the method.

The team first asks whether rules would do. Plan tier and company size already split accounts into four groups with different sales treatment, so a rule-based segmentation is ready at once.

But the product team suspects that accounts of the same size want very different things, some wanting control and approvals, others wanting speed with no setup. No attribute captures that, so the team runs a data-driven study on top of the rules.

It writes a 10-item needs battery, which sets the floor at 700 completes. It adds three descriptors and a holdout question on intent to upgrade, and fields by email to a random sample of account admins. Repeated analysis reproduces four segments. Three are stable, and one appears in only some of the reruns, so the team drops it from targeting and reports it as unresolved.

One stable segment is small but shows the highest upgrade intent, with an interval clear of the others, and it concentrates in accounts with many seats. That descriptor makes it reachable. The team builds a three-question typing tool, stores the segment as an attribute, and offers a new approvals workflow to half of that segment first. The other half is the holdout.

Build a Typing Tool

A typing tool is a short set of questions, or a rule on attributes you hold, that assigns a new person to a segment without rerunning the full study. Build it before the readout, because it lets sales, product, and marketing use the segments in the next study, the next campaign, and the next sprint.

Test it on respondents held out of the analysis, and report how often it assigns them to the segment the full analysis did.

Run a Holdout Test Before Rolling Out

Give one randomly chosen group within a target segment the segment-specific treatment, and give a matched group the standard treatment. Compare the outcome. This is the guide’s recommended design, not a sourced protocol, and it’s the most direct answer to the question the study can’t answer by itself: whether segmentation changed performance.

In Sprig: the Synthesize Agent themes open-text answers, which helps describe a segment in its members' own words. Themes group responses, not respondents, so theming isn’t segmentation. A theme mentioned by 30 percent of respondents isn’t a segment holding 30 percent of the market.

Running the Analysis with Claude or ChatGPT

The segmentation itself runs in a connected AI client or a statistics tool, against data you’ve exported. This chapter covers only what’s specific to segmentation.

Theme creation and the recode loop belong to the advanced cross-tab guide, comparing segments you already have belongs to the simple cross-tab guide, and validating AI output in general belongs to the guide on analyzing survey data with AI without sacrificing rigor.

What the Analysis Produces

Two files. The first is a stability table with one row per candidate number of segments, plus a list of every exclusion and every disagreement between methods. The second is the segment evaluation table, with one row per segment giving its count, share, holdout outcome with two independently computed intervals, and a reachability flag.

The second table works the same way whether the segments came from rules or from clustering, so it feeds the scorecard in the previous section either way.

Prompt One: Segmentation With Stability

Paste this into a client that can run code, with the file attached. Fill the bracketed values before you send it.

What this prompt does: fits a segmentation to a needs battery and tests
whether the solution is stable.
What it returns: a recommended number of segments, segment sizes, a
stability table, and a list of every exclusion and disagreement.

File: [FILENAME, default the exported study CSV]
Segmentation variables: [COLUMN LIST, default the battery columns I name]
Don't cluster on any other column.

Use code to calculate this, not estimation.

Rules:
1. Exclude any respondent with a blank or non-response on any
   segmentation variable, and report how many you excluded. Blank is
   not zero.
2. Separately exclude any respondent whose answers across the battery
   have zero variance, and report that count.
3. If fewer than [MIN RESPONDENTS, default 70 times the number of
   segmentation variables] respondents remain, stop and report the
   shortfall. Zero remaining counts as under the threshold.
4. Method: k-means. Standardize each column to mean 0 and unit variance
   only if [STANDARDIZE, default NO], and state which you did. Fit
   solutions for k from [K MIN, default 2] to [K MAX, default 8], which
   is a search range and not a recommendation. For each k, run
   [RESTARTS, default 30, the number used in the 2014 simulation] random
   restarts and keep the best solution by within-cluster sum of
   squares.
5. For each k, run [BOOTSTRAPS, default 100] pairs of bootstrap
   resamples, which is a compute budget and not a threshold. Cluster
   each resample, use each solution to assign every case in the FULL
   original dataset to a segment, and compute the Adjusted Rand Index
   between those two sets of assignments. Report the mean and the
   spread across pairs. Don't compute the index between the bootstrap
   partitions directly.
6. Report the variance ratio criterion for each k, and BIC for the
   latent class or mixture model in rule 8. Don't use the elbow
   method.
7. Any segment with fewer than [MIN SEGMENT, default none, stop and ask
   me before running] respondents is reported as "too small to
   profile". A segment with zero members is also "too small to
   profile", not omitted.
8. Recompute the segmentation for the recommended k by a genuinely
   different method, a latent class or Gaussian mixture model on the
   same variables, and compare the two partitions with the Adjusted Rand
   Index.
9. If the two methods agree less than [AGREEMENT BAR, default none, stop
   and ask me before running], report the disagreement and show both
   partitions' segment sizes. Don't reconcile it silently and don't
   choose one.
10. State whether the evidence looks natural, reproducible, or
    constructive, and why. If the evidence can't separate them, say so.
11. If any rule can't be applied to the data provided, say so instead
    of substituting another approach.
12. Report the library names, versions, and random seeds you used.

Save as segmentation_stability.csv, one row per k, plus
segmentation_notes.txt listing every exclusion and disagreement.

Prompt Two: The Segment Evaluation Table

Run this after prompt one, or after assigning rule-based segments, with a segment column added to the file.

What this prompt does: builds the segment evaluation table for a
finished segmentation, rule-based or data-driven.
What it returns: one row per segment with count, share, holdout outcome
with two independently computed intervals, and a reachability flag,
plus a list of exclusions and disagreements.

File: [FILENAME, default the exported study CSV]
Segment column: [SEGMENT COLUMN, default a column named "segment"]
Holdout outcome column: [OUTCOME COLUMN, default none, stop and ask me]
Reachability columns: [DESCRIPTOR COLUMNS, default every column that
begins "Attributes_", already coded to numbers or categories, and
verify the prefix in my file]

Use code to calculate this, not estimation.

Rules:
1. Exclude any respondent with a blank or non-response on the segment
   or outcome column, and report how many you excluded. Blank isn't
   zero.
2. If fewer than [MIN RESPONDENTS, default none, stop and ask me]
   respondents remain after rule 1, stop and report the shortfall. Zero
   remaining counts as under the threshold.
3. Any segment with fewer than [MIN SEGMENT, default none, stop and ask
   me] respondents, including a segment with zero members, is reported
   as "too small to evaluate" and not interpreted.
4. For each segment, report the count, the share of the sample, and the
   mean holdout outcome with a 95 percent interval.
5. For each reachability column, report the share of the segment
   holding its most common value beside the share among everyone
   outside the segment, each with a 95 percent interval. Report the
   intervals and leave the "reachable" label to me.
6. Recompute every segment's holdout interval a second way by
   bootstrapping [ITERATIONS, default 2000] resamples of respondent
   rows, a compute budget and not a threshold, and compare it with the
   analytic interval.
7. If the two intervals' endpoints differ by more than [TOLERANCE,
   default none, stop and ask me], report the disagreement and show
   the inputs. Don't choose the answer that looks more reasonable.
8. If any rule can't be applied to the data provided, say so instead
   of substituting another approach.

Save as segment_evaluation.csv, sorted by holdout outcome descending,
plus segment_evaluation_notes.txt listing every exclusion and
disagreement.

Setup and Loading Your Data

Export the study as CSV and choose the numerical value format, not response text, because clustering and interval arithmetic can’t consume label strings. Confirm the numeric coding direction of every item before you run anything. Attributes arrive as text strings, so code them to numbers or categories before they become reachability columns.

Keep the labeled export beside the coded one, since the labels are your codebook. Separate the battery columns from everything else before you start, because the whole discipline of this analysis depends on clustering only on the segmentation variables.

In Sprig: the CSV export runs up to 950,000 rows, and its download link expires after one week. The MCP connector returns up to 1,000 responses per call, and the documentation lists no way to page past that, so use the CSV for a battery sized from the table in the sample-size section, such as 1,400 completes.

Attributes export only when you include user attributes, and the export documentation doesn’t mention group membership as a column, so don’t assume one exists.

Reading the Output

Read the too-small rows first, then the intervals. Two segments whose holdout intervals overlap haven’t been shown to differ. Read the stability table next. If the recommended number of segments has a high mean index but a wide spread across pairs, the solution is fragile in a way the average hides.

Then read the disagreement list. A disagreement between k-means and the latent class model isn’t a failure of the prompt. It’s the finding, and it tells you the structure is a function of the method.

What to Verify Before Reporting

Run the stability procedure twice on different random seeds and confirm that the recommended number and the segment sizes land in the same place. Recompute one segment’s mean from the raw rows by hand. Check the model’s exclusion counts against your own count of blanks and straight-liners. Then walk the validation checklist from the analysis section.

You remain responsible for which segments you target. The model does arithmetic. It doesn’t decide what the arithmetic means.

Pitfalls

Five failure modes matter for this analysis.

Accepting the first partition. Asking a model to "find the segments" and taking the first answer reproduces the failure behind the reliability critique below. Always ask for stability.
Non-determinism. Thinking Machines Lab (September 2025) reported 80 unique completions out of 1,000 at temperature zero on an open-source model. Whether your client behaves the same is untested here, so run the prompt twice and compare.
Lost in the middle. Liu and colleagues (TACL, 2024) found models use information in the middle of long inputs less well. As practice advice, not a finding from the paper, keep instructions at the start or end and split large sets.
Sycophancy. Sharma and colleagues (ICLR 2024) documented it, and OpenAI withdrew a model update in April 2025 for being overly agreeable. Never ask a model to confirm a segment you already expect.
Personal information. Open-text answers contain personal details nobody asked for. Whether a vendor trains on them depends on which plan tier you use, not on the vendor alone.

The Critique You Should Know

Segmentation carries three serious published critiques, and each aims at a different step. The counter-evidence is real, and part of it comes from the same research programs.

The Reliability Critique

Most segmentations are constructed, not discovered. Ernst and Dolnicar (Journal of Travel Research, 2018, "How to Avoid Random Market Segmentation Solutions") reviewed 78 segmentation studies published between 2010 and May 2016. Of those, 53 were data-driven, and they concluded that approximately 92 percent were at risk of having presented a random solution. Forty-nine of the 53 didn’t control for sample randomness.

Starting points were the sharpest gap. About half of the 53 studies, 30, used algorithms that depend on starting points, such as k-means, neural gas, and bagged clustering. Only three specified how many starting points they used, with 5, 10, and 500 random restarts.

The audit doesn’t show that any particular study’s segments were wrong. It shows that the studies didn’t report the checks that would tell you. That’s the gap this guide’s stability protocol closes.

They also describe three situations, building on Dolnicar and Leisch (2010). In natural segmentation, distinct groups exist. In reproducible segmentation, they don’t, but repeated runs converge on the same partition. In constructive segmentation, the analyst is choosing one of many possible groupings.

Across 32 tourism datasets, 72 percent were reproducible, 22 percent constructive, and only two datasets, 6 percent, contained naturally occurring segments. Those figures come from tourism data and shouldn’t be quoted as a universal rate.

The Targeting Critique

The Ehrenberg-Bass tradition argues that targeting a segment is often the wrong strategy altogether. Uncles, Kennedy, Nenycz-Thiel, Singh, and Kwok (Journal of Advertising Research, 2012) put the finding in their title: in 25 years, across 50 categories, user profiles for directly competing brands seldom differ. If the buyers of your brand look like the buyers of your competitors', segmenting them is an expensive way to reach fewer category buyers.

A brand tracking survey that records who buys which brand is one place to check whether yours do.

Fennell and colleagues point the same way from a different method. The descriptors they tested predicted who used a category, not which brand people chose.

The critique has limits of its own. It concerns the profiles of brand buyers in established categories, so it argues against targeting on demographics more than it argues against needs-based segmentation. It still sets a bar. Before you target a segment, check that its members really differ from the category’s buyers on something you can act on.

The Implementation Critique

Dibb and Simkin (2009) followed a European utilities company through a major segmentation project, classified the barriers it met in implementation, and derived practitioner rules to reduce them. The lesson is that projects often fail after the analysis. A segmentation gets presented, admired, and never changes a budget, a roadmap, or a message.

The Counter-Critique

The strongest counter to the targeting critique comes from the same literature that undermines demographics. Rossi, McCulloch, and Allenby (1996) found that targeting built on purchase history could greatly improve direct marketing profitability. Even a single observed purchase occasion produced about 50 percent more net couponing revenue than blanket couponing. Targeting paid when it rested on purchase behavior.

That answers the targeting critique for behavioral bases specifically, not for segmentation in general. It’s a couponing result, so it’s a weaker reply to the brand-profile evidence than to the demographic evidence, and it supports this guide’s rule to segment on needs or behavior only by analogy.

The reliability critique has a constructive answer from inside its own research program. Dolnicar and Leisch (2017) show that a solution with good overall quality criteria can lack attractive individual segments. Segment-level stability can find attractive segments the global criteria miss. That argues for evaluating segments one at a time rather than abandoning the method.

Where This Guide Lands

Segmentation is a design decision, not a discovery. Treat the segments as constructed unless the stability check says otherwise, segment on behavior or needs, and require a holdout test before a segmentation changes a budget. If you can’t run the holdout test, start with a rule-based segmentation this quarter, and earn the data-driven version once the organization has shown it will act on segments at all.

Common Mistakes

Thirteen mistakes recur, and each has a tell you can check for. Most share a cause. They treat the clustering output as the end of the study when it’s the start of the evaluation.

Segmenting on demographics. The tell is a profile table where segments differ on age and income and on nothing anyone can act on. Move age and income to the descriptor block and rebuild the basis from needs or behavior.
Putting descriptors into the basis. The tell is a suspiciously clean profile on the variable you most wanted to discuss. Rerun with the descriptor removed from the basis and see whether the segments survive.
Fielding a long battery on a small sample. Twenty basis items on 300 people is 15 respondents per variable, under a quarter of the recommended 70.
Running the clustering once. The tell is a methods note with no stability statistic and no count of reruns. The fix is the stability protocol, reported with the number of restarts.
Choosing the number of segments to match a slide template. Five because the last agency delivered five isn’t a method. Let stability and the number of treatments you can run set the count, and write down how many treatments that’s before you look at the solutions.
Presenting a constructed segmentation as a discovery. Calling segments the market’s natural groups when the data sit in the constructive regime is the claim that makes segmentation work distrusted later. Say which regime your evidence supports.
Building segments nobody can reach. A segment defined only by attitudes, with no descriptor or attribute that identifies its members, can’t be targeted. Check reachability before you name anything.
Rolling out without a holdout test. A segmentation deployed to everyone at once can never show whether it helped. Keep a control group inside at least one target segment from the first day.
Setting quotas on basis variables. A quota on a need statement fixes part of the answer in advance. Put quotas on descriptors only, and report segment sizes as findings only when no quota touched the basis.
Naming segments before reading the profiles. A name arrives with a story, and the story shapes how the team reads the numbers. Read the profiles first.
Treating open-text themes as segments. Themes group answers and segments group people, and the advanced cross-tab guide covers theming itself.
Calling a customer study a market segmentation. Segments from your own active users describe people who already chose you. That’s a useful population and not the market.
Ending at the readout. A segmentation that stops at the deck never reaches a campaign or a roadmap. End at the typing tool, because that step puts the segments into every later study.

If you adopt one habit from this list, make it this one. Before any segment gets a name, it has to pass the scorecard, and you write the scorecard’s pass conditions down before the data arrive.

Synthetic Respondents

Synthetic respondents aren’t usable for extracting or sizing segments. Segmentation is a claim about how answers vary together. That’s exactly what synthetic data gets wrong.

Bisbee and colleagues (Political Analysis, 2024) compared real survey respondents with synthetic responses generated by a language model. They found that synthetic data distorted relationships between variables, and that it understated variance.

Deflated variance is fatal for clustering in particular. An algorithm run on synthetic answers will find cleaner, more separated segments than real people produce. That error is invisible in the output. The one defensible use is pretesting a battery for wording and coverage before it goes to field. Never use synthetic respondents for any segment estimate.

If you need an early read on a battery, pilot it with a small group of real respondents instead.

How Sprig Supports This Method

Sprig fits the fielding and targeting ends of a segmentation study, and it doesn’t perform the segmentation itself. The platform documentation lists no clustering, latent class analysis, or factor analysis, so extraction, stability checks, and segment evaluation run in a connected AI client or a statistics tool.

For the rule-based branch, attributes set through the SDK, CSV import, Google Tag Manager, a customer data platform, or the public API define segments you already know. Those segments become targeting and filtering criteria for later studies.

For the data-driven branch, the platform fields the battery through in-product, email, and panel delivery, then exports a respondent-by-variable file with attributes attached. That’s the shape a clustering script needs. After the analysis, you can load the segment assignment back as an attribute through one of those routes, as the attributes documentation describes, so later studies can target each segment.

A practical split of labor follows. Do the design and fielding in the platform, and do the analysis in a notebook or a connected AI client. Bring the result back as an attribute so the next study can target it, and keep the analysis files with the field dates so a rerun can match them.

None of this replaces a researcher’s judgment on basis selection, segment count, or targeting. You remain responsible for all three.

Alternatives and Adjacent Methods

Several methods answer a neighboring question better than a segmentation does, and the right choice often comes first.

Rule-based segmentation on attributes fits when you already know the variables that drive treatment. It’s cheaper, reproducible, and immediately actionable.

Persona research fits when a design team needs to picture the people in a segment. Personas sit on top of a segmentation and add no population claim of their own.

Cross-tab analysis across existing segments fits when the groups are known and the question is how they differ.

Conjoint or MaxDiff fits when the question is which attributes drive choice or which features matter most within a segment.

Product analytics fits when the basis you want is observed behavior you already log. Behavioral segmentation from logs avoids self-report entirely.

Jobs-to-be-done interviews fit before the battery is written. They generate need statements in customers' own words, and they’re a qualitative method outside the survey category.

A customer needs assessment fits when the question is which unmet need to prioritize rather than who has it.

If you don’t yet know how customers differ, run a segmentation. If you know the groups and want to compare them, cross-tab. If you know the groups and want to pick a feature, price, or message for one, use the instrument built for that choice, and then come back to segment only if the answer varies by group.

Frequently Asked Questions

What Is a Market Segmentation Study?

A market segmentation study is research that divides a market into groups of buyers who differ enough to warrant different products, messages, or prices. It uses basis variables such as needs or behavior to form the groups, and descriptor variables such as demographics to describe and reach them. It’s the first step of segmentation, targeting, and positioning.

How Do You Run a Market Segmentation Study?

How Do Rule-Based and Data-Driven Segmentation Differ?

Rule-based segmentation assigns customers to segments using variables you already hold, such as plan tier or company size. Data-driven segmentation uses survey data and a clustering method to find groups. Rule-based segments are reproducible and immediately actionable, while data-driven segments can reveal distinctions you didn’t suspect but need a larger sample and a stability check.

How Many Respondents Do You Need for a Segmentation Study?

Plan for at least 70 completed responses per segmentation variable, from a 2014 simulation study by Dolnicar and colleagues. An 8-item needs battery needs about 560 respondents, and a 20-item battery needs 1,400. Count segmentation variables, not questionnaire length, and cut variables before you cut sample.

Should You Segment on Demographics or Needs?

Segment on needs or observed behavior, and use demographics to describe the segments afterward. Fennell and colleagues found that demographic and general psychographic variables predicted product use but weren’t useful for predicting relative brand preference. Purchase-history targeting beat blanket couponing in the work of Rossi, McCulloch, and Allenby, though that study didn’t pit it against demographics.

How Many Segments Should a Segmentation Have?

No published source supports a specific number of segments. Choose the count from how stable each solution is across repeated runs and from how many distinct treatments your organization can run. Use the variance ratio criterion, the Bayesian information criterion (BIC), or the gap statistic to compare candidates, and don’t rely on an elbow plot.

How Do You Know Whether Your Segments Are Real?

To validate a market segmentation, rerun the clustering on resampled data and score whether the same segments reappear, using the Adjusted Rand Index, a chance-corrected measure of agreement between two groupings. Then confirm the segments differ on a holdout outcome. Ernst and Dolnicar found natural segments in only 6 percent of 32 tourism datasets, so treat segments as constructed until the checks say otherwise.

What Is the Difference Between K-Means and Latent Class Analysis?

K-means groups respondents by distance, so scaling and random starting points affect the result. Latent class analysis fits a statistical model, offers model selection through the Akaike and Bayesian information criteria (AIC and BIC), and handles mixed measurement levels. Prefer it for mixed batteries. Use k-means with many restarts and a stability check for a clean single-scale battery.

What Is a Typing Tool in Market Segmentation?

A typing tool is a short set of questions, or a rule on known attributes, that assigns a new person to a segment without rerunning the full study. It lets sales, product, and marketing apply the segmentation in later campaigns and studies. Test it on respondents held out of the analysis before relying on it.

What Are the Types of Market Segmentation?

The six common types are demographic, geographic, firmographic, psychographic, behavioral, and needs-based, named for the variable they group on. Behavioral variables have the most direct support in the evidence cited here, and needs-based segmentation is the closest survey equivalent. Demographic and geographic variables usually work better as descriptors that help you reach each segment.

Should You Run Factor Analysis Before Cluster Analysis?

Running factor analysis before clustering is a common recipe, but Dolnicar and Grün argue that it’s not generally the best procedure for identifying homogeneous groups. Cut redundant items at the design stage instead, which also lowers the sample you need.

What Is the Difference Between Market Segmentation and Personas?

A market segment is a sized group with a membership rule that can be tested and shown to be wrong. A persona is a narrative describing one invented person, built on top of a segment to help a design team. Run the segmentation first, and build personas only for the segments a team will design for.

Can You Run a Segmentation Study In-House?

Many teams can. A rule-based segmentation needs no model, because customers are assigned using variables you already hold. A data-driven segmentation needs someone who can run a clustering method in a statistics tool or a connected AI client, rerun it to check stability, and build the segment evaluation table. You remain responsible for basis, segment count, and targeting.

Can Sprig Run a Market Segmentation?

The platform can field the survey and target rule-based segments, but it doesn’t perform data-driven segmentation. No clustering or latent class analysis is documented, so extraction runs in a connected AI client or statistics tool against the exported data. Segments stored as attributes then target every later study.

How Long Does a Market Segmentation Study Take?

No published source gives a standard duration. The timeline depends on instrument design, the pilot, the field window, and the analysis reruns. Panels publish no fielding turnaround, so get a written feasibility estimate before you commit to a date. Budget analysis time separately, because the stability check and the holdout test take longer than a single clustering run.