---
url: "https://questionstar.com/blog/market-research-methods/"
title: "Market research methods: segmentation and positioning"
description: "The main market research methods, how a survey divides your customers into meaningful segments, and where your offer stands against alternatives. With examples."
lastmod: "2026-11-13"
type: "blog"
---

# Market research methods: segmentation and positioning

*How do I use a survey to find out which customer groups my market consists of and where my offer stands against the alternatives?*

The main market research methods, how a survey divides your customers into meaningful segments, and where your offer stands against alternatives. With examples.

A survey shows you what no sales report contains: what your customers need, how they see your offer and what they intend to do. You gain two things from it. First, segments, that is, groups of customers who need similar things: for example, households that cook from scratch two to four days a week, as opposed to those that cook almost every day. Decide what you will divide your customers by before you look at the results. Second, the position of your offer: for the same set of attributes, you ask how important each one is to the respondents and how well your offer and the alternative deliver on it. A survey is one of several market research methods. These methods are classified by research objective (exploratory, descriptive or causal research), by data source (secondary data that already exist or primary data that are newly collected) and by methodology (qualitative or quantitative).

You need very little statistics for this: a cross-tabulation, that is, a table that shows for each group how many respondents gave which answer, and a comparison of means between two groups. The calculations in this article are set in frames marked as examples. Each frame opens with its result in one sentence. If you don't want to follow the numbers, read that sentence and move on.

This article answers five questions in turn:

1. [What are the main market research methods, and what does a survey contribute?](#methods)
2. [Why does a segment have to be defined before you evaluate it?](#define-first)
3. [Which questions do you need to segment customers with a survey?](#questions)
4. [How do you analyze segments, and when is a segment usable?](#analysis)
5. [How does a survey show where your offer stands against the alternatives?](#positioning)

Straight to the examples and to putting it into practice:

- Worked examples you can check: [Who wants the meal kit?](#worked-example) (400 customers, three segments) and [Vegetable box versus supermarket](#worked-example-position) (six attributes compared)
- [From the answers to the result in four steps](#your-data)
- [Analysis in QUESTIONSTAR](#questionstar), [checklist](#checklist) and [frequently asked questions](#faq)

Segmentation and positioning are standard topics in marketing courses, and the slides on both are available in the [course](/principles-of-survey-research/chapter-6/slide-232.html) on this site. There, both topics can easily look like something for large market research departments with specialized software. But the core can be done with an ordinary online survey, even in a small company: with a handful of well-chosen questions, a cross-tabulation and a group comparison. The hard part is less the calculation than the order in which you proceed. That order is what this article is about.

## What are the main market research methods, and what does a survey contribute?

Market research links a company to its customers and its market through information. The information is used to identify opportunities and problems and to develop and evaluate actions (Malhotra and Birks 2007, p. 6). So the starting point is not a method but a decision that has to be made: a new offer, a different price, a different way of addressing customers. My course slides put it this way: "Results have no practical value if the research problem is only vaguely defined" ([Marx 2026, slide 9](/principles-of-survey-research/chapter-1/slide-009.html)).

The following classification shows you which route fits which question and what you can clarify before you set up a survey. The methods can be classified from three angles: by objective, by data source and by methodology ([Marx 2026, slides 13](/principles-of-survey-research/chapter-1/slide-013.html), [14](/principles-of-survey-research/chapter-1/slide-014.html) and [15](/principles-of-survey-research/chapter-1/slide-015.html)).

| Angle | Type | What it is for | Typical routes |
|------|------|------|------|
| Objective | exploratory | understanding the problem and forming hypotheses | smaller surveys, focus groups, interviews |
| Objective | descriptive | describing and measuring: who, what, when, where, how | larger surveys, observation |
| Objective | causal | testing cause and effect: what if? | experiments, A/B tests, consumer panels, each with a suitable research design |
| Data source | secondary | using data collected earlier for your own question | statistics, publications, databases, your own sales data |
| Data source | primary | collecting data that do not yet exist | surveys, interviews, observation, experiments |
| Methodology | qualitative | understanding behavior and the reasons for it, few people, not representative | interviews, focus groups, smaller surveys |
| Methodology | quantitative | counting and measuring, many people, statistical analysis | larger surveys, observation |

*Author's own table based on Marx 2026, slides 13, 14 and 15, and Malhotra and Birks 2007, pp. 69, 73, 79 (objectives), pp. 94, 100 (data sources) and p. 152 (methodology).*

The three angles don't exclude one another: an online survey of 400 customers is descriptive, primary and quantitative at the same time, and five conversations with customers beforehand are exploratory, primary and qualitative. The individual routes have articles of their own: [survey methods](/blog/survey-methods/) (online, by telephone, face to face, by mail), [qualitative methods](/blog/qualitative-methods/) such as interviews and focus groups, the [A/B test in a survey](/blog/survey-experiments/) and the [panel survey](/blog/panel-survey/).

**What a survey contributes.** Sales data show what customers have done: who buys, what, how often and for how much (Malhotra and Birks 2007, p. 100). What customers expect from an offer, how they perceive it and what they intend to do is not recorded at any checkout. In the classification on my slides, these are the characteristics that can't be observed: benefit perceptions, attitudes and perception, preferences, motives and intentions ([Marx 2026, slide 234](/principles-of-survey-research/chapter-6/slide-234.html)). This is what a survey is for.

**What it can't do.** Respondents can't or won't provide every piece of information. They are often not aware of their own motives, so asking why doesn't always get you an accurate answer (Malhotra and Birks 2007, p. 266). Ask about behavior and about ratings instead. Comparing the groups then gives you pointers to possible reasons, but it doesn't prove them.

For a small company, this adds up to a route in five steps. It starts with what already exists: secondary data are easy to access, comparatively inexpensive and quickly obtained, even though they rarely provide all the answers (Malhotra and Birks 2007, p. 96). Exploratory conversations then help to define the problem more precisely before anything is measured (Malhotra and Birks 2007, p. 71).

## What is a market segment?

Customers don't all want the same thing. If you build an offer for "the average customer", you build it for someone who may not exist in that form at all. Market segmentation addresses this problem. It refers to "dividing the 'relevant market' into groups of consumers who are internally homogeneous and externally heterogeneous" ([Marx 2026, slide 232](/principles-of-survey-research/chapter-6/slide-232.html)). In plain words: within a group, people want similar things, and between the groups, different things. A segment is one such group.

You can divide customers by very different characteristics. My slide sorts them by two questions: can the characteristic be seen from the outside, and is it related to the product ([Marx 2026, slide 234](/principles-of-survey-research/chapter-6/slide-234.html))?

|  | General | Product-specific |
|------|------|------|
| Observable | cultural, geographic, demographic and socioeconomic characteristics | user status and usage situation, usage frequency and intensity, brand loyalty |
| Not observable | psychographic characteristics, values, personality and lifestyle | benefit perceptions, attitudes and perception, preferences, motives and intentions |

*Based on Marx 2026, slide 234.*

Age and place of residence are easy to collect but often say little about what someone buys. The expected benefit often says more. As early as 1968, Haley called it the most relevant segmentation base, because the benefits people seek in a product are the basic reasons why true market segments exist (Haley 1968, cited in [Marx 2026, slide 235](/principles-of-survey-research/chapter-6/slide-235.html)). On the other hand, you can't tell the expected benefit by looking at someone. This tension runs through every segmentation, and it comes up again under [When is a segment usable?](#usable).

## Why does a segment have to be defined before you evaluate it?

The survey has run, 400 responses are in, and the obvious question is: which customer group is the most interesting one for us? So the search begins: by age, by place of residence, by household size, by the way the customers found you. At some point a group turns up that stands out. Of the twelve customers from the neighboring town who came through a recommendation, nine want the new offer. That is 75%. Among all respondents it is 38%. This looks like a segment.

Why isn't that enough? Think of a marksman who fires many shots at a barn wall without aiming. The bullet holes are scattered across the wall, and somewhere a few of them happen to lie close together. Afterwards he paints the target around that spot. He always hits the bull's-eye. That says nothing about his skill, because a target painted after the shot can't be missed. For a hit to mean something, the target has to be up before the shooting starts.

Something similar applies to customer groups. If you inspect the answers first and then draw a group around the results that stand out, you will always find a group that looks distinctive. This tells you no more about whether those customers really differ from the rest than the painted target tells you about the marksman's skill. The reason: responses vary from person to person simply by chance. If you split 400 respondents into many small groups, some will have a high share of interested customers and others a low share by chance alone, just as some bullet holes happen to cluster on the wall. And there are many possible groups.

> **Example.** Six background questions with four answer options each yield 264 groups, and on average about 13 of them stand out by chance alone, even if in truth none of them differs from the rest of the respondents.
> 
> Six background questions with four answer options each give 24 simple groups and another 240 as soon as you combine two characteristics at a time, 264 in total. Suppose that in truth none of these groups differs from the rest of the respondents, and each is tested at the usual threshold of 5%: then on average about 13 "significant" groups are still to be expected (264 times 0.05 = 13.2, more on this under [Many tests, many false positives](/blog/statistical-significance/#multiple-testing)).

The group from the neighboring town is the target painted around the bullet holes. The nine out of twelve may reflect a genuine difference or simply chance. The same data can no longer tell the two apart.

The second reason concerns not the timing but the characteristic. Often a group is described by the result itself: "the interested ones", "the dissatisfied ones". Then the same question that divides the respondents is the one you want to evaluate the groups on afterwards, and the comparison goes round in a circle: all you learn about the interested ones is that they are interested. How many of them there are can be estimated. Who they are remains open, and so you can't address them in a targeted way. The characteristic you divide by and the thing you want to evaluate the groups on, the outcome, therefore have to be two different questions. And for any characteristic you can't identify just by looking at someone, whether a need, an attitude or an intention, also ask about characteristics that help you identify and reach the group in the market: household size, place of residence, ordering channel, and, for business customers, industry and number of employees.

These two reasons determine the order of the analysis. Before the analysis, you decide what you divide by and where the boundaries between the groups lie, ideally while you are still writing the questionnaire. Only then do you check whether the groups differ in what you want to know. This gives every question one of three roles:

| Role | What it does | When it is fixed | Example |
|------|------|------|------|
| Segmentation variable | divides the respondents into groups | before data collection, including the boundaries | How often do you cook from scratch in a normal week? |
| Outcome | what you evaluate the groups on | before data collection | interest in a new offer |
| Descriptor variables | help you identify and reach a group | included in the questionnaire before data collection, analyzed after the groups are formed | household size, place of residence, ordering channel |

This is how descriptive market research works in general: it is characterized by the prior formulation of research questions and hypotheses (Malhotra and Birks 2007, p. 73).

There is another side to this: searching is allowed and often useful. There are even methods that do exactly that. Cluster analysis forms groups from the data themselves. There, the groups are suggested by the data and not defined in advance (Malhotra and Birks 2007, p. 672). But the result of a search is a hypothesis and not yet a finding. To return to the picture: put up the painted target and shoot again. In practice this means testing the group you found on data that were not involved in the search: on the next survey, or on one half of the responses that you set aside at random before the search (Malhotra and Birks call for the same check for every cluster solution, 2007, p. 682). If the difference shows up there again at a similar size, that supports the group. If it doesn't, the original hypothesis has not been confirmed for the time being.

In short: in the approach described here, you define the groups by need, behavior or situation before the analysis, then compare the outcome across those groups. A group that only stands out when you search through the results is a hypothesis to begin with, which you test on new data. And a group that is described only by the outcome itself can't be identified in the market.

## Which questions do you need to segment customers with a survey?

The three roles give you the structure of the questionnaire. For segments and for the positioning chart further down, you need five building blocks:

| Building block | Role | Example question | Question type |
|------|------|------|------|
| 1. Behavior, situation or need | segmentation variable | "On how many days of a normal week do you cook with fresh ingredients?" (5 to 7 days, 2 to 4 days, 1 day at most) | single choice, categories that don't overlap |
| 2. What you are deciding about | outcome | "How likely would you be to order the meal kit?" (definitely not, probably not, maybe, probably, definitely) | scale with five points |
| 3. Importance of attributes | need in detail | "When you buy vegetables, how important is ... to you?" (1 = not at all important to 5 = very important) | matrix, one row per attribute |
| 4. Rating of your own offer and of the alternative | perception | "How well does our vegetable box deliver on ...?" and "How well does the supermarket deliver on ...?" (1 = very poorly to 5 = very well) | two matrices with the same rows |
| 5. Characteristics that help you identify the group | descriptor variables | household size, place of residence, ordering channel, and for business customers industry and company size | single choice |

Four notes on this:

- **The segmentation question belongs in every questionnaire you want to segment with, early on and as a required question.** Without it, a response can't be assigned to any segment. Write the categories so that every person fits into exactly one. The boundaries you draw here are the boundaries of your segments.
- **Work out beforehand whether every segment will be large enough.** As a rule of thumb, you need at least 20 to 25 cases per subgroup if you want to draw conclusions about the group (Jacob et al. 2019, p. 72). Malhotra and Birks make the general point: a sample has to be larger if the data are analyzed at the subgroup or segment level than if the analysis is limited to the total sample (2007, p. 409). If you expect only one person in ten to fall into a segment, it takes 200 to 250 responses in total to get 20 to 25 cases in that segment on average. The actual number can turn out lower, so plan with a reserve. How to select specifically by group is covered in the article on [sampling methods](/blog/sampling-methods/).
- **Ask about importance and rating separately,** first all the importance questions, then all the ratings. Someone who gives both for each attribute one right after the other may let the first answer influence the second (Martilla and James 1977, p. 79).
- **The list of attributes is drawn up before the survey,** from conversations with customers and from what you know about your offer (Martilla and James 1977, p. 79). What is not on the list can't appear in the results. I come back to this under [limits](#limits).

How to label scales and how many scale points make sense is explained in the article on the [rating scale](/blog/rating-scale/). The structure and order of the questions are covered in [questionnaire design](/blog/questionnaire-design/), the trial run before the launch in the [pretest](/blog/questionnaire-pretest/) article.

## How do you analyze segments?

The analysis consists of two simple comparisons. Which one fits depends on the outcome.

- **The outcome has categories** (interested or not, customer or not): then it goes into a [cross-tabulation](/blog/cross-tabulation/) together with the segment, with the segments in the columns. You compare the [column percentages](/blog/cross-tabulation/#percentages), that is, the shares within each segment. The [chi-square test](/blog/cross-tabulation/#chi-square) tests whether the differences are larger than you would expect from random sampling variation. [Cramér's V](/blog/cross-tabulation/#cramers-v) measures the strength of the relationship, from 0 (no relationship) to 1 (perfect relationship).
- **The outcome is a rating on a scale** (importance, satisfaction): then you compare the means of the segments. With two segments, the [independent-samples t-test](/blog/the-t-test/) checks whether the gap between the two means is larger than chance would lead you to expect. Independent means: the two groups consist of different people. With three or more segments, you use [analysis of variance](/blog/statistical-tests/#several-groups) or limit yourself to a single comparison of two segments that you named in advance.

What matters is that the number of these comparisons stays small: one segmentation variable, one outcome, plus the attributes you selected in advance. Anything you discover in the data beyond that, you report as exploration.

### Worked example: who wants the meal kit?

> A vegetable box provider divides 400 customers into three segments by cooking frequency before the survey: interest in the meal kit is 26.0% among those who cook almost daily, 50.0% among those who cook several times a week and 35.0% among those who rarely cook.
> 
> The company and all the numbers are invented. The provider delivers vegetable boxes by subscription in its region and is considering offering a meal kit as well: ingredients and recipes for three meals a week. Before the survey, it decides what to divide by: the number of days in a normal week on which the household cooks from scratch. The hypothesis behind it: people who cook every day anyway don't need recipes, and people who almost never cook won't order a meal kit either. "Interested" means anyone who ticks "probably" or "definitely" on the intention to order. That, too, is fixed beforehand. Also named in advance is one comparison on importance: the households that cook several times a week against those that cook almost daily. According to the hypothesis, it matters less to the daily cooks that someone takes the planning off their hands.
> 
> **Frequencies** (segments in the columns):
> 
> | Meal kit | almost daily (5 to 7 days) | several times (2 to 4 days) | rarely (1 day at most) | Total |
> |------|------|------|------|------|
> | interested | 39 | 85 | 28 | 152 |
> | not interested | 111 | 85 | 52 | 248 |
> | Total | 150 | 170 | 80 | 400 |
> 
> **Column percentages** (each segment = 100%):
> 
> | Meal kit | almost daily | several times | rarely | all |
> |------|------|------|------|------|
> | interested | 26.0% | 50.0% | 35.0% | 38.0% |
> | not interested | 74.0% | 50.0% | 65.0% | 62.0% |
> | Base | 150 | 170 | 80 | 400 |
> 
> In the middle segment, the households with two to four cooking days, one in two is interested. Among those who cook almost daily, it is just over one in four. The gap is 24 percentage points, and to those who rarely cook it is 15 percentage points. Can that be chance? The chi-square test checks this across all three segments. Its result is the p-value, here p < 0.001: differences at least this large would occur in fewer than one in 1,000 such surveys if in truth the three segments did not differ. Besides the p-value, you read off three things:
> 
> - **Test statistics for the report:** chi-square = 19.86 with 2 degrees of freedom (the article on [cross-tabulation](/blog/cross-tabulation/#chi-square) shows where both come from).
> - **Strength of the relationship:** Cramér's V is 0.22. Interest is noticeably related to how often people cook, but it is far from being determined by that alone.
> - **Enough cases in every cell of the table:** this is checked using the expected frequencies, that is, the numbers that would be in the cells if all segments were interested equally often: [none should be below 5](/blog/cross-tabulation/#expected-below-5). Here the smallest is 30.4.
> 
> In addition, the importance of two attributes, measured from 1 (not at all important) to 5 (very important), as the mean per segment:
> 
> | Attribute | almost daily | several times | rarely |
> |------|------|------|------|
> | The planning is done for me (recipes, quantities) | 2.94 | 3.80 | 3.35 |
> | Freshness | 4.60 | 4.59 | 4.50 |
> 
> In the comparison named in advance, the middle segment is 0.86 scale points above the almost-daily cooks on "the planning is done for me". The independent-samples t-test gives p < 0.001, so a gap of this size can hardly be explained by chance alone (t = 6.99, calculated in Welch's version, which doesn't assume that the spread is equally large in both groups). Mean importance ratings for freshness range from 4.50 to 4.60 across the three segments, so they are practically the same.
> 
> The result supports the hypothesis and sharpens it. The meal kit is above all an offer for households that cook regularly but not every day, and for whom it matters that someone takes the planning off their hands. Freshness is the most important thing to everyone, but it is no use for telling the segments apart: its means are almost the same in all three segments.

In the worked example, the greatest interest is in the middle segment: the households that cook from scratch two to four days a week. The following examples refer to them.

Two things can be taken away from an analysis like this in general. First, a grouping defined in advance answers a clear question with a few comparisons named in advance. Second, the most important attribute is not necessarily the one on which segments differ. My course slides call this the "Benefit Segmentation Paradox": the variables that separate segments and the variables that customers demand most are usually not the same. "The 'drivers' often do not vary between segments at all and have no discriminating power (e.g. price, quality)" ([Marx 2026, slide 238](/principles-of-survey-research/chapter-6/slide-238.html)). For segmentation, what counts is what separates. For the offer, both count.

## When is a segment usable?

A difference in the table doesn't yet make a segment you can work with. My course slides name six criteria that determine "the efficiency and economic viability of a segment solution" ([Marx 2026, slide 233](/principles-of-survey-research/chapter-6/slide-233.html)):

| Criterion | Meaning according to slide 233 | How to check it with the survey |
|------|------|------|
| Identifiability | consumers can be identified on the basis of easily measurable variables | Can membership be established with one or two simple questions? |
| Substantiality | segments must be large enough to amortize investments | share of the segment among the respondents, projected to customers or the market |
| Accessibility | targeted communication is possible | Does the segment differ in the descriptor variables (location, ordering channel, industry)? |
| Stability | over the period of planning, implementation and effect of the measures | Are the segments of similar size in a second survey, and do they differ in the outcome again? |
| Behavioral relevance | uniform response to segment-specific measures | Do the segments differ in the outcome? |
| Actionability | meaningful and helpful in formulating the marketing mix | Does the difference have consequences for the offer, the price or the communication? |

For a first segmentation, three checks are enough: size, reachability and difference in need.

**Size.** The share in the sample is an estimate with a range of uncertainty. The confidence interval shows how wide that range is. It is explained in the article [Confidence interval](/blog/confidence-interval/).

> **Example.** The middle segment makes up 42.5% of the respondents, which projected to 2,400 customers is about 1,020 households, of which about half, roughly 510, are interested in the meal kit.
> 
> The estimates in detail, each with its 95% confidence interval (normal approximation):
> 
> | What is estimated | Estimate | Range |
> |------|------|------|
> | Share of the middle segment among the respondents | 42.5% (170 of 400) | 37.7% to 47.3% |
> | Households of this segment among 2,400 customers | about 1,020 | about 900 to 1,140 |
> | Interested in the middle segment | 50.0% (85 of 170) | 42.5% to 57.5% |
> | Interested among those who rarely cook | 35.0% (28 of 80) | 24.5% to 45.5% |
> 
> For the 80 who rarely cook, the margin of error is already plus or minus 10.5 percentage points. The projection assumes that the respondents were selected at random from all customers, that it was not mainly certain customers who responded, and that no quota limited the size of the segments.

The smaller a segment is in the sample, the less precise every statement about it becomes. And when respondents express interest, that doesn't yet mean they will buy. So read projections from survey data with caution.

**Reachability.** This is where the third role from the questionnaire helps. You compare the segments on the descriptor variables, again with a cross-tabulation: do the households of a segment live in the city more often, are they more likely to order through the app, are there children in the household more often? The clearer the profile, the more targeted the communication. Malhotra and Birks recommend the same for groups from a cluster analysis: describe them with variables that were not used to form the groups, such as demographics, product usage or media usage (2007, p. 682).

Here the tension from the beginning returns: what can be seen from the outside makes a segment easy to identify but often says little about the need, and the expected benefit is hard to get at. That is why a combination of the characteristics covers the criteria best ([Marx 2026, slide 236](/principles-of-survey-research/chapter-6/slide-236.html)). The three roles from the questionnaire are such a combination.

**Difference in need.** A segment is worth treating separately only if it needs something different from the others. In the example it does: the segments differ in their interest in the meal kit and in how important it is to them to have the planning done. If two segments are also close together in practical terms on all the outcomes defined in advance, you can combine them. A non-significant difference alone is not enough for that, especially with small segments.

## Where does your offer stand? Measuring positioning with a simple chart

Once you know who your offer is for, the next question follows: how do these customers see it compared with what they would otherwise buy? According to my course slides, positioning aligns all marketing activities with the preferences of potential customers, "taking into account the competing products". The objective is that "the actual characteristics as perceived by customers are brought into alignment with the desired target characteristics" ([Marx 2026, slide 240](/principles-of-survey-research/chapter-6/slide-240.html)).

The decisive word is "perceived". The next slide shows a taste test of two well-known cola brands. In the blind test, 51% of the participants preferred one brand and 44% the other. When the same people knew what they were drinking, 23% preferred the first and 65% the second ([Marx 2026, slide 241](/principles-of-survey-research/chapter-6/slide-241.html)). The drink had stayed the same. The only thing that had changed was the knowledge of the brand. So what counts for the position of an offer is how customers see it, and the most direct way to find out is to ask. The [overview of customer surveys](/blog/customer-surveys-overview/) shows which kinds of customer surveys there are for this.

The simplest way to measure this is to compare the importance of each attribute (the target) with perceived performance (the actual). You need building blocks 3 and 4 from the [questionnaire](#questions):

1. **Attribute importance:** How important is each attribute?
2. **Your offer's performance:** How well does your offer deliver on each attribute?
3. **The alternative's performance:** How well does the alternative deliver on the same attributes? The alternative is what respondents would otherwise use. Ask them what it is rather than assuming it.

For each attribute, you calculate three means and the gap between your offer and the alternative. Plot both in a chart, importance on the horizontal axis and the gap on the vertical axis, and two lines divide the picture into four fields. The horizontal line is at zero: above it your offer is ahead, below it behind. The vertical line separates the more important attributes from the less important ones. In this article it sits at the mean importance across all the attributes asked about. A different boundary named in advance is just as possible, because where the lines go is a matter of judgment (Martilla and James 1977, p. 79).

|  | Your offer is ahead | Your offer is behind |
|------|------|------|
| More important attribute | This is your position. You can build on it in your communication. | This is where your offer is vulnerable and needs improvement. |
| Less important attribute | an advantage, but not a main argument | secondary |

The basic pattern, importance against performance, is known as importance-performance analysis (Martilla and James 1977, pp. 77-79). The comparison with an alternative is added here. The same gap can arise when both offers are rated well and when both are rated poorly, so always read the rating itself alongside the gap.

### Worked example: vegetable box versus supermarket

> In the middle segment (170 respondents), the vegetable box is ahead of the supermarket on freshness (+1.1) and regional origin (+2.1), and behind it on value for money (-0.8) and free choice of contents (-1.6).
> 
> Respondents in the middle segment rated six attributes three times, each from 1 to 5. The alternative is the supermarket: most of them named it as their substitute, and everyone is able to rate it. The means are invented:
> 
> | Attribute | Importance | Vegetable box | Supermarket | Gap | Field |
> |------|------|------|------|------|------|
> | Freshness | 4.6 | 4.5 | 3.4 | +1.1 | more important, ahead |
> | Regional origin | 4.1 | 4.7 | 2.6 | +2.1 | more important, ahead |
> | Value for money | 4.0 | 3.1 | 3.9 | -0.8 | more important, behind |
> | Choosing the contents yourself | 3.9 | 2.8 | 4.4 | -1.6 | more important, behind |
> | Home delivery | 3.6 | 4.2 | 3.0 | +1.2 | less important, ahead |
> | Little packaging | 3.0 | 4.3 | 2.2 | +2.1 | less important, ahead |
> 
> The mean importance of the six attributes is 3.87, and that is where the vertical line runs. So "more important" means: above 3.87. "Ahead" means: the gap to the alternative is positive.
> 
> You read the chart field by field. Freshness and regional origin are important to the segment, and the box is ahead: this is the position it already has in the eyes of these customers. Value for money and free choice of contents are important too, and here the supermarket is ahead: these are the points where the offer is vulnerable. On delivery and packaging the box is clearly better, but both matter less to the segment. For the meal kit, this gives a pointer: fixed recipes take away even more choice, and that is precisely where the box is already behind. A choice between several recipes could make up for that.

### The limits of the simple chart

The chart is quick to make. That makes it all the more important to know what it can't do.

- **The lines are a convention.** Where "more important" begins is a matter of judgment. The value of the method lies in the relative levels, not the absolute ones (Martilla and James 1977, p. 79). The two authors often consider the middle of the scale a workable boundary and move the lines when low ratings are absent. If the line sits at the mean of the attributes asked about, as it does here, it depends on the list: in the example, "choosing the contents yourself" at 3.9 is just above the line at 3.87. If a seventh, very important attribute had been asked about, it would be below it. Martilla and James consider means acceptable as long as they are consistently close to the medians, that is, to the values that split the answers into two halves of equal size (ibid.).
- **Stated importance often does little to distinguish segments.** In the example, five of the six importance values lie between 3.6 and 4.6, all above the middle of the scale. And what is similarly important to everyone doesn't separate the segments (see the paradox above). Offers can still differ clearly on the same attribute, as with freshness in the example. One addition is derived importance: you check which individual ratings are most closely related to the overall judgment (see [key driver analysis](/blog/correlation-vs-causation/#driver-analysis)). If customers have to trade attributes off against each other, for instance price against choice, conjoint analysis is the appropriate method.
- **The list determines the picture.** Whoever specifies the attributes has to find all the essential ones, and that is hard. The result depends on the attributes chosen (Malhotra and Birks 2007, p. 698). The same goes for the alternatives: which ones you include in the comparison determines which differences become visible (Malhotra and Birks 2007, p. 696).
- **Customers are not the market.** If you survey only your own customers, you survey people who have already decided in favor of the offer. Their judgment is probably more favorable than that of non-customers, so the lead is more likely overestimated than underestimated. How non-customers judge is something only other respondents can show (see [Finding survey respondents](/blog/finding-survey-respondents/) and [Representative sample](/blog/representative-sample/)).
- **Small gaps are not a finding.** Means vary from sample to sample. Because the same people rate both offers, the paired t-test checks whether a gap is larger than chance would lead you to expect (see [The t-test](/blog/the-t-test/#gepaart)). With six attributes that is six tests, so interpret borderline results cautiously.
- **The chart is not a perceptual map.** A perceptual map depicts several providers at once in an attribute space, and the distances between them correspond to their perceived similarity ([Marx 2026, slide 242](/principles-of-survey-research/chapter-6/slide-242.html)). That takes methods of its own, see the next section.

## Which methods go further?

For segmentation and positioning, there are methods that do more than a cross-tabulation and a comparison of means. You don't have to be able to run them, but you should know that they exist and what they are for. Cluster analysis divides respondents into groups that are as uniform as possible on the basis of several characteristics (Malhotra and Birks 2007, p. 671). Multidimensional scaling represents perceptions and preferences spatially, and the result is the perceptual map (Malhotra and Birks 2007, p. 694). Conjoint analysis determines from the evaluation of complete offers how important individual attributes are (Malhotra and Birks 2007, p. 704). It requires a dedicated experimental design when the questionnaire is developed and can't be calculated afterwards from ordinary questions. Factor analysis condenses a long list of attributes into a few dimensions beforehand (Malhotra and Birks 2007, p. 646, see [Latent constructs and multi-item scales](/blog/latent-constructs-and-multi-item-scales/)). How conjoint analysis and perceptual maps are built is shown in chapter 6 of the [course](/principles-of-survey-research/chapter-6/slide-242.html).

## How to calculate this with your data

You get from the answers to the result in four steps. The right-hand column walks through them once using the meal kit example, with the numbers you read off at each step and carry over into your report or thesis:

| Step | What you do | In the meal kit example |
|------|------|------|
| 1. Compare segments | Create a cross-tabulation: the segmentation question in the columns, the outcome in the rows. Read off the column percentages and the p-value of the chi-square test | Cooking frequency in the columns, intention to order in the rows, "probably" and "definitely" grouped as "interested". Read off: 50.0% of those who cook several times a week are interested, 26.0% of the almost-daily cooks and 35.0% of those who rarely cook, p < 0.001 |
| 2. Compare needs | Read off the means of the attributes selected in advance for each segment. Check the comparison of two segments named in advance with the independent-samples t-test | "The planning is done for me": mean 3.80 among those who cook several times a week, 2.94 among the almost-daily cooks, a gap of 0.86 scale points, p < 0.001 |
| 3. Estimate size | Read off the segment's share of all respondents and determine its range | Those who cook several times a week: 170 of 400 respondents, that is 42.5%, range 37.7% to 47.3%. Project this to all customers only if the respondents were selected at random from them and it was not mainly certain customers who responded |
| 4. Determine the position | Limit the analysis to the segment, read off three means per attribute, calculate the gap and draw a scatter plot | Only the 170 who cook several times a week. Freshness: importance 4.6, vegetable box 4.5, supermarket 3.4, gap +1.1. The same for the other five attributes. In the chart, importance on the horizontal axis, the gap on the vertical axis, one line at 0 and one at 3.87, the mean importance. Also state in the report who was surveyed and which alternative was offered for comparison |

This is how the result of step 1 might read in the report: "Half of the households that cook from scratch two to four days a week are interested in the meal kit (85 of 170), compared with just over a quarter of those that cook almost daily (39 of 150). Interest differs significantly across the three segments (chi-square = 19.86, 2 degrees of freedom, p < 0.001, Cramér's V = 0.22, n = 400)." The n stands for the number of respondents. Steps 1 and 2 are described in detail under [Who wants the meal kit?](#worked-example), step 3 under [When is a segment usable?](#usable) and step 4 under [Vegetable box versus supermarket](#worked-example-position).

You calculate the cross-tabulation, the chi-square test, the t-test and the means directly in QUESTIONSTAR's reports, see the next section. This also applies to the paired comparison of your offer with the alternative, attribute by attribute. The reports do not compute the methods that go further: analysis of variance for three or more segments, cluster analysis, multidimensional scaling, conjoint analysis and factor analysis. There are two ways to run these analyses.

**With a statistics program.** Export the responses as an SPSS, CSV or XLSX file and run the analysis in SPSS, R or the free program JASP. What to watch out for is described in the article on statistical tests under [How to calculate this with your data](/blog/statistical-tests/#your-data).

**With an AI assistant.** If QUESTIONSTAR is connected to an AI assistant via the [MCP interface](/help/integrations/mcp-ai/), for example in ChatGPT or Claude, the assistant can read your survey responses and run the analysis. Here, the assistant performs the analysis, not QUESTIONSTAR. Name the method and the questions explicitly, for example: "Use cluster analysis to group respondents into two to four clusters based on their answers to the six importance questions. Show the mean ratings and the size of each cluster, and check whether the solution is similar across two random halves of the data." Treat the result as a suggestion for the next survey. Anything that goes into a thesis, a publication or an important decision should be checked in a statistics program.

On data protection: the interface does not hand names, email addresses and similar contact data of your respondents over to the assistant. Free-text answers can still contain personal information. Mark such questions as "Private question" in the survey's privacy settings. With that setting switched on, they are not passed to the assistant either ([help article on the MCP interface](/help/integrations/mcp-ai/)).

## Market research with QUESTIONSTAR

You create cross-tabulations, means and t-tests in QUESTIONSTAR's reports, without exporting the responses:

- **Cross-tabulation:** The "Crosstab" report shows chi-square, degrees of freedom, the p-value and Cramér's V. You switch on column percentages and expected frequencies in the settings. If an expected frequency is below 5, the report points it out. Answer options can be grouped in the report, for example "probably" and "definitely" as "interested".
- **Means and t-test:** You add descriptive statistics such as the mean and the standard deviation in the reports. The t-test is available for independent samples (two segments) and paired for two questions of the same questionnaire. You need the paired t-test for the positioning chart: the same people rated your offer and the alternative, so each person's two ratings belong together as a pair. In the report, for each attribute you select the row of one matrix question against the same row of the other, for example "freshness" for the vegetable box against "freshness" for the supermarket. The second route is the [MCP interface](/help/integrations/mcp-ai/): there, every matrix row is a variable of its own, and the paired t-test runs on the QUESTIONSTAR server.
- **Filters:** With a filter, you analyze one group specifically. For the positioning chart of a segment, set the filter to that segment and read off the 18 means: three question blocks with six attributes each.
- **Quotas:** With quotas, you limit the number of responses from individual groups. You need to recruit specifically from groups that are missing. What you can set is a maximum number per answer option ([Features](/features/)). If you limit a segment by quota, its share can no longer be projected from the sample.
- **Export and chart:** You export the responses as an SPSS, CSV or XLSX file. You draw the chart of importance and gap in a spreadsheet as a scatter plot: for each attribute, the importance and the difference between the two ratings.

The reports don't show confidence intervals. You calculate the range of a share with the calculator in the article [Confidence interval](/blog/confidence-interval/). All the analyses mentioned are included in every plan, including the free one.

**Compare segments right in the report**

Set up your survey and compare your segments in the "Crosstab" report, in the free plan too.

[Start for free](/order/?code=free)

## Checklist: segments and position with a survey

- Define the decision, review existing data and talk to a few customers before writing the questionnaire.
- Choose a segmentation variable based on need, behavior or situation, and set the group boundaries before data collection.
- Define the outcome before data collection, as a separate question and with the rule for who counts as "interested".
- Include descriptor variables that help you identify and reach each segment.
- Plan for at least 20 to 25 cases per segment.
- Ask about importance and rating separately, using the same list of attributes for your own offer and the alternative.
- Compare the segments in the columns of the cross-tabulation (column percentages, chi-square test, Cramér's V), and check each segment's size, reachability and difference in need.
- Report groups that only stood out while you were searching the data as hypotheses, and note them for the next survey.
- State with the positioning chart who was surveyed and which attributes and which alternative were included.

## Frequently asked questions about market research, segmentation and positioning

### What are the main market research methods?

Market research can be classified from three angles: by objective (exploratory, descriptive, causal), by data source (secondary data, which already exist, and primary data, which are newly collected) and by methodology (qualitative or quantitative) (Marx 2026, slides 13 to 15).

### What is market segmentation?

Market segmentation is the division of a market into groups of customers who are similar within each group and different from one group to the next. Each of these groups is called a segment. The aim is to tailor the offer and the communication to the groups instead of treating all customers the same (Marx 2026, slide 232).

### Which characteristics are used to segment customers?

Observable characteristics such as age, place of residence or usage frequency, and unobservable ones such as the expected benefit, attitudes or intentions. General observable characteristics make a segment easy to identify but often say little about buying behavior. The expected benefit says a lot about it but is hard to get at. That is why a combination is recommended (Marx 2026, slides 234 and 236).

### Why should a segment be defined before the analysis?

Because every data set contains groups that stand out by chance. If you search through many groups and pick the most conspicuous one, you can no longer use the same data to check whether the difference is real. A grouping defined in advance answers a clear question with a few comparisons named in advance. Groups that only stand out during the search are hypotheses for the next survey.

### How large does a segment have to be in the survey?

As a rule of thumb, you need at least 20 to 25 cases per subgroup if you want to draw conclusions about the group (Jacob et al. 2019, p. 72). For precise shares you need considerably more, because the smaller the segment, the wider the range.

### How do I measure positioning with a survey?

With three blocks of questions about the same attributes: how important each attribute is, how well your own offer delivers on it and how well the alternative delivers on it. From the means you get, for each attribute, the importance and the gap to the alternative. Important attributes on which you are ahead are your position, and important ones on which you are behind are the areas for improvement.

### Do I need a cluster analysis for segmentation?

No. If you define in advance one characteristic to divide by, a cross-tabulation and a comparison of means are enough. Cluster analysis is worthwhile if you want to group by several characteristics at once. It forms the groups from the data, so you should test its result on new data or on a second half of the responses (Malhotra and Birks 2007, pp. 672, 682).

## Sources

Haley, R. I. (1968). Benefit Segmentation: A Decision-Oriented Research Tool. Journal of Marketing, 32(3), 30-35. Cited in Marx (2026), slide 235.

Jacob, R., Heinz, A. and Décieux, J. P. (2019). Umfrage. Einführung in die Methoden der Umfrageforschung [Surveys: an introduction to the methods of survey research] (4th ed.). Berlin/Boston: De Gruyter Oldenbourg. In German. p. 72 (minimum number of cases per subgroup).

Malhotra, N. K. and Birks, D. F. (2007). Marketing Research. An Applied Approach (3rd European edition). Harlow: Pearson Education. p. 6 (definition), pp. 69, 71, 73, 79 (exploratory, descriptive and causal research), pp. 94, 96, 100 (primary and secondary data), p. 152 (qualitative and quantitative), p. 266 (limits of the survey), p. 409 (larger samples for analysis at the subgroup or segment level), p. 646 (factor analysis), pp. 671, 672, 682 (cluster analysis), pp. 694, 696, 698 (multidimensional scaling, attribute ratings), p. 704 (conjoint analysis).

Martilla, J. A. and James, J. C. (1977). Importance-Performance Analysis. Journal of Marketing, 41(1), 77-79. p. 79 (asking separately, list of attributes, position of the lines, medians).

Marx, P. (2026). Principles of Survey Research. Online course, chapter 1, slides 9, 13, 14 and 15, and chapter 6, slides 232, 233, 234, 235, 236, 238, 240, 241 and 242. QUESTIONSTAR. [Course, chapter 6](/principles-of-survey-research/chapter-6/slide-232.html).

**Ready for your next survey?**

Create questions, collect answers and compare segments right in the report. Start for free, no credit card required.

[Start for free](/order/?code=free)

May your target always be in place before the first survey response arrives.

**Related articles:**

- [Cross-tabulation: how to read and test a contingency table](/blog/cross-tabulation/)
- [Customer surveys: an overview](/blog/customer-surveys-overview/)
- [Sampling methods: random, quota and cluster sampling](/blog/sampling-methods/)

*Dr. Paul Marx*

Author: Dr. Paul Marx
Dr. Paul Marx is the founder and Managing Director of QUESTIONSTAR and was a professor of marketing at the University of Siegen.

## Related articles

- [cross-tabulation](/blog/cross-tabulation)
- [customer-surveys-overview](/blog/customer-surveys-overview)
- [sampling-methods](/blog/sampling-methods)

*Author: Dr. Paul Marx — see [About](/about).*