# Market Segmentation Survey in India: From Segmentation Variables to Segments You Can Act On

How to run a market segmentation survey in India: pick bases, cluster with k-means or latent class, profile personas, then test segments for stability and use.

Published 2026-07-03 · Updated 2026-10-12
Canonical: https://hercules.works/blogs/market-segmentation-survey-india

A segmentation survey asks a sample of buyers two kinds of questions: the variables you will group them on (needs, attitudes or behaviour) and the variables you will describe them with (age, city tier, NCCS, media). A clustering method such as k-means or latent class analysis forms the groups. Before anyone writes personas, rerun the analysis on resampled data to see whether the same segments come back, and check that each one is big enough, reachable and different enough to act on. Hercules Works can field the questionnaire; the clustering and its checks are steps you should be able to see.

## Segments Are Made by Your Choices, So Make Them Deliberately

A segmentation study looks like one analysis, but it is a chain of decisions: which variables define a segment, which method groups people, how many groups to keep and what to call them. Change any one and the segments change.

Below are the main bases with Indian examples, the questionnaire split, two clustering methods, a made-up four-segment result, and the two tests a usable segmentation must pass: stability and actionability.

## Decide First: Pre-Set Groups or Groups Found in the Data

Some segmentations are defined before fieldwork. You split buyers by city tier, NCCS class or heavy and light use, and the survey measures how the groups differ. Media and sales teams already use those labels, so the result is easy to target.

Others are found in the data. Everyone answers the same needs or attitude battery, and a clustering method groups people who answer alike. This can reveal groups nobody has named yet, but the result depends on the method and the sample.

Many studies combine the two: data-driven segments built on needs, then described with the pre-set labels so the business can find them.

## Segmentation Bases and Where Each One Falls Short

| Base | Indian examples of variables | Good for | Watch for |
|---|---|---|---|
| Demographic and socio-economic | Age, gender, household size, city tier, NCCS class | Media buying and distribution plans | People of the same age and class often want different things |
| Behavioural | Purchase frequency, pack size, channel (kirana, modern trade, quick commerce), brand repertoire | Sizing revenue and loyalty programmes | Shows what people do, not why; self-reported frequency is approximate |
| Attitudinal | Agreement with statements about health, price, tradition or convenience | Messaging and positioning | Rating scales pick up how a person uses scales as well as what they think |
| Needs-based | Importance of benefits such as "ready in ten minutes" or "tastes like home" | Product and portfolio decisions | Needs that everyone rates high do not separate anyone |

For socio-economic class, Indian research has long used NCCS, which Research Live describes as "based on ownership of consumer durables and vehicles". The Market Research Society of India adopted a newer system, ISEC, which "considers the educations of the highest educated female and male adults, and the occupation of the chief earner" ([Research Live, 21 February 2024](https://www.research-live.com/article/news/india-adoptsnew-socioeconomic-classification/id/5122883)). Pick one and state which, because the two will not sort households identically.

Psychographic segmentation is often listed as a fifth base. It groups people by values, lifestyle and personality traits rather than by what they own or do. In practice it sits next to the attitudinal row, because both rely on rating statements. If you ask for psychographic segmentation, name the statements you will use and how the segments will be labelled, because a set of segments nobody can describe in one sentence will not be used by the business.

## Split the Questionnaire: Segmenting Questions and Profiling Questions

Keep two lists. **Segmenting variables** go into the clustering. **Profiling variables** stay out of it and are used afterwards to describe each segment and find it in the market. Customer profiling surveys and persona work mostly live in the second list.

| Block | Purpose | Example question |
|---|---|---|
| Screener | Define who is in scope | "In the last month, have you bought packaged breakfast cereal, oats or ready-to-cook upma or poha?" |
| Needs battery (segmenting) | Grouping | "How important is each of these when you choose a breakfast product?" a list of benefits, each on a 5-point scale |
| Attitudes (segmenting or profiling) | Grouping or describing | "I would rather cook breakfast from scratch than use a packet" (agree/disagree) |
| Behaviour (profiling) | Sizing and targeting | Frequency, brands bought, where bought, pack size |
| Demographics (profiling) | Finding the segment | Age, city, NCCS or ISEC questions, household size |
| Media (profiling) | Reaching the segment | Apps used daily, TV channels watched, languages read |

Write needs statements that people could plausibly disagree about; one that almost everyone rates "very important" separates nobody. Randomise the order of statements within the battery.

## K-Means or Latent Class: Choosing the Clustering Method

**K-means** puts each person in the group whose average profile is closest to theirs, measured as a distance. You choose the number of groups in advance and run it for several values. Because it works on distances, variables on different scales usually need standardising first, and the algorithm starts from random points, so run it many times and keep the best solution.

**Latent class analysis** treats segment membership as an unobserved variable and estimates each person's probability of belonging to each class. Jay Magidson and Jeroen Vermunt list its advantages over traditional cluster analysis as "probability-based classification", variables that "may be continuous, categorical (nominal or ordinal), or counts or any combination of these", and covariates that "can be used for cluster description". In their bank example, the model chosen "was the one that had the lowest BIC statistic" ([Magidson and Vermunt, *A Nontechnical Introduction to Latent Class Models*](https://www.statisticalinnovations.com/wp-content/uploads/Magidson2002.pdf)). Note that Magidson's firm, Statistical Innovations, sells latent class software.

If your segmenting variables are all ratings on one scale, k-means is the simpler option to run and to explain. If they mix ratings, yes/no answers and counts, or you want membership probabilities to flag people who sit between segments, the advantages Magidson and Vermunt list point to latent class.

## Worked Example: A Four-Segment Breakfast Study (Made-Up Data)

These figures are invented to show what a segment profile looks like. Imagine 900 urban buyers of packaged breakfast products rated five needs from 1 (not important) to 5 (very important), and a clustering run produced four groups.

| | Rushed Fuelers | Home-Taste Loyalists | Value Planners | Health Seekers |
|---|---|---|---|---|
| Share of sample | 30% (270) | 26% (234) | 24% (216) | 20% (180) |
| Ready in under 10 minutes | **4.7** | 2.9 | 3.6 | 3.8 |
| Tastes like home-cooked food | 3.1 | **4.8** | 3.5 | 3.0 |
| Low price per serving | 2.8 | 3.4 | **4.6** | 2.6 |
| High protein or fibre | 3.3 | 2.7 | 2.9 | **4.7** |
| A brand I already trust | 3.6 | 4.1 | 3.2 | 3.9 |
| Profiling: Tier 1 city residents | 68% | 41% | 47% | 72% |
| Profiling: mostly buy online | 54% | 18% | 26% | 49% |

Read the bold cells first: each segment is defined by the need it rates highest relative to the others. Then read the profiling rows, which were not used to form the groups. Rushed Fuelers and Health Seekers lean towards Tier 1 and online buying; Home-Taste Loyalists do not. That difference is what a media or channel plan can act on.

Check what the smallest segment can support. Health Seekers have 180 respondents, so a share near 50% inside that segment carries a 95% margin of error of about ±7.3 points (1.96 × √(0.25 ÷ 180)). If you need to compare that segment by city tier, the bases get smaller again, so size the sample for the smallest group you plan to cut, not for the total.

## Test 1: Do the Same Segments Come Back?

Sara Dolnicar and Friedrich Leisch point to two sources of randomness in any segmentation: the sample, and the algorithm, which often uses random starting points. Their example: cluster 500 respondents, survey another 500, cluster again, and "we can either get a very similar or a very different solution" ([Dolnicar and Leisch, preprint of *Marketing Letters* 21(1), 2010](https://epub.ub.uni-muenchen.de/10960/1/tr063.pdf)).

Their remedy is the bootstrap. Draw many resamples of your respondents, each the same size as the original, rerun the segmentation on each, and measure how often pairs of solutions agree using the adjusted Rand index. They report that 200 bootstrap samples "worked well" in their examples. Many values close to 1 suggest reproducible segments; many close to 0 suggest the data have no structure and the segments are being constructed.

They also describe three outcomes: natural segments that exist in the data, reproducible segments built on some structure, and constructive segments imposed on unstructured data. Constructive segments can still be useful, they argue, but the manager should know that the split is one of many possible. Ask for this label in the report. Run the bootstrap for each number of segments you are considering; the authors say this comparative reproducibility "can validly be used" to guide the choice of how many segments to keep.

## Test 2: Can the Business Use Each Segment?

The paper notes that general criteria proposed in the literature for judging segmentation solutions include "measurability, accessibility, substantiality, differentiability and actionability". Turn them into questions for each segment before you sign off.

| Criterion | Question to ask of each segment | Breakfast example check |
|---|---|---|
| Measurable | Can we size it and find its members again with a few questions? | Build a short classifier from the needs battery and test it on a holdout sample |
| Accessible | Can we reach it through channels we use? | Health Seekers: 49% mostly buy online in the made-up table |
| Substantial | Is it big enough to justify its own product or message? | Multiply the segment share by category buyers in your target cities |
| Differentiable | Would it respond differently to a different offer? | Value Planners should react to price-pack changes more than others; test it |
| Actionable | Do we have a product, price or message to give it? | If no team will own Home-Taste Loyalists, the segment is description only |

A segment that no channel can reach, or that no team will act on, is a finding rather than a target. Keep it in the report as context and do not build a persona for it.

## What a Segmentation Survey Cannot Tell You

It cannot tell you the segments are permanent. Needs can shift with prices, new formats and life stage, so re-run the needs battery from time to time rather than treating the first solution as fixed.

It cannot assign every customer in your database to a segment unless you ask them the classifying questions or build a model from data you hold on everyone. Magidson and Vermunt make the point that when membership is based on survey data, "a large amount of classification error is almost always present for non-surveyed customers".

Nor does a profile table replace people. A persona written from averages describes nobody exactly, so read that segment's open-ended answers before a team adopts it.

## Running This Study on Hercules

Our part is fieldwork and the first cut of analysis: you write the questionnaire, and studies can be targeted by NCCS (A/B/C), age and geography ([panel methodology](/panel-methodology)). The panel page does not list a clustering or segmentation module, so ask to see the method, the number of runs and a stability check before you accept segments, or export respondent-level answers and run the bootstrap yourself.

The breakfast example sized at 900 completes is a Pro bundle: 900 × ₹120 = ₹1,08,000 ([pricing](/pricing)). Everyone in it uses a smartphone, which is how our panel reaches people, so a segment of offline, low-connectivity buyers will not show up at all. Before buying the bundle, use the Free plan, which includes 30 responses, to see whether the needs statements read clearly on a phone.

## Frequently asked questions

### What is the difference between segmentation and customer profiling?

Segmentation divides buyers into groups that want different things. Profiling describes a group, or your customers as a whole, by who they are and how they buy. In a segmentation study, profiling variables are kept out of the clustering and used afterwards.

### How many segments should I have?

There is no correct number in advance. Compare solutions with different numbers of segments on how well each reproduces across bootstrap resamples and whether every segment passes the actionability checks. For latent class models, information criteria such as BIC are also used.

### Should I use k-means or latent class analysis?

K-means is simpler when all segmenting variables are ratings on the same scale. Latent class analysis handles mixed variable types and gives each person a probability of belonging to each segment, which helps with people who sit between groups.

### Can I segment on NCCS alone?

You can, and it is easy to target. But people in the same NCCS class often differ in what they need from a product, so NCCS usually works better as a profiling variable alongside a needs-based segmentation.
