Most research projects go wrong long before the data arrives. The problem is loosely stated, a method is chosen because it is familiar, and the analysis is improvised afterward. A clear process prevents that, and it gives you a place to put each method you will meet along the way.
This guide walks through the five steps in order. At each step it points to the method that belongs there and to a longer article on it, so you can treat this page as a map and read deeper where your project needs it.
Key takeaways
- The market research process has five steps: define the problem, develop the research plan, collect the data, analyze it, and present the findings.
- Defining the problem comes first and shapes every later choice, including whether new research is needed at all.
- The method follows the objective: MaxDiff ranks lists, conjoint sets prices and designs products, and segmentation finds groups that differ.
- Write the analysis plan before fielding, so you decide which comparisons matter before the data can tempt you.
What are the stages of a market research project?
There are five stages, and each one produces something the next one needs.
- Define the problem and the research objectives. The output is a one-page statement of the decision, the questions it depends on and what you already know.
- Develop the research plan. The output is a design: approach, population, sample, method, questionnaire, timeline, budget and an analysis plan.
- Collect the data. The output is a clean dataset, with quality checks run during fielding rather than after.
- Analyze the data. The output is findings: tables, tests, segments and models tied to the objectives.
- Present the findings. The output is a recommendation, the evidence behind it and a way to measure what happens next.
The process is a loop more than a line. The report from one project becomes the “what we already know” for the next, which is why storing and comparing studies matters, a point we return to in the last step.
What is the first step in the market research process?
The first step is defining the problem. That means writing down the decision the research will inform, the questions that decision depends on, and what is already known. It is where most studies are won or lost, because a vague problem produces a long questionnaire and a report nobody can act on.
A good test is to list the possible answers and write down what you would do for each. If every answer leads to the same action, you do not need the study. If you can name an action for each, you have a decision worth researching.
The article on the first step in the marketing research process walks through this with a worked example, including how to separate a symptom (“sales are down”) from the problem behind it, and how to write objectives as questions that data can answer.
Should you reason from deduction or induction?
This choice belongs at the end of step one and the start of step two, because it decides what kind of research you design. Deductive research starts with a hypothesis and tests it, such as “younger buyers weight software above range.” Inductive research starts with observation and builds explanations, through interviews and open-ended questions that surface what buyers actually talk about.
Neither is better in general. If you can state what might be true, test it with a structured survey. If you cannot, explore first. Many programs do both in sequence: inductive work to form the hypotheses, deductive work to test them at scale.
The longer article on approaching market research with deduction or induction explains how to tell which one your question needs and how the two combine.
How do you develop the research plan?
Step two turns the objectives into a design. You choose between primary research (new data) and secondary research (existing data), define the population, decide how many people you need and how you will reach them, and pick the method. You also write the questionnaire, and, importantly, you decide the analysis plan before fielding.
The analysis plan is a list of the tables and tests you expect to produce. Writing it early exposes questions you forgot to ask and questions you will never use. It also keeps you honest: if you decide which comparison matters after seeing the data, a chance difference can look like a finding.
The method follows the objective, not the other way around. The table below pairs common objectives with the method that usually fits.
| If the objective is to… | Consider | Typical output |
|---|---|---|
| Rank a list of messages, benefits or features | MaxDiff | Importance scores for each item |
| Design a product or set a price | Conjoint analysis | Value of each feature level and price point |
| Find groups of people who differ in needs or behavior | Segmentation | A set of segments with profiles |
| Understand what people talk about, unprompted | Interviews or open-ended questions | Themes and hypotheses |
| Check whether something has changed | Tracking across waves | Trends with significance tests |
When do you use MaxDiff?
Use MaxDiff when you have a list of items and need to know which matter most. Respondents see a small set of items, say four, and pick the most and the least important or appealing. They repeat this over several screens with different sets, and the choices are turned into a score per item.
It works better than a rating scale for this job because respondents cannot call everything important. Forcing a “least” choice produces separation between items that rating scales often leave bunched together.
Here is an illustrative example, not real data. Suppose 200 respondents each answer 3 screens, so there are 600 screens, and each of four product benefits appears on every screen. The score for each benefit is (times chosen best minus times chosen worst) divided by times shown.
| Benefit | Shown | Best | Worst | Best minus worst | Score |
|---|---|---|---|---|---|
| A: Saves time | 600 | 240 | 60 | 180 | 0.30 |
| B: Lower price | 600 | 150 | 120 | 30 | 0.05 |
| C: Better support | 600 | 120 | 180 | -60 | -0.10 |
| D: Wider range | 600 | 90 | 240 | -150 | -0.25 |
The best column sums to 600 and so does the worst column, as it must with one best and one worst pick per screen. Benefit A leads clearly, and D trails. The full method, including how to design the sets and estimate scores, is in the guide to MaxDiff analysis.
When do you use conjoint analysis?
Use conjoint when the decision is about trade-offs, most often between features and price. Instead of asking people what they value, you show them complete product profiles and ask which one they would choose. Because each profile bundles features and a price, respondents must give something up, as they do in a shop, and the analysis estimates how much each feature level is worth to them.
Profiles are built from attributes and levels. Suppose a snack has three pack sizes, three prices and three flavors. That gives 3 x 3 x 3 = 27 possible products. You do not show all of them to every respondent; a design shows each person a manageable subset, and the model fills in the rest.
Conjoint is the method for configuring products and setting prices, and it is the wrong tool for ranking a long list of messages, which is a MaxDiff job. The article on conjoint analysis and how it is used covers attributes, levels and what the output looks like.
When do you use segmentation?
Use segmentation when the average hides the answer. If one group loves a concept and another rejects it, the overall score looks lukewarm and points to the wrong decision. Segmentation splits the market into groups worth treating differently.
There are four main types: demographic (age, income, household), geographic (region, urbanicity), psychographic (attitudes, values, needs) and behavioral (usage, loyalty, occasion). Attitudinal segments are richer, and behavioral segments are easier to target. The strongest segmentations combine both and are checked against outcomes the business can observe, such as purchase or churn.
Segmentation touches every step. You plan for it in step two by asking the questions that will define segments. You build it in step four, with methods such as clustering. You use it in step five by reporting results by segment. The article on the four main types of segmentation compares them and shows when each one fits.
How do you collect and analyze the data?
Steps three and four are the most mechanical, and the plan from step two does most of the work.
During collection, watch quality while the survey is in the field, not after. Look at completion time, straight-lining (the same answer to every item in a grid), duplicate responses and open-ended answers that make no sense. Remove what you must, and write down what you removed and why.
During analysis, build the planned tables first. A crosstab shows a result for each group, such as purchase intent by age band, and a significance test tells you whether a gap between groups is bigger than chance would explain. Then explore what the plan did not anticipate, and label those results as exploratory. Segments, MaxDiff scores and conjoint outputs can be added as variables, so that every other question can be cut by them.
How do you present findings and act on them?
Lead the report with the decision and your recommendation, then show the evidence. Tie each finding to an objective from step one. If a finding answers no objective, it is probably interesting rather than useful, and it can go in an appendix.
State the limits plainly: the sample, the margin of error and anything the data cannot say. Then define what you will measure next, so the project can be judged on what happened after the decision rather than on the report itself.
How do you compare research over time?
Many questions are about change: did awareness rise after the campaign, did satisfaction slip after the price increase? Answering them means comparing a new wave of research with an earlier one, and the comparison is only fair if the two waves were measured the same way.
Keep questions, scales and sample definitions consistent. Note every change and the wave in which it happened. Test movement before explaining it, because a small shift can be sampling noise.
Look for sustained direction across several waves rather than a single-wave jump, and store studies where they can be compared, not in separate files. Connecting findings across studies and waves lets a team reuse what it already knows instead of starting over. If you follow brand health or competitors continuously, you can track what changed and why.
The guide to comparing past and present survey results lists the checks to run before you call a difference real.
How do you run this process in practice?
Whatever tools you use, the process needs the same capabilities: crosstabs with weighting and significance testing, charts for the findings and a way to share results with the people who decide. Halo Reports on mTab provides crosstabs and banner tables with weighting and significance testing, charts and interactive report pages; see Halo Reports for details.
What are the common mistakes?
- Choosing the method first. “We need a conjoint” is a solution to a problem you have not stated. Write the objectives, then pick the method.
- Deciding the analysis after the data arrives. Without a plan, you will find patterns that are noise. Write the table list in step two.
- Using one method for another’s job. Rating scales rank poorly, MaxDiff cannot set prices and conjoint cannot rank thirty messages. Match the tool to the objective.
- Reporting only the average. If segments disagree, the average misleads.
- Changing the wording between waves. A reworded question or a new scale breaks the comparison with the past.
What should you do next?
Start with step one on a live project. Write the problem statement on a page: the decision, its owner, three to five objectives and any hypotheses. Then use the table above to pick the method that matches each objective, and read the matching article before you write the questionnaire. If the page is hard to write, that is useful information, and it is far cheaper to learn it now than after fieldwork.