How Do You Define a Research Question? From Business Problems to Verifiable Research Questions

Many surveys fail not because of insufficient sample size or unsophisticated statistical methods, but because they go astray from the very first sentence. Statements like "understand why users churn" or "evaluate opinions on the new proposal" sound like research tasks, yet they neither clarify which decision the results should support, nor limit whose experiences are under examination, nor specify whether the need is causal clues, population proportions, or evidence of causation. If teams write questionnaires directly based on these, they often end up with a set of un-actionable attitude percentages.

To turn business needs into research questions, use the five-column derivation method: the decision to be made, the facts not yet known, the subjects and contexts with relevant experiences, the evidence sufficient to reduce uncertainty, and the boundaries the conclusions cannot cross. These five columns are not a writing format but a project initiation checklist. If any column is vague, the questions, sample, or analysis are likely to miss the point.

First, Distinguish Four Seemingly Similar Sentence Types

Research usually starts from a business outcome but cannot stop there. The four sentence types below exist at different levels; conflating them is a common cause of research scope creep.

A business outcome is not a research question, and a research question is not a single line in a survey. A research question typically requires multiple types of evidence to answer jointly: behavioral logs show where people stop, interviews reconstruct their understanding and obstacles at the time, and structured questionnaires estimate the distribution of different obstacles within the target population. Conversely, if an item cannot be linked back to a research question and a decision, it should not remain in the survey just because it "might be useful later."

The Five-Column Derivation Table: Moving from "Wanting to Know" to "How to Judge"

Column one is the decision. Avoid phrases like "optimize the experience." Instead, specify the decision-maker, the options, and the timeline—for example, "the product owner needs to prioritize among three candidate changes for the next version." If the outcome won't change any action regardless of the results, the project is at best information gathering and shouldn't consume a formal research budget.

Column two is the unknown. Inventory what is already answered by logs, customer service records, sales feedback, and historical studies. If the retention decline is already localized to the setup phase via logs, then asking users "where did you leave?" has limited value; what is truly unknown might be what they misunderstood, why they didn't seek help, or what conditions might have encouraged them to continue. The American Association for Public Opinion Research's Best Practices for Survey Research explicitly recommends assessing at the outset whether the objectives are specific, whether existing data can answer the question, whether other methods are more appropriate, and whether a single survey is sufficient.

Column three is the subjects and context. "Users" is almost never a directly sampleable population. You need to specify roles, key behaviors, time windows, and necessary exclusion criteria. For example, "administrators who created a workspace for the first time in the past eight weeks, completed registration, but did not publish their first survey within seven days" is easier to qualify than "churned users" and also reduces errors from long-term memory recall.

Column four is the evidence. First write down what evidence is necessary for the decision, then choose the method. If you need to know what obstacles exist, you need narrative accounts of experiences with context. If you need to estimate how many people each obstacle affects, you need a clearly defined target population and structured measurement. If you need to prove a specific change will improve retention, self-reports or cross-sectional correlations are insufficient; consider randomized experiments, quasi-experiments, or at least behavioral data with a temporal sequence.

Column five is the boundary. Write down what this study cannot answer before collecting data. Conclusions drawn from newly registered administrators cannot automatically be extrapolated to mature customers. Proportions from a convenience sample cannot masquerade as overall market share. A correlation such as "people who experienced difficulties are more likely to churn" cannot independently prove that difficulties cause churn. Boundaries are not disclaimers at the end of a report; they are design conditions that determine sampling, wording, and the strength of conclusions.

A Full Derivation Example: Don't Compress a Retention Problem into a Single Cause Question

Suppose a collaboration software company notices a decline in new customer retention, and the initial request is "send out a survey to see why users stopped using it." If taken literally, the team might list options like price, features, difficulty, no immediate need, etc., and let the churned users select all that apply. The problem: options come from internal guesses. Respondents might use "price" to summarize an overall lack of value. And surveys cannot reliably reconstruct which specific task they struggled with.

The research brief after the five-column derivation could be:

• Decision: At the version review meeting in three weeks, select one area—initial setup, team invitation, or sample guides—for priority redesign.
• Unknowns: Logs show exits are concentrated on the configuration page, but we don't know what users intended to accomplish, how they interpreted the fields, or whether the exit was a temporary pause or a judgment that the product wasn't a fit.
• Subjects and Context: Team administrators who registered for the first time in the past eight weeks, had a specific work task, entered the configuration page, but did not publish content within seven days; excluding internal tests and trials without a real task.
• Evidence: Task path logs, purposefully sampled interviews based on exit nodes, a short survey measuring the candidate obstacles identified in interviews across a qualified sample, and behavioral validation in the next version.
• Boundaries: The first two phases are for discovering mechanisms and estimating self-reported obstacles; they cannot directly prove that a specific UI change will improve retention.

Based on this, the primary research question can be: "Among new administrators with a real task who entered initial setup but did not publish content within seven days, what incorrect expectations or task obstacles did they form during configuration, and how are these obstacles distributed among populations at different exit points?" A subsequent validation question would be: "Will changes addressing the highest-priority obstacle make it more likely for similarly qualified new customers to complete their first publication within seven days?" The first question suits qualitative and descriptive evidence; the latter enters causal validation. By separating these, the team no longer expects a single survey to simultaneously explain mechanisms, estimate scale, and prove the effectiveness of a redesign.

Verbs in Research Questions Determine the Evidence You Need

The verbs in your questions can help check the alignment between question and method, but they cannot be mechanically mapped to a single tool.

• "Understand, discover, reconstruct" usually point toward processes, language, and underlying mechanisms. Interviews, observations, open-ended tasks, or document analysis are more appropriate, producing themes, pathways, and hypotheses to be tested rather than population percentages.
• "Estimate, distribution, compare" require a clear population, sampling frame, measurement definitions, and uncertainty. Structured questionnaires can answer how many in the sample choose an option, and only cautiously infer to the population if sampling and weighting conditions are met.
• "Predict" requires testing performance on data not used for modeling and distinguishing predictive accuracy from causal explanation. A variable predicting churn doesn't mean changing it will reduce churn.
• "Cause, increase, decrease" ask a counterfactual question: if everything else were comparable and only this factor changed, would the outcome differ? Prioritize randomized controlled designs; when randomization isn't possible, clearly state the identification assumptions of the quasi-experiment and potential confounders.

The UK Government's Magenta Book evaluation guidance targets public policy, but its principle applies equally to commercial research: align questions with the intervention logic, available data, and actual decision points. The U.S. Census Bureau, when introducing its survey methodology research, also lists cognitive interviews, focus groups, usability testing, field observations, and experiments as different methods for improving questionnaires and data quality. The rationale for combining methods should come from evidence gaps, not from what a platform happens to offer.

Four Elimination Tests to Block Bad Questions Before Project Kickoff

1. Indifferent Decision Test: Hypothesize the results are A, B, and "no significant difference." If none of these outcomes would change your choice, resources, or next step, the question lacks practical decision value.
2. Unknowable Test: Did respondents directly experience the phenomenon, and can they reasonably recall it? Asking customers to explain why other customers left, to accurately predict purchasing six months ahead, or to recall their feelings about a button from a year ago often yields speculation, not evidence.
3. Preconceived Conclusion Test: If data contradicting the team's expectations can always be explained away by "users didn't understand the question" or "the sample is wrong," the research cannot refute the original belief. AAPOR also explicitly warns against using surveys to advocate for a predetermined position. Specify at the outset what results would lead the team to abandon their initial hypothesis.
4. Method-First Test: After stripping out method words like "conduct a survey" or "interview twenty people," can the team still articulate which specific uncertainty needs to be reduced? If not, it's a procurement or execution requirement, not a research question.

Passing these elimination tests doesn't mean the design is complete. It merely signifies the question is worth pursuing. Next, you still need to verify accessibility to the target population, measurability of core concepts, whether the sample can support the planned subgroup analyses, and whether ethics, privacy, time, and budget permit.

A One-Page Project Brief Should Specify What Remains, Not How Many Questions

The qualified output of a kickoff meeting isn't a fully drafted questionnaire; it's a one-page agreement that both business and research sides can reiterate: background facts and existing evidence, the decision to be made, one primary question with a limited set of sub-questions, the target population and exclusion criteria, the evidence needed for each question, the proposed methods and their sequence, conclusion boundaries, owners, and decision date. Adding a "Research Question – Evidence – Method – Decision" matrix can quickly reveal orphaned questions and evidence gaps.

Within the survey team, the approved research brief can serve as input for AI-generated questionnaires or interview guides. Then, check each item against its purpose: Which research question does it serve? What variable will it create? How will it be analyzed? Delete questions without a corresponding rationale. If a critical question lacks evidence, add methods, not filler questions. AI here handles initial drafting and iterative acceleration; it cannot replace the research lead in determining the level of evidence, nor should it automatically upgrade interview insights into population-level conclusions.

Finally, have someone who wasn't involved in the project read only this one page and answer four questions: Who is making what decision this research will help? What is the most critical unknown? Which people and evidence can address it? What conclusions are explicitly off-limits? If the person can only summarize it as "understanding user needs," hold off on distribution. An executable research question isn't about how professional it sounds; it's defined by its capacity for contrary results to be honestly reported, for different results to lead to different actions, and for its readers to know where the evidence ends.

Sources for Research Question Frameworks

• American Association for Public Opinion Research: Best Practices for Survey Research
• HM Treasury: The Magenta Book — Central Government guidance on evaluation
• U.S. Census Bureau: About Survey Methodology