Learn · methods

How to design a qualitative research study that actually works

Most qualitative research doesn't fail in the interview room. It fails two weeks earlier, when the study design was vague, the topics were too broad, and nobody thought clearly about what "done" would actually look like. The design is the highest-leverage part of any qualitative study. A well-designed study with a mediocre interviewer will outperform a poorly designed study with an excellent one.

Start with a research question, not a topic list

The most common design mistake is treating research objectives as a list of topics to cover rather than questions to answer. Topics are a means to an end. The end is answering a specific question about a specific thing you need to understand.

The difference in practice:

Topic-based thinking: "We want to understand the onboarding experience."

Question-based thinking: "We need to understand where users lose confidence in the onboarding process and what's driving drop-off at the document upload step."

The first framing produces a study that covers a lot of ground shallowly. The second produces a study designed to answer a specific question with enough depth to act on the findings.

Before you design anything, write the research question in one or two sentences. Be specific enough that you can evaluate, at the end of the study, whether you answered it. If you can't write that sentence, the research isn't ready to design yet.

Scope your topics correctly

Once the research question is clear, break it into topics: the individual areas of inquiry the study will cover. Each topic becomes a section of the interview, with its own depth setting and its own resolution criteria.

Three principles govern good topic scoping.

Narrow enough to be covered. A topic should be resolvable in 5 to 10 minutes of focused conversation. "The purchase journey" is not a topic. It's a research program. "The moment of commitment: what made the participant decide to proceed and what almost stopped them" is a topic. It's narrow enough to explore substantively in a single conversation section.

Broad enough to matter. The flip side of over-scoping is under-scoping, creating topics so narrow that they produce thin, technical answers rather than genuine understanding. "How long did the checkout process take?" is a data point, not a research topic. "What made the checkout feel trustworthy or untrustworthy" is a topic with real evidence behind it.

Sequenced logically. The order of topics matters. A study that opens with abstract strategic questions before establishing context will get weaker answers than one that starts with concrete recent experience and moves toward meaning. Start specific and personal, then move to broader patterns and interpretation. Participants warm up. They give better answers later in a session than earlier.

Most well-scoped studies have between three and six topics. More than six is a signal that the research question is too broad. Fewer than three usually means the study is a quick check rather than a full study, which is fine, but should be designed as such.

Match depth to what each topic actually needs

Not every topic requires the same depth. One of the most common design errors is applying the same interview structure to every topic regardless of what you're trying to learn from it.

Quick checks. Confidence ratings, simple reactions, first impressions. These need one question and one follow-up. If you're spending more time than that, you're over-investing in a topic that doesn't require depth.

Problem and solution exploration. Understanding friction, unmet needs, or the gap between what users expect and what they experience. These need 3 to 5 turns: establish the situation, surface the problem, understand what they've tried, get to the specific moment where it breaks down.

Behavioural deep dives. Understanding the motivations, mental models, and decision logic behind a behaviour. These need 6 to 10 turns: specific instance first, then the context, then the reasoning, then the underlying belief or value that explains the behaviour.

Narrative and journey mapping. Understanding a sequential experience from beginning to end. These need careful chronological structure: anchor the participant at the start of the story, move through it in sequence, and resist the urge to jump to the interesting parts before you've established the full context.

Concept and hypothesis validation. Testing a specific idea before committing to it. Present the stimulus neutrally, get the immediate reaction before any evaluation, then probe the reaction. Don't lead with the evaluation question. You'll contaminate the data.

When you're designing a study, assign a depth type to each topic explicitly. It forces a decision about what you actually need from each section and prevents the pattern of treating every topic as though it requires the same effort.

Write resolution criteria before you launch

This is the step most researchers skip, and it's the one that matters most for AI-conducted research.

Resolution criteria are the specific conditions that need to be met for a topic to be considered substantively covered. Before the study goes live, you should be able to answer: what does a completed, well-covered response to this topic look like?

Not abstractly. Specifically.

Weak resolution criteria: "The participant has discussed their experience with onboarding."

Strong resolution criteria: "The participant has described at least one specific moment where they felt uncertain, named what caused that uncertainty, and indicated what would have resolved it."

The strong version tells the interview exactly what it needs to learn. The weak version gives it no guidance at all. It's just an instruction to keep asking about onboarding until something vague has been said about it.

Strong resolution criteria do two things. First, they force you to think clearly about what evidence you actually need before you go into the field, which often surfaces gaps in the research question itself. Second, they give the interview engine a clear signal for when a topic is done and it's time to move forward.

Check for dead ends before you launch

A dead-end topic is one where the conversation has nowhere to go after the opening question. The participant answers, and the only options are to probe the same question again (diminishing returns) or to move on (thin coverage).

Dead ends usually happen in one of two ways.

The topic is a closed question dressed as an open one. "Did you feel confident during the process?" produces a yes or no. "Walk me through what happened when you got to the document upload step" is a topic with legs. It anchors on a specific moment and invites a narrative response.

The topic has no follow-up path. Every topic should have at least one natural follow-up for the most common responses. If the participant says the experience was fine, what does "fine" mean to them? If the participant says it was bad, what was the first thing they noticed? Before launching any study, read through your topics and ask: if the participant gives a two-sentence answer to the opening question, what do I ask next? If you can't answer that, the topic is a dead end.

How Sofi uses your study design

When you write a research brief in Fieldwork, Sofi generates a complete study structure from it: topics, objectives, depth settings, interview logic, and resolution criteria for each section. This is not a template. It's an interpretation of your brief, what you said you needed to learn translated into a study structure designed to surface that evidence.

Sofi also checks her own output before you see it. She flags potential structural problems: topics that could produce dead-end conversations, resolution criteria that are too vague to evaluate, depth settings that don't match what the topic requires. You get a design that's been stress-tested before it reaches participants.

But Sofi's output is a starting point, not a finished design. The researcher's job is to review the structure against their actual research goals before anything goes live. When reviewing the generated study, look for four things:

Topics that are too broad. If a topic could take 30 minutes to cover properly, it needs to be broken into two or three narrower topics.

Resolution criteria that are too abstract. "Participant understands the feature" is not a resolution criterion. "Participant has described a specific scenario in which they would use the feature and what they'd expect to happen" is one.

Missing depth calibration. Check that the depth assigned to each topic matches what you actually need from it. A quick reaction topic shouldn't be set up for a 10-turn deep dive.

Dead-end topics. Read each topic and ask: if the participant answers the opening question in one sentence, what does the interview do next? If the answer isn't obvious, the topic needs to be redesigned.

The review takes 10 to 15 minutes for most studies. A study that's well-designed before launch produces findings that are immediately useful. A study that launches vague produces findings that are interesting but inconclusive, and sends you back into the field.

What this looks like in practice

A product manager at an e-commerce company wants to understand why users on the Growth plan aren't using the bulk export feature, which was expected to be a primary driver of upgrade behaviour. She writes a brief: the research question is what's blocking users from running their first bulk export, and she wants to understand whether the barrier is awareness, confidence, or perceived value.

Sofi generates a study with four topics. The PM reviews it and notices that the third topic, on perceived value, has resolution criteria that are too abstract. She rewrites them to require the participant to name a specific use case where the export would save them time. The study goes live.

Eleven sessions complete over two days. The findings are clear: users know the feature exists. They've seen it. They're not starting it because they don't understand what format the output comes in, and they're not willing to run their first export on live data in case they do something wrong. Neither of those barriers showed up in the analytics. Both are fixable with documentation changes, not product changes.

A checklist before you launch

[ ] The research question is written in one or two specific sentences
[ ] Every topic in the study connects directly to that question
[ ] Each topic is narrow enough to be covered in 5 to 10 minutes
[ ] Each topic has a depth setting matched to what you need from it
[ ] Each topic has specific, observable resolution criteria
[ ] No topic is a dead end: there's a forward path for the most common responses
[ ] The topic sequence moves from specific and concrete to broad and interpretive
[ ] There are no more than six topics
[ ] Screening criteria match the participants you actually need

If any of these fail, fix the design before you launch. The field will not fix a bad design. It will just produce a lot of data that doesn't answer the question.

Frequently asked questions

How many topics should a qualitative research study have?

Most well-scoped studies have between three and six topics. Fewer than three usually indicates a quick check rather than a full exploratory study, which is fine but should be designed accordingly. More than six typically indicates the research question is too broad or the study is trying to do too much in a single round of fieldwork.

What's the difference between a research topic and a research objective?

An objective is what you need to learn: the question the study is designed to answer. A topic is the area of conversation designed to surface evidence for that objective. One objective might be served by two or three topics. Topics are the instrument; objectives are the goal.

How do I know when a topic is done during an interview?

A topic is done when the resolution criteria have been met: when you have the specific evidence you defined in advance. Without resolution criteria, you'll either move on too early because the participant said something about the topic, or stay too long because the conversation is interesting but not producing new evidence.

What should I do when my study produces findings I didn't expect?

Unexpected findings are usually the most valuable ones. If a pattern emerges that wasn't anticipated in the study design, update the design for the next wave of research: add a topic that probes the unexpected finding directly, or run a targeted follow-up study. Don't try to interpret an unexpected pattern from data that wasn't designed to surface it.

How do I write a good research brief for AI study generation?

Be specific about what you need to understand, not just the topic area. Include the context (what decision is this research supporting?), the participant profile (who do you need to talk to and why?), and the key unknown (what do you not know that this research needs to answer?). The more specific the brief, the more precisely calibrated the generated study structure will be.

What's the most common reason studies produce shallow findings?

Vague resolution criteria, almost every time. Researchers define topics well but don't specify what a substantively covered response actually looks like. The interview has no way to know when enough is enough, so it moves on too early or loops on the same topic without producing new evidence. Writing strong resolution criteria before launch is the single most impactful thing you can do for study quality.

Related on Fieldwork

Last updated: 2026-04-10