TL;DR: Generative Research
- Generative research happens before you build – it uncovers user needs, mental models, and problems worth solving, rather than testing whether a solution works.
- It’s mostly qualitative: in-depth interviews, ethnographic observation, diary studies, and contextual inquiry are the most common methods.
- The biggest mistake teams make is skipping generative research and going straight to usability testing – then wondering why their product doesn’t resonate.
- Remote generative research works well, but async methods (diary studies, video journals) often surface richer, more candid data than live video calls.
- AI-assisted research platforms have made exploratory user research more accessible for smaller teams – running structured interviews without the usual recruitment bottleneck.
What Is Generative Research and Why It Matters in UX
Generative research is the practice of exploring a problem space to understand people – their lives, behaviors, motivations, and frustrations – before you’ve committed to building anything specific.
It’s not about validating a prototype. It’s not testing whether button A or button B converts better. It’s the earlier, messier, more important work of figuring out what to build in the first place.
The name comes from what it produces: it generates insights, hypotheses, and directions. Which is different from evaluative research, which tests specific things you’ve already made.
In UX and product development circles, you’ll also hear it called discovery research or exploratory user research – they’re largely the same thing, though there are small distinctions worth noting (more on that in the FAQ).
Why so many teams skip it
Generative research takes time. It’s open-ended by design, which makes it feel uncomfortable for teams used to shipping against tight deadlines. There’s no clear output like a conversion rate or a usability score. You’re sitting with ambiguity for a while – and most product teams hate that.
So they skip straight to usability testing, or worse, they run a quick survey and call it “research.”
Teams who skip generative work tend to build solutions that are technically functional but miss the actual problem. The product works, users just don’t care about it.
That’s an expensive mistake to find out about post-launch.
Generative vs Evaluative Research: What’s the Difference
Generative research is often discussed alongside evaluative research – but the two serve different purposes and belong at different stages of the product cycle. Generative work explores problems and possibilities; evaluative work tests solutions against real users.
For a full comparison of both approaches – when to use each, how to sequence them, and common mistakes when mixing them – see our guide to generative vs. evaluative research.
How to Conduct Generative Research Step by Step
There’s no single right process here, but most successful generative research follows a recognizable shape.
Step 1: Define what you’re trying to learn (not what you want to prove)
This is where most teams go wrong. They come in with hypotheses dressed up as research questions: “We want to understand why users love our onboarding flow” when they actually mean “we want confirmation that our onboarding flow is good.”
Generative research requires genuine curiosity. You need to be willing to find out that your assumptions are wrong.
Write down your learning goals – not your business goals. “We want to understand how [user type] currently handles [task] and what frustrates them about it” is a learning goal. “We want to prove our product fits this market” is not.
Step 2: Choose your method based on what you’re exploring
More on individual methods in the next section, but the short version: the method should follow the question, not the other way around.
Exploring how people feel about something over time? Diary study. Want to see behavior in its natural context? Ethnographic observation. Trying to surface mental models quickly across many people? In-depth interviews, probably with a thematic analysis approach baked in from the start.
Step 3: Recruit the right participants
This is the friction point most teams dread – and for good reason. Recruiting takes time. You’re looking for people who actually represent your target user, not just people who are available and willing.
A few things that help:
- Write a screener survey with behavioral questions, not demographic ones (“How often do you [specific behavior]” beats “What is your age”)
- Over-recruit slightly – plan for 20% no-shows
- 5 participants per distinct user segment is a reasonable floor for qualitative work, though more gives you more confidence in themes
- Aim for variety within your target segment, not homogeneity
Recruitment platforms like User Interviews can help with finding participants, though they add time and cost. Some teams use their own customer base. Others use professional panels.
Step 4: Design your discussion guide (not a script)
A discussion guide for generative research is not a questionnaire. It’s a loose framework that keeps you oriented without turning the conversation into an interrogation.
Structure it around themes, not questions. Within each theme, have 2–3 opening questions you can use to start, then let the conversation lead you.
Build in silence. When someone finishes answering, don’t rush to the next question. Silence prompts elaboration, and elaboration is where the interesting stuff lives.
Step 5: Run the sessions
A few things that trip up less experienced researchers:
Don’t ask leading questions. “How frustrating is it when X happens?” is leading. “What happens when X occurs?” is not.
Follow the participant, not your guide. If they go somewhere unexpected and interesting, follow them. You can always return to your guide.
Take sparse notes during the session. Full note-taking pulls your attention from listening. Jot keywords only, then synthesize afterward.
Record with permission. Transcripts are far easier to work with than memory.
Step 6: Analyze and synthesize
This is where most guides go vague. “Identify themes” is not enough direction.
A practical approach: affinity mapping. After your sessions, pull all your notes and observations into individual items (one insight per sticky note, physical or digital), then group similar items until patterns emerge. Don’t force categories – let them form naturally.
Tools like Dovetail or Miro work well for this. So does a physical wall if you have the space.
The output of synthesis isn’t a report – it’s a set of insight statements that your team can act on. “Users feel anxious about X because they don’t know Y” is an insight statement. “Users mentioned X 14 times” is a data point.
Best Generative Research Methods to Discover User Needs
In-Depth Interviews
The workhorse of generative UX research. One-on-one conversations, 45–90 minutes, designed to surface how people think and behave rather than what they say they want.
In-depth interviews work because they create space for stories. People reveal more in narrative than in answers to direct questions. A good interviewer asks “tell me about a time when…” and then mostly listens.
The limitation: what people say they do and what they actually do often diverge. Interviews capture perception, not always behavior.
Ethnographic Observation (Contextual Inquiry)
Watching people in their actual environment – at their desk, in their kitchen, at the store – rather than asking them to recall or describe their behavior.
Contextual inquiry combines observation with light interviewing: you watch, then ask about what you just saw. “I noticed you opened a spreadsheet just then – can you walk me through why?”
This method surfaces behaviors people don’t even think to mention because they’re automatic. It’s slower to run and analyze, but the quality of insight tends to be higher.
Diary Studies
Participants document their own experiences over a period of time – days, weeks, sometimes longer – through photos, voice notes, video clips, or structured prompts.
Diary studies are especially good for capturing behavior that’s distributed over time (not something you can observe in a single session) or behavior that’s private or context-dependent.
The downside: participant dropout. The longer the study, the harder it is to keep people engaged. Incentive structures and lightweight submission methods (WhatsApp-style voice notes, for example) help.
Card Sorting
With card sorting, participants organize topics or concepts into groups that make sense to them. Used to understand how people mentally categorize information – especially useful for information architecture decisions.
Open card sorts (participants create their own categories) work well for generative phases. Closed card sorts (you give them the categories) are more evaluative.
Focus Groups
Often overused and misused. Focus groups introduce social dynamics that can suppress minority opinions – people tend to converge toward the loudest voice in the room. They’re also poor at capturing individual mental models.
That said, they can be useful for generative work if you’re exploring social or cultural dimensions of a problem, where group dynamics are part of what you’re trying to understand.
Use them deliberately, not as a cheaper alternative to individual interviews.
Fly-on-the-Wall Observation
Pure observation with no interaction – you watch, you don’t ask. Works well in settings where your presence asking questions would change behavior (customer service counters, waiting rooms, etc.).
Real Examples of Generative Research in Product Development
IDEO’s hospital cart redesign
One of the more cited examples from the product design world: IDEO researchers spent time shadowing nurses in hospitals to understand what they actually needed from a medication cart. They didn’t start with a design brief – they started by watching how nurses moved, what they grabbed for, where they stored things, and what created friction. The resulting cart design came from those observations, not from a survey about what features nurses wanted.
That’s generative research in its clearest form: understanding behavior in context before committing to a direction.
Slack’s pre-launch discovery
Before Slack went public, the team ran extensive interviews with potential users about how they communicated at work – what tools they used, where information got lost, what was frustrating about existing systems. They weren’t testing Slack. They were understanding the problem. The product direction that emerged came from those conversations.
Spotify’s discovery research for podcast features
Spotify has published research on how they used generative methods – particularly diary studies and in-context interviews – to understand how and why people listened to podcasts. The insight that listening was often tied to specific activities (commuting, cooking, exercising) shaped how they designed discovery and recommendations.
When Generative Research Is the Wrong Call
Most guides don’t say this out loud: sometimes you don’t need generative research.
If your product is already live and you’re trying to understand why conversion dropped, that’s an evaluative problem, not a generative one. Analytics, session recordings, and targeted usability tests will answer it faster.
If you’re working in a domain you know extremely well – and your team includes people who’ve spent years as the user you’re designing for – you probably have enough tacit knowledge to skip some of the early generative work and move faster.
If your timeline is genuinely 48 hours and you need to ship a fix, a quick guerrilla test beats a two-week discovery sprint.
Generative research is most valuable when you’re entering a new market, designing for a user type you don’t know well, or when your existing product isn’t landing and you’re not sure why.
How AI Is Changing Generative Research (Without Replacing It)
This is worth acknowledging because the landscape has shifted noticeably in the last two years.
Traditionally, generative research was slow because recruitment was slow. Finding the right participants, scheduling sessions, managing no-shows – that process often took longer than the research itself.
AI-assisted platforms have started to change this. Tools that generate synthetic personas and conduct structured interviews at scale – like Articos – have made a certain type of exploratory research accessible to teams that previously couldn’t afford the time or budget for it. Rather than waiting weeks to schedule participants, teams can run structured discovery interviews in under an hour.
This doesn’t replace contextual inquiry or deep ethnographic work. It can’t observe behavior in context or surface the kind of unexpected, ambient detail that comes from spending time with real people in their environment. But for early hypothesis generation – understanding how a specific user type thinks about a problem before investing in a full research programme – AI-assisted approaches have become a practical starting point.
The broader shift is one toward user research without recruitment barriers, which matters especially for agencies and consultants who need to show research-backed thinking on projects that don’t have the budget for a traditional research engagement.
Start a free generative research study with Articos →
A Ready-to-Use Question Bank for Generative Research Interviews
Most guides give you principles for asking questions. This is a concrete list you can adapt directly.
Opening (to establish context and rapport)
- Tell me a bit about what your role involves day-to-day.
- Walk me through what a typical [day / week / project] looks like for you.
- What does success look like in your role?
Exploring behavior and habits
- Can you describe the last time you had to [task related to your topic]?
- Take me through how you typically handle [process]. Start from the beginning.
- What do you do first? Then what?
Probing for friction and workarounds
- What’s the most annoying part of that process?
- Have you ever had to work around something that wasn’t quite right for what you needed?
- Is there something you wish you could do differently?
Surfacing motivations
- Why does [thing they mentioned] matter to you?
- What would happen if you couldn’t do it that way?
- Who else is affected when [thing goes wrong]?
Closing
- Is there anything I haven’t asked about that you think is important?
- If you could change one thing about how you currently handle [topic], what would it be?
- Is there someone else I should probably talk to about this?
FAQs: Generative Research
Mostly, yes – the terms are often used interchangeably. The subtle distinction some practitioners make: “discovery research” tends to imply a project-specific phase (the discovery phase of a product sprint), while “generative research” is a broader methodology that can happen at any stage when you need to understand users from scratch. If someone says discovery research, assume generative. If they want something more specific, they’ll tell you.
Open-ended, behavioral, and retrospective questions work best. Ask people to narrate past experiences (“tell me about the last time you…”) rather than to predict future behavior (“would you ever…?”). Future-looking questions produce unreliable answers – people describe what they think they’d do, which rarely matches what they actually do. See the question bank above for specific examples you can use directly.
Yes – and for some methods, remote actually works better. Diary studies benefit from remote setups because participants can capture moments in context as they happen (a voice note on a commute, a photo in a store) rather than trying to reconstruct them later in a lab. Video interviews lose some of the ambient context of in-person sessions, but for most generative work the tradeoff is worth it for the scheduling flexibility. Contextual inquiry is harder to do remotely – though screen-sharing sessions offer a partial substitute for digital workflows.
Write a screener that focuses on behavior, not demographics. “Do you regularly [specific behavior]?” filters for the right people more reliably than age or job title. Recruitment platforms like User Interviews or Respondent give you access to pre-vetted panels. Your own customers are a legitimate starting point if you’re researching existing users – just be aware that customers skew toward people who already like you. Always over-recruit by 20–25% to account for no-shows.
Qualitative, almost always. The point of generative research is depth and nuance – understanding why and how, not how many. Small sample sizes (typically 5–15 participants per user segment) are standard. If you’re collecting survey data at scale during a generative phase, it’s usually to orient the qualitative work – understanding the landscape before you go deep – not to replace it. Quantitative data tells you what is happening; generative research tells you why.
