User research methods are the ways a product team finds out what people need and how they actually use a product. The main ones are interviews, field observation, diary studies, usability tests, card sorting and tree testing, surveys, analytics and A/B tests. Pick one by the question you need answered: talk to people to learn why, watch them to learn what they really do, and count when you need to know how many.
Few teams need all of them. A round of interviews before design, a few small usability tests during the build and decent analytics after launch is more research than most products get.
Three ways to tell the methods apart
Nielsen Norman Group's overview of 20 UX research methods sorts them along three axes. The first is attitudinal versus behavioral: what people say versus what they do, and the two disagree more often than you'd expect. Someone who tells you they always compare prices may still click the first result. The second is qualitative versus quantitative. Interviews and moderated tests explain why something happens and how to fix it; surveys, analytics and A/B tests answer how many and how much. The third is context: is the participant using the real product in their own setting, a task you scripted, a prototype, or nothing at all?
The same article maps methods to stages. Generative methods (interviews, field studies, diary studies) help you decide what to build, formative ones (usability tests, card sorting, tree testing) improve a design while you make it, and summative ones (analytics, A/B tests, benchmarks, surveys) measure what you shipped. NN/g's table lists surveys in both the first and the last group: early on to check how common a problem from your interviews is, after launch to track how the product is received.
Which method answers which question
Method | What it tells you | When to use it | Participants |
|---|---|---|---|
User interviews | Why people do what they do, what they use today, what gets in their way | Before design, and whenever you target a new group | Start with 5 or 6, add more until you stop hearing new things |
Field study or contextual inquiry | How the work is really done, including workarounds | Before designing for work you don't do yourself | A handful of people or sites |
Diary study | Behavior that unfolds over days or weeks | Habits, repeated tasks, long processes | Recruit extra, people drop out |
Moderated usability test | Where people get stuck in a flow, and why | From the first clickable prototype onward | 5 per round, several rounds |
Quantitative usability test | Success rates and task times you can compare | Benchmarking before and after a redesign | About 40 |
Card sorting | How people group and name your content | Designing navigation | About 15 |
Tree testing | Whether people can find things in a proposed menu | After card sorting, before visual design | A few for insight, more for numbers |
Survey | How common an attitude or behavior is | After interviews tell you what to ask | About 385 for a ±5 point margin |
Analytics and session recordings | What people do, at scale | Once the product is live | Everyone who uses it |
A/B test | Which of two versions does better on one metric | Live product with steady traffic | Often tens of thousands per variant |
Focus groups are missing on purpose. They show you what people say in front of each other, which can help with marketing language. They are a weak way to find usability problems or to predict what anyone will do.
Interviews: where most teams should start
An interview is a one-to-one conversation, usually 45 to 60 minutes, with someone who has the problem your product is meant to solve. Its job is to understand that problem and how people deal with it today. It is a bad tool for getting your idea approved, because people are polite and will say they like it.
The most useful rule for interviews is to ask about specific things that already happened. "Tell me about the last time you…" gets you a story with details you can check. "Would you use a feature that…?" and "How much would you pay?" get you guesses, and people are poor at predicting their own behavior. Rob Fitzpatrick's book The Mom Test is built around this rule and is worth an evening.
Use a semi-structured guide: five to eight topics with a few open questions each, and permission to follow the participant somewhere you didn't plan. Run each session with two people, one asking and one taking notes, and record it if the participant agrees. Here is the core of a guide for, say, a tool that helps clinics build staff schedules:
- Walk me through how this month's schedule got made. Where did you start?
- What did you use? A spreadsheet, paper, a system, messages to the team?
- What was the hardest part last time? What happened?
- When something changed at the last minute, what did you do?
- Who else is involved, and who checks or approves the schedule?
- Have you tried anything to make this easier? Why did you stop using it?
- Is there anything I should have asked and didn't?
Recruit people who have done the task recently: existing customers, people who contacted support, your sign-up list, LinkedIn, professional communities. In business software the buyer is often not the daily user, so ask your contact to introduce you to the people who do the work. Pay for their time if you can, and use a three- or four-question screener so you don't spend an hour with the wrong person.
How many interviews? NN/g's Maria Rosala argues there is no golden number: start with a small representative sample of 5 or 6, analyze as you go, and stop when new sessions stop adding new themes. A narrow question with a similar group of people can get there quickly. A broad study across a diverse population may need 20 to 30. The famous five-user rule comes from usability testing and doesn't tell you how many interviews to run.
The usual mistakes: pitching your solution halfway through, asking leading questions ("Wouldn't it be easier if…?"), treating a compliment as evidence, and letting the most articulate participant set the conclusions.
Watching the work: field studies and diary studies
People are poor narrators of their own routines. They skip the steps they no longer notice and the workarounds they're slightly embarrassed about. Contextual inquiry, from Hugh Beyer and Karen Holtzblatt's work on contextual design, fixes this by putting you next to the person while they do the real task. They act as the expert and you as the apprentice, asking about what you see as it happens. It pays off when the work is complex, spread across several tools or people, or done somewhere you can't picture from a desk, like a hospital ward or a warehouse floor.
Existing records are worth studying too. If the product will run on data people already keep (an HR export, the spreadsheet a team plans with, a year of support tickets), ask for a sample early, anonymized if necessary, and read it the way you'd read interview notes. Odd categories, duplicate entries and manual overrides point to rules the software will have to handle, and each one is a good question for your next session.
Diary studies handle what one session can't: behavior that happens over days or weeks, like a monthly close, the first month with a new app, or a task that comes up at unpredictable times. Participants log short entries when the moment happens, often a photo and two lines from their phone, over anything from a few days to a month or longer. Prompt them at the right moments, keep each entry small, and recruit more people than you need, because some will stop logging.
Usability testing: watching people try the design
A usability test gives a participant realistic tasks and watches them attempt each one while thinking aloud. Phrase tasks as goals ("You need to swap Thursday's shift with a colleague") and avoid naming the interface ("Click the Swap button"). Then stay quiet. The useful moments are the hesitations, the misreadings and the places where someone gives up. A clickable prototype is enough, and testing before any code is written is the cheapest time to find out a flow doesn't work.
Jakob Nielsen's case for testing with five users still holds for this kind of study. Across the projects he and Tom Landauer analyzed, a single user revealed about 31% of the problems, so five users find about 85%. The recommendation that follows is easy to miss: run three rounds of five and fix things between rounds, instead of one study with fifteen people. Two caveats come with it. If you have distinct user groups, test each one: three or four people per group when there are two groups, three each when there are more. And the rule covers finding problems. It says nothing about measuring them.
When you do need numbers, for example to show that a redesign improved task success, you need a much larger sample. NN/g's Raluca Budiu and Kate Moran recommend 40 participants for most quantitative studies, which gives roughly a 15% margin of error at 95% confidence on a success rate.
Remote moderated sessions over a video call work for most software. Unmoderated tools such as Maze, Lyssna or UserTesting are faster, but you can't ask a follow-up question, so keep them for well-defined tasks. Test mobile apps on the participant's own phone when you can.
Card sorting and tree testing: getting the structure right
If your product has a lot of content or settings, its navigation is a research question of its own. In a card sort, participants group topics written on cards into categories, either naming the groups themselves (open), sorting into groups you defined (closed), or a mix. In Nielsen's card-sorting analysis, results from 15 users correlated at 0.90 with those from a much larger sample, against 0.75 for five users, which is why 15 is the usual target. Nielsen himself warns against designing a menu purely from the similarity scores; use them as input.
A tree test checks the result. Participants see only a text version of the proposed hierarchy and try to find things in it ("Where would you change your notification settings?"). You learn which tasks fail and where people looked first. Optimal Workshop and Lyssna both run card sorts and tree tests.
Surveys, analytics and A/B tests: counting
A survey tells you how common something is among your users. It is a poor way to find out why, and an unreliable way to predict behavior. Write the questions after your interviews, so you know what's worth counting, and ask about recent facts ("In the last month, how many times did you…?") rather than intentions.
For sample size, the standard formula for a proportion is n = z² × p(1 − p) / e². With 95% confidence (z = 1.96), the worst case p = 0.5 and a margin of ±5 points, that comes to about 385 responses from a large population. The formula assumes a random sample. A link in your newsletter reaches your most engaged users, so the real uncertainty is larger than the math suggests.
Analytics shows what people do at scale: funnels, drop-off points, which features get used. In Google Analytics 4, Path exploration (under Explore) shows the routes people take. Session recordings fill in some of the why; Microsoft Clarity does this for free, though visitors from the EEA, the UK and Switzerland must give consent before it can show full replays. Plan your tracking events before launch, because adding them later means weeks of missing data.
A/B tests compare two versions on one metric, and they need traffic. With a 3% conversion rate, detecting a lift to 3.3% at the usual 95% confidence and 80% power takes about 53,000 visitors in each variant. Most young products don't have that. Without the traffic, test the change with five users, ship it and watch the metric, or use your first release as the experiment, which we wrote about in MVP development.
In tools for experts, feedback collected inside the product and tied to one specific output often tells you more than a survey. A thumbs-down on a single result shows which output was wrong; a survey a month later only shows that someone was unhappy. We built an exercise-program generator for a sports physiotherapy clinic, where the reasoning behind each program lives in the coaches' heads. Coaches rate each generated program with a thumbs up or down, swap exercises for alternatives and regenerate. In 2026 we're adding stored comparisons between the AI's program and the coach's, so the logic can be checked against past athletes.
A small study you can run in two weeks
- Day 1. Write down the decision the research has to inform ("Do we build shift swaps before leave requests?") and three to five questions. If no answer would change what you do, skip the study.
- Days 2 to 4. Recruit six to eight people who fit, with a short screener, expecting one or two no-shows. Book 45-minute sessions.
- Days 3 and 4. Write the interview guide or test tasks and pilot them on a colleague. The pilot always finds a confusing question.
- Days 5 to 8. Run the sessions with a note-taker. Spend ten minutes after each one writing down what surprised you, while it's fresh.
- Days 9 and 10. Put every observation on its own note, group them, and write the findings with the evidence behind each one.
- Share one page. What you learned, how sure you are, and what changes in the plan. Link the recordings if your consent form covers internal sharing.
When the next decision comes up, run another one. The screener, the consent form and the guide carry over, so later rounds take less setup.
Recordings, consent and privacy
Interviews and tests collect personal data: faces, voices, names, and sometimes screens full of other people's data. Under the GDPR and UK GDPR you must tell participants, at the time you collect their data, what it's for, how long you'll keep it and who it will be shared with. In practice:
- Use a short, plain consent form for the recording, and get it signed before you press record.
- Say who will see the recording: your team only, or your client as well.
- Set a retention period, delete raw recordings when it ends, and keep anonymized notes.
- Check where your video, transcription and AI tools store data, and whether the vendor can use it for training.
- If participants handle sensitive data (patients, finances), ask them to use test data or keep those screens out of the share. Health information is special category data and needs more care.
Can AI stand in for participants?
Not for findings. When NN/g compared synthetic users with real ones, a real learner had finished three of seven online courses while the synthetic user claimed to have completed every course it mentioned. Synthetic users praised discussion forums that real participants found contrived, and seemed to care about everything equally. The authors' advice is to use them for desk research, for generating hypotheses and for piloting an interview guide, and never to present their output as research with real users.
AI is more useful after the sessions: transcripts, a first pass at tagging, pulling quotes for a theme. Check the themes against the transcripts yourself, and check the privacy points above before you upload a recording anywhere.
From findings to decisions
A finding should have three parts: what you saw, the evidence, and what it means for the product. A made-up example: "Four of six bookkeepers export invoices to Excel to chase late payments, because the app has no overdue view. We should build that view before the reporting module." Notice that the finding names a decision. One that doesn't is trivia.
Rank findings by how often people hit the problem and how badly it blocks them, then take them into the backlog with your usual prioritization, which we cover in our post on product management frameworks. How research feeds wireframes and prototypes is in our product design process post. Keep the raw notes somewhere searchable. Six months later, someone will ask why a screen works the way it does.
Frequently asked questions
What is the difference between generative and evaluative research? Generative research helps you decide what to build. Evaluative research checks a design or a shipped product, so it covers both the formative studies you run while designing (usability tests, tree tests) and the summative ones after launch (benchmarks, A/B tests). It is a wider term than usability testing.
What if we can't get access to the people who use the product? This is common in B2B and healthcare, where a buyer or an IT department sits in between. Ask your contact for half an hour with two people who do the work, and tell them exactly what you'll ask. Meanwhile, use what you already have: support tickets, notes from sales and onboarding calls, and the product's own analytics. Treat all of it as leads to check in a real session, and don't fill the gap with synthetic users.
Can developers or product managers do user research? Yes, for interviews and usability tests, with a written guide and a couple of practice sessions. Bring in a specialist for large quantitative studies, for sensitive groups such as children or patients, or when a decision is expensive enough to justify it.
Is user research the same as market research? They overlap. Market research asks who will buy and at what price. User research asks what people need and how they use the product. A product needs both, and they often share interview participants.
If you have research findings and need a team to turn them into working software, write to us.
Related Articles
WMS vs TMS: What Each Does and Which One You Need First
A WMS runs what happens inside the warehouse; a TMS runs what happens after the truck leaves. Where they meet, which to buy first, and when you need neither.
Read More: WMS vs TMS: What Each Does and Which One You Need FirstSaaS Pricing Models: How to Choose One and Bill for It
The main SaaS pricing models, when each one fits, and the decision to make first: your value metric. Plus what your pricing asks of the billing code.
Read More: SaaS Pricing Models: How to Choose One and Bill for ItSaaS Metrics: What to Track, How to Calculate It, What Good Looks Like
The core SaaS metrics with formulas and worked examples, what recent benchmarks say is normal, and the definition mistakes that quietly distort MRR, NRR and LTV.
Read More: SaaS Metrics: What to Track, How to Calculate It, What Good Looks Like
