The eighth interview wraps up on Thursday afternoon. Somewhere in eight hours of recordings, one participant said she downloads every large request “to check it properly,” and another described the same detour without ever naming it. User interview analysis is the work of recognizing those two moments as one finding, counting how many participants share it and deciding what it means for the product.
Done well, the analysis gives your team a short list of findings with the evidence attached. Below, we walk through the method, follow one quote from the transcript to insight, and show how to tell when you have interviewed enough people.
TL;DR
- Write your research questions down before coding so every code ties back to a decision.
- Code what each segment is about first, then add your interpretation once patterns repeat.
- Report findings as counts (“5 of 8 participants”) with the limit stated alongside.
- Use AI to draft codes and summaries, then check every claim against the cited moment in the session.
Make Four Decisions Before User Interview Analysis Starts
A few choices made before you open the first transcript save hours later. They also make the analysis easier to defend, because anyone reviewing it can see what you were looking for and how you worked.
| Decision | Options | How to choose |
| What the analysis must answer | The study’s research questions, written as questions | Keep them visible while you code; record unexpected findings separately for review. |
| When to analyze | After each session, in batches, or once at the end | Debrief within an hour of each session, then code in batches of three or four |
| How codes are created | Deductive (start from a list tied to your questions) or inductive (let codes come from the data) | Use deductive codes for known topics and stay open to inductive codes for anything unexpected |
| Who codes | One researcher, or two who compare | Use two coders for high-stakes studies and agree on code definitions first |
Transcript quality matters as much as these decisions. If you ran the user interviews as moderated live sessions on Userlytics, each recording can be transcribed with timestamps and split by speaker, and you can correct misheard product names inline before anyone starts coding.
Run Your User Interview Analysis in Six Steps
The steps below follow the thematic analysis approach that Nielsen Norman Group describes for qualitative user research, adapted for a typical round of 6 to 12 interviews. The order matters: each step narrows the data so the next one has less to sort.
1. Read Every Transcript Before You Code
Read each transcript from start to finish without labeling anything, even the sessions you moderated yourself. Memory favors the vivid sessions, and a first read puts the quieter participants back on equal footing.
Keep a short list of surprises as you go. Those notes often become the first candidate codes, and they show you where your expectations were wrong before you start sorting.
2. Label Each Segment With a Descriptive Code
Go back through each transcript and tag every relevant segment with a short label naming its topic, such as “approval steps” or “budget check.” NN/g compares a code to a keyword in an article: shorthand that lets you pull every related segment together later.
Keep a codebook as you work. Give each code a one-line definition and an example quote, so that a second coder, or you three weeks later, applies it the same way.
3. Add Interpretive Codes Where Meaning Repeats
Once the same descriptive codes keep appearing, add a second layer that records what you think the segment means. “Budget check” describes the topic; “doesn’t trust the in-tool total” is your reading of why the participant does it.
Mark whether each segment reports behavior or shows it. A participant describing a workaround and a participant performing it on screen are different kinds of evidence, and you will weigh them differently in step 5.
4. Cluster Codes Into Candidate Themes
Move the coded segments into groups that share a meaning, then name each group with a sentence that states the pattern. A theme such as “Approvers verify over-budget requests outside the tool” says more than a label such as “Trust.”
Do this step with your team where you can. Affinity mapping works well here: according to NN/g’s guide to affinity diagramming, a small cluster still deserves attention, and a note that fits two clusters can be duplicated rather than forced into one.
5. Test Every Theme Against the Full Data Set
Treat each theme as a hypothesis and go back to the transcripts to test it. Count how many participants contributed to it, and search for segments that contradict it.
Review any participant whose sessions do not fit the pattern before you set them aside. An outlier may describe a different role, a different workflow, or an early signal the rest of the group has not hit yet. Exclude a session only when the recording clearly shows the participant was not engaging with the questions, and record why.
6. Write Insight Statements With the Evidence Attached
Turn each surviving theme into a statement with four parts: what participants do or believe, how many of them, what it means for the product, and what the evidence cannot tell you. Link each statement to two or three timestamped quotes.
A theme describes the data. An insight tells a product team what to change or what to find out next, and that is the version stakeholders will act on.
Follow One Quote Through a Complete Interview Analysis Example
The interview analysis example below is illustrative, a model study written to show the method, and it reports no real results. Picture a team interviewing eight procurement managers about how they approve purchase requests in an internal tool.
One manager says: “When a request comes in over budget, I don’t approve it in the tool. I download it, check it against our spreadsheet, then come back and click approve.”
| Stage | What the researcher records |
| Raw segment | The quote above, with participant ID and timestamp |
| Descriptive code | Approval outside the tool; spreadsheet cross-check |
| Interpretive code | Does not trust the budget figure shown at approval |
| Evidence type | Reported behavior (described in words, with no on-screen action) |
| Candidate theme | Approvers verify over-budget requests outside the tool |
| Theme test | 5 of 8 participants describe a version of the check; 2 say they approve in the tool; 1 never handles over-budget requests |
| Insight statement | Five of eight procurement managers leave the tool to verify over-budget requests because the approval screen does not show the remaining budget. Showing remaining budget at approval is the first change to test. The interviews tell us the workaround exists; a usability test will show whether the new screen removes it. |
Notice what the table preserves. Anyone who doubts the insight can follow it back to the five participants, the exact quotes, and the two managers who approve in the tool are counted openly instead of disappearing from the readout.
Short on time for analysis? Userlytics’ UX Consultants can plan the study, moderate the sessions, annotate the recordings, and deliver a findings report with recommendations. Explore UX Consulting
Match the Interview Analysis Method to the Study
Thematic analysis is the backbone, but the way you sort and display the data can change with the size of the study and the question it answers. NN/g’s thematic analysis guide describes three common ways to work: analysis software, journaling, and affinity diagramming. A fourth format, the participant-by-theme matrix, helps most when you need to count and compare.
| Interview analysis method | How it works | Best fit |
| Coding in analysis software | Transcripts are tagged in a tool that logs every code and pulls coded segments together on demand | Larger studies, several coders, findings you will revisit later |
| Journaling (memos) | You highlight the transcripts and write notes on your reasoning as you go | Solo researchers who need a record of how they reached each theme |
| Affinity diagramming | Quotes and observations go on physical or digital sticky notes that the team clusters together | Building shared understanding in a workshop |
| Participant-by-theme matrix | A grid with participants as rows and themes as columns, each cell holding the supporting quote or a blank | Counting support for each theme and comparing segments |
Build a Participant-by-Theme Matrix
The matrix is the quickest way to see whether a theme is broad, narrow, or split by segment. Add a column for each participant attribute that matters to the study, such as role, team size, or region, and fill each cell with a short quote and timestamp.
| Participant | Role | Verifies outside the tool | Wants budget shown at approval | Delegates small approvals |
| P1 | Regional manager | Yes, 12:40 | Yes, 14:05 | No |
| P2 | Category buyer | No | Yes, 21:10 | Yes, 08:30 |
| P3 | Regional manager | Yes, 06:15 | No mention | No |
Read the columns for breadth and the rows for each person’s full story. If every “Yes” in a column comes from the same role, report the finding for that role only. The rows above continue the illustrative procurement study and report no real results.
Check Whether Your User Interviews Have Reached Saturation
Saturation is the point where new interviews add little or no new information for your research questions. Researchers often claim it without showing it, so a simple, reportable check strengthens any readout.
Researchers Guest, Namey, and Chen proposed one in a 2020 paper in PLOS ONE. You count the unique codes in your first few interviews (the base size, usually four to six), then measure how many new codes each following run of two or three interviews adds as a share of that base. Once a run falls at or below a threshold you set in advance, such as 5% new information, you have evidence of saturation.
Here is the check applied to the illustrative procurement study, with a base size of four and runs of two:
| Run | New codes in the run | New information (new codes divided by 30 base codes) |
| Interviews 1 to 4 (base) | 30 unique codes | Base set |
| Interviews 5 and 6 | 4 | 13% |
| Interviews 6 and 7 | 2 | 7% |
| Interviews 7 and 8 | 1 | 3%, at or below the 5% threshold |
Following the paper’s notation, you would report saturation at interview 6, written “6+2” to show that two more interviews confirmed the drop. The same paper summarizes earlier work by Guest and colleagues in which 70% of 114 themes appeared within the first six of 60 interviews, and 92% within the first 12.
Two limits apply. Common themes surface early, so a pattern shared by a small subgroup can still be missing at saturation. Multi-market studies also need more: a cross-cultural study cited in the same paper found fewer than 16 interviews per site were enough at each of four sites, while 20 to 40 were needed to identify themes shared across sites.
Separate What Participants Said From What They Did
User research interviews collect what study participants report about their behavior, thoughts, and feelings. As NN/g’s introduction to user interviews points out, self-reported data has limits, and finding out what users actually do requires observing them or collecting behavioral data.
Run a quick pass over your themes before you report them and sort each one into a category:
- Beliefs and attitudes: what participants value, fear, or expect. Interviews are a strong source for these.
- Reported behavior: what participants say they do. Treat these as leads to confirm.
- Predictions: what participants say they would do with a new feature. Weight these lightly.
- Observed behavior: what participants did on screen during the session. These carry the most weight.
For any theme that rests on reported behavior, plan a usability testing round with realistic tasks. In the procurement example, five managers described leaving the tool; a task-based session on a prototype with remaining budget shown would tell the team whether they still do.
Use AI for Interview Transcript Analysis and Keep the Judgment
AI can take on the time-consuming mechanical parts of interview transcript analysis: transcribing, first-pass tagging, searching across sessions, and drafting a summary. The judgment calls stay with the researcher, including which codes matter, whether a theme holds, and what the team should do next.
Here is how that split works on the Userlytics platform:
| Analysis task | What the platform does | What you still check |
| Prepare transcripts | Transcriptions are timestamped, split by speaker, searchable, and editable inline | Product names, acronyms, and misheard terms before coding |
| First-pass coding | Userlytics’ AI Annotations feature flags key moments with timestamps, summaries and tags, labeled “AI-Generated” and fully editable | Rename, retag, or delete annotations that do not match your codebook |
| Apply your own codes | Custom, color-coded tags set at the project level work across every session | One definition per tag, shared with every coder |
| Ask questions across sessions | The AI Insights feature answers questions about the whole study and cites the session and timestamp behind each observation | Open the cited moment and confirm the participant said what the summary says |
| Analyze other languages | AI Transcript Translations translates transcripts into 15 supported languages, synced to the video | Reset to the original text before quoting, since the Help Center notes translations may contain minor inaccuracies |
| Draft the readout | Qual Report builds an Executive Summary, Key Findings, Strategic Areas and Recommendations once every session is complete and processed | Compare its findings with your tested themes and decide which to keep |
A practical rule: no AI-drafted finding goes into a readout until a researcher has opened at least one cited moment for it. The citations make that check fast, which is why they matter more than the summary itself.
Report User Interview Analysis Findings Stakeholders Will Act On
Stakeholders rarely read a full analysis. They read the first page, watch a clip or two, and decide, so a consistent format for each finding does a lot of the work.
Write each finding in the same shape so readers can scan them quickly:
Finding: [What participants do or believe], reported by [X of Y participants] ([segment, if the pattern is limited to one]). Why it matters: [The product decision it affects]. Evidence: [Two or three timestamped quotes or clips]. Limit: [What the interviews cannot show, and the method that would].
Then structure the readout itself around the decisions it should change:
- Lead with the decision. Open with the two or three changes the findings support, then show the evidence for each.
- Give counts instead of percentages. “5 of 8 participants” tells the reader both the pattern and the sample size; “63%” from eight people implies precision the study does not have.
- Let participants make the case. One short clip of a participant describing the workaround does more than a paragraph of summary. Collect the strongest moments into highlight reels organized by theme.
- State the limits up front. Say who you interviewed and who you did not. When we published our study with Sage on the AI-assisted buyer journey, the report stated its sample at the start: 60 prospective buyers split evenly across the United States, the United Kingdom and France (20 per market).
- Close with the next question. Name what the interviews could not answer and the method that will, such as a usability test for reported behavior or a survey to size a pattern.
If your interviews are one input among several, the broader guide to UX analysis covers how qualitative findings sit alongside task success rates, time on task, and other behavioral metrics.
Avoid These Five Interview Analysis Mistakes
Most weak readouts trace back to a handful of shortcuts. Check your draft against this list before it goes to stakeholders.
| Mistake | How it shows up | Fix |
| Analyzing from memory | Findings come from the sessions the team watched live | Read every transcript before coding (step 1) |
| Percentages from small samples | “75% of participants want an export” from eight interviews | Report “6 of 8 participants” and treat the pattern as direction |
| Making a theme look unanimous | Participants who disagree drop out of the readout | Count dissent openly and review outlier sessions before setting any aside |
| Treating predictions as behavior | “Participants would use this feature” becomes a roadmap item | Tag it as a prediction and test it with realistic tasks |
| Passing on AI summaries unchecked | A finding has no quote or timestamp behind it | Open the cited moment and confirm the verbatim before reporting |
Your Next Step After the Final Interview
Before your next round of interviews closes, set up the analysis first. Put the research questions at the head of a shared codebook, agree on code definitions with anyone else who will code, and decide your saturation threshold in advance.
Then run user interview analysis on the first four transcripts as your base set. You will know within the next two or three sessions whether new themes are still appearing, and your readout will start with evidence already attached.
Run your next round of interviews with transcripts, AI Annotations, and Qual Report in one place. Book a free demo
FAQs
Research teams tend to raise these questions once user interview analysis is underway.