Before You Run an Affinity Mapping Session: A Worked Example From 8 User Interviews

    November 17, 2026
    8 min read
    By Emil Thornovik
    Share this article

    Mira holds a yellow note that says "I export to Excel because I don't trust the approval status", looks at the wall, and puts it next to two others about checking numbers twice. Nobody told her where it goes. That small decision, repeated 296 times by four people over an afternoon, is affinity mapping: grouping observations by what they mean, from the bottom up, until the groups start explaining something.

    The method goes back to the KJ method developed by the Japanese anthropologist Jiro Kawakita in the 1960s. Product teams now run it on Miro or FigJam boards instead of paper, and the hard part has not changed: it only works if the notes going onto the wall are good. What follows is one illustrative study, walked through from preparation to the six themes it produced.

    1. Prepare the study so the notes are worth clustering

    The example is an illustrative B2B SaaS study. A product team selling invoice-approval software ran 8 interviews, 45 minutes each, with finance managers at companies of 50 to 500 people. The research question was narrow: why do customers keep approving invoices outside the product, by email and spreadsheet, after paying for it?

    Before any note gets written, the team needs a few things in place. Skipping them is the most common reason a session ends with 40 vague groups and no decision.

    • One research question, written at the top of the board, so every cluster can be tested against it.
    • Transcripts or detailed notes for every interview, ideally within 48 hours, while memory still helps.
    • A rule for what counts as one note: one observation, quote or behaviour, never a summary of a whole answer.
    • A participant code on every note, such as P3, so you can trace a theme back to how many people raised it.
    • Three or four people from different roles: at least one designer, one product manager, one engineer or support lead.
    • A timebox: 2 to 3 hours for the clustering, with a break before naming themes.
    • A parking area on the board for notes that do not fit anywhere yet.

    The note rule matters more than anything else on that list. "P5 is frustrated with approvals" is an interpretation. "P5: 'I approve it in the app, then email my boss the PDF anyway'" is data. Only the second kind can be clustered honestly, because two people can look at it and agree on what it says.

    2. Write and cluster the notes in rounds

    The team wrote notes from transcripts the day before the session. Eight interviews produced 296 notes, an average of 37 per participant, which is a typical range for 45-minute sessions with a focused guide. Each note carried a participant code and, where useful, a timestamp to find the original quote.

    The session ran in a fixed sequence, and the order is where most of the value comes from:

    1. Spread all notes randomly on the board, mixed across participants, so no one sees interview order.
    2. In silence for 25 minutes, everyone moves notes next to similar ones; anyone can move any note, including someone else's placement.
    3. Allow notes to be duplicated if they genuinely belong in two places, but mark the copy.
    4. After the silent round, talk through only the clusters that people moved back and forth more than twice.
    5. Split any cluster with more than about 25 notes; it usually hides two different ideas.
    6. Write a label for each cluster as a sentence in the user's voice, such as "I don't trust the status I see", never a single noun like "Trust".
    7. Group related clusters into themes and name each theme as a finding that answers the research question.
    8. Count participants per theme and record which notes did not fit at all.

    Silence in round 2 sounds like a gimmick. It is the step that stops the most senior person in the room from setting the categories in the first five minutes. Talking after the first pass means arguments happen about specific notes, which is productive, instead of about abstract categories, which is not.

    The first pass produced 31 clusters. After discussion and splitting, the team merged them into 6 themes, with 22 notes left in the parking area. That leftover pile is worth reading again at the end; in this study, 5 of those 22 later formed a small seventh insight about mobile approvals.

    3. Turn clusters into themes with a strength score

    Six themes look tidy on a board. The question that follows is which ones matter, and here a little arithmetic helps keep the team honest. The team used two numbers per theme:

    theme_strength = participants_who_raised_it / total_participants
    note_share     = notes_in_theme / total_notes_clustered
    flag a theme as "loud" if note_share > 2 x theme_strength / number_of_themes
    

    Worked example. The theme "I don't trust the approval status I see in the app" contained 64 notes from 7 of the 8 participants. Theme strength is 7 / 8 = 0.875. Note share is 64 / 274 clustered notes, about 0.23. The theme "The approval rules don't match how our company signs off" contained 41 notes from 6 participants: strength 0.75, note share 0.15. A third theme, "Notifications arrive at the wrong time", had 38 notes but came from only 2 participants: strength 0.25, note share 0.14.

    The third theme is the trap. With 38 notes it looks almost as big as the second, yet nearly all of it came from two talkative finance managers at one type of company. The loud-theme flag catches it: 2 × 0.25 / 6 is about 0.083, and its note share of 0.14 is well above that. The team kept the theme but labelled it as a hypothesis for a follow-up, not a finding. The first two themes went into the roadmap discussion.

    One limit deserves its full weight. With 8 participants, strength numbers are rough indicators, not statistics. The difference between 6 and 7 people out of 8 means nothing on its own. The score is a guard against counting notes as if they were people; it is not a replacement for judgement about how severe each problem is.

    4. Check the map before you share it

    A map that looks finished can still mislead. Before presenting, the team ran through a short review, and two of its checks changed what they reported:

    • Every theme answers the original research question, or it moves to a separate "other findings" list.
    • Every theme label is a sentence a participant could have said, not a team opinion.
    • Each theme has at least 3 participants behind it, or it is marked as a hypothesis.
    • At least one quote per theme is copied word for word, with a participant code.
    • Clusters that contradict each other are shown side by side, not merged away.
    • The parking area has been read once more and either sorted or explicitly discarded.
    • The board shows which roles were in the room, so readers know whose eyes did the grouping.

    The contradiction check turned out to be the most useful. Two participants loved the approval rules engine and wanted more of it; three found it too rigid. Merging them into "rules engine feedback" would have hidden the real finding, which was that company size predicted the reaction: the two fans worked at the largest companies in the sample.

    That finding would never have survived a merged cluster, and it changed the next step. Instead of redesigning the rules engine for everyone, the team planned 4 more interviews with companies above 300 employees to test whether size, or the number of approval layers, was the real driver. Presenting the map with that open question visible took 10 minutes longer and kept the team from redesigning a feature its largest customers liked.

    Questions teams ask about affinity mapping

    How is affinity mapping different from thematic analysis? Thematic analysis is a broader qualitative method with defined phases and often a codebook, used heavily in academic research. Affinity mapping is a faster, visual, team-based way to reach similar themes. Many research teams use affinity mapping for a first pass and a codebook when the study needs to be repeated or audited.

    Can one researcher do affinity mapping alone? Yes, and many do, but the method was designed for groups, and a solo map tends to confirm what its author already believed. If you work alone, ask one colleague to review the clusters and challenge two or three of them before you present.

    Should quantitative data go on the map? Keep survey numbers and analytics off the notes themselves; they bias the clustering toward whatever is countable. Bring them in after the themes exist, as evidence for or against each one.

    Run an affinity mapping session when you have 5 or more qualitative interviews on one clear question and a few hours with a mixed team. For 2 or 3 interviews, a simple debrief document is faster. For 30 or more, tag notes in a research repository such as Dovetail first, and map only the tagged highlights.

    Related Articles