How to Synthesize 3-8 Product Discovery Interviews Without Overclaiming

    August 23, 2026
    8 min read
    By Arden Vaskerby
    Updated November 19, 2026
    Share this article

    Hear the same complaint in two interviews and it starts to feel like a law of the market. But with five conversations, four people agreeing tells you less than it seems. The real share of people who hold that view could be anywhere from about two in five to almost all of them. In that gap, teams build the wrong thing with full confidence.

    The risk is highest in B2B, because people at the same customer sit at different points of one process. A sales engineer feels the delays. A compliance analyst (GRC) worries about approving the wrong answer. An IT leader may find neither urgent. Treat them as one "user" and you average away the very differences that decide what to build.

    A synthesis is an argument a skeptic can audit

    A good summary of interviews is an argument that anyone can check. A skeptic should be able to see, line by line, who supported a claim, who added a condition, who disagreed, and how far the claim can honestly go. With three to eight interviews, the summary does not tell you how common something is. It sorts claims into assumptions for a prototype, a narrower test, or no build.

    Three formats can do this job, and each fails in its own way with so few interviews. A written memo explains a process well, but disagreement gets lost in smooth prose. An affinity map (sticky notes grouped by theme) shows repeated situations and shared words. But when only five people were interviewed, its sticky-note counts start to look like market size. An evidence matrix takes longer and turns into noise if claims are vague. Yet it is the only format that lets a team still see disagreement after the workshop ends.

    The choice follows the decision. If the synthesis only needs to brief a new designer on how a workflow runs, a memo is enough. If it will commit engineering capacity or change positioning, build the matrix and keep the memo as the cover note. A memo may call approval delays a priority; the matrix has to show which roles said so, why, and where the pattern breaks. Without it, the slide deck becomes the only surviving research record.

    Evidence cards stop notes from laundering into requirements

    A matrix is only as honest as the notes behind it. Turn notes into small evidence cards: one observation, one person, one moment in the conversation. "Needs better reporting" fails that test, because it mixes a problem, a solution and a guess in one line. A usable card preserves behavior: "I5 exports a status list before the weekly deal review because account owners cannot tell which security requests await approval."

    Each card gets a fixed ID. Then anyone can question a claim without rereading transcripts or relying on the researcher's memory. Every card has the same fields, and each field prevents one specific mistake:

    • Interview ID and role stop distinct jobs collapsing into one imagined user type.
    • Situation and trigger keep a rare event from being read as recurring work.
    • Observed behavior holds the line against a feature request standing in for what someone actually did.
    • Consequence records the lost deal, delay or cost that makes the behavior matter.
    • Interpretation keeps the researcher's own guess in a separate field.
    • Source pointer keeps the claim open to later review.

    Keeping what people did apart from what you think it means does most of the work. "The analyst copied an answer from a prior questionnaire" is evidence. "The team needs an AI answer generator" is a product hypothesis. The same line runs through the interview itself: what someone did last Tuesday, or already paid for, carries weight; what they say they would do with a product that does not exist yet carries very little. Cards built that way stop a passing comment from becoming a committed roadmap feature.

    Plan the time honestly. A rough guide: a 45-minute interview gives 10 to 25 usable cards, and turning one transcript into cards takes about an hour. So five interviews cost about four hours of talking plus five hours of card writing before any matrix exists. That is why you should only make cards for moments that matter for the next decision, not for every remark.

    Five B2B users, one matrix

    A filled-in example makes the trade-offs concrete. Consider an illustrative discovery set for a vendor-security response workspace: a tool that helps teams answer customer security questionnaires, reuse approved evidence and route sensitive responses for review. The five interviewees share a commercial workflow from different positions:

    • I1, sales engineer: loses deal momentum awaiting approved security answers.
    • I2, GRC analyst: verifies reused answers match current controls and policies.
    • I3, security operations lead: maintains evidence files and fears stale documents entering customer responses.
    • I4, IT director at a smaller firm: uses an existing review ticket queue and feels limited pain.
    • I5, revenue operations lead: compiles sales-leadership status updates when security requests block deals.

    "Everyone wants automation" would be the lazy summary. I1 wants speed, I2 control, I3 evidence freshness, I5 visibility, and I4 questions whether the category needs another system at all.

    Claim I1 I2 I3 I4 I5 Confidence label
    C1: Delays come from finding and approving reusable answers, not drafting every response Supports Supports Partial: stale evidence concern Contradicts: ticket flow is enough Supports Directional
    C2: An approved answer library is a credible first-release value proposition Supports Supports with review history Supports with expiry controls Qualifies: useful only above a request volume Supports Directional
    C3: Review status must be visible outside security Supports Partial Partial Contradicts Supports Tentative
    C4: CRM integration belongs in the first release Does not raise it Does not raise it Does not raise it Does not raise it Supports Tentative
    C5: Automatic answer generation should replace reviewer approval Rejects for sensitive answers Rejects Rejects No view Rejects Strong in sample

    The labels describe this set, not statistical results. "Strong in sample" means direct support from the relevant roles with no material disagreement. It does not mean every target account behaves this way. "Directional" means a plausible mechanism with a boundary or a contradiction. "Tentative" means one role raised it, or support is too thin for a confident roadmap choice.

    The labels change decisions. C2 can justify a prototype answer library with approval history and evidence expiry. C4 cannot justify an integration project: one revenue operations lead asking for CRM sync is a lead for the next round of interviews, not a release scope. C5 becomes a guardrail. Every generated draft enters a reviewer-controlled approval path before a customer sees it.

    The interviewee who disagrees draws the segment line

    Averaging I4 away would throw out the most useful row in the matrix. I4 shows where the product stops being useful: with few requests and a ticket queue that works, a separate workspace adds little. Hide that, and your marketing will attract buyers who never really start using the product.

    Record contradictions in their own field and question each one:

    1. Does the interviewee differ by role, company size, maturity or workflow ownership?
    2. Is the disagreement about the problem, its frequency, its consequence or the proposed solution?
    3. Does it lower confidence, or define a segment where the claim does not apply?
    4. What evidence would distinguish the competing explanations?

    More interviews help only when they target that boundary. In this example the next three conversations should go to firms with low questionnaire volume and mature ticketing. Suppose I4's firm handles about 10 questionnaires a quarter while I1's handles 60. If the next three low-volume firms all say a ticket queue is enough, the team has a working segment rule: the workspace earns its place somewhere above a few dozen questionnaires a quarter, with the exact threshold still to be tested. One contradiction turned into a sharper segmentation hypothesis.

    Five interviews can name a problem but cannot size it

    The Nielsen Norman Group draws the line cleanly. A handful of participants is fine for identifying what is broken, because a real problem one person hits is a real problem. The same handful says nothing reliable about how many people share it.

    The math is blunt. Say four of five people support a claim. The share of all customers who agree could be anywhere from about 38% to 96% (a 95% Wilson interval). With one of five, it could be anywhere from about 4% to 62%. Even five of five only tells you it is probably above 57%. None of these ranges can size a market or show which of two features more people want.

    Saturation research points the same way. In the widely cited study by Greg Guest, Arwen Bunce and Laura Johnson, How Many Interviews Are Enough? (Field Methods, 2006), analysis of sixty interviews showed thematic saturation arriving within the first twelve, with the basic elements of the main themes already visible after about six. That is good news for a team running three to eight conversations. Six interviews can surface the main patterns well enough to name a claim. It is also a warning. New themes stopped appearing at twelve in a group of very similar people. A B2B set with several roles and company sizes needs more interviews before nothing new comes up.

    A fair reader might conclude that small-sample interviews are nearly worthless: if five conversations cannot size anything, why not run a survey of 300 and skip the matrix? For "how many" questions, such as pricing tiers, feature ranking or segment size, that is the right call. For "why" and "how", the survey comes too early. A survey can only count answers to options the team already thought of. Interviews are how the team discovers that "faster answers" means speed to a sales engineer and control to a GRC analyst, before writing survey options that would blur the two. The usual sequence is interviews to name the claims, then a survey or a prototype metric to size them.

    Mention counts and merged requests are how overclaiming sneaks in

    "Five out of five mentioned X" often signals weak qualitative reasoning. The count may reflect the interviewer's prompt, the wording of the discussion guide, or a broad complaint that means something different to each role. Counts help auditability. They cannot replace context.

    Merging pain with solution is the second shortcut. "Users asked for a dashboard" can conceal three needs: sales wants a blocker list, GRC wants a view of review workload, and leadership wants revenue-risk visibility. One dashboard could serve all three poorly. Preserve the job and the consequence before grouping requests.

    Source drift is the third. A statement moves from slide to planning document to requirement and loses its interview ID along the way. Every material claim needs source IDs, a confidence label and a boundary. If it cannot be traced, it is an assumption, not evidence.

    When to stop interviewing and decide

    Keep the matrix small enough for a product trio to review in one sitting, usually five to eight claims. Include only claims that could change the next decision: target segment, problem priority, workflow sequence, value proposition, scope boundary or risk guardrail. A giant theme catalog documents research without helping anyone choose.

    Ownership is split on purpose. The researcher owns evidence fidelity and confidence labels; the product manager owns the decision. Design and engineering review workflow and feasibility assumptions. Commercial teams may challenge account fit, but must not rewrite evidence to fit a sales narrative.

    Three review passes catch most errors:

    1. Evidence pass: every cell points to a real card and keeps its meaning in paraphrase.
    2. Claim pass: compound claims are split, role differences exposed, contradictions marked.
    3. Decision pass: the team states what to build, test, defer or stop pursuing, and what evidence would reverse the call.

    A stopping rule keeps discovery from becoming a habit without an exit. If two consecutive interviews in a segment add no new card to any claim in the matrix, stop interviewing that segment and decide. If they keep adding contradictions, the segment definition is wrong, and more interviews of the same kind will not fix it.

    The decision log matters as much as the matrix. "Prototype C2; defer C4; test whether low request volume limits the value of a dedicated workspace before narrowing the target segment" is a decision. "Explore the findings further" is not.

    After prototype testing, add evidence rows rather than rewriting old conclusions. The team then sees whether behavior confirmed the interview pattern, exposed a faulty interpretation or revealed a segment the original five conversations missed. That is where the frequency question finally gets an answer, and the matrix records who was right.

    Related Articles