Differential Diagnosis for Customer Interviews
A step-by-step guide to run customer interviews like a clinician: define hypotheses, set priors, run targeted probes, and update beliefs with each call.

This guide is for working PMs who run discovery and customer interviews. If your calls confirm what you already believe or yield noisy stories, borrow a clinician’s playbook. Use differential diagnosis to set up competing hypotheses, ask targeted probes, and update your belief after each interview.
In brief
- List alternative explanations for the user’s behavior, not just one.
- Assign rough priors before you talk.
- Design short, high-variance probes that move your belief.
- Treat each interview as a test. Log the evidence and update with an informal likelihood ratio.
- Iterate across diverse users. Rule in, rule out, or reprioritize bets.
Why borrow from medicine
Clinicians do not jump to one cause. They run a differential diagnosis. They list plausible options, estimate a pre-test probability, run targeted tests, and update until they can rule in or rule out. Discovery can work the same way.
| Typical interviews | Diagnostic interviews |
|---|---|
| Start with a favorite story or pet theory. | Start with a list of competing hypotheses. |
| Many broad questions; little signal per minute. | Few targeted probes; high signal per minute. |
| Update at the end, if at all. | Update after each probe with simple evidence ratios. |
| Confirmation bias goes unchecked. | Pre-committed priors and rules reduce bias. |
Treat each interview as one diagnostic test. Your goal is to move probability, not collect quotes.
The steps at a glance
Define alternatives
List 4–6 plausible causes for the behavior you see.
Set rough priors
Assign pre-test probabilities based on base rates and past data.
Design probes
Write 3 short tests that would shift odds sharply.
Run the test
Conduct the interview. Log exact observations.
Update and iterate
Apply informal likelihood ratios. Reprioritize, then repeat.
Step 1 — Define competing hypotheses
Write down 4–6 explanations that could fit the same symptom. Example: why did users churn? Price sensitivity. UX friction. Missing capability. A competitor offers clearer value. Implementation burden on their side. You do not need perfect coverage. You do need real alternatives that would change your roadmap if true.
Tip: Phrase each hypothesis as a clear, falsifiable statement. Example: “Users leave because they cannot complete core task X without help.”
Step 2 — Assign pre-test probabilities
Give each hypothesis a rough prior. Use base rates from your analytics, win/loss notes, support logs, and market norms. These are not precise. They are a commitment to what you believe before you hear a new story.
Write the priors down next to each hypothesis. You will adjust them after each call. Keep the scale simple. Percentages or odds both work as long as you stay consistent.
Step 3 — Design short, high-variance probes
Open questions have their place. But you need probes that move probability. In medicine, tests have likelihood ratios. In interview practice, you can mimic that. Choose tasks or questions where a “yes” would strongly support one hypothesis over the others, and a “no” would count against it.
| Probe | What it reveals | If result supports |
|---|---|---|
| Cost-justification vignette: “Imagine your manager asks you to justify renewing at the same price. What do you say?” | Ability to defend value vs price pain. | Clear defense → not price; struggles → price more likely. |
| Task completion with think-aloud on a core flow. | Real friction vs reported friction. | Hesitation/errors → UX friction; smooth, fluent → not UX. |
| “How would you fix it?” freeform redesign. | What they miss most and who they compare you to. | Requests mirror a rival → competitor; asks for a gap → missing feature. |
Probe design checklist
- One hypothesis per probe. Keep it short.
- Write accept/reject signals before the call.
- Avoid leading words. Prefer “show me” over “would you”.
- Timebox to 3–5 minutes each.
Step 4 — Run the interview as a test
Open with context. Then run your three probes. Ask follow-ups only to clarify observable facts: what they did, what they saw, what they decided. Capture direct observations, not interpretations. Example: “Hesitated 12 seconds on field labels. Asked for help.”
After the call, log the result of each probe as Supporting, Neutral, or Against for each hypothesis. Keep it simple and consistent.
Step 5 — Update with informal likelihood ratios
You do not need a formula to benefit from likelihood ratios. Think in odds. If a probe strongly favors H1 over H2, bump H1 up and H2 down. If the signal is weak, do not move much. Write the updated belief and the reason next to it.
End each round by deciding: do we have enough to rule in, rule out, or reprioritize? If not, design the next probe and find a user who can produce high contrast on it.
Worked example — churned-user triage
Context: a user churned last month. We want to understand why. We list four hypotheses:
- H1 Price sensitivity
- H2 UX friction in a core flow
- H3 Missing feature that matters
- H4 Competitor offers clearer value
Priors (pre-test): H1 10%, H2 70%, H3 15%, H4 5%.
Probes we prepare:
- Cost-justification vignette.
- Task completion with think-aloud on the core flow.
- “How would you fix it?” prompt.
Observations from one interview:
- On the vignette, they give a clear, specific argument for value at the current price. Little hesitation.
- On the task, they complete steps fluently and comment that the flow is “fine.” No visible stumbles.
- On fixes, they describe a workflow that mirrors what a rival product is known for and say they are “now using something else.”
Update: Probe 1 counts against H1. Probe 2 counts against H2. Probe 3 strongly supports H4 over H3 and H2. After updating, our belief shifts from H2 being most likely (70%) to H4 being most likely (about 80%).
Decision: mark H2 as unlikely for this user, and prioritize a deeper look at competitive positioning and the specific workflow they value. Queue two more interviews with recent churns who chose known rivals to test this pattern.
Mistakes to avoid
- Confirmation bias: writing probes that can only confirm your favorite idea. Write accept and reject signals first.
- Overfitting one interview: a strong signal can guide the next probe, but never generalize from a single call to the whole cohort.
- Leading questions framed as tests: if your wording nudges the answer, your likelihood ratio is fake.
- Ignoring base rate: a rare behavior that you observed once should not outweigh a common pattern in your data.
- Stopping too early: if two hypotheses remain plausible, keep testing across diverse users.
Quick replication plan
- Choose 5 hypotheses for one job-to-be-done or behavior.
- Set priors. Write a one-line base-rate note for each.
- Draft 3 high-variance probes. Define what supports and what goes against for each hypothesis.
- Run 8 interviews with diverse users. Keep each to 25–30 minutes.
- Update after every 4 interviews. Merge or drop hypotheses that stay weak.
- Synthesize with a simple scorecard: final probabilities, top evidence for/against, decision (rule in/out/reprioritize).
You can speed up the roll-up with Interview synthesis once you have clean notes and probe outcomes.


