You are about to book customer calls, and the risk is that every one of them ends with polite enthusiasm and no decision. Before the first call, write your questions down and mark each one as a fact about the past or an opinion about the future. Cut the opinions. That single edit changes more than any script.
Here is a hypothetical to show why. Consider a team building clinic software. Every clinic owner they interview says reporting is painful and describes it in detail, so reporting ships in the first release. Weeks later the usage data shows people open tomorrow's schedule each morning and rarely open reports. Nobody mentioned the schedule because it was not a complaint. It was routine. That gap is the default output of a conversation where one person visibly hopes for an answer and the other can see it.
What should a customer discovery interview produce?
A customer interview should produce evidence about what someone already did, not a verdict on the product you hope to build. Facts cost the speaker effort to retrieve, and that effort is the signal.
Alex Osterwalder wrote in Strategyzer in January 2018 that "asking for opinions rather than facts is the single biggest error in customer discovery interviews." His alternative to "what do you think the challenge is?" is "when is the last time you struggled with this particular challenge?"
An opinion is free. "Yes, I'd use that" costs the speaker nothing and is forgotten within the hour. My test for every question is whether someone else could check the answer. "How do you handle that today?" can be checked. "Would you switch to a tool that did it for you?" cannot be checked by anyone, including the person answering.
If you sell B2B SaaS, apply a second filter. The person who feels the pain and the person who releases the budget are often different people. Callidus, the clinic platform I built on React, Firebase and Stripe Connect, has six roles in its case study, from receptionist and practitioner to owner and super admin. This suggests that in a product like it, the person who lives in the schedule is not the person who signs the invoice, and you need to interview both. Ten enthusiastic end users and zero approvers give you a consistent story about a purchase that never happens.
Which Mom Test questions work for SaaS?
The questions that work ask about the customer's life and specific past events, never your idea. Rob Fitzpatrick's The Mom Test is built on three rules: talk about their life instead of your idea, ask about specifics in the past instead of generics about the future, and talk less than you listen. The book's site says it is used as core curriculum at Harvard, UCL, Seedcamp, Microsoft Ventures and Shopify.
Rules are easy to nod at and hard to hold under pressure, so here are the substitutions written down.
| What you asked | What it actually measures | What to ask instead |
|---|---|---|
| Would you use this? | Politeness | Walk me through the last time you dealt with this. |
| How much would you pay? | Imagination | What are you paying for it now, and who signed that off? |
| Do you like the idea? | Rapport with you | What have you already tried in order to fix it? |
| Is this a big problem? | Willingness to agree | What did it cost you the last time it happened? |
The third rule is the one that breaks people. Silence in a call feels like a fault you must repair, and the fastest repair is to start explaining your product. Every second you spend explaining teaches the other person which answers please you.
Problem interviews and solution interviews are different jobs
Run them on separate days, with separate people. The moment you show a mockup, a research conversation becomes a demo, and everything said afterwards is colored by having seen your work.
A problem interview has no artifact in the room. You map how the work is done now, what the workaround costs, who else is involved and what happened the last time it broke. None of that depends on your product existing, which is why the findings survive a pivot and are the research worth doing before you scope a minimum viable product.
A solution interview needs the artifact: put it in front of someone, stay quiet and watch where they stall. A June 2026 post on SaaS Validation argues that interviews capture stated preference, and that "the gap between what a customer said they wanted in an interview and what they actually did with the prototype is where the most valuable product information lives."
That does not make problem interviews worthless. They are good at establishing what already happened, which is a fact. They fail when you ask them to predict. Keep them in their lane.
How do you run a willingness to pay interview?
Ask what they pay for the current fix, who approved that spend and what happened the last time a similar request was refused. Each of those has already happened, so none requires the person to imagine a future purchase.
Structured methods have a place. Perspective AI's 2026 guide to willingness-to-pay research says the Van Westendorp Price Sensitivity Meter, introduced in 1976, triangulates an acceptable price range with four questions, and advises 150 to 300 interviews for a full pricing study or 20 to 30 for a single-question pilot. Perspective AI sells interview software, so read that as a vendor's guidance. A gauge tells you the range, not why the top of the range exists. For that, run the conversation in this order:
- Establish the current spend on the problem, including the salary cost of whoever does it manually today.
- Find out who signs. Ask for their title, then ask what that person rejected most recently.
- Ask for a commitment rather than a number. A signed letter of intent, a discounted pre-payment or an introduction to that approver all cost the customer something real.
- Watch the following week. A warm "send me pricing" that gets no reply is a data point, and a more honest one than the number said on the call.
Consider pricing a product like BookBed, which its case study lists at €9 per month for up to twenty units. In a hypothetical interview, an owner who already pays a booking channel a commission every month has a live number to hold that price next to. An owner with no reference point gives you a compliment instead of an answer. That is an illustration, not a result.
Replace "what is your biggest challenge?"
I would drop that question. Everyone has a rehearsed answer, the one they give at conferences, and it comes loose from their actual week. Ask what they did on the most recent Monday instead. This is a recommendation, and the rehearsed answer has no version of Monday.
How many customer interviews should you run before you build?
Twenty conversations with real buyers before you write code is still a sound target. Jason Lemkin's 20 interview rule on SaaStr, published in October 2014, splits them into phases: interviews 1 to 5 to understand the market from the buyer's side, 6 to 10 to confirm the pattern, and 11 to 20 to refine the pitch. He stresses actual decision-makers, not friends. If you are choosing tooling meanwhile, the stack decisions for a SaaS MVP can wait until the pattern is visible.
What a cancelled customer knows that your survey does not
A churned customer can describe the decision, while a cancellation survey only captures the reason. A decision has a date and a trigger, and often another vendor inside it. Reasons are what people type into a text box to get past it. If you track churn rate, these calls tell you why the number moves.
Adam McQueen's February 2025 churn interview guide on Klue says "75% of tech businesses evaluate alternatives before renewing," and cites no source for it, so treat it as directional. The mechanism is plausible, though: by the time someone cancels they have often been living with a comparison for a while. Ask what the other tool showed them that yours did not, and when they first went looking. Frame the call as learning, not winning them back. It stops being research the moment you offer a discount.
Should you use AI-moderated interviews?
Use them for breadth once problem interviews have told you what to ask, not instead of them. Perspective AI's 2026 report, published in May 2026, claims 81% of research teams and 51% of customer success teams use AI customer research, with completion rates above 85% against 22% for long-form surveys. The public page refers to "our 2026 sample" without stating a respondent count, and the publisher sells AI-moderated interviews. The figures may be right, but I can't verify them, so I would not build a plan on them. Async interviewing suits a question you already know how to phrase. It is weaker at the moment a human hears something odd and abandons the script to chase it.
Open your billing dashboard, filter to accounts that cancelled in the last ninety days and pick three you would still be glad to have back. Send each two sentences asking for twenty minutes to learn what went wrong, with no pitch attached. If all three name the same trigger and it is not on your roadmap, what will you change this quarter?