How to Run Customer Discovery Interviews

Pick one or two personas. HBS recommends defining a few target personas, each with its own situation and motivations, and focusing on one or two primary ones (HBS Rock Center). Distinguish between the person who uses a product and the person who pays for it, since in many businesses they are different people with different problems.

Decide how many. HBS guidance, read on 2026-10-17, suggests that in the early stages the right number is about five interviews per persona, run with a consistent structure (HBS Rock Center). Five per persona is a first pass. Keep going while each interview still teaches you something, and stop when you can predict what the next person will say.

Find people. Start with warm introductions and work outward:

• People in your network who match the persona, or who know someone who does.
• Online communities, forums and groups where the persona discusses the problem.
• Short cold emails or messages that ask for advice about a problem, not time to pitch a product.
• Events and meetups where the persona gathers.

The Frank Thoughts newsletter, summarizing The Mom Test, puts the goal well: "the point of cold calls is to stop having them." Once someone responds, use them to reach others in the same market (Frank Thoughts). The NSF I-Corps hub at Desert Pacific suggests ending every interview by asking whether there is someone else you should talk to and whether they would introduce you, since a warm introduction is more likely to lead to another interview than a cold approach (Desert Pacific I-Corps).

Frame the request around learning. "I am researching how small clinics handle scheduling and would value twenty minutes of your experience" gets more yeses than "Can I show you my app?" and produces a better conversation.

The Mom Test reduces good interviewing to three rules, summarized by the research platform Koji: talk about their life instead of your idea, ask about specifics in the past instead of hypotheticals about the future, and talk less and listen more (Koji).

HBS turns the same ideas into a list of what to do and what to avoid (HBS Rock Center):

Past behavior is reliable because it already happened. Koji's guide explains that a customer cannot embellish what they did last month as easily as they can construct a rosy hypothetical future (Koji). Compare two questions:

• Weak: "Would you use a tool that automated your invoicing?"
• Strong: "Tell me about the last time you sent invoices. How long did it take? What went wrong?"

The first invites a polite yes. The second produces a story with times, tools and frustrations you can check against other interviews.

HBS adds a point about what to listen for: when you ask open questions, you are looking for emotion and frustration, because the emotional response is the signal (HBS Rock Center). A customer who sighs before describing a monthly process is telling you more than one who calls it "a bit annoying."

Keep the guide short. Three or four core questions, plus follow-ups, are enough for a twenty to thirty minute conversation.

Get comfortable with silence. When a customer stops talking, wait a few seconds before asking the next question. People often add the most useful detail after a pause, such as the workaround they are embarrassed about or the cost they had not mentioned. Filling every silence with your own comment is one of the quickest ways to turn an interview into a pitch.

1. Context (2 minutes)
   "Tell me about your role and what a typical week looks like."

2. The last time (10 minutes)
   "Tell me about the last time you [dealt with the problem area]."
   Follow-ups: What happened next? Who else was involved?
               How long did it take? What did it cost?

3. What they have tried (5 minutes)
   "What have you tried to fix this? What happened?"
   "What do you use today? What do you like and dislike about it?"
   "Have you paid for anything to solve it?"

4. Priority and decision (5 minutes)
   "Where does this rank among the problems you deal with?"
   "If you wanted to change how this works, who would need to agree?"

5. Close (3 minutes)
   "Is there anything I should have asked but didn't?"
   "Who else should I talk to about this?"

Two follow-ups do most of the work: "Can you give me an example?" and "Why?" Use them whenever an answer is vague.

Notice that the guide never describes your solution. In early discovery, that is deliberate. Once you have confirmed the problem, a later round of interviews can show a mockup or prototype, but the questions should still focus on how it fits what the customer already does.

Three kinds of answers feel like data but are not.

• Compliments. "That sounds amazing" tells you about the person's manners, not their needs. Koji's summary of The Mom Test calls compliments social lubricant and says they are worth nothing unless accompanied by a commitment of time, money, reputation or a referral (Koji).
• Fluff. Generic claims ("I always..."), future promises ("I would definitely...") and hypotheticals ("I might..."). The Frank Thoughts newsletter describes fluff as statements with a weak qualifier and recommends pushing back on it (Frank Thoughts). Bring the conversation back to a specific past event: "When was the last time that happened?"
• Your own pitch. If you find yourself explaining your idea, you have stopped learning. Stop, apologize and ask another question about their experience.

When the problem is confirmed and you do show an early product, look for two signals that the research platform Looppanel describes from the book. Commitment means the person gives up something they value, such as time, reputation or money. Advancement means they move to the next step of your real sales process (Looppanel). Ending a meeting with "That's great, keep me posted" is neither. Ending it with "Let's set up a trial with my team next week" is both.

The jobs to be done approach offers a useful structure for interviews with people who have already bought or abandoned a product. It treats customers as "hiring" products to make progress toward a goal, and focuses on the moment they decide to switch.

User researcher Nikki Anderson, writing for dscout, describes recruiting three groups: people who have just started using a product, people who have just stopped, and people who have not yet used it. Her interviews center on the "switch event" and ask what other solutions people considered, what they were trying to find and what they were trying to solve (dscout). She warns that asking someone directly why they switched produces a surface answer, and that the deeper motive comes out through stories.

Commoncog, describing its own team's use of the method, recommends working backward from the purchase and quotes the opening question the method's authors suggest for every interview: "Why did you buy it and where were you when you bought it?" The article suggests using environmental details, such as the weather or a trip, to help people place events on a timeline accurately (Commoncog). It also notes that a badly run interview cannot be rescued in analysis: if you did not capture the timeline, the gaps stay empty.

This format works best once you have customers or a clear competitor to study. For a product that does not exist yet, interview people who have switched between existing alternatives in the past few months in your space.

Interview in pairs. HBS recommends a discovery team of two or three, where one person asks questions and the others observe and take notes (HBS Rock Center). The Frank Thoughts newsletter suggests focusing notes on the big beats, such as pain points, follow-up tasks and introductions, and on strong emotions (Frank Thoughts).

Write it up the same day. Looppanel's summary of the book warns against the "learning bottleneck", where what was learned stays in one founder's head instead of being shared promptly and faithfully with the team. It suggests a sortable spreadsheet with key signals in their own columns (Looppanel). A simple structure:

Look for patterns, not quotes. After each batch of five or so interviews, count how many people described the same problem unprompted, what they use today, and who decides. A single vivid story can mislead. A problem that half the interviewees raise without prompting is worth pursuing.

Use AI for preparation and sorting, not for the conversations. The NSF I-Corps hub at Great Lakes suggests using AI to draft open-ended questions, generate candidate personas and rehearse interviews by asking a model to play the customer and then critique your questions. It also lists the limits: you still need to interview real customers to validate any idea an AI suggests, AI output can be false, and you should not paste sensitive or personal information into these tools (Great Lakes I-Corps).

Decide what changes. End each batch by writing down which assumptions were confirmed, which were contradicted, and what you will test next. Discovery that does not change a decision was only conversation.