How to Run Customer Interviews Without Leading the Witness

TL;DR. A leading question plants its own answer, so the respondent agrees with you instead of telling you what they did. You avoid it by asking about specific past behavior, keeping the answer out of the question, and staying quiet after they finish. The hardest moment is the follow-up after a thin answer: paraphrase it and you lead them, accept it and you lose the signal. The fix is the non-directive probe, where you reflect their exact words or ask for the last real time it happened. Below are the four patterns that quietly lead, the rewrites that fix them, and how to recover when you catch yourself doing it.

What a leading question actually is

A leading question is one that carries its own answer. You put the conclusion inside the question, and the respondent does the polite thing and confirms it. "Wouldn't it be useful to see all your invoices in one place?" is not a question. It is a statement wearing a question mark, and "yes" tells you nothing about whether the person has ever struggled with invoices.

Nielsen Norman Group draws the line plainly: a leading question subtly prompts the respondent toward a particular answer, and it skews your data without anyone noticing. The damage is quiet. People rarely push back on a friendly assumption, so you walk out with a notebook full of agreement and call it validation.

The Mom Test names the same failure from the question side. Never put the answer inside the question. The moment you do, you stop learning what the person believes and start measuring how agreeable they are. That is why "a compliment is not data" is the rule the rest of this article keeps returning to. A compliment is usually just the respondent accepting the answer you handed them.

The four patterns that look safe but lead

Most leading questions do not look like leading questions. They look careful. Here are the four that slip past you, each with a fix grounded in a Mom Test or JTBD rule.

1. The loaded value word

You drop a word that pre-judges the answer. "How frustrating is it when the report runs slowly?" assumes frustration and assumes the report runs slowly. The respondent now rates a feeling you assigned them.

• Bad: "How frustrating is the current onboarding process?"
• Good: "Walk me through the last time you onboarded a new client. What happened?"

The rule: ask about past behavior, not opinions. A neutral question lets the person reach for frustration on their own, or not reach for it at all, which is itself the data.

2. The answer-in-the-question

You name the outcome you are hoping for and ask them to confirm it. "Would a shared dashboard save your team time?" hands them the benefit and the verdict in one breath.

• Bad: "Would a single shared dashboard save your team time?"
• Good: "How does your team check project status today? Who looks, and when?"

The rule: never put the answer inside the question. You want the workflow they actually run, not their guess about a feature you just described. Maze makes the same point about survey design: as soon as you suggest the answer, you are testing compliance, not preference.

3. The hypothetical

You ask the person to predict their own future. "Would you use an app that did this?" feels like research, but the answer is a wishlist, not a behavior. People are generous and wrong about what they will do.

• Bad: "Would you pay for a tool that automated this?"
• Good: "What are you using to handle this right now, and what did the last one cost you?"

The rule: talk about the last real project, not a hypothetical. The Mom Test treats hypothetical fluff as one of the three bad-data types for exactly this reason. A future "yes" is free. A past purchase is paid for. For the deeper version of this move, the switch interview walks the timeline of the moment a person actually changed tools, which is where real motivation lives.

4. The compliment fish

You ask a question whose only job is to collect praise. "What do you think of the idea?" invites a compliment, and you take the compliment as a green light.

• Bad: "Do you think this is a good idea?"
• Good: "When did this problem last cost you something, and what did you do about it?"

The rule: treat a compliment as noise until a behavior backs it. Enthusiasm about your idea tells you the person is nice. Evidence that they already spend money or time fighting the problem tells you the idea might matter. That gap is the whole argument behind our manifesto on why a compliment is not data.

The non-directive probe: follow the flinch

The four patterns above are the easy part. You can scrub them out of a discussion guide before the call. The hard moment comes live, in the follow-up, when someone gives you a thin answer like "the reporting is kind of a pain."

You have two bad instincts here. The first is to paraphrase: "So it's slow and clunky?" Now you have replaced their vague word with two specific words of your own, and they will nod, and you will write down "slow and clunky," and none of it came from them. The second bad instinct is to accept it and move on, which leaves the most interesting thing they said unexamined.

The non-directive probe is the third path. You follow up on the vague answer without paraphrasing and without dropping it.

• Echo their word as a question. They say "a pain." You say "a pain?" and stop talking. The word is theirs, so the elaboration is theirs too.
• Ask for the last instance. "When was the last time reporting was a pain? Walk me through that day." You trade the abstraction for a story you can verify.
• Ask what they did next. "And then what did you do?" Behavior fills in where the adjective was vague.

This is what we mean by follow the flinch, not the script. The flinch is the small hesitation or the loaded word that signals there is something underneath. Your job is to probe it with their language, not yours. The silence after you echo a word is uncomfortable, and that discomfort is doing the work. Resist filling it.

Recovering mid-interview after you've led

You will catch yourself leading. Everyone does, especially late in a long day of calls when fatigue makes the script feel safer than the silence. The recovery is simple and you can do it out loud.

When you notice you handed over the answer, neutralize it on the spot. If you just asked "wouldn't that be easier?", follow immediately with "actually, ignore that. How do you handle it now?" Respondents do not mind. They read it as you being careful.

If they have already agreed with a leading question, do not bank the agreement. Go get the behavior that would prove or break it. "You said a shared view would help. When did the lack of one last cause an actual problem?" Either a real story arrives, in which case you have data, or it does not, in which case the earlier "yes" was the compliment you suspected.

The meta-skill is noticing in real time, and that is genuinely hard for a tired human holding a discussion guide. It is also the part that decays first at volume, which is the bottleneck most agencies feel.

How an AI interviewer enforces it

This is where Lùc comes in, and it is worth being precise about what the machine does and does not do.

A discussion guide cannot enforce itself. The four patterns and the non-directive probe are rules a person has to apply live, under social pressure, on every single answer, across every interview in a project. Discipline that holds on interview three tends to slip by interview thirty. The tools competitors point at this problem mostly help you write a cleaner guide up front. Strella, for instance, markets "unbiased interview guides," which is a real and useful thing. But a clean guide is the easy half. The follow-up is where interviews are won or lost, and the follow-up happens after the guide stops talking.

Lùc enforces the craft in the moment the question is asked. It reads the answer before it asks the next thing. When a respondent gives a vague word, it echoes the word instead of paraphrasing. When a respondent offers a compliment, it does not bank it; it goes looking for the behavior underneath. It asks about the last real instance rather than the hypothetical future, and it runs that same discipline on interview thirty exactly as it did on interview three, in parallel, while the researcher spends their judgment on synthesis and the conversations that actually decide something.

That is the vision we are building toward in closed beta: not faster interviews, because faster bad interviews are still bad interviews, but interviews where the discipline holds at volume. Discipline is the product.

Frequently asked questions

How do I run a customer interview without leading the respondent?

Ask about specific past behavior instead of opinions or predictions. Keep your hoped-for answer out of the question, stay quiet for a beat after the respondent finishes, and when an answer is vague, probe it by reflecting their exact words rather than your interpretation. Treat any compliment as noise until a concrete behavior backs it. The simplest test: if the person could answer "yes" without ever having faced the problem, your question is leading.

How do you follow up on a vague answer without leading the witness?

Use a non-directive probe. Repeat the respondent's own vague word back to them as a question ("a pain?") and then stop talking, or ask for the last specific time the thing happened and walk through it ("when did reporting last cost you an afternoon?"). Do not paraphrase, because a paraphrase swaps their meaning for yours and they will agree with the version you wrote. The goal is to make the elaboration come from their language and their behavior, not from your guess about what they meant.

What makes a customer interview question biased?

A question is biased when it makes one answer easier to give than the others. Four common ways: you load it with a value word that pre-judges the feeling ("how frustrating is..."), you name the answer you want and ask them to confirm it ("wouldn't a dashboard help?"), you ask about a hypothetical future instead of the real past ("would you use..."), or you fish for a compliment instead of a behavior ("do you like the idea?"). The fix for all four is the same move: ask what the person actually did the last time the situation came up.

---

Want to run disciplined interviews at the volume a project needs? Join the Lùc closed beta waitlist. And if you are deciding which method to lean on, read the Mom Test versus Jobs-to-be-Done breakdown and how many customer interviews are actually enough.