How to write a discussion guide that gets honest answers

A discussion guide is not a questionnaire. A questionnaire collects answers to questions you already thought of. A discussion guide sets up a conversation that can surprise you. The difference decides whether your research tells you something new or politely confirms what you walked in believing.

Whether a human moderator runs the session or an AI moderator does, the same principles apply. Here is what separates guides that produce honest, usable answers from guides that produce noise.

Start from the decision, not the questions

Before writing a single question, write down the decision this study should inform. “Should we build the integrations page before the mobile app?” is a decision. “Learn about user needs” is not. Every section of your guide should trace back to that decision; if a question cannot change what you do next, cut it.

A practical test: imagine the two most extreme answers a participant could give. If both would leave your plan unchanged, the question is decoration.

Ask about the past, not the future

People are unreliable narrators of their future behaviour and surprisingly precise narrators of their recent past. “Would you use a weekly summary email?” invites politeness. “Tell me about the last time you needed to catch up after a week away” invites a story, and stories contain the friction, workarounds, and priorities you are actually looking for.

The most honest answer a participant can give you is a description of something they already did.

One question per question

Compound questions get half-answers. “How do you plan your week and what tools do you use?” will get you a tool list and no planning story, or the reverse. Split them. Short questions also give a moderator, human or AI, clean openings for follow-ups instead of forcing it to untangle which half the participant answered.

Let the follow-ups do the digging

Good guides are shallow on questions and deep on probing. The laddering technique climbs from what someone did, to the consequences it had, to why those consequences mattered. That third rung is where the useful material lives, and it almost never comes out on the first ask.

This is worth encoding explicitly. In GhostClick, every question carries a follow-up depth from 1 to 7 that caps how far the AI moderator will probe; two is a sensible default, and the higher depths are for laddering studies that need to climb from behaviour to values. If you moderate live, write your probe budget in the margin instead. Either way: decide in advance where you want the moderator to dig, or it will dig evenly everywhere and exhaust your participant.

Kill the leading questions

Any question that contains its own preferred answer will find it. Watch for these patterns:

• Embedded praise: “How much easier is the new flow?”
• False binaries: “Do you prefer speed or reliability?” when the honest answer is a third thing.
• Social pressure: “Most people find onboarding simple. How was it for you?”

Neutral phrasing feels flat when you write it. That flatness is the point; the energy should come from the participant’s answer, not your question.

Set the tone on purpose

A guide about expense reports and a guide about fertility apps should not be moderated in the same voice. Decide whether the conversation should feel professional, friendly, neutral, empathetic, or direct, and write the guide in that register so the moderator inherits it. Consistency matters more than the choice itself: a warm opening followed by clinical probing reads as insincere.

Pilot it before you trust it

Every guide contains one question that reads clearly to you and means nothing to a participant. You find it by running the interview, not by rereading the guide. Do a full dry run as if you were a participant, answer everything yourself, and note where you hesitate: hesitation marks ambiguity. Then fix the guide and run it once more. An hour of piloting routinely saves a study.