The secret behind the quality: 70% of insights live in the follow-up question

Watch a skilled interviewer at work and you will notice the questions on their script do almost none of the heavy lifting. The script gets the conversation started. What it never does is get to the reason. The reason lives one question deeper, in the thing the interviewer asks after they hear the first answer. In our experience, about 70% of the insight in an interview comes from that follow-up, not from the question you planned in advance.

I am the CTO at Nava Insights, and that 70% is the single fact that shapes most of what my team builds. If the follow-up is where the value is, then the engineering problem is not "ask good questions." It is "listen well enough to ask the right next one, in real time, every time."

Why the scripted question is only the surface

A scripted question is a guess. A good guess, written by someone who knows the research, but a guess about what matters, made before you have heard a word from the person in front of you.

So it gets you the surface. Ask "what did you think of the onboarding?" and you get "it was a bit confusing." That sentence is true and nearly useless. It does not tell you what was confusing, when the person got lost, what they expected instead, or whether they would have given up if they were not being paid to continue. Every one of those is a follow-up, and every one of them is where a decision actually gets made.

This is the line surveys cannot cross. A survey can ask the planned question beautifully. It cannot hear "a bit confusing" and ask "confusing how, exactly, walk me through the moment you got stuck." It has no idea what was said, so it cannot respond to it. A dashboard can tell you how many people were dissatisfied. It cannot tell you why, because the why was never a checkbox.

The scripted question gets you the answer. The follow-up gets you the reason, and the reason is the whole point.

That is the gap between counting opinions and understanding them. The follow-up is the why behind the number, and it is why a real conversation will always beat a form.

The craft of a good probe

Before I get to the engineering, it is worth being honest that probing is a craft, and most of it is judgment. A great human interviewer does several hard things at once, mostly without naming them.

• They listen to what was actually said, not to what they expected to hear, and they probe the specific words the person used.
• They decide when to dig and when to move on, because probing everything is exhausting for the participant and probing nothing is pointless.
• They stay neutral. They ask "what made you say that" rather than "so it was frustrating, right," because the second version puts the answer in the participant's mouth.
• They sit with silence instead of rushing to fill it, because the most honest thing a person says often comes right after an uncomfortable pause.

Those four behaviors are the difference between an interview that produces evidence and one that produces noise. They are also genuinely difficult, which is why good qualitative interviewers are rare and expensive, and why a study of eight to fifteen of them traditionally costs thousands and takes weeks.

Doing it in real time, at scale

Here is the engineering problem, stated plainly. To probe well, the AI moderator has to understand the answer the instant it lands, decide whether it is worth digging into, and form a neutral, specific next question, all inside the natural rhythm of a spoken conversation. If it pauses too long to think, the moment is gone and the participant feels the seam. If it probes everything, the interview drags and people drop out. If it leads, the data is contaminated before you ever read it.

So the moderator is doing a few things at once. It tracks what the conversation still needs to cover, so it can recognize when an answer has opened a door worth walking through versus when it should let a topic close and move on. It anchors follow-ups to the participant's own language, because probing the actual words is what keeps a question neutral and specific rather than generic and leading. And it knows when to stop, because a good probe has a floor and a ceiling: enough to reach the reason, not so much that you are interrogating someone over a minor point.

Staying neutral is the part I care about most, because a leading follow-up does more damage than no follow-up at all. "Why did that bother you" quietly assumes the person was bothered. "What was that like for you" does not. We hold the moderator to the second kind. The goal is to draw the reason out of the participant, never to hand them one and watch them agree.

The reason we can run this at all is that voice gives the model far more to work with than a text box would. People say more out loud, they say it more spontaneously, and the hesitations and emphasis in a spoken answer are themselves signals about where the real feeling sits and where a probe will pay off. A richer input makes a smarter follow-up possible.

Why this is the whole game

Everything downstream depends on this one thing being good. The analysis, the sentiment read, the behavioral archetypes, the answer NavaGPT gives you when you ask why a segment churned: all of it is only as good as the conversation it came from. A shallow interview produces shallow data, and no amount of clever analysis recovers a reason that was never drawn out in the first place. Quality of conversation drives quality of data. There is no shortcut around it.

And because every insight Nava surfaces is traceable to the specific quote and transcript position it came from, you can check the work. When the system says a reason is common, you can read the actual follow-up exchange where each person said it, in their own words. Nothing is a black box. That traceability is also a forcing function on us, because there is nowhere for a weak probe to hide when you can click straight through to the moment it happened.

So when you read your next study, do not skim the planned questions. Read the follow-ups. That is where the work was done, and that is where the 70% lives. Getting that one part right, at scale, in real time, in a way you can audit afterward, is most of what we mean when we talk about quality.