Turning video research into a customer interview script

The short answer: turn every claim your research produced into a past-behaviour question, delete every question the research already answered, and keep only the items where a plausible answer would change what you build next week. That usually leaves eight to twelve questions, which is a real interview script rather than a wish list.

Most founders treat research and interviews as substitutes: do enough desk work and you can skip the awkward calls, or skip the desk work and let interviews teach you everything. Both are expensive in different ways. Interviews without preparation spend their first twenty minutes on vocabulary and context that fifteen public videos would have handed over for free. Research without interviews produces a document nobody can defend, because every claim in it is second-hand.

What each half is actually good at

The division of labour stops being subtle once you name it. Public discussion is excellent at breadth and poor at specificity. A live conversation is the reverse. Trying to get breadth from six interviews is slow and biased by whoever answered your message; trying to get a specific budget figure out of a public video is guesswork.

That last row is the one worth dwelling on. Anything people find slightly embarrassing is systematically absent from public material — a limit covered in more detail in what AI video research gets wrong. Interviews exist to cover exactly that blind spot, so a script that does not probe it is wasting the format.

Step one: invert every claim into a behaviour question

A research pass leaves you with claims. Each claim carries an implicit prediction about behaviour, and the question you want is the one that tests the prediction without announcing it.

The left column produces agreement, which feels like validation and predicts nothing. The right column produces stories, which can be wrong but cannot be merely polite. If someone cannot recall a single instance of the problem you researched, that absence is your finding — and it is one of the early signals covered in how to disqualify a SaaS idea early.

Step two: cut everything the research already settled

This is the step people skip, and it is where the time saving lives. If your corpus established that practitioners in this space universally use a particular tool, do not spend interview minutes asking which tool they use. Ask instead what it fails to do — the part public content under-reports, because complaining about a beloved tool costs a creator audience goodwill.

Run the whole draft through one filter: for each question, imagine the two most likely answers, then ask whether your next week differs between them. If it does not, the question is conversational filler. A first draft that survives this filter at fifty percent is doing well, and the surviving half is nearly always the concrete half.

Step three: the shape of the script itself

Twelve questions is the practical ceiling for a thirty-minute call, and the ordering matters more than the wording. Open with the last concrete instance, widen into workflow, narrow onto the contested claims, and only at the very end mention what your research suggested.

Opening, two questions. Establish the last time the problem occurred and what the person did about it. No hypotheticals, no product mention. If the opening does not surface a recent instance, the rest of the script is being answered from imagination, and that is worth noting on the transcript.

Workflow, three to four questions. Have them narrate the current process step by step, including the parts they consider too boring to mention. Boring steps are where most durable products live — the same reason instructional content is such a productive source, as in finding micro-SaaS ideas in tutorial videos.

Contested claims, three to five questions. One per disputed finding, each phrased as a past event. This is the core of the call and the reason you ran the research pass at all.

Close, one to two questions. What did you expect to exist and could not find, and who else should I talk to. Only after that do you describe what you are working on, if at all.

The corpus also tells you who to call

The unglamorous bottleneck in this process is finding people to talk to, and a research pass quietly solves part of it. Working through fifteen sources surfaces the channels, communities, meetups and forums where the segment congregates — which is a recruiting list you did not have to buy.

Two approaches work. Practitioners who published detailed content about the problem are unusually willing to talk about it, because they have already demonstrated they enjoy the subject. And the communities that keep appearing in the corpus are where a polite request for twenty minutes lands better than a cold message, since you can reference the specific discussion that brought you there.

Both routes bias toward the vocal end of the market, which is worth naming in the notes rather than pretending away. It is the same sampling skew that affects the corpus itself, and the correction is the same: deliberately include a few people who are not publishing about this, even though they are harder to reach.

Closing the loop back into the research

After five or six calls, some interview answers will contradict the corpus. That is the useful outcome rather than a failure of the research: it means the public record was skewed toward a particular kind of source, and the correction belongs back in the document.

In practice that means keeping the research artifact live rather than treating it as a deliverable. Update the contested claims with what the interviews found, mark which ones are now settled, and re-run the corpus when the vertical moves — the discipline described in monitoring a niche with recurring research. A corpus that says what is discussed, paired with interview notes that say what was done, is a considerably stronger basis for a build decision than either alone.

What the research half costs

Interviews cost calendar time, which is the expensive input. As of August 2026 the research half runs $19 a month for 25 videos and 2 projects, $59 for 80 videos and 8 projects, or $199 for 250 videos, 20 projects and 3 seats — the full breakdown is on the pricing page. The relevant comparison is not tool cost against zero; it is tool cost against the four orientation interviews you no longer need to schedule.

If you are working a vertical you do not know at all, run the corpus first and the calls second; the sequencing argument is in a one-week research plan for an unfamiliar vertical, and the fuller validation sequence sits in the SaaS idea validation checklist.

Closing thought

A good interview script is mostly deletions. Research earns its place by telling you which questions to delete, and the ones left standing are the few where a real person’s real week is the only available evidence.