The best interviews you have watched did not read questions off a card. The interviewer listened, caught a word, asked what happened next, and waited for the real answer. A Cast interview works that way, inside rules that keep it from ever steering the person. This post covers those rules, the research each one rests on, and the other half of the interview: the AI user, who answers from a defined life and nothing else.
A map of topics, not a list of questions
Every Cast study has sixteen core topics, each with one fixed question: the person's week, the last time they dealt with the problem, what set it off, what it cost, how often it happens, how they handle it, what they put in place themselves, a time they tried something different, who decides, how they tell whether it is going well, the one thing they would change, their first reaction to the idea, what would stop them, what would have to change on their side, what they would need to see, and finally what we should have asked. The study's goals add topics of their own, and a pricing study asks for the person's own price points before it says any price. The whole interview takes about 25 minutes. The order is a funnel, general before specific and the person's life before any idea, following Schwarz (1999) (يفتح في علامة تبويب جديدة) on how earlier questions change the reading of later ones.
The wording of each topic is fixed and identical for every Ideal Customer Profile (ICP) in a study, because wording alone moves answers: Loftus and Palmer (1974) (يفتح في علامة تبويب جديدة) changed one verb in a question and changed what people remembered. If two groups heard different questions, a difference between them could be the wording and not the people. Nothing before the idea names it, who is behind it or what we hope to hear.
Two checkpoints, where the Director reads everything
Between the checkpoints, code asks the topics in order. An answer that comes back thin, very short or general where a real episode was asked for, gets one instant follow-up from a short list written for that topic, such as asking for the last time it happened or what they meant by their own word. Otherwise the conversation moves on.
Then, just before the idea is shown, the Director stops and reads the whole conversation so far. It writes three to five follow-ups at once, each pointing at the answer it comes from: a vivid moment worth walking through, a vague answer where a concrete one was asked for, two facts that cannot both be true (both quoted word for word), a thread the person raised that no topic covers, or an example of something they mentioned a few questions earlier. Each is asked in turn. At the end, after the idea, the price and the objections, it stops again and writes two to four more, free to go back to anything said before or after the idea. The last topic is the one experienced researchers close with: what should I have asked you that I didn't?
A checkpoint may open with a recap, and the recap has a rule: one sentence, twenty words or fewer, in the person's own words, and none at all when it adds nothing. "You said the bike sat for a week. What happened on the day it came back?" Code counts the words and checks that they are the person's own; a recap that fails is dropped and the question is asked bare. The rule comes from reflective listening in motivational interviewing, which Miller and Rollnick (يفتح في علامة تبويب جديدة) describe as restating the person's words so they can correct them, selectively and without judgment, and from a plainer concern: a transcript should never make you wade through the interviewer's summary of what you just read.
The deepening moves are the ones with the best evidence. Asking for a specific episode, not a generalization, is Flanagan's Critical Incident Technique (1954) (يفتح في علامة تبويب جديدة). Walking through what was going on just before the episode is context reinstatement from the cognitive interview, which a meta-analysis of 65 experiments (يفتح في علامة تبويب جديدة) found produces many more correct details with no rise in invented ones. Asking why something matters, only after a concrete instance exists, is laddering, from Reynolds and Gutman (1988) (يفتح في علامة تبويب جديدة). Reconstructing the timeline from first thought to decision is the jobs-to-be-done switch interview.
Hunting for what is different
A checkpoint is not written in a vacuum. The Director also receives two short inputs that the AI user never sees. The first is what sets this AI user apart from the rest of their group, drawn from the record by code: a role or a company size few others share, or a distinctive detail of their life. The second is a few lines of what the other AI users in the same group have already said on the same topics. The Director is told to prefer the moments where this person differs from what the group has covered well, and to chase what is missing. It never tells an AI user what others said, never compares them with anyone, and code rejects any follow-up that borrows four words in a row from another interview. The goal is a set of interviews that adds up to more than one interview repeated eight times.
A time it went the opposite way
One kind of follow-up exists to find what makes the difference. A contrast question sets two of the person's own moments side by side: a good week and a bad week, a time it went the opposite way, before and after a change they described, what they did and what they would do, or how they handle it and how someone they mentioned handles it differently. It uses only moments the person told us, in their own words. Contrast is also where mixed feelings go, because a person who values something and resents it is not contradicting themselves. Only facts that cannot both be true, a number, a date, how often, who did what, are put back to the person as something to square.
Curious, never leading
The Director may persist on the concrete, put two facts side by side by quoting both verbatim, and ask for the evidence behind a claim. It may not put an answer in the question, name a number first, voice an opinion or disbelief, characterize the person's situation, use a value word the person did not use first, or push toward a purchase. Each of those is a code check that runs on every follow-up before the AI user sees it, and a follow-up that fails is dropped. The research is clear on why: Dell and colleagues (2012) (يفتح في علامة تبويب جديدة) found people about two and a half times more likely to prefer what they thought the interviewer built; Tversky and Kahneman (1974) (يفتح في علامة تبويب جديدة) showed that any number becomes an anchor; and Sharma and colleagues (2024) (يفتح في علامة تبويب جديدة) showed that language model assistants tend to match what a user implies, so an interviewer that pushes an AI user gets agreement, not truth. The Director never sees an AI user's stance label or its record, and it never uses the study's hypothesis wording, so it cannot steer toward the answer it expects even by accident.
The AI user's side
An AI user plays one defined person and nothing else: no tools, no web search, no facts beyond its own life. The richer the record, the truer the answers. Park and colleagues (2024) (يفتح في علامة تبويب جديدة) built agents from two-hour interviews with 1,052 people and found they matched those people's survey answers at 83% of the people's own two-week consistency, ahead of agents built from demographics alone.
Long conversations have a weakness, though. A model asked about its own life tends to invent a detail and then tell it a little differently later, and Abdulhai and colleagues (2025) (يفتح في علامة تبويب جديدة) found that off-the-shelf models drift from their assigned personas and contradict what they said earlier. So each AI user gets two supports. Before the interview, a life story is written for them from their record alone: the people in their life, the plain figures they would know, and four to six key episodes with fixed details of when it happened, who was there, what it cost and what they did next. Where the record is vague, the story settles on one plain fact inside it. The AI user retells its story instead of inventing one. The story never mentions the idea and never changes the person's view of things, so an unconvinced person stays unconvinced.
During the interview, every answer is also checked against a running fact sheet. Whenever an AI user gives a concrete detail about itself for the first time, a person, a number, a date or a tool, it joins a list, and the list is shown to the AI user just before every later question: this is what you have already told this interviewer, keep it as you said it. A detail told twice is told the same way.
AI users are also told to correct a question that presumes something they did not say, to disagree when their own situation disagrees, to say "I don't know" rather than invent, to estimate like a person rather than compute, and to change their mind only for a reason. None of this makes an answer shorter. No AI user is told to be brief and nothing in the interview trims what they say, because a full, specific answer is what the follow-ups feed on. Afterward, a Critic from a different model family reads every line against the record, and every quote in the findings is verified word for word against the transcript by code.
Variation is the proof
Real people disagree, contradict themselves and have outliers. A synthetic cast that does not is a caricature. Bisbee and colleagues (2024) (يفتح في علامة تبويب جديدة) found that model-generated survey answers matched real averages but clustered far too tightly, and Cheng, Piccardi and Yang (2023) (يفتح في علامة تبويب جديدة) showed how persona simulations slide into exaggerated, flattened portraits. So Cast builds variety into the cast, with Fuzziness and a floor of skeptics of at least one in four, and then the checkpoints go looking for who differs. A cast that cannot disagree with itself is not telling you anything.
What a Cast interview can give you is what the evidence supports: rankings, directions, themes, the words people use and the reasons behind them. What it cannot give you is a percentage of a market. Anything about price or purchase intent is labeled directional and uncalibrated, and every count in a report is computed by code.
The house rules, with their sources, sit on the Script step of every study. The interview is where all of it is enforced, on every turn.
