Last updated: 17 September 2026
Quick Answer: A semi-structured interview follows a written guide of open questions but lets the interviewer probe, reorder and follow what the respondent actually says. The CDC's field epidemiology manual puts a typical topic guide at 10 to 15 broad questions with bulleted probes underneath. The guide sets the agenda; the interviewer decides the path.
Easy to state, hard to execute. Weller and colleagues, in PLoS One in 2018 across 28 interview datasets, found probing and prompting mattered more than the number of interviews, with ten interviews producing 95% of the most salient ideas when the interviewer listed exhaustively. What matters is how much room the guide leaves, and who may use it.
How Much Latitude Should the Guide Leave?
Enough that the interviewer can follow an unexpected answer, and not so much that two interviewers produce incomparable datasets. Harvard Library frames the middle position: the researcher holds more control over topics than in an unstructured interview, and, unlike a structured questionnaire, sets no fixed range of responses. How much latitude an AI moderator takes from the same guide is part of when AI-moderated interviews are reliable.
Designs differ less in their questions than in who holds each decision, and comparability is what you buy with structure and sell when you add latitude.
| Design | Who decides the wording | Who decides the order | Who decides when to probe | What you can compare afterward |
|---|---|---|---|---|
| Structured | Guide, fixed | Guide, fixed | Nobody; probes are scripted or absent | Counts across the whole sample |
| Semi-structured | Guide sets it, interviewer may rephrase | Interviewer, inside the guide's sections | Interviewer, against a written probing intent | Themes across the sample, plus the fixed core questions |
| Unstructured | Interviewer, live | Interviewer, live | Interviewer, live | Cases, not counts |
Which Probe Types Belong in the Guide?
Four, and writing them down separates a probing plan from improvisation. Generic encouragement produces generic answers, so probe text for anything nuanced belongs in the document rather than the moment, and a guide that names the type leaves less to whoever is in the room. How probe order changes when the room holds a group rather than one respondent is covered in focus group questions and their order.
- Elaboration. "Tell me more about that." University of Florida IFAS Extension calls this a good neutral prompt; it is the default because it adds no content of its own.
- Echo. Repeat the respondent's own word back, verbatim, then "and then what happened?" Using their word rather than a synonym is the whole technique; a synonym substitutes your category for theirs.
- Clarification. "I am not sure I understand." Admitting the gap draws a second, plainer explanation, and beats offering a guess for confirmation.
- Specification. "Which store, and was anyone with you?" Concrete detail separates a remembered event from a self-theorized account of behavior.
Laddering, a moderately structured method for surfacing means-end associations, goes further: ask why the previous answer matters, three times, climbing attribute to consequence to value, then stop, because a fourth rung usually produces a slogan.
A Worked Semi Structured Interview Guide, Annotated
A guide is a short spine of open questions, a probing intent under each, and a stated rule for when to depart. The CDC's field epidemiology manual describes the shape as a limited number, 10 to 15, of broad open-ended questions followed by bulleted points to facilitate optional probing. Below is one section, from a detergent-switching study.
A Real Guide Section
SECTION 3. The switch itself. Target 6 to 8 minutes.
Main question. "Walk me through the last time you bought laundry detergent and picked something different from what you usually get."
Planned probes, in this order, stopping when the account is concrete: 1. Elaboration. "What happened just before that?" 2. Specification. "Which store, and was anyone with you?" 3. Clarification. "I want to be sure I follow. What made that trip different from the one before?" 4. Echo. Repeat their own word, then "and then what?"
Ladder, only once the probes produce a reason rather than a story: 1. "What does that get you that the old one didn't?" 2. "And why does that matter on a normal week?" 3. "What would it mean if you couldn't get that?"
Leave the guide here if any of these three appear: 1. They describe switching back. A return is a stronger signal than the switch. 2. They name somebody else as the decider. 3. They mention price and cleaning performance in the same breath.
Follow whichever appears and skip Section 4; screening already captured what it asks.
Do not ask. "What do you like about the new one?" It assumes liking, offers no route to indifference, and returns a verdict where this section needs an account.
What the Annotation Adds
It names the probe type rather than trusting the interviewer to pick one, caps the ladder, and writes the departure rule down instead of leaving it to the room. Pilot it before fielding, which Kallio and colleagues number phase four of five. Two interviewers using this page will not run identical sessions, but their data sits in one table.
How Does Mode Change Who Answers and What They Say?
Mode decides who can take part and how candid they are once they do. A guide built for a 45-minute video session excludes anyone without a stable connection, a private room, or the confidence to be filmed, and none of those absences is random.
On candor, Pew Research Center randomly assigned 3,003 respondents to phone or web and found a mean difference of 5.5 percentage points across 60 questions, concentrated where somebody might want to look better to another person.
Mode can also widen the pool. WhatsApp-native interviewing runs inside the app, with no link and no install, and handles text, voice notes and code-mixed speech in one conversation, reaching respondents a browser session cannot. Outbound AI phone interviewing reaches people who use a smartphone for nothing but calls. Given the client's brief, Alchemic designs and tailors the discussion guide before fielding, publishes 57+ languages, and runs fieldwork across 14 markets including the USA and the UK.
When Is a Semi Structured Interview the Wrong Tool?
Whenever you need numbers you can project, or you do not yet know what to ask. Three situations call for something else, and each one is cheaper to catch in the brief than in the transcripts, where the remedy is a second study.
- You need counts with an error bar. A fixed instrument on a powered sample gives projectable percentages; a design whose stimulus varied by respondent cannot.
- You do not yet know the question. A guide only probes what somebody thought to put on it; exploratory work surfaces what you missed. See qualitative research methods.
- A listener would suppress the answer. Pew's gap above is the evidence; a self-administered mode sometimes beats any interview, however skilled.
Sample sizing is contested too. Hennink and Kaiser, in Social Science and Medicine in 2022, found saturation between 9 and 17 interviews, mostly in homogeneous populations with narrow objectives, thinner ground than the number's popularity suggests.