{"site":{"name":"Koji","description":"AI-native customer research platform that helps teams conduct, analyze, and synthesize customer interviews at scale.","url":"https://www.koji.so","contentTypes":["blog","documentation"],"lastUpdated":"2026-09-20T17:13:30.924Z"},"content":[{"type":"documentation","id":"20f06f40-465d-4a54-ac61-246f58c1eb58","slug":"repair-clarification-research-interviews","title":"When Participants Ask What You Meant: Repair in Research Interviews (2026)","url":"https://www.koji.so/docs/repair-clarification-research-interviews","summary":"Other-initiated repair occurs about once every 1.4 minutes in conversation across 12 languages, so a 20-minute interview contains roughly fourteen unscripted clarification exchanges that never appear in the discussion guide. Speakers prefer the most specific repair type available, and the most specific type -- a restricted offer -- supplies a candidate answer and asks for confirmation, silently converting an open question into a yes-or-no one while the transcript still shows the original open question. The fix is a pre-written clarification per question: a verbatim restatement, one scripted gloss, and an explicit refusal to narrow.","content":"Your interview guide contains fifteen questions. Your study actually asked several thousand, because every time a participant said *sorry, what do you mean?* a moderator improvised an answer, and that answer -- not the written question -- is what the participant responded to. Conversation researchers call this **repair**, it happens roughly once every 1.4 minutes in ordinary talk, and it is almost never written down, scripted, or audited.\n\nThe short version: **script one clarification per question before you field, and never repair by offering a candidate answer.** The rest of this article explains why those two rules are the ones that matter.\n\n## Repair is constant, universal, and invisible in your guide\n\nThe most comprehensive study of the phenomenon analysed conversation in 12 languages from 8 language families -- deliberately chosen across typological profiles, including languages with no historical contact -- looking for a system by which a listener signals trouble and the speaker fixes it.\n\nThey found one everywhere. In 48.5 exhaustively sampled hours of conversation they identified 2,053 cases, which works out to \"a repair initiation about once every 1.4 minutes across all languages.\" Counting only independent sequences, 95 percent of repair initiations occur within 4.13 minutes of the previous one. Their summary is that no five minutes go by in any of the languages without a repair sequence occurring.\n\nNow put that base rate against a research study. A 20-minute interview contains roughly fourteen opportunities for the participant to signal that something did not land. Run 200 interviews and you are looking at something on the order of 2,800 unscripted micro-exchanges, each one a small renegotiation of what a question meant -- none of which appear in the discussion guide, and most of which are handled differently by each moderator, or differently by the same moderator at 9am and at 5pm.\n\nConversational base rates will not transfer exactly to a research interview; participants reading a question on screen, or talking to a system they expect to be literal, behave differently from friends on the telephone. Treat fourteen-per-interview as an order of magnitude, not a forecast. The order of magnitude is the point.\n\n## Three ways to ask what you meant, and only one is safe\n\nThe same study found that all 12 languages use the same three practices for initiating repair, and it ranks them by how much information they contain about the trouble:\n\n| Type | What it does | Example | Mean length |\n| --- | --- | --- | --- |\n| Open request | Signals a problem but leaves open what or where it is | *Huh?* | 3.7 characters |\n| Restricted request | Asks about one specific component of the trouble | *Who?* | 10.4 characters |\n| Restricted offer | Offers a candidate for what was said and asks for confirmation | *she had a boy?* | 13.0 characters |\n\nPeople systematically choose the most specific type they can manage rather than the cheapest. The authors note this is \"a non-trivial result: the locally egocentric behaviour for a person performing other-initiated repair would be to choose the simplest form possible (e.g., 'huh?') and let the other do most of the work.\" Instead, speakers do the work themselves. They describe the principle as revealing \"an element of altruism in how people initiate repair in conversation.\"\n\nFor ordinary conversation, that cooperative instinct is a feature. For a research interview, the most specific type is the dangerous one, and this is the whole trap:\n\n**A restricted offer is a candidate answer wearing a clarification costume.**\n\nWhen a participant says *what do you mean by workflow?* and the moderator replies *like, is it the approval steps that slow you down?*, the moderator has offered a candidate. The participant now confirms or denies a specific proposition instead of answering an open question. The interview has silently converted an open_ended question into a yes_no one -- and confirming is by far the cheaper move, for the same structural reason covered in [acquiescence bias](/docs/acquiescence-bias) and in the analysis of [response timing and hedged answers](/docs/dispreferred-responses-interview-timing).\n\nAnd here is the part that makes it a measurement problem rather than a moderation problem: **the transcript still shows the original open question.** Your guide says you asked *walk me through your workflow*. The analyst codes the response as an answer to that. Nobody in the analysis chain can see that the participant was really answering *is it the approval steps?* unless they read every transcript line by line, which at 200 interviews nobody does.\n\n## The repair script: the artifact almost nobody writes\n\nThe fix is unglamorous and takes about twenty minutes for a fifteen-question guide. For each question, write down in advance:\n\n1. **A verbatim restatement.** The same question, read again, unchanged. This is the first response to any open request, because *Huh?* usually means the participant did not hear it, not that they did not understand it.\n2. **One scripted gloss.** A single pre-agreed sentence defining the term the participant is most likely to query -- written before fieldwork, by the person who owns the research question, not improvised by whoever is moderating.\n3. **An explicit refusal to narrow.** A closing line such as *whatever that word means to you is what I am after* that hands the interpretation back rather than supplying one.\n\nThen apply one rule during fieldwork: **repair by restating or by glossing, never by offering a candidate.** If a participant pushes for an example, give an example from a domain that is not the study domain, or decline. The moment you name a plausible answer, you have scored your own question.\n\nWhere a genuine ambiguity shows up repeatedly -- three or four participants querying the same word -- that is not a moderation problem, it is a defect in the instrument, and the right response is to fix the question and note the change rather than keep patching it live. This is closely related to the premise problem: a term people keep querying often turns out to carry an unstated assumption, which is the subject of [presupposition in interview questions](/docs/presupposition-interview-questions).\n\n## What repair costs, and why participants do it anyway\n\nOne more finding is worth carrying into study design. Repair is efficient: the two-utterance repair sequence is, on average, about as long as the single utterance it is fixing. The disruption is small, and participants pay it willingly because the alternative -- answering the wrong question -- is worse for both parties.\n\nThat has a direct implication for how you should feel about repair in your data. **A study with visible repair sequences is healthier than one without.** Zero clarification requests across 200 interviews does not mean your questions were clear. It much more likely means participants were guessing at your meaning and answering anyway, which is the same failure with no observable trace. Repair requests are your instrument telling you where it is ambiguous, for free.\n\nSo the goal is not to eliminate repair. It is to make sure every repair is handled identically, and that the handling is recorded.\n\n## How Koji handles repair\n\nThree things about the platform apply directly.\n\n**The clarification is scripted and identical.** Koji's AI moderator works from your study brief and the written question set, so a participant who asks what a term means gets a consistent answer rather than whatever the moderator on shift improvises. The same gloss reaches participant 3 and participant 197. That is the difference between a documented instrument and an undocumented one, and it is the same mechanism described in [interviewer variance and moderator drift](/docs/interviewer-variance-moderator-consistency).\n\n**Every repair is in the transcript.** A human moderator's clarification often happens as an aside, off the recording, or in a side comment that never makes it into the artifact of record. In a Koji interview, both the participant's request and the AI's response are turns in the logged conversation, which means the analyst can see exactly which question each answer was answering -- and can search for the ones that got queried repeatedly.\n\n**Probing is open rather than candidate-offering.** Koji's AI follows up on what the participant actually said instead of running a fixed list of follow-ups, so the natural repair behaviour is restatement and open pursuit rather than *is it X?* The mechanics of that are in [probing questions in user interviews](/docs/probing-questions-user-interviews), and the quality of each resulting conversation is scored automatically on a 1-to-5 scale, with only conversations scoring 3 or above consuming a credit.\n\nChoosing question types deliberately also reduces the surface area for trouble. Koji supports six -- open_ended, scale, single_choice, multiple_choice, ranking, and yes_no -- and the most repair-prone construction is an open_ended question resting on a piece of internal vocabulary. Where the concept is genuinely closed, a single_choice or ranking question makes the answer space explicit and removes the ambiguity rather than glossing it mid-interview. The [structured questions guide](/docs/structured-questions-guide) covers when each type is the right call.\n\n## Frequently asked questions\n\n### What is repair in a research interview?\n\nRepair is the conversational system by which a listener signals that something did not land and the speaker fixes it -- a participant asking *sorry, what do you mean?* and the moderator responding. It occurs about once every 1.4 minutes in ordinary conversation across all 12 languages studied, which means a 20-minute interview will typically contain a dozen or more of these unscripted exchanges.\n\n### Why is offering an example a bad way to clarify a question?\n\nBecause it converts an open question into a yes-or-no one. The most specific repair type, a restricted offer, supplies a candidate answer and asks for confirmation -- and confirming is structurally cheaper than disagreeing. The transcript still shows your original open question, so nobody downstream can see which question was actually answered.\n\n### What should a moderator say when a participant asks what a question means?\n\nRestate the question verbatim first, since a request for clarification often just means it was not heard. If that does not resolve it, deliver a single pre-written gloss, then explicitly hand interpretation back with something like *whatever that means to you is what I am after*. Never name a plausible answer.\n\n### Is it bad if participants ask for clarification often?\n\nNo -- the opposite. Repair requests tell you exactly where your instrument is ambiguous. A study with zero clarification requests across hundreds of interviews more likely means participants guessed at your meaning and answered anyway, which is the same failure with no visible trace.\n\n### How do I make clarifications consistent across a large study?\n\nWrite one clarification per question before fielding: a verbatim restatement, one scripted gloss, and a line that refuses to narrow. With Koji, the AI moderator delivers that scripted clarification identically to every participant and logs both the request and the response in the transcript, so consistency is a property of the study rather than of who was on shift.\n\n### Does repair happen in text interviews too?\n\nYes, though the pattern differs -- participants can re-read a written question, so hearing problems disappear while vocabulary problems remain. The candidate-offering trap is identical in both modalities, and because Koji logs every turn in text and voice interviews alike, you can audit which questions triggered clarification requests in either format.\n\n## Related Resources\n\n- [Structured questions guide](/docs/structured-questions-guide) -- picking among the six question types, and when a closed type removes an ambiguity instead of glossing it.\n- [Presupposition in interview questions](/docs/presupposition-interview-questions) -- why a term participants keep querying often carries an unstated premise.\n- [Probing questions in user interviews](/docs/probing-questions-user-interviews) -- following up on what was said without supplying a candidate answer.\n- [Acquiescence bias](/docs/acquiescence-bias) -- why confirming a candidate is structurally cheaper than rejecting it.\n- [Response timing and hedged answers](/docs/dispreferred-responses-interview-timing) -- what a pause before an answer does and does not tell you.\n- [Handling difficult interview participants](/docs/handling-difficult-interview-participants) -- field patterns for vague answerers and participants who resist the question.\n","category":"Research Methods","lastModified":"2026-09-20T03:24:31.239085+00:00","metaTitle":"Repair in Research Interviews: When Participants Ask","metaDescription":"Participants ask for clarification about once every 1.4 minutes. Why offering an example scores your own question, and how to script repair first.","keywords":["participant misunderstood the question","other-initiated repair","clarification bias research","interview clarification script","candidate understanding interviews","moderator improvisation","interview guide clarity"],"aiSummary":"Other-initiated repair occurs about once every 1.4 minutes in conversation across 12 languages, so a 20-minute interview contains roughly fourteen unscripted clarification exchanges that never appear in the discussion guide. Speakers prefer the most specific repair type available, and the most specific type -- a restricted offer -- supplies a candidate answer and asks for confirmation, silently converting an open question into a yes-or-no one while the transcript still shows the original open question. The fix is a pre-written clarification per question: a verbatim restatement, one scripted gloss, and an explicit refusal to narrow.","aiPrerequisites":["Experience running or designing moderated interviews"],"aiLearningOutcomes":["Recognise the three repair initiator types and which one is unsafe in research","Write a repair script of restatement, gloss and refusal for each question","Read clarification requests as instrument diagnostics rather than noise","Audit transcripts for questions that were silently narrowed mid-interview"],"aiDifficulty":"intermediate","aiEstimatedTime":"12 min"}],"pagination":{"total":1,"returned":1,"offset":0}}