{"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":"2026-09-22T10:58:54.241Z"},"content":[{"type":"documentation","id":"6c080489-68d4-4331-8c02-42c33186b6b6","slug":"epistemic-territory-interview-questions","title":"Epistemic Territory in Interviews: Ask People About What Is Theirs to Know (2026)","url":"https://www.koji.so/docs/epistemic-territory-interview-questions","summary":"Every question places the participant on a knowledge gradient, and answers come back in the same confident grammar whether or not the participant has access. Heritage's K+/K- distinction and Heritage and Raymond's work on agreement show how to design questions within the participant's own territory, grade evidence as did, saw, was told or think, remove interviewer assessments, and avoid known-answer questions.","content":"**Every interview question assigns the participant a position on a knowledge gradient -- someone who knows, or someone who is guessing -- and the answer comes back in the same grammar either way.** Conversation analysts call this *epistemics*: the study of who has the right to know what, and how talk displays it. The practical rule for researchers is short. **Ask people about what is theirs to know.** A participant's own experience, their own last decision and their own workflow are in their territory. Other people's motives, their company's future behaviour and *what users like you want* are not -- and the participant will still answer, fluently and confidently, with nothing behind it.\n\nThis guide covers where the idea comes from, how to map the territory of each question before you ask it, and how to read the agreement tokens that tell you whether a participant is confirming something from their own access or just going along with yours.\n\n## The research behind epistemic territory\n\n### Territories of knowledge\n\nThe linguist Akio Kamio proposed in 1997 that speakers treat information as falling inside or outside their *territory*, and choose their wording accordingly: direct statements for what is theirs, hedged or reported forms for what is not. John Heritage built on this in two 2012 papers in *Research on Language and Social Interaction* (45(1)), *Epistemics in action* and *The epistemic engine*, arguing that information imbalances are what drive conversational sequences in the first place.\n\n### K+ and K-, status and stance\n\nHeritage's key distinction is between two things that usually travel together and sometimes do not:\n\n- **Epistemic status** is a person's actual access to a domain -- K+ if they know more, K- if they know less. A customer is K+ about how they used your product yesterday. You are K+ about what the product was designed to do.\n- **Epistemic stance** is how a turn *presents* the speaker's knowledge. A question presents the asker as K-. A declarative presents them as K+. *You use the export feature weekly?* and *You use the export feature weekly.* ask for different kinds of response.\n\nThe research problem is the gap between them. A question can position a participant as K+ in a domain where their status is K-, and they will usually accept the position rather than disclaim it.\n\n### Agreement is not all the same\n\nHeritage and Geoffrey Raymond's 2005 paper *The terms of agreement* (Social Psychology Quarterly, 68, 15-38) studied what happens after someone makes an assessment. Going first with an assessment carries a small claim to authority. The second speaker can simply agree, or can use practices that mark their agreement as **independent** -- based on their own access, not on the first speaker's lead. That difference is visible in the words, and it matters enormously when the first speaker is the interviewer.\n\n## Mapping the territory of a question\n\nBefore a question goes into a guide, ask whose territory the answer lives in.\n\n| Question | Whose territory | What you get back |\n|---|---|---|\n| *Walk me through the last time you exported a report.* | Participant's -- first-hand, recent | Specific, checkable detail |\n| *Why did you pick this plan?* | Participant's, if they chose it | A reason, possibly rationalized, but theirs |\n| *Why did your company choose us?* | Often not theirs -- a buyer they may never have met | A plausible story in the grammar of fact |\n| *What do the rest of your team think of it?* | Others' -- reported at best | Secondhand opinion, filtered through the participant |\n| *Would people like you pay for this?* | Nobody's -- a prediction about a population | A guess delivered as a claim |\n| *What should we build next?* | Yours, being handed to them | Feature requests that answer your job, not theirs |\n\nThe last four rows are not bad questions to *want answered*. They are bad questions to ask *this person*. The fix is either to find the person whose territory it is -- the economic buyer, the teammate, a sample of the population -- or to reframe the question into the participant's own territory: not *what does your team think?* but *what has your team said to you about it, and when?*\n\n### Grade findings by access\n\nOnce you think in territories, it becomes natural to tag each piece of evidence by how the participant came to know it:\n\n1. **Did** -- first-hand experience (*I exported it twice last week*).\n2. **Saw** -- witnessed (*I watched my manager try and give up*).\n3. **Was told** -- reported (*finance told me it was the price*).\n4. **Think** -- inferred or predicted (*I guess most people would*).\n\nA finding supported mainly by level 4 answers from twelve participants is not twelve pieces of evidence about the thing. It is twelve pieces of evidence about what participants imagine. Reporting the access level alongside the quote is one of the cheapest rigor upgrades available to a research team.\n\n## When the interviewer goes first\n\nInterviewers constantly make small assessments without meaning to: *that sounds frustrating*, *so the setup was pretty slow?*, *a lot of people find that confusing*. Each one positions the interviewer as having a view, and gives the participant something to agree with.\n\n### Three ways to say yes\n\nHeritage and Raymond's work gives you a way to read the reply. Compare three responses to *So setup was pretty slow?*:\n\n- **Yeah.** A minimal agreement. It accepts the interviewer's framing without adding access of its own. Weak evidence -- it may be politeness.\n- **It was slow, yes.** A confirmation that repeats the claim in the participant's own words before agreeing. The participant is asserting the matter as theirs, not merely going along.\n- **Oh, it was terrible -- it took us three weeks.** An upgraded, independent assessment with detail the interviewer did not supply. Strong evidence: the participant has shown their own access.\n\n**The strength of an agreement is readable from its form.** A transcript full of *yeah* after leading assessments has less evidence in it than its word count suggests, because the participant has been ratifying the interviewer's view rather than reporting their own. The better fix is not better reading but better asking: *How long did setup take?* has no assessment in it to agree with.\n\n## The known-answer problem\n\nThe reverse trap is asking a question the participant knows you already know the answer to -- their plan tier, their company size, what they told the sales team last month. In ordinary conversation, a K- question about something the asker is plainly K+ on is odd, and participants respond to the oddness: they give short answers, they wonder what the *real* question is, or they treat it as a test.\n\n### Confirm instead of asking\n\nResearchers who have looked at this in qualitative interviews recommend acknowledging it -- *I know we covered this earlier, but could you tell me again...* -- which resolves the oddness. A better fix is to not ask at all: state what you know and invite correction (*You moved to the annual plan in March -- is that right?*), which puts the participant in the K+ position on the part that is genuinely theirs: whether the record is accurate.\n\n## Designing a guide around territory\n\n- **Open in the participant's territory.** Start with recent, first-hand, concrete behaviour. It is the easiest ground to be expert on and warms up detail for everything after it.\n- **Move out of it explicitly.** When you need a second-hand view, say so: *You may not know this directly, but what have you heard about why the decision was made?* The permission to not know makes a *don't know* possible, and makes a real answer more trustworthy.\n- **Route population questions to populations.** *Would people pay for this?* belongs in a survey of the people, not in one person's mouth.\n- **Remove your own assessments.** Replace *that sounds frustrating -- was it?* with *what was that like?* so any assessment in the transcript is the participant's.\n- **Do not ask what you already know.** Confirm it instead.\n\n## Common mistakes\n\n- **Treating fluency as access.** Participants answer out-of-territory questions just as smoothly as in-territory ones.\n- **Counting agreement tokens as evidence.** *Yeah* after your own assessment is ratification, not testimony.\n- **Asking one person about a population.** A prediction about *people like me* is a guess with a sample size of one.\n- **Asking known-answer questions.** They make the interview feel like a test and shorten the answers.\n- **Reporting quotes without access level.** *Did*, *saw*, *was told* and *think* are four different grades of evidence.\n\n## How Koji applies epistemic question design\n\nAn interview guide can be written with territory in mind, but it is the live follow-up where epistemic drift happens -- the unscripted *so that was frustrating?* that plants an assessment. That is where an AI interviewer can be more disciplined than a tired human moderator on their eighth session.\n\n- **Consistent, non-assessing follow-ups.** Koji's AI interviewer runs voice and text interviews with the same follow-up discipline in every session, asking *what happened next?* and *how long did that take?* rather than offering its own evaluations for participants to agree with.\n- **Probing for access.** When a participant makes a claim about someone else, Koji's AI follow-up questions can ask how they know -- *did you see that, or did someone tell you?* -- so the access level ends up in the transcript.\n- **Context instead of known-answer questions.** Putting what you already know about the audience into the study brief lets the AI interviewer confirm facts rather than quiz participants on them.\n- **Population questions go to structured items.** Koji's [structured questions](/docs/structured-questions-guide) support six types -- open_ended, scale, single_choice, multiple_choice, ranking and yes_no. A willingness-to-pay or preference question asked as a scale or single_choice item across the whole sample gives you a population answer instead of asking each person to guess one.\n- **Analysis that keeps the quote.** Koji's automatic analysis links every theme to the underlying quotes, so you can check whether a finding rests on *did* answers or *think* answers before you present it.\n\nTraditional survey platforms like SurveyMonkey or Qualtrics fix the wording but cannot follow up on a claim to ask how the participant knows it. Moderated interviews can, but depend on the moderator remembering to. Koji gives you the follow-up of an interview with the consistency of a survey, at a fraction of the time.\n\n## Frequently asked questions\n\n### What is epistemic territory in an interview?\n\nIt is the domain of knowledge a participant has a genuine right to speak about -- their own experiences, decisions and workflows. Questions inside that territory get answers backed by access. Questions outside it, such as other people's motives or predictions about a market, get answers in the same confident grammar with nothing behind them.\n\n### What do K+ and K- mean?\n\nThey are John Heritage's labels for relative epistemic status: K+ is the more knowledgeable party in a domain and K- the less knowledgeable. Questions present the asker as K- and the recipient as K+. Research problems arise when a question positions a participant as K+ in a domain where they are actually K-.\n\n### Why is yeah weak evidence in an interview?\n\nBecause a minimal agreement after the interviewer's own assessment ratifies the interviewer's view rather than reporting the participant's. Heritage and Raymond showed that speakers who have independent access tend to mark it -- by confirming in their own words or upgrading with new detail. Those forms are much stronger evidence than a bare yeah.\n\n### How do I avoid asking questions participants cannot answer?\n\nMap the territory of each question before you ask it. If the answer belongs to someone else, find that person or reframe the question into what the participant has directly seen or been told. If it is a prediction about a population, ask the population with a structured question, as Koji's scale and single_choice types allow.\n\n### What is a known-answer question and why does it cause problems?\n\nIt is a question the participant knows the interviewer can already answer, such as their plan tier. It makes the exchange feel like a test and tends to produce short answers. State what you know and ask the participant to confirm or correct it instead.\n\n### Should I grade interview evidence by how the participant knows it?\n\nYes. Tagging quotes as did, saw, was told or think separates first-hand evidence from reports and guesses. A theme supported mostly by think answers is evidence about what participants imagine, not about what happened. With Koji, every theme links back to its quotes, which makes that check quick.\n\n## Related Resources\n\n- [Structured questions in AI interviews](/docs/structured-questions-guide) -- sending population questions to scale and single_choice items.\n- [Presupposition in interview questions](/docs/presupposition-interview-questions) -- what your question asserts before anyone answers.\n- [What participants mean but never say](/docs/conversational-implicature-interview-answers) -- reading the answer side of the exchange.\n- [Backchannels and continuers in research interviews](/docs/backchannels-continuers-research-interviews) -- the listener's minimal tokens.\n- [Avoiding leading questions](/docs/avoiding-leading-questions) -- keeping your assessments out of the transcript.\n- [Expert interviews guide](/docs/expert-interviews-guide) -- when the territory you need belongs to a specialist.\n","category":"Interview Techniques","lastModified":"2026-09-22T03:26:50.848907+00:00","metaTitle":"Epistemic Territory: Interview Question Design","metaDescription":"Participants answer out-of-territory questions as fluently as in-territory ones. How to design questions by who has the right to know.","keywords":["epistemic territory","interview question design","k+ k- epistemics","heritage raymond","agreement tokens","known-answer questions","evidence grading"],"aiSummary":"Every question places the participant on a knowledge gradient, and answers come back in the same confident grammar whether or not the participant has access. Heritage's K+/K- distinction and Heritage and Raymond's work on agreement show how to design questions within the participant's own territory, grade evidence as did, saw, was told or think, remove interviewer assessments, and avoid known-answer questions.","aiPrerequisites":["Experience writing interview guides"],"aiLearningOutcomes":["Map the epistemic territory of each interview question","Grade evidence by access level","Read minimal versus independent agreement","Replace known-answer questions with confirmations"],"aiDifficulty":"intermediate","aiEstimatedTime":"12 min"}],"pagination":{"total":1,"returned":1,"offset":0}}