{"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":"2026-10-12T15:11:04.773Z"},"content":[{"type":"documentation","id":"8c119c55-d389-4b8a-a22e-64ae388ace0b","slug":"tracing-report-findings-to-evidence","title":"Tracing Report Findings Back to the Evidence","url":"https://www.koji.so/docs/tracing-report-findings-to-evidence","summary":"Koji reports link every claim to the interviews behind it. Underlined claims in the summary, finding cards (with a percentage, an N of M count and a View supporting responses button), clickable charts for all six structured question types, recommendations and personas all open an evidence panel beside the report. The panel shows the clicked text, a computation line generated from the data, and the supporting interviews by participant number with highlighted quotes. Clicking a row opens the full transcript scrolled to the exact message. Shared reports show the evidence panel but not full transcripts. Reports leave out previews, test runs and the lowest-effort interviews, and generating or refreshing a report uses 5 credits.","content":"To check where a finding in a Koji report comes from, click it. Every underlined claim in the summary, every finding card's **View supporting responses** button, every bar, slice and scale cell in a chart, and every recommendation opens an evidence panel beside the report. It lists the interviews behind that claim, each with the participant's own words highlighted, and one more click opens the full transcript, scrolled to the exact message. You never have to take an AI summary on trust: the people behind each sentence are one click away.\n\nThis guide covers where the evidence lives in a report, how to read the panel, and a quick routine for checking a finding before you share it.\n\n## Why traceable findings matter\n\nAI analysis is fast, and that speed is only useful if people believe the output. Stakeholders who weren't in the interviews will ask the same question every researcher has heard: \"How many people actually said that?\" A summary that can't answer it gets politely ignored.\n\nTrust in AI output is still low. In the KPMG and University of Melbourne global study [Trust, attitudes and use of artificial intelligence: A global study 2025](https://doi.org/10.26188/28822919), which surveyed over 48,000 people in 47 countries, only 46% said they were willing to trust AI systems. Research teams feel that gap directly. The fix isn't a disclaimer. It's showing the work.\n\nTraditional survey tools give you a chart and stop there. If 34% picked \"too expensive\", you know the number but not the story behind it. Koji keeps the number and the story together: every figure in a report is linked to the interviews it was counted from, so you can read why people said what they said, in their words.\n\n## Where the evidence lives in a report\n\nA Koji report opens on the **What did we learn?** tab. Evidence links run through the whole page.\n\n### Highlighted claims in the summary\n\nThe summary at the top of **What did we learn?** is written as a short story of what you found. Claims that rest on specific interviews are underlined. Hover over one and it highlights; click it (or press Enter when it has keyboard focus) and the evidence panel opens with that sentence as its description. Screen readers announce each claim together with how many interviews support it.\n\nThe same clickable claims appear in finding-card bullets, recommendation descriptions and the **From the follow-ups** notes under structured questions.\n\n### Finding cards\n\nEach finding card shows:\n\n- **A prevalence ring** with a percentage and an \"N of M\" count: how many participants support the finding out of the interviews in the report.\n- **A label** that reflects priority and reach: **Critical risk**, **High-impact finding**, **Notable pattern** or **Emerging signal**.\n- **The headline and key points.**\n- **View supporting responses**, which opens the panel with every interview behind the finding.\n\nCards are sorted by how many participants support them, so the most widely shared findings come first.\n\n![Koji high-impact finding card backed by 26% of participants with a link to supporting responses](/email/shots/shot-finding.jpg \"A finding card: the share of participants behind it, and one click to their answers\")\n\n### Charts\n\nEvery chart in **Findings by question** and the appendix is clickable:\n\n- **Bar charts** (themes, single choice, multiple choice and ranking questions): click a bar to see the interviews counted in it. Theme bars also show the phrases participants actually used, since related phrasings are grouped into one theme.\n- **Pie charts** (yes/no questions and NPS breakdowns): click a segment or its legend item.\n- **Scale charts** (ratings, NPS, CSAT): click a score to see who gave it, and why.\n- **Notable quotes**: click a quote to see it in context.\n\nEach chart also carries a sample-strength badge, from **Insufficient sample** to **Strong sample**, so you can see at a glance how much weight a split can bear. Koji supports six structured question types (open-ended, scale, single choice, multiple choice, ranking and yes/no), and each one gets its own chart; see the [structured questions guide](/docs/structured-questions-guide).\n\n### Recommendations\n\nOn the **What can you do?** tab, each action shows a supporter count and a confidence label (**High confidence**, **Moderate confidence** or **Low confidence**). Hover the label to see \"N of M participants\". Expand a recommendation to see the theme it's **Grounded in**, a representative quote, and **View supporting responses**.\n\n### Personas\n\nOn the **Personas** tab, each persona's participant count is a button. Click it to open the interviews behind that persona, and click the **In their words** quote to open the interview it came from.\n\n## Reading the evidence panel\n\nOn a desktop screen the panel opens to the right and pushes the report aside, so the claim and its evidence stay side by side. On smaller screens it opens full screen.\n\nTop to bottom, the panel shows:\n\n1. **Description:** the exact text or chart label you clicked.\n2. **Computation details:** a plain line explaining how the figure was counted, for example \"Drawn from 11 interviews the analysis identified as supporting this claim.\" This line is generated from the data, not written by the AI, so it always matches the count.\n3. **Supporting interviews (N):** one row per interview, labelled by participant number (#7, #12), with the supporting quote highlighted. Participant names don't appear in these rows.\n\nFor choice questions in the **Question Coverage** section of the appendix, the panel groups people by the option they picked under **Responses by answer**, so you can compare the reasons behind each choice.\n\n![Koji report summary with a side panel of supporting participant quotes behind one finding](/email/shots/shot-evidence-panel.jpg \"Click a claim and the quotes behind it open beside the report\")\n\n### Opening the full transcript\n\nClick any row to open that interview inside the panel, without leaving the report. It has two tabs, **Insights** and **Transcript**. When the evidence points to a specific message, the panel opens on **Transcript** and scrolls straight to it, with the supporting phrase highlighted. Use **Back** to return to the list.\n\nReading the turn before and after a quote is the fastest way to check it wasn't taken out of context. It also shows you how the AI interviewer's follow-up questions drew the answer out, which is often where the most useful detail sits.\n\n## A five-minute routine for checking a finding\n\nBefore you put a finding in a deck or a roadmap, run through this:\n\n1. **Check the count.** Look at the \"N of M\". Two people out of forty is a lead worth following, not a conclusion.\n2. **Read the quotes, not just the headline.** Do the highlighted words actually say what the finding claims?\n3. **Open two transcripts.** Pick one that agrees strongly and one that's borderline. Read the surrounding turns.\n4. **Look for the counter-signal.** Check the chart for the same question: did a meaningful group say the opposite?\n5. **Mind the sample badge.** If a split is marked **Limited sample**, report it as directional.\n\n![Checklist for verifying an AI research finding: check the count, read the quotes, open two transcripts, look for the counter-signal and mind the sample size.](/visuals/checklist-for-verifying-ai-research-findings/checklist-for-verifying-ai-research-findings.png \"Run this before a finding goes into a deck or a roadmap\")\n\n## Which interviews a report includes\n\nA report counts completed interviews that have been analysed. Previews and test runs never appear, and the lowest-effort interviews are filtered out by the [quality gate](/docs/how-the-quality-gate-works) (you aren't charged for those either). When interviews were filtered out, the **Participants** card shows \"+N filtered out\", so the \"M\" in every \"N of M\" matches the interviews actually in the report.\n\nGenerating or refreshing a report uses 5 credits. **Regenerate** stays disabled until new qualifying interviews arrive, so you won't spend credits on an identical report. Each refresh creates a new version, and the evidence links always point at the interviews in that version.\n\n## Evidence on shared reports\n\nWhen you publish a report, people with the link see the same clickable summary, finding cards, charts and recommendations, and they can open the evidence panel with its quotes and computation details. They can't open full transcripts: those stay inside your workspace. Quotes in recommendations and notable quotes show the participant's display name if one was collected, otherwise an anonymous participant ID, so think about what you collect in your intake form before you share widely. See [Publishing and sharing reports](/docs/publishing-sharing-reports).\n\n## Tips for presenting traceable findings\n\n- **Quote the count with the claim.** \"11 of 42 participants (26%) said permission mapping slowed rollout\" lands better than \"many users\".\n- **Link, don't screenshot.** A published report lets a sceptical stakeholder click through to the quotes themselves.\n- **Pair a number with a voice.** One verbatim quote next to a percentage makes both more believable.\n- **Say what the sample can't tell you.** Naming the limits makes the rest of your findings more credible, not less.\n\n## Related Resources\n\n- [Structured Questions Guide](/docs/structured-questions-guide)\n- [Reading Your Research Report](/docs/reading-your-research-report)\n- [Generating Research Reports](/docs/generating-research-reports)\n- [Publishing and Sharing Reports](/docs/publishing-sharing-reports)\n- [How the Quality Gate Works](/docs/how-the-quality-gate-works)\n- [Viewing Interview Transcripts](/docs/viewing-interview-transcripts)\n","category":"Analysis & Synthesis","lastModified":"2026-10-12T04:14:11.355003+00:00","metaTitle":"Trace AI Research Findings to the Evidence | Koji Docs","metaDescription":"Every finding, chart and recommendation in a Koji report opens the interviews behind it. See supporting quotes, open transcripts and check findings in a click.","keywords":["traceable ai research findings","ai report citations","verify ai insights","supporting quotes research report","evidence behind research findings"],"aiSummary":"Koji reports link every claim to the interviews behind it. Underlined claims in the summary, finding cards (with a percentage, an N of M count and a View supporting responses button), clickable charts for all six structured question types, recommendations and personas all open an evidence panel beside the report. The panel shows the clicked text, a computation line generated from the data, and the supporting interviews by participant number with highlighted quotes. Clicking a row opens the full transcript scrolled to the exact message. Shared reports show the evidence panel but not full transcripts. Reports leave out previews, test runs and the lowest-effort interviews, and generating or refreshing a report uses 5 credits.","aiPrerequisites":["A Koji study with completed interviews","A generated research report"],"aiLearningOutcomes":["Open the evidence behind any claim, finding, chart or recommendation","Read the evidence panel and its computation line","Jump from a quote to the exact moment in the transcript","Check a finding in five steps before sharing it","Know what viewers of a shared report can see"],"aiDifficulty":"beginner","aiEstimatedTime":"8 minutes"}],"pagination":{"total":1,"returned":1,"offset":0}}