Finding What a Customer Said Across Your Interview Transcripts

TL;DR: Koji does not have a search box over your transcripts, and there is no cross study search inside the app. What it has instead is analysis that happens per interview, a report where every claim links to the quote behind it, a responses grid, readable transcripts, CSV and JSON export, and an MCP connector so Claude, Cursor or any MCP assistant can read your studies and answer questions across them. In practice you find the moment a customer said the thing by following a claim to its evidence, not by typing keywords.

---

Why Finding Things Again Is the Hard Part

The dirty secret of customer research is that most of the value of a past interview is locked away because nobody can get back to it. A PM asks "did anyone mention onboarding friction last quarter?" and the honest answer is "maybe, let me re-read 40 transcripts."

The usual fix is a search box, which sounds right and mostly is not. Keyword search only finds the words you guessed. Semantic search finds things you cannot verify. And either way you land in a list of fragments with no idea which conversation mattered or whether the participant was in your segment.

Koji attacks the same problem from the other end. Rather than making you go looking, it analyses every interview as it lands and puts the finding in front of you with its evidence attached. Getting back to a specific moment then means following a claim down to the quote, which takes a click and cannot mislead you, instead of guessing a phrase.

---

What You Actually Have

Six surfaces, and it is worth knowing which one answers which kind of question.

1. Per interview analysis

Every completed interview is analysed on its own, immediately. That produces a quality score from 1 to 5, structured answers for each question, the themes in that conversation with the quotes that support them, a summary and a sentiment. No tagging backlog, no taxonomy to maintain, and no week of a researcher's time before anything is findable. It works from your first interview.

2. The report, which opens with the answer

The research report leads with key takeaways, then themes with their evidence counts, then a section per question. Before you go hunting for a quote, read the theme list. A theme that appears in 14 of 15 interviews and one that appears in 3 are different findings, and the count tells you which you are looking at.

Most "did anyone ever mention X" questions are answered here in ten seconds, because if enough people mentioned X it is a theme with a name.

3. The evidence panel: claim to quote to transcript

This is the part that replaces search for most day to day work.

Claims in the report carry citations. Click one and an evidence panel opens next to the text with the participant quotes that support that claim, and a link straight into the full transcript where it was said. The charts work the same way, so a bar in a distribution leads back to the people inside it.

The practical effect: you never have to trust a summary and you never have to search for its basis. "Six participants raised concerns about pricing" is one click from the six quotes and one more from the whole conversation around any of them. That is the workflow that actually replaces re-reading.

4. The responses grid

The study's Responses view is a grid: one row per respondent, one column per question. Rows sort by date or by quality score, and the view downloads as CSV.

Use it when your question is "what did everyone say to question four", which is most comparative questions. Reading down one column is faster than any query you could have typed, and you see the whole distribution rather than the top matches.

5. Readable transcripts

Every interview can be read end to end, with its structured answers and the follow-ups the AI interviewer asked in place. The evidence panel drops you here with the surrounding conversation intact, which is the context a search result fragment always lacks.

6. Personas and question coverage

The report groups participants into personas and shows how well each question in your brief was actually covered across the interviews. Personas answer the segment shaped questions, "did enterprise users see this differently", without you filtering anything by hand. Question coverage tells you where the evidence is thin before you build an argument on it.

---

Across Studies: Ask an Assistant Over MCP

A report belongs to one study. When the question spans several, that is the job for the Koji MCP connector.

Connect Koji at https://www.koji.so/api/mcp/mcp over OAuth in Claude, Claude Code, Cursor or any Model Context Protocol client. The assistant then has four reading tools:

Then ask in plain language:

• "List my studies and tell me which ones touched pricing."
• "Read the churn study and the win-loss study and tell me where they disagree."
• "Has any participant in the last year asked for a shared workspace? Quote them."
• "Summarise the onboarding friction finding for a leadership update."

The assistant answers from the studies and transcripts it actually retrieved. Any quote it gives you can be opened in Koji and read in context, which is the check worth doing before a quote goes into a deck.

This is also the answer to "can Claude or Cursor read my Koji transcripts". Yes, through the connector, with your account's access and nothing more.

---

Export When You Want the Raw Text

Sometimes the right tool really is your own. Koji exports interviews and responses as CSV and JSON, so you can load a corpus into a spreadsheet, a notebook or whatever you already use. See Exporting Research Data.

The public report link is the other way out. Publish the report and the link works for anyone you send it to, no account needed, and it is the thing to paste into Slack or Notion when someone asks what you learned.

---

Common Workflows

"Did anyone ever mention X?"

Open the study report and read the theme list first. If X is a real pattern it has a name and an evidence count. Click the theme, read the quotes in the evidence panel, open a transcript if you want the context. If the question spans studies, ask an assistant over MCP.

Theme validation

You think pricing transparency is a theme. Check the evidence count and read the quotes behind it. Real themes have density and the quotes sound like different people saying the same thing. Thin themes have three quotes that only resemble each other if you squint.

Drafting a PRD with real voice

PRDs grounded in actual quotes get more buy-in than PRDs full of paraphrase. Find the theme that matches the problem, open its evidence panel, take three to five quotes from distinct participants, and lead the "why now" section with them. Each one is already attributable to a real conversation.

Bug triage

A support ticket describes confusion at checkout. Open the relevant study, look for a checkout theme, check its evidence count and sentiment. Three participants with the same friction is a top of queue bug. One is an edge case.

Pre-launch risk check

Before shipping, read the reports for the studies that touched that area, or ask an assistant over MCP to do it across all of them. The forgotten concern from six months ago is usually sitting in a theme with a low evidence count that nobody clicked.

---

Habits That Make Research Findable

1. Capture the attributes on the way in. Import respondents with personalised links carrying role, company or plan tier, or ask in the screener. Personas and comparisons are only as good as what you collected before the interview.
2. Use structured questions where the answer is quantifiable. Structured questions give typed values, so a distribution is a fact rather than an impression. Keep the open follow-up to explain the number.
3. Let the AI follow up. The probing follow-ups are where the most useful sentences usually appear, because that is where a vague answer gets specific.
4. Trust the quality gate. Only interviews scoring 3 or above on the 1 to 5 scale are billed and carried into the report, so weak conversations do not quietly become findings. See Understanding Quality Scores.
5. Keep one study per question. A study with a clear brief produces a report you can actually return to. A sprawling study produces themes nobody trusts.
6. Publish the report. A published report link is the cheapest institutional memory there is, and it is what colleagues will find when they go looking.

---

How This Compares to Tag-Based Repositories

Repository tools built on manual tagging make search useful only after someone spends real time tagging a corpus, and the taxonomy drifts the moment two people maintain it. Traditional survey tools do not offer transcript level retrieval at all.

Koji's difference is that the analysis is automatic and happens per interview: themes, structured answers, quality scores and quotes are produced as each conversation finishes. There is nothing to tag before the report is useful, and the evidence behind any claim is always one click away.

The honest trade is that there is no keyword box and no cross study search inside the app. If your instinct is to type a phrase and scan matches, the equivalent here is to open the report, follow the claim to its evidence, and use an assistant over MCP for anything that crosses study boundaries.

---

Related Resources

• Reading Your Research Report - Themes, evidence counts, personas, question coverage and the evidence behind each claim
• Koji MCP Integration Overview - Let Claude, Cursor or any MCP assistant read your studies and transcripts
• Exporting Research Data - CSV and JSON export of interviews and responses
• Viewing Interview Transcripts - Reading a single conversation end to end
• Structured Questions Guide - The six question types Koji supports
• Understanding Quality Scores - The 1 to 5 score behind the quality gate
• Understanding Themes and Patterns - What the analysis extracts from each interview