Why twenty interviews turn into mush
The problem isn't the volume but the form it's stored in. After twenty conversations you have twenty documents, each with its own logic, its own phrasing, its own order of topics. Reading them end to end is impossible: by the fifteenth you don't remember the third, and your overall impression is assembled from the last two plus one especially vivid one.
That's precisely why discovery findings so often match the hypotheses people walked in with. Not because the hypotheses were right, but because memory keeps what confirms and drops the rest. Coding is the procedure that gets in the way of that.
If the decision is already made and you're looking for arguments, this isn't research but evidence-gathering, and twenty interviews will only make it costlier. The honest test: state in advance what result would make you abandon the idea. If no such result exists, you can skip the research.
Three steps from transcripts to decisions
- Quotes. From each transcript you pull verbatim fragments where the person describes their own experience: what they did, what they hit, how it ended. Not opinions about your product — descriptions of reality.
- Codes. Each quote gets a short label for its substance. A code isn't the topic of conversation but the type of observation: “works around the system by hand”, “doesn't trust the numbers”, “hears about problems from the customer”.
- Patterns. Codes that recur across people independently add up to a regularity. One code from one person is an anecdote. The same code from seven people out of twenty is what the whole exercise was for.
What counts as a quote
The main filter is the past tense. People predict their own behaviour badly and remember their own actions well. “I'd use that” isn't data; “last time I exported it to a spreadsheet and added it up by hand” is.
“Yes, automated reports would be handy. I think we'd use something like that if it existed.”
“Last Monday I exported three reports, merged them into a spreadsheet by hand and spent about two hours on it. Same every week.”
The first quote is polite and commits to nothing: agreement that convenient beats inconvenient. The second contains frequency, duration and a specific action — from it you can price the problem and check whether your idea addresses that particular one. A good rule: if a quote has no past-tense verb in it, it probably isn't data.
Coder: assemble a pattern
Assign codes to the quotes and see what adds up. Below are fragments from a notional study; in real work there would be two hundred. All computed in your browser.
Threshold rule: a code becomes a pattern when it appears across at least three different people. Two is a coincidence that is very tempting to read as a trend.
How many interviews are enough
The answer isn't a number but a signal: interviews stop producing new codes. That usually happens between the twelfth and the twentieth conversation within a single segment — but if the segments differ, each has its own count, and twenty interviews spread across five segments give you nothing about any of them.
A practical technique: after each conversation, note how many codes were new. While two or three are new, keep going. When two interviews in a row yield zero new codes, you're done with that segment.
Four ways to ruin the research
- Leading questions. “Surely doing that by hand is inconvenient?” — the person will agree out of politeness. Ask how they did it last time instead.
- Coding during the conversation. While you're inventing a label you aren't listening. Code from the transcript, afterwards.
- Codes built around the hypothesis. If the labels were invented in advance and describe your idea, you'll find exactly that. Codes must come out of the quotes, not the other way round.
- A single coder. A one-person coding reflects one person's picture of the world. At minimum, have a second person code the same five transcripts independently and compare the overlap.
Checklist before drawing conclusions
How to present the result
The output of discovery isn't a forty-page report nobody opens. The form that works is shorter: a list of patterns, each with how many people confirmed it and two or three verbatim quotes. The quotes are not optional — they are what survives retelling and what stops a finding from decaying into a platitude.
Give the contradicting evidence its own section. Not out of scrupulousness but out of practicality: it's the first thing you'll be asked for, and its absence reads as a sign that the research went looking for confirmation rather than an answer.
Whisperer transcribes the interviews and assembles structured notes from them, so the step of “turn the recording into text you can pull quotes from” stops being work. The coding stays human: deciding what the substance is — that is the research.