Founders & Small BusinessFree
Customer Discovery Synthesis for Founders
Every theme comes back with the number who raised it unprompted and the number you actually asked, so a belief can read untested.
River's Customer Discovery Synthesis reads each interview on its own first, then counts every theme twice across the set. One count is the people who raised it before you named it. The other is the people you actually asked. In the worked example, 14 interviews with commercial cleaning contractors produced 196 coded extracts across nine themes, and 41 of the 126 theme-by-participant cells were never covered at all. That blank third is what a single percentage hides.
Unlike a transcript summariser, which compresses each call into a paragraph and then looks for repeated words, this keeps your interview guide beside the transcripts and joins the two. Prompted agreement gets its own column rather than being pooled with unprompted mention. Eight people agreeing when asked is a different finding from four people bringing it up first. Each belief you wrote down beforehand comes back confirmed, disconfirmed, untested or not testable at this sample.
Written for the founder deciding whether to build, and for the operator who has run a dozen conversations and cannot say what any of them proved. Run it after the interviews, before the roadmap, and before the offer and packaging design it should feed. Where the question is what competitors already do badly, the competitor and market scan reads their customers' public reviews instead. Prospects predict what they would do, and reviewers report what they did, so the two answer different halves of the same positioning question.
A theme count needs two denominators
A count is only worth as much as the thing it was counted from. Compressing each conversation into a paragraph before you count anything loses the context, nuance, and detail that makes that customer story unique, so what you end up counting is paragraphs. Code the extracts instead, then hold two columns open. Ask 14 people about a cost and 8 of the 10 you raise it with will agree, because agreeing is cheap and the question supplied the answer. Report that as 12 of 14 and you have tripled a real finding of 4.
The 14 cleaning contractors show what that does to a ranking. The rota file that corrupts when two supervisors edit it drew 34 mentions, 17.3% of every coded extract, which puts it second of the nine themes. Three people raised it, which puts it eighth. One of them talked about it for 20 minutes. Willingness to pay looks stronger still, at 5 of the 6 people asked, or 83%. Then the denominator arrives: 5 of 14, none of them unprompted, and 8 never asked at all.
Which makes how-many-interviews the wrong question. Researchers who separated the two kinds of saturation found the range of themes settled at nine interviews while 16 to 24 were needed to understand them. In that study the first interview alone produced 53% of the new codes. Interview one feels conclusive because it is, about the list. Here the last new theme arrived at interview 10, and three themes were still gaining detail after interview 11. So the next four conversations belong to those three themes rather than to a fresh sample.
How it works
Send the conversations
Transcripts, recordings or your own notes. Say who each person was and when you spoke to them.
Write the beliefs down
What you thought was true before interview one, in your own words, however embarrassing that reads now.
Add what you asked
The guide, and whether every participant got the same questions. Coverage is computed from this.
Read the verdicts
Both counts per theme, a verdict per belief, and the questions worth asking next.
What you get
- Every theme counted twice, once for unprompted mentions and once for the people you asked
- A verdict on each belief you wrote down first: confirmed, disconfirmed, untested or not testable
- The coverage grid, showing which participants were never asked about which theme
- Prompted agreement in its own column, so a question that supplied its answer never counts as evidence
- Verbatim extracts under every theme, attributed to a participant, a role and a segment
- The next four questions, aimed at the themes still moving rather than at more interviews
Common questions
How many interviews do I need before this is worth running?
Eight or nine is enough to see the range of themes, and it is not enough to understand any of them. Run it there anyway. The useful output at that point is the coverage grid: it tells you which themes only two people were ever asked about. Those gaps are what the next round of conversations is for.
What is the difference between disconfirmed and untested?
Disconfirmed means you asked and the answer came back against you. Untested means the question never really got put. A belief asked of 6 of 14 people, always after you had pitched, is untested no matter how many said yes, and calling it confirmed is the mistake this exists to catch.
Everyone I spoke to was encouraging. Does that mean anything?
On its own, no. Encouragement is what a stranger offers when the alternative is an awkward silence, and it survives coding badly. It produces no extract about a specific past event, no number, and no name of a thing they already pay for. Those three are what the themes get built from, so praise mostly falls out.
I never wrote a hypothesis down. Can I still use it?
Yes, and you write them now, before reading the transcripts back. Reconstructing what you believed after you know what people said is how a disconfirmed belief quietly becomes one you never held. State them from memory, mark them as reconstructed, and the verdicts stay honest enough to act on.
Will it tell me what to build?
It tells you which themes have enough participants behind them to justify building for, and which ones are one loud interview. The judgement stays yours, and the sheet is what you defend it with: extract counts, participant counts, and the segment each participant came from, so a roadmap argument stops being a memory contest.
What do I do with the people who said they would pay?
Treat it as the weakest column on the board and go get a real price signal instead. What they already spend on the problem is evidence, a signed pilot is better, and once you have customers the price increase communication plan is where the pricing question actually gets settled. The first pilot arrives as somebody else's paper, so read it with the customer contract review.
Customer Discovery Synthesis for Founders
Fill in the form and your workspace opens with the work already underway.