Feedback synthesis
Feedback synthesis turns a pile of support tickets, app reviews and NPS comments into ranked themes, each carrying the pieces of feedback it came from and a count worked out from them. It reads what is there and refuses what it cannot read, rather than summarising a mess into a confident paragraph.
- Asks for
- The feedback — pasted, or a CSV export from your support tool or app store
- What you are trying to decide, so the ranking answers that rather than nothing
- Gives back
- Themes ranked by weight of evidence, each with its pieces by number
- The count under each theme, worked out from the pieces rather than estimated
- Insight sentences, each citing the feedback it rests on
- The feedback it could not use, and why each piece was unusable
- Plan
- Pro and Max
- Use next
- Interview synthesis, Interview notes, Prioritisation, Survey
What it gives you
Theme 1 — Export silently produces an empty file. 23 pieces. #14 #31 #44 #52 #58 #61 … (23 total, 19 from the last 30 days) #44 "clicked export, got a 0kb csv, no error. third time this week" #61 "the download works but theres nothing in it" Insight: the failure is silent, so people retry rather than report (#44, #61, #58). Unusable: 61 pieces. 48 were five characters or fewer; 9 were not in English.
How it works
Feedback arrives as hundreds of short, badly punctuated, partly duplicated fragments from people who had one bad morning. The usual summary of that is a slide saying customers want better performance, which is both true and useless: it names no change anybody can make and no number anybody can check.
Every piece gets a number as it goes in. Themes are built from those numbers, ranked by how many pieces sit under each one, and the count is arithmetic over the pieces rather than the model’s impression of what was common. A theme says how many of your fragments support it, so a loud single complaint stays a single complaint.
It also says what it could not use. Fragments too short to mean anything, in a language the run was not for, or saying only "great app" are listed as unusable rather than quietly absorbed into whichever theme they were nearest. That list is usually worth reading on its own: the proportion of your feedback that carries no information is a fact about your feedback form.
Who it is for
A PM holding a quarter of support tickets or a thousand reviews who has to decide what to fix, and needs the argument to survive someone asking how many people said it.
Where it stops
It weighs the feedback you gave it, and your feedback is not your users: the people who write in are the angry and the delighted, and a theme with thirty tickets behind it may still be thirty people out of ten thousand. It ranks by evidence, not by value — a cheap fix for a small theme can beat an expensive fix for a large one, and that call stays yours.
Questions people ask
What formats can I bring feedback in?
Paste it as text, one piece per line, or upload a CSV of the kind support tools and app stores export. Columns it recognises are used — date, rating, channel — and anything else is kept with the piece, so a theme can tell you that nineteen of its twenty-three tickets arrived in the last month. Nothing needs cleaning first; cleaning is part of what it does, and what it had to throw away is reported.
How is the count under a theme worked out?
By counting the pieces attached to it, in code, after the writing is done. The number is not the model’s estimate of how common something felt, which is the usual way a summary ends up saying "many users" about four tickets. Because the pieces are numbered you can open any theme and read every fragment that produced the figure, which is what makes the number worth putting in front of anybody.
Can it handle duplicate and near-duplicate feedback?
Yes, and it has to: one outage produces forty tickets saying the same thing in forty ways. Near-duplicates are grouped under one theme and counted once each, so the weight reflects how many people wrote in rather than how many different phrasings they used. Exact duplicates from a single sender are flagged, because one person filing the same complaint six times is a different signal from six people.
Run it on your own work
Feedback synthesis needs an account, because it writes against what you give it and remembers what you have told it before. It is part of Pro and Max. Or see what the plans cost first.