Find Out Whether the Group Was Actually a Group
Paste a transcript with speaker labels, say who moderated, and you get the share of words each participant contributed, who dominated, who was barely heard, and how much of the session the moderator filled.
The transcript is parsed and counted on your device. Nothing is uploaded, which matters when the session is unpublished research with named participants in it.
The Validity Problem You Can Actually Measure
Most of what makes a focus group good or bad is a judgment call. Whether the themes are real, whether a quote represents the room, whether people were honest: none of that is countable.
Participation balance is the exception. If one participant produced most of the words and another produced almost none, then the themes came from a couple of voices no matter how many people were in the room. That is arithmetic, and it either happened or it did not.
It matters because the write-up almost never says so. A findings deck reports what the group thought. It rarely reports that sixty percent of the transcript came from one person, and a reader has no way to tell from the themes alone.
Dominance and Silence Are Different Problems
Both show up as an unbalanced split, and they need different fixes.
A dominant participant pulls the discussion toward their framing. The others respond to what has already been said rather than to the question, which is why the second and third answers on a topic so often sound like variations of the first. The number to look at is their share against an even split: with four participants an even share is twenty five percent each, so somebody at sixty is taking more than twice their share.
A silent participant is the quieter failure. Nobody notices in the room, because the session felt busy. But a participant at three percent has not told you what they think, and the honest reading is that their view is unmeasured rather than absent or agreeing. Treating silence as consent is how a focus group produces false consensus.
This flags both: anyone over double an even share, and anyone under half of it.
The Moderator Count Is the One That Surprises People
Moderator share is reported separately and is excluded from the participant balance, because a moderator is supposed to talk less.
Past roughly a quarter of the words, the session is being led rather than facilitated. The symptom is that participants start answering the moderator instead of each other, and a discussion where everyone addresses the facilitator is really a sequence of short interviews happening in one room.
It is worth checking even when the group felt lively, because filling silence is the most natural thing a moderator does and the hardest to notice yourself doing.
What the Score Means
One number for how evenly the participant words were distributed, where a perfectly even group is a hundred and one voice taking everything approaches zero. It is the average distance from an even split, scaled, with a cap applied when the moderator dominates.
It is a measure of balance and nothing else. A perfectly balanced group can still be a bad group: wrong recruits, leading questions, a topic nobody cared about. A low score does not invalidate the session either, since some topics genuinely have one person with the relevant experience. It tells you how to read the findings, not whether to keep them.
What to Do With a Session That Came Out Unbalanced
A low score is not a reason to throw the session away. It is a reason to read it differently, and there are four honest responses depending on how bad it is.
Report it. The cheapest and most underused option is to put the split in the write-up. One line saying that one participant produced most of the discussion tells a reader exactly how much weight to give a theme, and it costs you nothing but the sentence. A findings deck that hides this is the one that gets quietly distrusted later.
Go back to the quiet participants. If two people barely spoke, their view is unmeasured rather than absent, and a short follow up with each of them separately usually recovers it. That is cheaper than running the whole group again and it removes the specific gap rather than averaging over it.
Weight the themes rather than the quotes. A theme raised by one person and echoed politely by nobody else is not the same as one three participants arrived at independently, even though both appear in the transcript. The per-person turn counts tell you which you are looking at.
Re-run only when the imbalance decided the outcome. Running a replacement session is expensive, and it is the right call when the dominant voice set the framing early and everything after that responded to them rather than to the question. The turn order in the transcript shows whether that happened.
Run It on Every Session, Not Just the Bad One
A single reading tells you about one group. Running the same check across every session in a study tells you something more useful, which is whether the imbalance is a property of the group or a property of the moderation.
One session at a low score is a group with a talkative participant in it. Four sessions in a row where the moderator sits above a quarter of the words is a habit, and it is the moderator’s habit rather than the participants’. The same goes for silence: if quiet participants show up in every session, the recruitment screener or the opening question is doing it, not the individuals.
That distinction matters because the fixes are completely different. A dominant participant is handled in the room on the day. A pattern across a study is handled before the next one, in how the discussion guide opens and how the moderator is briefed.
It is also the only way to know whether a change you made worked. Re-brief the moderator, run the next session, and compare the number rather than the feeling.
What This Does Not Do
It does not find themes, extract quotes, or score sentiment.
Those need something reading for meaning, and a tool that counted its way to a theme would produce confident nonsense. For that half, the meeting analyzer works from the recording and produces a transcript, chapters and action items, and you can ask it about themes directly.
It also counts words rather than speaking time. A participant who said little but said it decisively reads the same as one who was talked over, and only one of those is a moderation problem. If that distinction matters, listen to the section rather than trusting the count.
It Needs Speaker Labels
Every line needs a name and a colon. Labels like P1, P2 and Speaker 1 work, which is what most transcription tools produce for multi-speaker audio, and so do real names if your transcript has them.
Leading timestamps are stripped, a time in parentheses after the name is handled, VTT voice spans are read, and a line continuing the previous speaker attaches to them rather than counting as a new turn. A transcript with no labels cannot be split into shares at all.
Getting the Transcript
If you have the recording but no transcript, the transcript generator produces one with speaker labels, which is the input this needs. For the session itself, the meeting recorder captures remote groups with each participant on the record.
If your transcript has labels in an inconsistent shape, the transcript formatter normalizes them first, and the transcript cleaner strips timestamps and filler.
