To analyse focus group transcripts with AI, check the speaker labels first, then build the codebook from one or two groups yourself. Let AI apply that codebook to every turn in every group, and check each finding against the full exchange, counting distinct participants and groups and keeping the dissent. AI handles volume; the researcher decides what the room meant.
Asking an AI tool for themes works for a quick read of one session. It is not enough for a client study with six groups across two markets, where a misattributed line or a missing minority view changes the recommendation. This guide sets out a method an agency can defend, the best published evidence on how accurate AI coding of a focus group is, and how to report the result.
What makes focus group transcripts different from interview transcripts?
A focus group is a conversation, so its meaning sits in who said what, in reply to whom, and whether anyone pushed back. An interview transcript records one person's view; a group transcript records several views that shaped each other and were shaped by the moderator.
That is not a new problem created by AI. In a 2000 paper, Kidd and Parshall observed that far more had been written about organising and running groups than about analysing them, and that there was little guidance on how group talk differs from individual talk. AI makes the gap more visible, because a model given a transcript will happily treat it as a stack of separate quotes.
Before you ask an AI tool anything, decide what each claim in your report is about. There are three useful units:
- The participant: what one person said and believed. Useful for profiles and for counting how many people held a view.
- The group: what a session converged on, or split over. Useful for comparing segments or markets, where each group is one case.
- The exchange: how a view formed. Useful when the finding is about persuasion, hesitation or social pressure, for example a concern that only surfaced after someone else admitted it.
A theme can be true at one level and false at another. "Price was the main barrier" may describe four of six groups, a minority of participants and one heated exchange. State which one you mean.
How do you prepare focus group transcripts for AI?
Fix the speaker labels against the recording, keep the moderator's turns marked as moderator, and give every participant a stable code linked to their profile. Then remove the identifiers the analysis does not need and confirm that consent and the client agreement cover AI processing.
- Correct the speaker labels. Automated transcription is a starting point. Local government guidance on qualitative methods from Camden's JSNA is direct about this: automated transcripts need checking against the original recording. In a group, the expensive errors are not misheard words but misattributed turns, because one wrong label moves a view into the wrong segment.
- Split merged turns. Where two people talk over each other, the transcript often runs their words together under one label. Renaming the label does not fix it; split the turn.
- Mark the moderator. Keep the moderator's questions and probes in the transcript, labelled as moderator. They are context for every answer that follows, and you will leave them out of any count.
- Use stable participant codes. Replace "Speaker 3" with a code such as G4-P3 that links to the screener: segment, market, user status, whatever the study is cut by. The same code must mean the same person across the whole file.
- Mark the discussion guide sections. Note where each section of the guide and each stimulus starts. That lets you compare reactions to concept B across groups that saw the concepts in a different order.
- Remove identifiers you do not need. Names, employers, places and anything else that identifies a participant should come out unless the analysis depends on it.
- Check the permissions. The Camden guidance also notes that participants consent to specific uses, and that personal data in AI tools carries real risk. Confirm your consent wording and client terms cover this processing before the first upload.
Should AI or a researcher build the codebook?
Build the first codebook yourself, from one or two groups, then let AI apply it to the rest. The strongest published evidence on AI coding of a focus group found the model reliable at applying a human codebook and much weaker at inventing its own codes.
A codebook is the list of codes, each with a definition, inclusion and exclusion rules and an example quote, that analysts apply to the transcript. The study is a 2026 paper in JMIR Formative Research. Three human coders built a 49-code codebook from a one-hour focus group with four medical students, split into 48 participant turns. The researchers then asked one large language model to do two jobs.
| Job given to the model | What happened | Result |
|---|---|---|
| Apply the researchers' 49-code codebook to all 48 turns | 2,352 separate coding decisions | 96% mean agreement with the human coders; mean Cohen's kappa 0.71; 57 codes wrongly applied (2.4%) and 28 missed (1.2%) |
| Generate its own codes from the transcript | 137 codes produced | 31.4% matched a human code, 26.3% were reasonable alternatives, 42.3% were judged not reasonable; 21 of the 49 human codes (42.9%) were not found |
Cohen's kappa is an agreement score that adjusts for agreement that would happen by chance. The study used the common benchmark that a score above 0.60 is substantial agreement; 33 of the 49 codes met it.
The errors are the useful part. The authors report three recurring patterns: reading too much into tone, failing to recognise an idea that continued across turns, and confusing what a participant would do with what they had done. Each one is sharper in a group, where ideas are built across several speakers and people speculate about features they have not used.
Keep the limits in view. This was one small group, one model and one subject, and results will differ by model, prompt and transcript. But the direction matches what the authors recommend: humans build the codebook, AI applies it across the data and offers alternative codes as a second opinion. Treat the model's own codes as prompts to look again, not as the coding frame.
How do you make sure the AI reads every turn?
Give the model small pieces of transcript with the codebook each time, rather than a whole project in one go, and keep a record of which turns it has read. A sample of turns, or the passages that look most relevant to a question, is not the same as reading the full set.
The JMIR team did exactly this. They asked the model to judge each turn against each code separately, so that no part of the transcript or the codebook was skipped. It takes more calls than one long prompt, and it means every coding decision can be counted and checked.
Long inputs carry a known risk: language models can attend less carefully to material in the middle of a very long context, an effect we explain in our guide to context rot. In a six-group study that middle is where groups three and four sit.
Two habits help:
- Ask for a coverage record. For each group: how many turns there were, how many were read, and whether any were skipped or summarised.
- Run a known-view check. Before trusting the output, pick a minority view you know sits in the middle of a late session. If it does not appear in the findings, the analysis did not read everything.
How do you analyse group dynamics and dissent with AI?
Ask for positions, not just themes: for each theme, which participants agreed, qualified, disagreed or said nothing, in which group. Then read the exchange around every apparent consensus to see whether it formed independently or followed one confident voice or a leading probe.
A position grid makes this concrete. Ask the AI to fill one row per theme per group, with a quote and turn reference behind each name, and leave the final column for the researcher. An illustrative row:
| Theme and group | Positions | Not stated | How the view formed |
|---|---|---|---|
| Pack size is hard to judge, G2 | Agreed: P1, P4, P6. Qualified: P2. Disagreed: P5. | P3 | Raised by P4 after the first probe; P1 and P6 added their own examples |
Then apply five checks:
- Count people and groups, not mentions. One talkative participant can produce a dozen coded turns. Report distinct participants and distinct groups.
- Treat silence as silence. "Not stated" is not agreement. Keep it separate in the grid.
- Look at the order. If a view appears only after one participant states it confidently, note it. Agreement that follows a single voice carries less weight than the same view raised independently in several groups.
- Read the probe before the answer. If the moderator asked "would that put you off?", a run of nods is a response to the question, not a spontaneous concern.
- Separate would from did. "I would never pay that" and "I stopped buying it" are different evidence. The JMIR study found this distinction was one of the model's recurring errors.
Keep dissent in the readout. A minority view with a clear reason is often what the client most needs to hear.
How do you check AI findings before they reach a client?
Check every quote word for word against the transcript, read the turns before and after it, and confirm the speaker and group. Then check that the strength of each claim matches its count.
- Quote. Is it verbatim, and is it from the person and group the finding says?
- Context. Does the surrounding exchange support the reading, or does the next turn reverse it?
- Tone. Was the remark ironic, a joke, or a reply to a challenge? Overreading tone was one of the recurring errors in the JMIR study.
- Strength. Does "most participants" mean most of the people in the room, or most of the turns coded?
- Coverage. Did every group feed the finding, or did one session dominate?
How should you report focus group findings?
Report each finding with its spread across groups, the number of distinct participants, the dissent and a caveat, and keep it directional. Focus groups explain how people think about something; they do not measure how many people think it.
A client-safe finding reads something like: "Raised in four of six groups, by participants in both segments; two participants in group 5 disagreed because they buy the larger size for a family." That sentence tells the client how far the finding travelled, where it did not, and why.
If the client needs to know how common a view is, pair the groups with a survey and size it there. Our guide to platforms that combine quantitative and qualitative research covers how agencies join the two. For turning the checked findings into slides, see how to turn research data into a client presentation faster.
Is it safe to put focus group transcripts into an AI tool?
It can be, when the participants' consent and the client agreement allow it and the tool's data terms fit both. Check before the first upload, not after the analysis.
Questions to settle for any tool:
- Does the participant information cover analysis by an AI tool, and is the tool approved for personal data at your agency?
- Who is the controller and who is the processor, and is there a data processing agreement?
- Is your data used to train or improve the provider's models?
- Where is the data processed, which sub-processors are involved, and how is it deleted?
ESOMAR's 20 Questions to Help Buyers of AI-Based Services, launched in March 2024, gives a fuller checklist, including how research data may be used as training data.
For Meaningful, the answers are published. Our data processing agreement states that customer data is not used to train or improve any AI models, that core AI processing happens in the EU, and that customers can delete projects and data sources at any time. Our customer GDPR guide sets out that the agency is the controller and Meaningful the processor.
Which kind of tool fits which focus group job?
Choose by the job. A general AI assistant suits exploring one short, cleared transcript; qualitative data analysis software suits formal codebook work shared by several analysts; a multi-source research platform suits groups that must be read alongside surveys and other evidence for a client. Meaningful publishes this guide, so treat the table as an editorial view of fit, not a test.
| Focus group job | Tool type | Examples | Check in a trial |
|---|---|---|---|
| Explore one short session that is cleared for upload | General AI assistant | ChatGPT, Claude, Gemini | Quotes match word for word; speakers stay distinct |
| Formal, codebook-led coding shared across a team of analysts | Qualitative data analysis software | NVivo, ATLAS.ti, MAXQDA | Code definitions, coder comparison and export |
| Read groups alongside surveys, social and desk research for one client readout | Multi-source research platform | Meaningful | Every group read; each finding linked to the turns behind it, with a caveat |
Our guide to AI tools for interview transcripts compares the analysis options in more depth.
How does Meaningful analyse focus group transcripts?
Meaningful takes your own interviews and groups as speaker-labelled transcripts, as one of seven sources in a project. Agents read every response and transcript in small, focused batches, reconcile what they find, and link each finding back to its evidence, with caveats.
The other sources sit in the same project: surveys, AI-moderated surveys, uploads of past studies and documents, social, desk research and AI perception. That means a theme from the groups can be checked against a survey cut or a social conversation without copying between tools. Outputs are readouts, decks, dashboards and live share links, and you can export your data, transcripts and reports.
Meaningful costs €500 a month, excluding VAT, with unlimited seats (price checked on 12 November 2026). There is a three-day free trial.
The method in this guide still applies. Meaningful does not recruit your groups, design your sample or replace your agency's methodological judgement. Check the speaker labels before upload, decide the unit of analysis for each claim, and review the evidence behind every finding before it goes to the client.
Frequently asked questions
Can ChatGPT analyse focus group transcripts?
Yes. A general AI assistant such as ChatGPT can summarise a transcript, suggest codes and find quotes, provided participant consent and the client agreement allow the upload. For a study with several groups, plan how every turn will be read, keep the speaker labels intact and check each quote against the transcript before using it.
How accurate is AI at coding focus group transcripts?
In a 2026 study of one focus group, a large language model applying the researchers' own codebook agreed with human coders on 96% of decisions. When it invented its own codes, 42.3% were judged not reasonable and it missed 21 of the 49 human codes. The evidence favours a human-built codebook that AI then applies.
Should I remove the moderator from a focus group transcript before analysis?
Keep the moderator's turns in the transcript, clearly marked, and leave them out of any counts. The moderator's question explains what participants were responding to, and a leading probe can explain why a group seemed to agree.
How do I count agreement in a focus group?
Count distinct participants and distinct groups, not mentions, because one talkative participant can produce many coded turns. Record silence as not stated rather than as agreement, and read the exchange to see whether agreement formed independently or followed one voice.
Do I need consent to analyse focus group transcripts with AI?
You need a lawful basis and participant information that covers the processing, and the client agreement must allow it. Check whether your consent wording anticipated AI processing, remove identifiers the analysis does not need, and confirm the tool's data terms before the first upload. The agency is usually the controller.
How many focus groups can AI analyse at once?
It depends on the tool, the model and the length of the transcripts, and a long input is not the same as a careful read. Ask for a coverage record that shows every turn in every group was read, then test it with a minority view you know sits in the middle of a late session.
Can AI replace the qualitative analyst on focus groups?
No. AI can apply a codebook, retrieve evidence and organise positions across many groups. The researcher still builds the codebook, judges what the exchange meant, decides which findings matter to the client and owns the method.
What should you do next?
Take one completed study with at least two groups whose consent and client terms allow AI processing. Build the codebook from the first group, have AI apply it to the rest, fill a position grid and run the known-view check. Compare the result with the readout your team delivered. You can run that test in Meaningful during the three-day free trial.