Run a focus group with AI agents
Set up an AI-moderated focus group with multiple synthetic personas.
Focus groups simulate a discussion among multiple AI personas. You can combine participants with different backgrounds, have them discuss your research topic, and review the range of perspectives and debate that follow.
Before you start
Decide what you want the group to discuss before you begin. If you want grounded participants, make sure you already have respondents with digital-twin profiles from an earlier survey.
Steps
Open Focus Groups
From Home, filter your studies by AI focus groups, or click the AI focus group tile under Start a study to open the focus groups page. Click New Focus Group.Step 1: Describe your research question
In the Brief step, write what you want to learn. Click Draft the guide if you want AI to draft discussion topics and probe questions for you (uses a small number of credits). You can also continue and write them yourself.Step 2: Write a discussion guide
In the Guide step, add up to 5 discussion topics. For each topic, write one main question and optional probe questions (up to 5 probes per topic). Set a round limit for each topic to control how long the discussion goes. You can also attach a stimulus (title and content) to give every participant the same context.Step 3: Select participants
In the People step, choose 2-6 AI personas to join the discussion. Use existing personas from the AI Personas page or create new ones. To ground the discussion in real data, select respondents marked Real answers. These are digital twins built from answers to an earlier survey. Each participant is color-coded so you can identify them in the transcript.Step 4: Configure the session
In the Review step, open Session settings. Choose the AI model. Set a moderator style (neutral, challenging, or empathetic). Pick a group dynamic (natural, debate, or consensus). Set the maximum number of rounds (5-50).Review the transcript and summary
Watch the discussion in a color-coded, real-time transcript with topic dividers. When the session finishes, review the AI-generated summary with key themes, points of agreement and disagreement, notable quotes, and per-participant analysis.
Details and limitations
Focus groups complement synthetic testing by producing qualitative discussion instead of individual survey responses.
Each persona speaks from a different perspective shaped by demographics, attitudes, and background. The AI moderator decides who speaks next and guides the conversation using your discussion guide.
After the session, an auto-generated summary includes key themes, points of consensus and conflict, notable participant quotes, and a per-participant analysis showing key contributions and how each stance changed.
Stimulus support: Attach a stimulus (title and content) to a focus group session to give all participants shared context. The stimulus text is shown to every participant and the moderator, so the discussion stays grounded in a specific concept, ad, or product description.
Grounded respondents (digital twins): Instead of generic AI personas, you can add grounded participants: respondents from a previous survey who have a digital-twin profile. The AI persona mirrors that respondent's answers, so the discussion draws on what real people said.
Use focus groups early in research design to spot blind spots, test question wording, and uncover topics you may have missed.
Interpret the output as a simulation. Every contribution in this discussion is generated by AI, including a digital twin marked Real answers. Earlier human answers provide context; the generated statements are not new interviews with those people and are not quotes you can attribute to them. Use the discussion to develop hypotheses and questions, then validate decisions with real participants.
Check the result
You should have a completed focus group session with a transcript, topic dividers, and an AI-generated summary to review.
If something goes wrong
If the discussion feels too generic, check the People step and confirm you selected the intended personas or respondents marked Real answers.