How to run a focus group: a practical guide with AI prompts

A focus group is a moderated discussion with 6-9 participants who share experiences, opinions, and reactions related to a product, service, or concept. A trained moderator guides the conversation through a structured set of topics, encouraging participants to build on each other’s responses while managing group dynamics that could distort the data. Focus groups work best in early discovery research, where the goal is to map attitudes, vocabulary, and mental models rather than to evaluate specific designs.

What question does a focus group answer?

What do users think about a concept, feature, or brand before it exists as a product?
How do people talk about their problems and needs — what language and mental models do they use?
What range of attitudes and expectations exists within a target audience?
Where do users agree and disagree on priorities, and what drives those differences?
What emotional associations do people have with a product category or experience?

When to use

In early discovery, when you need to understand the range of user attitudes and expectations before committing to a direction.
When you want to observe how people react to concepts, messaging, or early-stage ideas and how they reason about them with peers.
When you need to map the vocabulary and mental models your users apply to a problem space — for example, before writing survey questions or designing information architecture.
When group interaction itself is valuable: hearing participants challenge, validate, or build on each other’s views reveals dynamics that one-on-one interviews miss.
When time or budget limits the number of research sessions you can run, and you need perspectives from multiple people within a single session.
When co-creating ideas or priorities with users in a workshop-like format.

Not the right method when you need to evaluate how people interact with an interface (use usability testing), when the topic is sensitive or personal and participants would self-censor in a group (use in-depth interviews), when you need statistically representative data (use a survey), or when you want to observe real behavior rather than reported attitudes (use contextual inquiry or diary study). Focus groups should never serve as the sole research method in a project — they generate hypotheses that require validation through other methods.

What you get

Session recordings (audio/video) and transcripts
Affinity-clustered themes from group discussions
Key insight document: areas of consensus, disagreement, and surprise
Participant quotes organized by theme
Mental model or vocabulary map showing how users frame the problem space
Prioritized concept rankings or preference data (if the session included a ranking exercise)
Recommendations for follow-up research based on emerging questions

Participants and duration

Participants: 6-9 per group. Fewer than 6 limits the range of perspectives and reduces the chance of productive disagreement. More than 9 makes it difficult for the moderator to ensure everyone participates and increases the risk of dominant voices taking over. If studying multiple segments, run a separate group for each segment rather than mixing them.

Number of groups: Run at least 2-3 groups per segment. A single session can produce idiosyncratic results driven by one participant’s personality or an unusual group dynamic. Patterns that appear consistently across groups are far more trustworthy than findings from a single session.

Session length: 90-120 minutes. Shorter sessions do not allow enough time for warm-up, core discussion, and any interactive exercises. Longer sessions cause fatigue and reduce the quality of later contributions.

Total timeline: 2-4 weeks (5-7 days for planning and recruitment, 3-5 days for sessions at 1-2 groups per day, 3-5 days for analysis and synthesis).

How to run a focus group (step-by-step)

1. Define the research objective and scope

Write down 2-4 specific questions this study must answer. Good objectives focus on attitudes, perceptions, or priorities rather than behavior. “What do potential customers expect from an AI writing assistant, and what concerns would prevent them from using one?” is a strong objective. “Do users like our homepage?” is too vague and better answered through usability testing.

2. Choose the format: in-person or remote

In-person focus groups allow the moderator to read body language, manage side conversations, and use physical materials (printed concepts, sticky notes, card sorts). Remote focus groups (via Zoom, Teams, or a dedicated platform like Recollective or Dscout) reduce geographic and scheduling barriers, can include screen-sharing exercises, and are easier to record. Choose based on the topic, the tools you plan to use during the session, and participant accessibility.

3. Write screening criteria and recruit participants

Define participants by behavioral criteria, not just demographics. If you are researching a meal delivery service, recruit people who have ordered food delivery at least twice in the past month rather than “adults aged 25-45.” Exclude people who work in the industry being studied — they respond differently from genuine users. Over-recruit by 20-30% to account for no-shows. Sources: customer databases, research panels (User Interviews, Respondent, Prolific), social media outreach. Incentives typically range from $75-150 for a 90-minute consumer session and $200-400 for B2B specialists.

4. Create the discussion guide

Structure the guide using the funnel technique: start broad, then narrow toward specifics.

Opening and warm-up (10-15 min): Introductions, ground rules (everyone’s opinion matters, no right or wrong answers, we record for analysis purposes only), and a low-stakes icebreaker. The icebreaker should relate loosely to the topic — not random fun facts. If researching travel apps, ask: “What was your last trip, and how did you plan it?”

Broad exploration (20-30 min): Open-ended questions about the problem space. “When you think about managing your finances, what comes to mind?” Let the conversation develop naturally. Prepare 2-3 follow-up probes per question.

Focused discussion (30-40 min): Present specific concepts, stimuli, or scenarios and ask for reactions. Use visual aids, prototypes, or written descriptions. Ask each participant to write their initial reaction before the group discussion begins — this reduces anchoring bias from whoever speaks first.

Interactive exercise (15-20 min, optional): Ranking, card sorting, dot voting, or a collaborative brainstorm using Miro, FigJam, or physical sticky notes. Exercises generate structured data and give quieter participants a non-verbal way to contribute.

Wrap-up (10 min): “What is the one thing you would want the team behind this product to know?” Thank participants and explain next steps.

5. Prepare logistics and the moderating team

Book the venue or set up the video call. Test all recording equipment (two devices for backup). Prepare printed materials or screen-sharing links. Assign roles: a moderator who leads the discussion, and an assistant who takes notes, manages the recording, and tracks which participants have not spoken. If the client or stakeholders will observe, set up a separate room or a muted video feed — observers in the same room change participant behavior.

6. Moderate the session

Open with ground rules and get verbal consent for recording. Maintain the conversation’s direction without suppressing tangents that reveal genuine concerns. Use these moderator techniques:

Call on quiet participants by name: “Maria, we haven’t heard your perspective on this — what’s your take?”
Redirect dominant speakers: “That’s a useful point, Alex. Let’s hear from others — does anyone see it differently?”
Ask for written responses before verbal ones on key questions to reduce anchoring and groupthink.
Use the “5-second rule”: after a participant finishes, wait five seconds before responding. Others often fill the silence with additional thoughts.
Probe for specifics: when someone says “It’s confusing,” ask “Can you describe what made it confusing? What were you expecting to happen?“

7. Debrief immediately after the session

Within 30 minutes of each session, the moderator and note-taker record their impressions: strongest themes, surprising moments, group dynamics (who dominated, who was unusually quiet, where consensus formed too quickly). This fresh perspective is invaluable during later analysis when transcripts alone can miss the energy and tone of the conversation.

8. Analyze and synthesize across groups

Transcribe recordings (use AI transcription tools for speed, but verify key quotes against the recording). Code transcripts by theme, noting which themes appeared across multiple groups versus only one. Pay special attention to points of disagreement — they often reveal unspoken assumptions or distinct user segments. Write insights as “observation + implication” pairs: “Participants consistently described AI-generated text as ‘soulless’ and associated it with spam email. This suggests that any AI writing feature must give users visible control over tone and output to overcome the negative framing.”

How AI changes focus groups

AI compatibility: partial — AI can accelerate the preparation and analysis phases of focus group research, but the live moderation of a group discussion remains a fundamentally human skill. Reading a room, managing interpersonal dynamics, and deciding in the moment whether to follow a tangent or redirect the conversation all require social intelligence that current AI cannot replicate.

What AI can do

Transcription with speaker identification: Tools like Otter.ai, Rev, and Trint convert multi-speaker recordings to text and label individual speakers, replacing hours of manual transcription for each session.
Discussion guide drafting: Given a research objective and participant profile, an LLM can generate a structured guide with open-ended questions, probes, and timing estimates. The moderator then refines based on experience and domain knowledge.
Thematic coding across sessions: AI can process transcripts from multiple focus groups, identify recurring themes, tag quotes by topic, and flag points where groups disagreed. This replaces the first pass of manual coding — the researcher still interprets meaning and context.
Sentiment and emotion detection: NLP tools can map emotional tone across a session timeline, revealing moments where the group’s energy shifted — useful for identifying which topics triggered the strongest reactions.
Report and presentation drafting: An LLM can turn coded themes and quotes into a structured findings report, saving 3-5 hours of formatting and writing per study.
Stimulus material creation: AI can generate concept descriptions, mock advertisements, or scenario narratives to use as discussion prompts during the session.

What requires a human researcher

Moderating the live session: Managing group dynamics — drawing out quiet participants, redirecting dominant speakers, reading body language, calibrating the pace of discussion — is the core skill that makes focus groups work. A poorly moderated session produces data that reflects the loudest voice, not the group’s actual range of views.
Interpreting group dynamics as data: Whether participants quickly reached consensus or debated intensely tells the researcher something about the strength and consistency of attitudes. AI analyzing a transcript cannot distinguish genuine agreement from social desirability or groupthink.
Making real-time decisions: Should you spend more time on an unexpected topic that participants care about, or stick to the guide? Should you call on someone who seems uncomfortable, or give them space? These judgment calls shape the quality of the data.
Validating AI-generated themes: AI can cluster quotes under headings, but only a human who was present in the room can tell whether a theme labeled “trust concerns” actually represents fear of data exposure, skepticism about product quality, or distrust of the company behind it.

AI-enhanced workflow

Before AI tools became capable of multi-speaker transcription and thematic coding, analyzing three focus group sessions of 90 minutes each meant roughly 15-20 hours of transcription work and another 8-12 hours of manual coding and synthesis. With AI-assisted transcription (same-day turnaround) and a first-pass coding tool like Dovetail or MindCoder, the same analysis compresses into 4-6 hours. The researcher’s time shifts from mechanical processing to interpretive work — examining why themes differ between groups, identifying subtle disagreements beneath surface-level consensus, and connecting findings to product strategy.

The preparation phase benefits as well. Instead of spending half a day drafting a discussion guide from scratch, a researcher can generate a solid draft with an LLM in 15 minutes, then invest an hour sharpening the questions based on their knowledge of the participants and the product domain. Stimulus materials — concept descriptions, scenario cards, ranking exercises — can also be drafted rapidly, leaving more time for pilot-testing the flow.

Where AI has no role is in the session itself. A 90-minute focus group is a live social event, and its value depends entirely on the moderator’s ability to create an environment where participants feel comfortable disagreeing, elaborating, and revealing what they actually think rather than what they believe they should say.

Beginner mistakes

1. Using focus groups to test usability

Focus groups cannot reveal whether a user can complete a task in an interface. When shown a design as a group, participants discuss their opinions about how it looks or what they think it does, but nobody actually attempts to use it. If you need to know whether users can navigate a flow, run a usability test with one person at a time. Jakob Nielsen’s original work on this distinction is over two decades old, yet the confusion persists.

2. Letting one or two voices dominate the discussion

In any group of 6-9 people, a few will be more talkative and confident. Without active moderation, those individuals set the agenda and the quieter participants default to agreeing. The result is data that reflects 2-3 opinions, not 6-9. Counter this by asking participants to write their thoughts before discussing, calling on quiet members by name, and explicitly inviting dissent: “We’ve heard one perspective — does anyone see it differently?“

3. Running a single focus group and treating results as conclusive

One session produces one data point. The group dynamic, the mood of the day, and the composition of participants all influence the discussion. A finding that appears in 3 out of 3 groups is far more credible than one from a single session. Plan for at least 2-3 groups per segment from the start.

4. Mixing distinct user segments in the same group

Placing power users and first-time users in the same session creates an uneven dynamic. Experienced users set the frame for the conversation, and beginners defer to their expertise rather than voicing their own confusion or needs. Run separate groups for each segment so that participants can speak freely among peers with similar experience levels.

5. Skipping the written-before-verbal step

When the moderator asks a question and immediately opens the floor, the first person to speak anchors the entire discussion. Subsequent participants respond to that person’s framing rather than sharing their independent view. Having everyone write a brief answer before the verbal discussion begins captures unbiased individual perspectives and gives the moderator a record of each person’s initial reaction.

Tools

Recruitment: User Interviews, Respondent, Prolific, Ethnio (intercept-based recruitment)
Remote platforms: Zoom (breakout rooms for sub-groups), Recollective (async and live qual), Dscout (mobile-first diary + live sessions)
Collaboration and exercises: Miro, FigJam (virtual whiteboards for sorting, voting, and brainstorming exercises during remote sessions)
Recording and transcription: Otter.ai (real-time transcription with speaker labels), Rev (human-verified transcription), Trint
Analysis: Dovetail (coding, tagging, and insight repository), Atlas.ti (qualitative data analysis), MindCoder (LLM-assisted qualitative coding)
Note-taking: Notion, Google Docs (timestamped observation notes for the assistant moderator)
In-person logistics: one-way mirror facilities (e.g., Schlesinger Group), high-quality audio recorders (Zoom H6), wide-angle cameras

Example from practice

A fintech startup was developing a budgeting app for freelancers and wanted to understand how this audience thought about money management before designing the product. They ran 3 focus groups of 7 participants each, all freelancers who had been self-employed for at least a year.

During the sessions, a clear divide emerged. One segment treated budgeting as a monthly ritual tied to invoicing cycles, while another described it as reactive — they only checked their finances when something felt wrong, like a declined card or an unexpectedly low bank balance. Both groups agreed that existing budgeting tools were designed for salaried workers with predictable income, but they disagreed sharply on what would fix this: the first group wanted forecasting based on pending invoices, while the second wanted alerts that would interrupt them before a problem occurred.

The team used these findings to define two distinct user personas and designed the app with both workflows. At launch, the alert-based onboarding path had 2.3 times higher 30-day retention than the forecasting path, which validated the focus group insight that the reactive segment was larger and more underserved. Without the focus groups, the team would have built only the forecasting feature — the one that matched their own mental model as financially organized founders.