What is a synthetic focus group, and when does it beat a room of eight?

The phrase “AI focus group” is doing a lot of work in 2026. Vendors in three different businesses use it. This is what the synthetic kind is, and what it is not.

The definition

One competitor’s May 2026 guide to the category sorted “AI focus group software” into three lanes: AI moderating real respondents, synthetic personas replacing them, and AI assisting a human moderator. Only the middle lane is a synthetic focus group. In the first and third lanes the participants are real people and the AI is a moderator or a note-taker; the evidence about synthetic respondents does not apply to them, and their evidence does not apply here. When a tool says “AI focus groups”, ask which lane first.

In the synthetic lane, the participants are synthetic users: personas with a profile, a memory and, in the better systems, real material behind them. What makes it a focus group rather than several interviews is that the personas hear each other.

How it runs

1. 01Setup. Pick the personas, write the topic, choose a mode. Moderated: you ask, they answer. Autonomous: they discuss the topic among themselves for a set number of turns.
2. 02The round. Each persona answers in turn and sees what the others said. In autonomous rooms the group itself decides who speaks next, as a real room does.
3. 03State moves. Each reply updates the persona’s emotional state and trust, and moods spread between participants. A persona challenged by another may hold, hedge or shift.
4. 04Interjection. The moderator can drop a message into an autonomous run, pause it and resume. The room picks the interjection up.
5. 05Export and attach. The transcript exports, and the group attaches to a study so its turns become evidence in the report.

What it is good for

• Disagreement you can read. Two personas built from different material react to the same proposal from different places. Where they split is a hypothesis about your real segments.
• Positions under challenge. A persona’s first answer is a reaction; its answer after another persona has pushed back is a considered position. Both are findings.
• Rehearsing the real room. The moderator guide, the order of stimuli and the follow-ups can be tested before a real group is booked, which is one of the uses even the sceptics allow.
• The question a real room would not answer. A persona can be asked what it withheld. A real participant cannot.

The failure modes

Convergence. Assistant models are trained to agree, as the sycophancy note explains, and a room of agreeable personas agrees with itself. The first persona sets the tone and the rest align. NN/g’s finding that synthetic users “seem to care about everything” becomes, in a group, a room that likes everything together.

Flatness. Real groups have a dominant voice, a quiet dissenter and someone who changes their mind. Prompted personas produce eight polite paragraphs of similar length. The variance that a group is for is the variance a language model smooths away.

Theatre. The over-correction. Prompted to be critical, personas perform disagreement without a reason behind it. That is no more informative than agreement and harder to spot.

A synthetic room that reaches consensus in two rounds has measured the model’s politeness. A room that splits, and can show you why, has measured something about your buyers.

Design rules

• Ground the personas in different material. Disagreement comes from different sources, not from different names. A procurement lead built from procurement emails and an analyst built from analyst notes disagree for real reasons.
• Let trust be earned. Personas whose trust starts neutral and moves with what is said have a reason to resist the moderator and each other.
• Read the trace, not the transcript. A transcript shows who agreed. The trace shows who withheld what, whose trust fell, and which beliefs fired. In a group, the withheld thoughts are usually the finding.
• Keep it small enough to read. Two to four personas for a working session; eight when you need a room and have the time to read it as one.
• Write the limitation. Two personas across four messages is a rehearsal. The report should say so, and in Perplica the compiled report does.

How Perplica runs one

Two to eight of your personas, a title and a topic. Moderated: you ask, with attachments if you like, and every persona answers in turn, replies arriving live with each persona’s mood and trust. Autonomous: up to fifty turns; pause, interject as moderator, resume. Moods spread between participants. Export the transcript as Markdown or text, or attach the group to a study so its turns become evidence. Our pricing playbook uses a moderated room as its fourth step, and the report it produced shows what a split room looks like on paper: one persona rejecting a bundle as a tactic, another separating the part with real value from the part that was overhead.