A synthetic focus group, and what it can tell you
A synthetic focus group shows your work to panelists a model plays from written documents. None of them is a real person. In Klingbar, a crew's marketing persona runs one on your box, from a cohort you build on klingbar.com out of a public library of 197 versioned panelists.
Used well, it tells you which objections your copy invites before anyone real has seen it. It cannot tell you how many people share them.
A panelist is a document someone wrote
A synthetic respondent here is not a model improvising a demographic bracket. It is a panelist: an authored document describing one invented person — their situation, what they pay for and resent paying for, whom they trust, how they talk, and the objections they tend to raise — plus the ad-targeting fields a media plan would use. For the length of one answer, the model is asked to hold to that document.
The library holds 107 consumer and 90 business panelists. Consumer panelists carry an Esri Tapestry segment, business panelists a 2022 NAICS code, and every one scores all 19 marketing channels. A published panelist never changes in place: an edit is a new version, so a finding from March can be read against the exact text its panelists were given.
Browse the panelist library — no account needed.
panelist_id: b2c-beh-podcast-superuser-067
tier: behavioral
version: 1.0.0
ad_targeting:
channel_propensity:
youtube: high
google_search: high
meta_facebook: low
# … and the other 16 channels
taxonomy:
esri_tapestry:
segment: "3C"
---
## Voice & Communication Style
## Common Objections & Pushback
- "Is the host actually using this?"
Abridged from one document in the public library: an invented person, like every panelist there. Browse the library.
What it can and cannot tell you
A synthetic focus group decides what to try next. Something else decides what is true.
Good for
The objections your copy invites. A panelist written to distrust something will say so, in its own words, before a customer does.
Which draft to drop. It will not rank four drafts reliably, but it often cuts one for a reason nobody on the team had noticed.
Whether your questions work. A question that draws only agreement is usually a badly framed question.
Words worth testing. Phrases panelists used unprompted are candidates for your own copy.
Not for
How many. It cannot size a segment, establish how common a reaction is, or give you a percentage anyone should defend.
Whether anyone will pay. Willingness to pay belongs to buyers and to an experiment.
What its author did not know. A panelist answers from its document, and a person with their own assumptions wrote the document.
Customer evidence. A synthetic quotation presented as a customer's voice is a fabrication, however carefully the panelist was written.
Where synthetic reactions go wrong
Put panelists in one conversation and they converge. Ask a leading question and a model hands back the answer it implies. Ask a model to summarize twelve transcripts and it can report a consensus the transcripts do not contain. Each of these is a property of the method, not of one vendor, so the way a Klingbar focus group runs is built around them.
One is not fixed by anything: synthetic respondents sound more certain than people do. Discount the confidence, keep the content.
- converging
- Every panelist answers in its own call and never sees another panelist's words.
- leading
- The same core questions every round; the persona may add at most two of its own.
- summary drift
- The score and the counts come from each panelist's closing lines, not from the summary.
- edits in place
- Panelists and cohorts are versioned, and every transcript records a hash of the exact text its panelist read.
- overconfidence
- Nothing. Read "I would never pay for that" as "I would hesitate".
How a crew runs one
Build a cohort on klingbar.com
Choose six to sixteen panelists from the library, weight and tag them, add a block of product context, and write the sentence that says who the cohort is shaped like. Every change is saved as a new version and rendered into the crew's prompt tree.
The persona runs the rounds on your box
The
marketingpersona'sfocus-grouptask shows the creative to each panelist separately, revises it between rounds, and runs at least three and at most five. How a round runs.Read the record
Every round's transcripts and synthesis appear under the crew on klingbar.com: a 0–100 score, what landed, the objections, and the panelists' own language.
Let real people decide
Creative that earns it goes to a card. Where your app runs Statsig and reports conversions the way the persona expects, an approved launch card then runs it as an experiment, and real traffic answers the question the panelists could only raise.
Hypotheses, not evidence
A cohort is shaped like your buyers. It does not stand for them, and no number it produces describes a market. Use a synthetic focus group to shape the question, find the objections, and cut weak drafts; then take the survivors to real people and a real experiment.
Wherever you report a finding, say it was synthetic: which panelists, which versions, which model played them. The guide Synthetic respondents vs real people walks through the published research on where the method holds and where it fails.
Rehearse before you ask real people
Klingbar is free during early access. You bring the box and your own engine logins; a focus group needs the marketing persona on a Claude Code login, which during early access we set up with you.