Test the words before the market reads them
Put headlines, body copy, and ad copy in front of a synthetic cohort, revise between rounds, and keep the words panelists reached for on their own. Then the final creative set goes to an approver as one card. Where your app runs Statsig and reports conversions the way the persona expects, an approved launch card runs the copy as an experiment, and real traffic picks the winner before you spend on media.
The welcome-email rewrite earned its card: 75/100 after three rounds, up from 50. "Book your first client tonight" landed. The "save ten hours a week" claim drew the most objections and is gone. Rounds are in the thread.
AP-41 — Final creative set: welcome email · cohort Solo salon owners v2Two arms: the current email as control, and the rewrite. Both are linked on klingbar.com.
Rounds on the words themselves
The stimulus can be a subject line and body, a landing page's headline and first screen, a paywall block, or ad copy drafted in markdown. Each panelist in the cohort answers the same fixed questions, alone. One of them asks which line the panelist would keep and which it would cut, which is the question message testing turns on.
After each round the persona reads the synthesis and rewrites the copy. For each change it notes which objection it answers and which line it kept because it landed, then runs the next round on the new draft. From round 2 every panelist sees its own previous reaction beside the revision and says whether its objection was answered. Three rounds at minimum, five at most.
- changed
- Subject line to "Your booking page is live — share it tonight"
- why
- Five panelists read "Welcome aboard" as a newsletter and said they would not open it
- kept
- "Book your first client tonight", which six panelists quoted back
Keep the words panelists used on their own
Each synthesis lists up to eight phrases panelists used without being prompted. These are candidates for your headline, not quotes from customers. A phrase a synthetic panelist reached for is a good thing to try on real people, and a bad thing to put in a testimonial.
Watch for your own words coming back. When a transcript is full of the copy's own vocabulary, the panelist is echoing the premise, and you have learned nothing about how a buyer would say it.
- stance
- positive 6 · mixed 3 · negative 1, from the panelists' own
STANCElines - their language
- "no-shows", "between clients", "my regulars", "the deposit thing"
- objections
- Wants to know whether deposits cost extra
Let the winner prove itself
After the first card, this half needs Statsig and a conversions source wired the way the persona expects, in your app's repository.
The final creative set, as one card
After the last round the persona posts one card naming the cohort and its version, with every arm's creative linked on klingbar.com. An approver approves it or denies it with a reason. The rounds before it ran unattended.
The launch card
The persona prepares the experiment in your app's repository, on a branch with a pull request, and pre-registers it: the primary metric, the guardrails, and the exposure cap. The registration is frozen once it is written. The launch card lists everything the launch will do, including the daily allocation moves and the stops it may make on its own. Once it is approved, the persona ships the code with the experiment switched off, checks that the new version is the one serving, and only then starts the experiment behind Statsig. If any step does not check out, it stops there and reports it as blocked.
Daily looks
Each day the persona takes one look at the cumulative counts. A person counts only once an arm's difference actually reached them, so an email counts when it is delivered, not when it is sent. The analysis, not the persona's own arithmetic, says whether a decision is reached. Stopping for harm to a guardrail needs no card.
A decision
Winner, equivalent, futile, cap, or harm. A winner needs at least seven days and a hundred exposures per arm, a probability of beating control of at least 0.90, and a small expected loss. The bar rises to 0.95 when the baseline rate is above 20%.
The rollout card
A winner moves to 100% only through its own card. Once it is approved, the persona rolls out, hard-codes the winning copy, and proposes a learning that quotes the low end of the lift rather than the average.
Owned traffic first
Ad copy can go through the rounds, but the experiment runs where you already have traffic: a landing page, an email, an in-app message. Ads are drafted, not placed: nothing leads from the rounds to a paid launch, and an idea that needs paid media ends as a question to you about budget. By the time you pay to show a message to strangers, it has already met your own visitors and a decision is on the record.
After each decided experiment, the cohort's page puts what the cohort predicted beside what the market chose. Over a few experiments, that record tells you how much weight a cohort's score deserves. Until then, treat a synthetic winner as a candidate. Concept testing covers the step before copy, and approval workflows covers how a card is decided.
Test the copy, then let traffic decide
Klingbar is free during early access. You bring the box, a Claude Code login, and the app the experiment runs in, wired for Statsig and conversions; we set up the marketing persona with you.
Request early access