THE METHOD
How do you test a landing page with real people?
You show real people who match your buyer your page for 5 seconds and ask what it does. JuryHour does that for you: an agent writes a challenger, Terac recruits the jury, code checks every answer, and the verdict comes from the database, not from a model's opinion.
What happens after I pay?
Eight steps, most of them automatic. Agents do the judgment work (writing, reporting); code does everything that involves money, limits or numbers.
- You pay through Whop. Your order gets a private status page you can watch.
- An agent researches and writes. It reads your page, looks up how your market talks, and writes two short cards: A, your message as faithfully as possible, and B, a challenger that uses only facts already on your page (Verdict). It also writes a one-line description of your buyer and a screening question.
- Terac recruits the jury. Terac is a research participant platform. Only people who pass the screening question can take your test, and Terac verifies their identity and pays them. Code checks that the panel fits the price before launching it.
- Real people judge. Each juror sees a card for 5 seconds, answers questions, then sees the other card. Which card comes first is random per juror, and nobody is told which is yours.
- Every answer is checked. Code rules first, then an AI grade. Work that fails is rejected with a written reason and not paid.
- The database counts. Scores and preferences are computed in SQL from approved answers only.
- An agent writes the report. Before it can be delivered, code checks every quote against the stored answers word for word and every number against the database.
- You get the verdict. Which version won and by how much, the top fixes with jurors' own words, and the winning copy ready to paste.
What do jurors see?
A short text card: headline, subheadline, three key points and a call to action, shown for 5 seconds and then hidden. That's about how long a stranger gives your page before deciding whether to keep reading.
In a Verdict, each juror sees both cards, one after the other. Which card comes first is chosen at random for each juror, so the order doesn't decide the winner. Quick Read shows only your card.
Jurors aren't told which card is the original, and they never see your site or who submitted it.
What exactly are jurors asked?
The same questions after every card, then a preference and an attention check. These are the exact words on the juror page.
| Question | Answer | Used for |
|---|---|---|
| In your own words, what does this product do, and who is it for? | Free text, at least 15 words | Understood it? (graded), quotes |
| How clear was it? | 1 (not clear at all) to 7 (very clear) | Clarity score |
| How likely would you be to try it? | 1 (not at all likely) to 7 (very likely) | Would-try score |
| Was anything confusing or unclear? (optional) | Free text | Fixes, quotes |
| Question | Answer |
|---|---|
| Which one interested you more? | The first description / the second description |
| Why? | Free text |
| Which of these was mentioned in the description you just saw? | Three options: one real point from the card and two made up (attention check, every tier) |
How do you know jurors actually tried?
Every answer passes code checks first and an AI quality grade second; only approved answers count or get paid. Rejections always carry a written reason.
- Time on task: at least 60 seconds. Faster submissions are rejected.
- Attention check: pick the point that was really on the card out of three.
- Minimum effort: at least 15 words describing what the product does, enforced on the page and again on the server.
- AI grading: an OpenAI model reads each answer against the cards, flags answers that are too short, off-topic or look pasted from an AI tool, and scores quality from 0 to 10. An answer needs a pass and a score of 5 or more. The same check decides whether the juror understood each card.
- A brake on rejecting: if more than 40% of a panel would be rejected, code stops rejecting and flags a possible bug on our side instead of blaming the jurors.
How are the numbers calculated?
In SQL, from approved answers only, rounded to two decimals. No model writes a number in your report; the agent copies them from the database, and code refuses the report if any number differs.
| Measure | How it's computed |
|---|---|
| Understood what it does | Share of approved jurors whose own description of the card was judged correct |
| Clarity (1–7) | Average of the "How clear was it?" answers |
| Would try it (1–7) | Average of the "How likely would you be to try it?" answers |
| Preferred it | Count of jurors who picked each card |
| Verdict | The card more jurors preferred, e.g. "B wins 12 of 15"; "No clear winner" when the counts are within 1 |
Every quote in a report is a juror's own words. The report stores which answer each quote came from, and delivery is blocked unless the quote matches that stored answer exactly.
What can't a verdict tell you?
A verdict is a directional signal from a small number of real people, not a sales guarantee. It's good at catching the big problems and picking the clearer of two messages.
- Sample size. 15 people won't detect small differences. Treat a 9–6 result as a lean, not a law.
- Audience match. Jurors qualify by answering a screening question. That narrows the panel but it's self-reported, so treat a narrow B2B audience match as approximate.
- Words, not design. Jurors judge your message as text. Layout, images, speed and price tables aren't tested.
- First impressions. 5 seconds measures whether your message lands, not whether someone would read your whole page or buy.
- Public pages only. Don't submit anything confidential.
What stops the agents from making expensive mistakes?
Agents decide; code enforces. Agents never touch money directly. They ask the Ops API, and code checks every request against fixed limits before anything happens.
- A jury budget per tier, and no panel launches below a 30% gross margin.
- Refunds can never exceed what you paid; refunds for lateness or failure are limited to one per order.
- Full refunds are automatic if your audience can't be recruited or too few jurors finish by the deadline; pro-rata if some are missing.
- Prices can only move within a fixed band, at most once a day per package.
- Every agent decision is logged with its reasoning.
More on how the company runs: About JuryHour.