Find the objection
before your customer does.
Put a headline, an ad or a whole landing page behind the glass. 4–8 simulated panelists read it cold and tell you what stops them. One of them will not like it, and that is the point.
One artifact in. A room full of reactions out.
You give it the copy, who the panel should represent, and what you want them to press on. It gives you every panelist's honest first read — and the objection you were hoping nobody would raise.
- 1
Paste the copy, or point at the page.
An ad, a headline, an email, a pricing page, a product description. Anything from 100 to 8,000 characters. Give it a URL instead and it reads the published page.
- 2
Say who the panel represents.
“Solo consultants who bill hourly.” “Procurement at a mid-market insurer.” The sharper the audience, the sharper the objections. Connect an Ounie brain and the panelists borrow what you already know about your customers, with citations.
- 3
Pick what they should press on.
Clarity, objections, pricing, positioning, or trust. The same copy tested for pricing and tested for trust comes back as two different rooms.
- 4
Read the room — including the person who hated it.
Every panelist scores it 1 to 5, says what they thought on first read, names what stops them, and offers the one change that would move them. At least one of them scores it low, and that disagreement is carried into the summary rather than averaged away.
Nobody was asked. That is the honest version.
These panelists are generated by a language model from the audience you describe. They are not people. No one was recruited, interviewed, surveyed or paid. There is no sample, so there is no margin of error and there are no percentages of respondents — you will not find any in the output, because we never produce them.
What a simulated panel is good at is the thing a solo founder at 11pm cannot do: reading their own copy as somebody who has never seen it, five different ways, in ninety seconds. Use it to find the objection you are blind to. Then go and talk to a real customer about it.
- Panelists are labelled by role, never by name — the API refuses a persona that reads as a person.
- Every response object carries
synthetic: true, and every export carries the notice in the file. - No demographics presented as findings, no sample sizes, no “% of respondents”.
These panelists are simulated by a language model. Nobody was surveyed, no real person is represented, and none of this is market research data.
The same copy, read five different ways.
A focus mode decides what the room is looking for. Run the same landing page for pricing and for trust and you get two genuinely different sets of objections — which is usually the fastest way to find out which one is actually your problem.
Objections
What stops them acting?
- Nothing tells me how long setup takes.
- I would have to move data out of a tool my team already likes.
- No mention of what happens if I cancel.
Illustrative output — your panel reacts to your copy
Before you ship the page
Run the hero and the pricing section separately. The hero usually scores fine and the pricing section is where the room falls apart.
Two versions, one room
Run both variants against the same audience and focus. You are not looking for the higher score — you are looking for which objections disappear.
Copy for an audience you are not
Selling to nurses, contractors, procurement, parents. The panel reads it as somebody who does not already know what you meant.
Inside an agent loop
Draft, panel, revise, panel again. The dissent line is a clean signal for an agent to act on, and the whole thing is one call.
A second opinion your agent can call.
A copywriting agent that never gets pushback writes confident, unread copy. Point it at this and the loop closes: draft, panel, read the dissent, revise. One call returns structured reactions with scores you can branch on — and a panel that came back unanimous returns a refusal, not a fake win.
// Any MCP host (Claude, Cursor, the AI SDK)
{
"mcpServers": {
"focus-group": {
"url": "https://focusgroup.ounie.com/api/mcp",
"headers": { "Authorization": "Bearer fcg_live_…" }
}
}
}
// …or the Ounie AI Team — the key rides the URL:
// https://focusgroup.ounie.com/api/mcp?api_key=ounie_live_…
// Tool call → run_panel
{ "artifact": "One tool. Every workflow…",
"icp": "Solo consultants who bill hourly",
"focus": "pricing", "panel_size": 6 }
→ { "panel_score": 2.8, "synthetic": true,
"dissent": "The freelancer would not even
reach the form — no price on the page.",
"personas": [ … ], "credits_spent": 8 }8 credits a panel. Nothing else.
One price, drawn from the same Ounie wallet every Ounie product shares. No subscription, no seats, no plan to pick.
4–8 panelists, scored, with objections and revisions. About 0.08 dollars.
Re-reading, comparing, exporting to CSV or Markdown, deleting. Free forever.
If nobody pushed back, the run is refused and refunded in full. We do not bill a non-answer.
A URL that returns no copy is caught before anything is reserved.
The ones the room always asks.
Are these real people?
No. Every panelist is generated by a language model from the audience description you give it. Nobody is recruited, interviewed, surveyed or paid, there is no sample, and the output never contains percentages of respondents or demographics presented as findings. Treat a panel as a structured second read of your own copy, not as evidence about a market.
Why does one panelist always dislike it?
Because a room where everyone agrees tells you nothing you did not already believe. The prompt requires at least one panelist to score the artifact 3 or lower with a concrete objection, and the code checks that it happened. If the panel comes back unanimous we run it once more, and if it is still unanimous we refuse the result and refund you in full rather than hand back a flattering non-answer.
So it will invent an objection even if the copy is good?
It will find the person the copy is not for and let them explain why. That is different from manufacturing a fault. Strong copy usually comes back with a high panel score, one low scorer who is out of scope, and objections that are genuinely about fit rather than quality — which is useful information about who you are excluding.
What does it cost?
8 Ounie credits per panel — 8 cents — drawn from the same wallet every Ounie product shares. Reading a finished panel, exporting it, comparing panels and deleting them are free forever. Re-running a panel bills again, because it is new work.
What does connecting an Ounie brain do?
It grounds the panelists. Instead of inventing customer context, they borrow what your brain already holds about your customers — their situations, constraints and vocabulary — and each panelist cites the page its context came from. It is optional: a panel built from the ICP description alone is a legitimate run, and the price is the same either way.
Can I test a page instead of pasting copy?
Yes. Give it a URL and it reads the published page's text. If the page renders its copy with JavaScript there will be nothing to read, and it tells you that rather than testing an empty artifact — and it does so before charging you anything.
Can an AI agent run this?
Yes, three ways. REST with a per-app key, MCP from Claude, Cursor or the Ounie AI Team, and a keyless x402 rail that takes USDC on Base with no Ounie account at all. Running a panel works on any of them; grounding it in your private brains needs your Ounie master key.
Is any of this stored?
Your panels and their panelists are stored on your account so you can re-read and compare them. Nothing is public, nothing is shared, and deleting a panel removes its panelists with it. Panels paid for over x402 are returned in the response and not stored at all — there is no account to store them against.
You have read your own copy too many times to see it.
Five strangers who have never seen it, in ninety seconds, for eight cents. One of them will tell you the thing you did not want to hear.