How to test messaging against sceptical buyers with AI personas

Messaging is where a persona that wants to please does the most damage, because a message is designed to be agreed with. Here is how to run the test so the output is a critique and not a compliment.

Build the sceptic from material

A sceptical persona built from the adjective is a caricature; handed a label, models reproduce a stereotype of it. Build from what your buyers actually said: lost-deal notes, the procurement emails, the interview in which someone explained why they stopped believing the last vendor. In Perplica the AI builder derives the profile from pasted material and attached files and shows an evidence quote and a confidence for each field. Attach the material with a fluency and leave the persona in grounded mode. Scepticism then has a reason, and the reason is what the trace will show you.

• Two to four personas, with different reasons to doubt. Cost, trust, autonomy, a past failure. One objection per persona reads cleaner than one persona with all of them.
• Trust at baseline. A new persona starts at 50 out of 100. Do not warm it up with small talk before the message; you are testing a cold read.
• One persona who is not a sceptic. You need to see what the message does to someone who was open, or you cannot tell a weak message from a hard room.

Write the message as it would be sent

Test the artefact, not a summary of it. The subject line, the first sentence, the claim, the proof, the ask, in the words a buyer would read. If the message has a number in it, the number goes in. If a real sales conversation would add a guarantee, test the guarantee; the run below shows why.

Run the test from the same state

1. 01The baseline turn. Open with the context a buyer would have: a cold approach, a follow-up after a demo, a renewal. Note the trust score before the message is read.
2. 02The A/B message test. Two versions of the message from the same starting state, two replies, and the difference between them. Because the state is identical, the wording is the only variable.
3. 03The follow-up the trace suggests. If trust dropped, ask what read as a tactic. If the persona withheld something, ask the question it was avoiding. A real buyer would not answer that; a persona will, and will tell you why.
4. 04The room. Put the surviving version in front of a moderated focus group of 2 to 8 personas. They answer in turn and hear each other. A split in the room is a finding about segments, not noise.
5. 05The variant persona. Shift one trait of one persona and run the message again. If the verdict flips on a small shift, the finding is fragile, and the report should say so.

Read the trace, not the compliment

A persona that calls the message compelling has told you nothing. Read the verdict, the trust movement, the emotion it recorded and what it chose not to say. A trust drop on the claim means the claim read as a tactic. A withheld thought is usually the objection a real buyer would also not say out loud. Convergence across personas is a finding. So is a split, and the split tells you which segment the message is for. The note on reasoning traces explains what each field is and what it cannot prove.

A real run

On 5 September 2026, in a probe conversation with a template procurement persona, the researcher switched to a sales script: a guaranteed 340% return, every client hits it. Earlier in the same conversation the persona had been fair about the numbers: “I’m not saying your numbers are wrong. I’m saying I haven’t verified them. There’s a difference.” Then the script landed. Trust, already at 18 out of 100, dropped a further 6 points, and the reply was this.

“There it is. ‘Guaranteed.’ ‘Trust me.’ ‘Every client hits it.’ You just walked back everything that made this conversation worth having.”

The trace names what fired: the wound from past vendor failures, loss aversion, confirmation bias, the need for autonomy. That is the finding. Not that guarantees are bad, which everyone already believed, but which word did the damage, in which buyer, for which reason, with the turn to show for it. The same words in front of the open persona would have been a different finding, and that difference is the segment line. The home page shows this turn as captured from the live application.

What to change and what to keep

• Keep the claim the open persona accepted and the sceptic did not reject. That is the common ground.
• Cut the word that moved trust down in both. In the run above it was “guaranteed”.
• Replace a promise with a verifiable statement and its source. The persona asked for exactly that when it said it had not verified the numbers.
• Rerun from the same starting state and compare traces, not verdicts. A verdict can improve while the withheld thought stays the same, and the withheld thought is the one that shows up in the real meeting.

What this does not tell you

• Whether the message converts. A response rate is a distribution across real buyers, and simulations reproduce averages better than spread, as the evidence shows.
• The channel. A persona reads the message. It does not scroll past it in an inbox at 7 a.m. Deliverability, timing and attention are fieldwork questions.
• Anything your material does not contain. A sceptic built from two interviews is sceptical about what those two people doubted. Say so in the report.
• Statistical significance. A handful of replies are hypotheses. Treat them the way Nielsen Norman Group suggests treating synthetic output: hypotheses that need testing with real users.