Build two to five personas from real buyer material, ask them a direct pricing question and follow up on what the trace shows, run the same question through a focus group so the personas react to each other, and compile a report where every finding cites the turn behind it. You leave with sharper hypotheses, the objections that fire first, and a document a stakeholder can audit, not a price. The price comes from fieldwork, which this makes cheaper and better aimed.
Pricing is where synthetic research is most tempting and most dangerous. Tempting because a real pricing study is slow and expensive. Dangerous because a persona that wants to please will accept any number you propose. Here is how to run it so the output is worth something.
Start with a question you can answer
“What should we charge?” is not a research question a simulation can answer; it is a decision. “Would procurement leads accept a 20% increase if it came bundled with a dedicated customer success manager and eight integrations?” is a question a grounded persona can reason about, because it is about reactions and justifications, not about the true market price. Write the question the way a study would: one decision, one audience, one proposed change. The study in Perplica is literally that sentence; everything attaches under it.
Build the personas from material
Skip this step and you have a prompted persona, and the pricing answer you get is the model’s opinion of what a procurement manager would say. Use what you already have: interview transcripts, win and loss notes, the procurement emails from the last renewal, the brief that describes the segment. In Perplica the AI builder derives a full profile from pasted material and attached files, showing an evidence quote and a confidence for each field, and attaches the material as knowledge the persona draws on. Build the roles that actually decide, not the ones that are easy to imagine: the procurement lead, the budget owner, the analyst who writes the recommendation.
- Two to five personas. Enough for disagreement, few enough to read every trace.
- Grounded mode, not open. You want answers that prefer your material and flag assumptions beyond it.
- Trust at baseline. Do not warm the persona up with flattery before the question. You are testing a cold reaction.
Design the test
- 01The direct question. Ask it plainly, once, to each persona in a one-to-one conversation: would this be a dealbreaker. Record the first reply before any justification is offered.
- 02The justification. Present the bundle or the reason for the change. Watch whether it reads as value or as a negotiating tactic; the trace will say which.
- 03The follow-up the trace suggests. If trust dropped, ask why the change felt like a tactic. If a persona withheld something, ask the question it was avoiding. This is where a synthetic test earns its place: you can ask the follow-up a real participant would refuse.
- 04The room. Put the same personas in a moderated focus group and ask the question again. Personas answer in turn and hear each other; the disagreements and the alignments are both findings.
- 05The variant. Shift one trait of one persona, more openness or more agreeableness, and ask again. If the answer flips on a small shift, the finding is fragile and you should say so.
Keep the instrument honest. A one-line question compresses nuance; the report will note it. Longer deliberative exchanges surface conditional acceptance that a yes or no hides. Classic human pricing methods such as price-sensitivity meters and staircase questions still belong in the fieldwork phase; this rehearsal tells you which price points and bundles are worth their sample.
Read the trace, not just the answer
The answer to a pricing question is the least interesting output. A persona that says “probably not” has told you what; the trace tells you why, and why is what transfers to real buyers. In each reply read the verdict, the trust movement, what was withheld and which beliefs or wounds fired. A trust drop on the bare question is defensiveness about being anchored. A trust drop on the justification means the bundle was read as a tactic. A withheld thought is often the real objection, the one a real participant would also not say out loud.
“Depends entirely on what’s driving it and what I’m getting in return” is a deflection. The trace behind it is the finding.
Compile the report
A study compiles into findings with a confidence, each supported by evidence items that cite the exact turn: a quote, a verdict, a withheld thought, a trust shift. Evidence that cannot cite a turn is dropped. The report also writes its own limitations, which for a pricing rehearsal should always include the sample size, the fact that personas are simulations, and the scenarios not tested. Edit the summary and claims in place, export it, and hand it to the person who has to sign off on fieldwork. Its job is to justify the next study, not to replace it.
A real run
Our own study asked: would procurement leads accept a 20% price increase? Two grounded personas, one conversation, one moderated focus group, five messages analysed. The compiled report found that a 20% increase is treated as a dealbreaker by default even when the persona will not say so outright, with a withheld thought as the receipt: his gut reaction was “yes, probably a dealbreaker” but he was not giving that leverage away. Trust dropped by 12 points on the bare question and a further 12 when the bundle was presented, which the persona read as unrequested features priced as a favour. One persona distinguished workflow-mapped integrations, which could be real value, from a dedicated customer success manager, which it called overhead dressed up as a benefit, and its trust rose 12 points for the distinction.
The report’s first limitation begins: “All personas are AI simulations, not real procurement professionals.” Its others note the two-persona sample, the sceptical skew, the single bundle tested and the one-line format. That is the standard: a finding you can defend, next to the reasons you should not over-defend it. The product page shows this report as captured from the live application.
What this does not tell you
- The price. Willingness to pay is a distribution across real buyers. Simulations reproduce averages better than spread, as the evidence shows; a spread is exactly what a pricing decision needs.
- Anything your material does not contain. A grounded persona reasons from what it was given. If no transcript mentions switching costs, the persona’s view of switching costs is an assumption, and a good one will flag it as such.
- Statistical significance. Five messages from two personas are hypotheses. Treat them as Nielsen Norman Group suggests treating synthetic output: hypotheses that need testing, and preparation for studies with real users, then the real sessions.
Can AI personas tell me what price to charge?
No. They can tell you how a grounded, sceptical persona reasons about a price change, which objections fire first, and which justifications are read as value or as tactics. That narrows the questions worth putting to real customers. The price itself is a fieldwork decision.
What material do I need for a pricing test?
Anything real about the buyers: interview transcripts, win and loss notes, procurement emails, briefs. The personas are built from it and constrained to it, so the test measures your buyers’ recorded reasoning rather than the model’s general opinions about price.
How many personas and how many questions?
Two to five personas covering the roles that actually decide, and a small number of direct questions with follow-ups. One-line questions compress nuance; ask, then probe what the trace shows. A focus group of the same personas surfaces disagreements a one-to-one interview does not.
Is a synthetic pricing test a replacement for a pricing study?
It is a rehearsal for one. Use it to sharpen the hypotheses and the bundles you take into fieldwork, and to draft the report structure. Then validate with real buyers before anyone changes a price list.