Synthetic persona
Also called: simulated buyer, AI persona, agent persona.
A synthetic persona is a simulated buyer, generated by a model and used in place of a human respondent. The term says nothing about quality, because it says nothing about sourcing. Two products can both sell synthetic personas while one invents its buyers from a prompt and the other builds them from a corpus of real buyer language. Sourcing is the whole distinction, and it is the question most vendors are vaguest about.
Three grades of sourcing
Every synthetic persona is built from something. That something decides whether it is useful, and the three options are not close in quality.
| Prompt persona | CRM / call-data persona | Evidence-compiled persona | |
|---|---|---|---|
| Built from | The sentence you just typed | Your own records, deals and call recordings | Those, plus sources outside your walls |
| Who it represents | Nobody in particular | People who already replied to you | Also people who never replied |
| Knows why you lost | No | Only where a loss was logged, and logged honestly | Yes. Competitors’ reviews say it in public |
| Sees people who never contacted you | No | No. They are not in your CRM | Yes. Community threads |
| Cites a source per finding | Nothing to cite | Sometimes | Yes, on every finding |
| Main failure mode | Invention | Survivorship | Corpus bias, reduced but still present |
Note the last row. The third column is the widest sample available, not a clean one. Any vendor claiming otherwise is overselling.
Why CRM-built personas feel rigorous and mislead
This is the trap, and it catches careful teams rather than lazy ones.
“We built our personas from our own call transcripts” sounds far more rigorous than a prompt, and it is. Real people said those words. The problem is which real people.
Everyone in your CRM replied to you. Everyone on a recorded call agreed to take it. Every closed-won record is a person you persuaded. Your first-party data is a census of people you already reached, which makes it the wrong sample for the job marketing actually has: reaching the people who did not respond.
So a CRM-built persona models your best-case buyer with real fidelity, then tells you your page works. It usually does. For them.
Where the non-responders leave a trace
They are not silent. They are just not talking to you.
- Competitors’ reviews. Written by people who evaluated your category and chose somebody else, explaining exactly why, in public. The closest thing to a transcript of the deal you lost.
- Community threads. People with the problem, comparing options, who have contacted no vendor at all and will not until they are ready.
- How AI assistants describe your category. Increasingly the first version of your pitch a buyer sees, and not written by you. See dark funnel.
WhyUser compiles from these alongside your own transcripts. The transcripts still matter. They are just not sufficient alone, and a vendor who stops there has built a very good model of your existing customers.
Four questions that separate the grades
Ask these of any vendor selling synthetic personas, including us.
- Name your sources. “Trained on B2B data” is not an answer. Which corpora, refreshed how often, and can I see the source on an individual finding?
- What share comes from outside my own systems? If the answer is none, you have a well-modelled version of the people you already reached.
- Run the same page twice. If the verdict moves, you cannot test an edit, because you can never separate your change from the model.
- What does it refuse to claim? A vendor who cannot name a limitation has not looked for one.
What no synthetic persona can do
- Replace talking to customers. It is built from customer evidence, so it sits downstream of it. Thin evidence in, confident nothing out.
- Give you an absolute conversion rate. Rank and direction hold up. A precise number does not, and one wrong magnitude sinks every correct finding beside it.
- Escape its corpus. Public reviews skew to the delighted and the furious. The quiet middle is under-represented, and a good system says so.
- Surprise you from outside its evidence. A real person occasionally says the thing nobody modelled. Worth paying a panel for.
Questions people ask about synthetic personas
What is a synthetic persona?
A simulated buyer generated by a model and used in place of a human respondent. The term says nothing about quality on its own, because it says nothing about what the persona was built from.
What is the difference between synthetic persona products?
Sourcing. Some invent the buyer from a prompt at the moment of asking. Some build it from your CRM and call recordings. Some compile it from those plus sources outside your walls, such as competitors reviews and community threads. The three are not close in quality.
Are personas built from my CRM and call data good enough?
They are rigorous about the wrong population. Everyone in your CRM already replied to you and everyone on a recorded call agreed to take it, so first-party data is a census of people you already reached. Marketing needs the people who did not respond, and they are not in there.
Where can I find the buyers who never responded?
In competitors reviews, written by people who evaluated your category and chose somebody else, and in community threads where people compare options without contacting any vendor.
Are synthetic personas accurate?
They are useful for rank and direction rather than absolute numbers, and only as good as what they were built from. The way to judge a vendor is to ask whether they publish a graded track record of predictions sealed before the outcome existed.