Synthetic personas are AI-generated customer profiles used to simulate market feedback: an AI model asked to "act as" a described customer type and react to a concept. They promise the speed of talking to your customer without the cost of recruiting one.
The appeal is obvious. The risk is subtle: a persona is a description, and a rich description of a person is not the same as a reliable model of how they'd behave. Lean too hard on demographic persona detail and you get answers that sound plausible and predict nothing.
The trap: persona colour vs. purchase signal
It's tempting to build synthetic personas out of demographics (age, gender, region, income) because that data is easy to write down. But those attributes are weak predictors of whether someone buys, and worse, language models don't reliably reproduce real subgroup differences. Cutting synthetic results by gender or ethnicity produces confident-looking numbers with no statistical basis.
We don't cut synthetic results by gender or ethnicity: those are persona flavour, not signal. What actually shapes a purchase decision is category usage and buying behaviour, so that's what we condition on.
The fix: condition on behaviour, measure by meaning
A synthetic read becomes reliable when two things are true:
- The respondent is defined by behaviour. How they shop the category, what they currently use, what they're willing to pay (the attributes that drive a real decision) rather than a demographic sketch.
- The answer is measured by meaning. Elicit a genuine free-text reaction and map it to a purchase-intent distribution by semantic similarity to calibrated anchors, instead of asking the model to invent a rating. This is the method validated in Maier et al. (2025).
This is why Reverb calls its respondents Echoes rather than personas. The word keeps the promise honest: an Echo reflects how a behavioural segment of your market would react; it's a directional instrument, not a synthetic human you should over-trust.
Using synthetic personas honestly
- Treat the output as a directional screen, and validate high-stakes calls with verified humans.
- Prefer behaviour-based segments over demographic personas.
- Watch for uncalibrated categories and non-Western markets: flag them rather than trusting them.
- Insist on a calibration score and a methods note, so you know how far to trust each read.
Used this way, synthetic personas stop being a novelty and become a real screening layer. For the broader picture, see what is synthetic market research and what are synthetic respondents.