Synthetic Audiences Are Not a Substitute for Talking to Customers

What a synthetic respondent actually is

A synthetic audience is a model producing text that resembles what a described person might say. It is not a sample. It has no experience of your category, no household budget, no bad week and no memory of the time your delivery arrived broken. What it has is a great deal of reading about people like the one it is imitating.

That distinction matters because the output is fluent, and fluency reads as insight. A transcript of twenty simulated interviews looks exactly like a transcript of twenty real ones, and it will arrive by Thursday, cost almost nothing, and tend to confirm the hypothesis that shaped the prompt.

Where it genuinely helps

There are real uses, and dismissing the whole category is as lazy as adopting it uncritically. The common thread is that the model works on the instrument, not on the conclusion.

• Pressure testing a discussion guide before you spend on recruitment, to find leading or confusing questions
• Generating a wide range of hypotheses to take to real people, rather than instead of them
• Rehearsing a stakeholder debrief so the awkward questions arrive before the room does
• Checking survey wording for ambiguity across different reading levels
• Summarising and coding transcripts you actually collected, with a person checking the codes

In each case a person still makes the judgement and the real evidence still comes from real people. The model is doing preparation work that would otherwise be skipped for lack of time.

Where it misleads

Simulation is weakest exactly where research earns its money: on the specific, the local and the new. It cannot tell you how customers in a particular Australian regional market feel about a delivery promise your competitor cannot match. It cannot tell you what changed in the last three months. It has no access to the things your customers know and have never written down.

It is also systematically agreeable. A model built to be helpful will find the merit in your concept, because the prompt described the concept and helpfulness means engaging with it. Real customers are indifferent, distracted and occasionally rude about ideas that have consumed a team for a quarter. That indifference is the most valuable signal available, and simulation removes it.

The deeper problem sits underneath. Stated preference research was already unreliable as a guide to behaviour. Simulating stated preference gives you a model of what people say they would do, one further remove away from what they actually do.

A simulated customer will never surprise you with something you did not think to ask, and the surprise is the entire reason to do research.

Keeping real contact in the process

The risk is not that a team consciously replaces fieldwork. It is that the cheap option quietly crowds out the expensive one, a quarter at a time, until nobody in the marketing function has spoken to a customer this year.

Guard it with a floor rather than a policy. A minimum number of real conversations each quarter, attended by the people making decisions rather than delegated to an agency, with the recordings watched instead of the summary read. For Australian teams the practical version is small and local: intercepts in store, calls with recent buyers, a morning sitting beside the service team. It needs very little budget, only calendar time that senior people are reluctant to give up.

How to report it honestly

If simulation informed a decision, label it. Say which findings came from real people, which came from a model and which came from your own reasoning. A research output that blends the three without marking them is not research. It is a document that will be quoted later with far more authority than it earned.