Audience forecasting with synthetic audiences

What synthetic audiences are.

A synthetic audience is a set of simulated respondents, generated by an AI model and given profiles that reflect a group of people you want to understand. You can put questions, messages or concepts to those respondents and read their reactions in minutes, without recruiting anyone.

Synthetic audience research uses these simulated respondents the way traditional research uses a panel or a focus group. The difference is that no person is answering. The model produces what a person with that profile might plausibly say, based on the data it was trained on and the information it was given.

That makes it a form of informed estimate. It can be a useful one, as long as you treat the output as a hypothesis to check. An answer from a simulated audience tells you nothing certain about what a real customer will do.

When synthetic audiences make sense.

They're most useful early, when a decision is cheap to change. Good uses include narrowing a long list of messages before paying for a test, finding objections you hadn't thought of, pressure testing a plan before a client presentation and preparing better questions for real interviews.

They're a poor fit when the decision is expensive or hard to reverse, such as a pricing change or a brand name, unless real research follows. They're also weak for a new kind of product that people have no reference point for, for small communities that are thinly represented in the model's training data, and for health or other sensitive topics where a wrong read could cause harm.

What the research says about accuracy.

Published comparisons between AI-generated answers and real surveys have mixed results. Simulated respondents can come close to real averages on some questions and miss badly on others, and they tend to show less variation than real people do. That last point matters for marketing, because the people who react strongly in either direction are often the ones you need to hear about.

Models also tend to be agreeable. Asked about a concept, a simulated respondent is more likely to find something nice to say, which can make a weak idea look stronger. And stated reactions are a limited guide even with real people, since what someone says about an ad and what they do when it appears in their feed can be quite different.

How we keep the read honest.

We design for these weaknesses from the start. We ground the simulated audience in real material whenever we can, such as your own research, customer interviews, reviews and public survey data, and we write down which sources shaped each profile. If we've got little to go on, we say so.

We give the model neutral questions and test the same idea in different wordings and orders, because a change in phrasing that flips the answer tells you the answer isn't stable. Where possible, we include something with a known outcome, such as a past message whose live performance you already have, to see whether the simulation would have called it correctly.

We report the results as directions and open questions. You won't see a simulated purchase intent score presented as a forecast of sales. Every finding comes with the real test that could confirm or kill it.

From a simulated read to real evidence.

The output of this work is a shorter list and better questions. Messages that hold up go into customer interviews, surveys or live ad tests, where real people and real budget give a stronger answer. We then compare what the simulation predicted with what happened, which tells you how far to trust it for your category next time.

Personal information doesn't belong in a simulation, so we work from aggregated and approved inputs.

How we measure this work.

The real measure comes later, when live results or real research can be set against the early read. We record what the simulation suggested, then compare it with what actually happened and write down where it was right and wrong.