What is a synthetic audience?

A synthetic audience is a group of simulated people: personas generated by a language model, each with a background, a situation and concerns, who read your material and react to it in place of real respondents. Teams use one to rehearse a decision before real people see it, to find the objections, the confusion and the differences between groups early. It is directional evidence, not a survey: nobody is recruited, nothing is sampled, and the results are not statistics about any population.

Written by MiroFish.co.uk, which runs synthetic audience studies, so read it with that in mind. The research summaries link to the papers themselves.

How a synthetic audience works

There are two broad designs.

One-shot respondents. Each persona answers on its own, as in a survey or an interview. This is quick and easy to compare across personas, but nobody hears anybody else.
Simulated discussion. The personas read the material and then react to each other over several rounds, as people do in a group or on social media. Objections can spread, arguments can win support and groups can split. MiroFish works this way: its agents post, comment and react on two simulated social platforms, run by the open-source OASIS framework.

In both designs the personas are generated by a language model from what you provide: the material, a description of the people you want to understand, and sometimes data about them.

What teams use one for

Testing a campaign message or a headline before the budget is spent.
Checking how customers might read a pricing change.
Rehearsing a launch and its positioning.
Stress-testing a public statement before it is published.
Deciding which questions to put to real customers in interviews or a survey.

What research says so far

Academic work on language-model respondents is recent and mixed, and worth reading before you rely on any synthetic result.

Argyle and colleagues (Political Analysis, 2023) conditioned a language model on the backstories of thousands of real survey respondents and found that it reproduced the response patterns of many subgroups, a property they call algorithmic fidelity.
Bisbee and colleagues (Political Analysis, 2024) found that synthetic survey answers matched the average answers of a real survey closely, but varied less than real people's answers, shifted with small changes in prompt wording and gave different results for the same prompt three months apart, and they raised serious concerns about using them as a replacement for human survey data.
Horton (2023) placed language models in classic economics experiments and found results qualitatively similar to the original studies with human participants.
Park and colleagues (2023) showed that generative agents with memory, reflection and planning produce believable individual and group behaviour in a simulated town.
The OASIS team (2024) simulated X-like and Reddit-like platforms with up to a million agents and reproduced social phenomena such as information spreading, group polarisation and herd effects.

Where synthetic audiences stop

They are not a sample. The personas are generated, not drawn from a population, so there is no margin of error and no way to say what share of real people would react a given way.
They inherit the model's assumptions. A language model's idea of a group can be stereotyped, out of date or wrong, and a persona built on it will be too.
They depend on their inputs. The material, the audience description and the model all shape the outcome, and two runs on the same material can differ.
They are not a forecast. A simulation shows what could plausibly happen under the assumptions you set, not what will happen.

Used for what it is good at, a synthetic audience finds objections, confusing wording and fault lines between groups quickly, and tells you what to check with real people. Used as a replacement for real research, it misleads.

Synthetic audience, survey or focus group?

Three ways to hear from an audience
	Synthetic audience	Survey	Focus group
Who answers	Generated personas	Real people, sampled	Real people, recruited
Typical time to results	Minutes to hours	Hours to weeks	Days to weeks
What it tells you	Likely objections, confusion and group differences, to explore	How many people think what, within a margin of error	Why people react as they do, in depth
Interaction between participants	Simulated, over rounds, in some designs	None	Real, with a moderator
Main risk	Plausible reactions that real people would not have	Leading questions and unrepresentative samples	Small groups and dominant voices
Best used	Before real research, to sharpen it	To measure	To understand

How to judge a synthetic audience tool

Can you see the evidence? Findings should lead back to what the personas actually said.
Are the limits stated plainly, or is it sold as prediction or as polling?
Can you describe the audience in your own words, and see who was simulated?
Do reactions develop over several rounds, or is each persona asked once?
Where does your material go, and which model providers process it?

Questions people ask

Is a synthetic audience the same as synthetic data?

Not quite. Synthetic data usually means generated records that imitate a real dataset. A synthetic audience is a set of simulated people who react to material; it may be built with or without real data about the people it stands for.

Can a synthetic audience replace market research?

No. It is useful before research, to find objections and sharpen the questions, and between rounds of research, to try changes quickly. A decision that depends on how many people think something needs real respondents.

How many personas does a synthetic audience need?

There is no standard number. More personas add variety, but because they are generated rather than sampled, adding personas does not make the result statistically representative.

What is the difference between a synthetic audience and synthetic users?

Mostly emphasis. Synthetic users usually stand in for users in product and UX research; a synthetic audience usually reacts to a message, a price or an announcement.

Sources

Start a free study