How do synthetic audiences work?
You've run into this problem before. Your team needs an answer by Thursday, but a traditional survey panel takes three weeks — so the decision gets made on intuition instead. What if there were a middle option: answers on demand, built from a model of your actual audience, that come back instantly and are almost as accurate as asking your customers in real life?
What is a synthetic audience?
A synthetic audience is a model of your customers, built from your customer data, so you can ask a question in plain language and get answers back right away.
The difference between a synthetic audience and an off-the-shelf LLM is that we simulate all of your customers and represent each of their answers distinctly, rather than as one garbled average. We measure our accuracy not just on how closely the audience answers like your customers, but on how closely the whole distribution of answers matches the spread of your real customers' answers. Put simply: we show you where your customers disagree, where they align, and the full range in between — so you can use that to inform a decision in any scenario.
How does a synthetic audience get built?
Stage one, we take your qualitative or quantitative data and build an audience modelled from the original responses. Every answer is rooted in ground truth from your dataset.
Stage two, we test it. We train the audience on 80% of the questions in your dataset, then test it on the remaining 20% it has never seen, and measure how closely its answers match those of the real respondents on the held-out questions.
Stage three, you ask. Ask in whatever format you need, all in one session — a multiple or single-choice survey, a focus group, a message test, or an image test. Answers come back in seconds, so you can read a surprising result, sharpen the question, and ask again without commissioning a new survey.
Stage four, we show the results in a form you can take straight into the boardroom — insight summaries, interactive graphs, and slides you can download to PowerPoint and share in minutes.
Stage | What goes in | What comes out | How you check it |
|---|---|---|---|
1. Build | Your real survey or first-party customer data | An audience modelled individual by individual from the real responses, outliers included | It's built from your data, so every answer is rooted in ground truth |
2. Test | 80% of the dataset to train, 20% held back to test | A score for how closely the audience matches answers it never saw | Two published measures: 1-MAE and NDAM |
3. Ask | A question in plain language — survey, focus group, message or image test | Answers in seconds, broken out by segment | Ask a follow-up; a model that's guessing breaks under one |
4. Read | The results of your questions | Insight summaries, interactive graphs and downloadable slides | Segment the output and read the spread, not just the average |
What data does a synthetic audience need?
To build a synthetic audience that simulates your customers, we need real quantitative data from the people you want to model. That primarily means survey data, or any nationally representative dataset. If none is immediately available, we can work with you to acquire the right data.
Once the audience is built, we run evaluations and hold-out testing together with you, so you trust your audience's answers and can reliably interpret the results.
What is the difference between quantitative and qualitative methods in our platform?
Quantitative tools in Electric Twin let you run survey questions and get segmented customer responses, just as you would in a traditional real-world survey.
Qualitative tools in Electric Twin let you run focus groups. You can take any customer segment from your survey and pull it into a focus group to debate why they answered the way they did — all in a single session. The full set of tools lives on one platform.
The general rule: if you need to know how many, run the structured version. If you need to know why, run the conversation.
How do you know the answers are right?
We run hold-out tests: we show the audience questions from your original dataset that it has never seen, and measure how closely its answers match the real ones.
We start by splitting your dataset in two. The first part builds and trains the audience. The second part is held out and never shown to the audience. We then test accuracy by asking those held-out questions and scoring how closely the answers match the real-world ones. Because those questions never appear in the training data, we know the audience is answering from the model, not from memory.
We report two accuracy figures. Up to 96% on 1-MAE, the industry-standard measure, and up to 92% on NDAM, our stricter measure that scores the shape of the whole distribution of answers rather than just how close the average landed.
Why two numbers? 1-MAE gets more generous as a question gains answer options, and a measure that flatters isn't one you should quote on its own.
The Times ran this protocol on its own data. It took a 10% holdout from a 642,000-subscriber base, built the audience from the rest, and scored the synthetic answers against the holdout. The result was a 92% NDAM match, against roughly 94% for the traditional methods already in use.
The full science, both accuracy scores and the published research, are on Electric Twin's accuracy page — and we break the measures down further in how accurate are synthetic audiences.
How is this different from asking ChatGPT?
ChatGPT trains on all the data it can find on the internet and gives you an answer averaged across all of it, to appeal to the majority. Ask it what your customer thinks and it hands you a composite of everyone on the internet.
On NDAM, GPT-5 reaches 65%. It's good at an about-right, generic answer, but it can't pinpoint what your specific customers said or give you a clear distribution of their responses. We score up to 92% on the same measure. There's more on this in how to use ChatGPT for market research and in Electric Twin compared with LLMs.
Distribution accuracy on NDAM, against a frontier model with no grounding in your audience, and against the rate at which real people agree with themselves. Source: Electric Twin accuracy methodology.
Why is even traditional research only 96% accurate?
Ask a person the same question twice and they'll give the same answer only about 96% of the time on 1-MAE. People are fickle — they misread questions and change their minds. So the ceiling for all research, traditional or synthetic, is around 96%, not 100%.
What can you actually use a synthetic audience for?
Job | What changes |
|---|---|
Questions too small for a study | A subject line, a product name, one pricing tier — currently decided by whoever's most senior in the room |
Audiences too niche to reach | A B2B segment of two hundred. High-net-worth individuals. Policyholders with exactly one claim. |
Repeat questions | Once fieldwork cost drops out, the same tracking question runs monthly instead of quarterly |
Segment contrast | Where two parts of one audience want opposite things — which an averaged result hides entirely |
Testing a shortlist of names, taglines or subject lines is covered in message testing with synthetic audiences, and publishers testing formats before they commission will find the detail in testing content before you commission it.
Where does this not work?
We recommend not using Electric Twin in these scenarios:
Audiences with too little data. Accuracy depends on the depth of real audience data behind the model. For very thin or newly formed audiences, primary research comes first — and a vendor who tells you otherwise is overselling.
Claims that require human respondents on record. Regulated claims, clinical studies and legally mandated consumer testing need real people who can be identified and counted. A synthetic audience doesn't replace those.
A replacement for your research programme. Electric Twin runs alongside surveys. Traditional research is the yardstick our accuracy is measured against, so periodic real-world validation isn't optional — it's what keeps the model calibrated. An audience that's never re-checked drifts away from the people it represents. We lay out how the two fit together in synthetic vs traditional research.
Frequently asked questions
How long does a synthetic audience take to answer a question?
In seconds. There's no recruitment, fieldwork or scheduling, so you can ask, read, sharpen the question and ask again.
Can a synthetic audience be built on our own customer data?
Yes — that's the preferred dataset. Electric Twin builds from customer-provided survey datasets as well as in-house nationally representative ones.
How often does a synthetic audience need re-validating?
Periodically, against fresh real responses. Audiences change, and re-validation keeps the model tracking the people it represents.
Is the accuracy measured on your data or ours?
Yours. Every onboarding takes your real data, builds the audience and runs the hold-out evaluation on it, so the figure you see is measured on your customers rather than an arbitrary benchmark.
Are synthetic audiences the same as AI personas?
No. A persona prompt asks a general-purpose model to act a part, with nothing to score it against. A synthetic audience is built from real data about real people and validated against held-out real answers. We go deeper in why generic LLMs aren't synthetic audiences and synthetic respondents, personas, panels: what's the difference.
