What is synthetic customer research? A plain-English guide

Synthetic customer research is the practice of using AI models to simulate how a group of customers might respond to a product, an offer, a price, or a message, instead of recruiting real respondents for every question. The models are built from real behavioural and survey data. The output looks like research: response distributions, ranked objections, comparisons between scenarios. The input is a decision you have not made yet.

The category has grown fast. Analysts at NielsenIQ describe synthetic respondents as predictive research models presented in the form of respondents, and Qualtrics now sells synthetic consumer panels alongside its traditional ones. That growth is why the word "synthetic" now needs a plain-English definition rather than a sales pitch.

How it works, in four steps

1. A model is built from real data. Purchase records, survey panels, behavioural logs, sometimes a company's own customer data. The model learns relationships between who someone is, what they have done, and what they chose.
2. A decision is framed as a scenario. "Repeat buyers see a three-pack at ₹1,499 while singles stay at ₹549." The scenario becomes the stimulus.
3. The model simulates responses for each audience. Not one answer, a distribution: what share would buy, would consider, would not, and the reasons that cluster around each.
4. Scenarios are compared. The value is rarely a single number. It is the gap between scenarios and the reasons behind it, which tells you what is worth testing for real.

What it is good at

• Exploration before commitment. Comparing five price points or three bundle configurations in an afternoon, before any spend.
• Surfacing barriers. Why a first-time visitor will not commit to three units before trying one. These objections are often the most useful output.
• Prioritising the real test. Deciding which single experiment would change your mind, so you run one instead of five.

Where it breaks

• Point estimates. A simulated "47% would buy" is a direction, not a forecast. Treat the decimal places as noise.
• Novel behaviour. If the decision has no analogue in the data the model was built from, the simulation is a guess dressed as a distribution.
• Unvalidated accuracy. Reviews of the category, including Greenbook's explainer and PyMC Labs' practical guide, agree on one point: accuracy depends on the audience, the task, the model, and the validation method. There is no category-wide accuracy number, and any tool quoting one without showing the comparison is quoting a marketing figure.

How to use it without fooling yourself

Ask three questions of any synthetic research, including ours.

1. What real data is underneath this audience? If the answer is "the model's general knowledge", weight the output accordingly.
2. Has this tool been scored against a decision like mine? Not a benchmark on a public dataset. A decision, with the outcome, in your category.
3. What real-world test does this point me to? If the output ends in a verdict rather than a test, it is being asked to do a job it cannot do yet.

Where Precheck stands

Precheck is a decision-intelligence platform that runs data-informed simulations of customer response. We say plainly on our own site that predictive accuracy has not been established yet. We are running simulations for a small number of large companies on data they have shared, and we invite teams to send a decision they already made so we can simulate it blind and compare the output with what actually happened. That comparison is the only honest path to an accuracy claim, and we would rather earn it than assert it.