Synthetic Customer Research: What a Census Misses

Methodology · 11 min read

TL;DR: A census-grounded panel can be drawn to match your customers' demographics. It cannot be drawn to be your customers, because no statistical office publishes who buys from you. The defining attribute of a customer base is an outcome, and the panel's grounding stops at the attributes a census measures. What comes back is a look-alike: the population subgroup your buyers were selected from, with the selecting removed. That makes prospect questions, category attitudes and instrument pretesting safe. It makes satisfaction, churn reasons, price paid and anything resting on product experience unsafe. This guide sets the boundary, names which customer questions survive, and shows where the panel earns its place around a customer study.

Can a synthetic panel represent your existing customers?

No. It can represent people who resemble them on measured attributes. A census-grounded draw reproduces the composition you specify, and your customer base is defined by something no census records: the fact that they chose you. What comes back is the look-alike population your buyers came from, with the selecting removed.

Synthetic customer research is the practice of putting questions about your own buyers to a generated panel instead of to the buyers. It is the fastest-growing claim in the category, and it is the one with the loosest boundary.

Start with what a draw actually does. You specify age, income, region, education and a category profile against national statistics, and the panel returns a sample matching that specification. That works because the statistics exist. Every attribute you calibrate on is one a statistical office publishes a distribution for.

Your customer base has no such distribution. It is not a slice of the population defined by an attribute. It is a residue left by a process: people saw your product, could afford it, preferred it to the alternative, and stayed. Two people with identical census attributes sit on opposite sides of that line every day.

So the honest description of what you get is a look-alike panel. It matches your customers on everything measured and differs from them on the one thing that made them customers.

Why is a customer base not a census category?

Because a census measures attributes people have, not choices they made about your brand. The U.S. Census Bureau lists the subjects the American Community Survey covers, and they are demographic, social, economic and housing characteristics. None of them is a purchase from you. The draw has no target to calibrate against.

The subjects page is worth reading once for this purpose alone. It sets out what the American Community Survey collects: age, sex, race, ancestry, education, employment, income, commute, housing tenure, and related characteristics.¹ The list is long and it stops well short of brand choice. The same holds for Eurostat and for every national statistical office behind a country panel.

That gap has a second consequence beyond the missing target. When you try to compensate by stacking constraints on the persona brief, asking for a 38-year-old urban graduate who has held a subscription for two years and switched from a competitor, the specification runs past what conditioning can carry. A 2026 arXiv study reports that language models simulate intersectional synthetic identities with an effective budget of roughly one to two dimensions.² Ask for six and you get one or two of them, plus fluent text that reads as though you got all six.

The error also moves. Writing in Political Analysis, Bisbee and colleagues found that synthetic responses can sit close to a human average while understating the variation around it, and that the size of the gap differs across subgroups rather than holding steady.³ A customer base is a subgroup, and it is exactly the kind whose offset you cannot look up.

This is a different constraint from the one covered in low-incidence audiences on a synthetic panel. That problem is rarity: whether an audience is documented well enough to represent. This one is not about rarity at all. A brand with 40 percent category penetration has an enormous customer base and still cannot have it drawn, because the defining attribute was never measured.

What does a look-alike panel actually get right?

Composition, category attitudes and prospect-side reasoning. A look-alike panel answers as the population subgroup your customers were drawn from, which is the right frame for people who do not yet buy from you. It drifts as soon as the question depends on having used your product, because nothing in the grounding encodes that use.

The distinction that matters is not customer versus non-customer. It is whether the answer requires experience your panel never had.

A question about how a category feels to a 45-year-old with two children and a commute is answerable from population grounding, because that is the kind of thing a national survey measures and a census-calibrated draw reproduces. A question about how your onboarding felt is not, because the only source for it is people who went through your onboarding.

Cross-domain benchmarking supports drawing the line there rather than at the category level. A 2026 arXiv benchmark of LLM-simulated survey responses reports that performance is domain-dependent rather than uniform, which is the useful finding: the method does not fail everywhere or work everywhere.⁴ A separate psychometric audit makes the failure mode explicit in its title, describing synthetic responses as plausible but not valid.⁵ Plausible is what you get when a model answers a question it has no grounding for. It reads correctly and it is not a measurement.

The table below sorts the common customer questions by which side of the line they fall on.

Which customer questions still need your own buyers?

Four classes. Anything resting on product experience, such as satisfaction or support quality. Anything resting on a transaction record, such as price paid or renewal behaviour. Anything resting on a remembered event, such as why someone churned. And anything that will be filed as evidence, where the record has to show real people were asked.

Experience questions fail first and most quietly. A persona asked to rate your product will produce a rating, a rationale and a distribution that looks reasonable. None of it was caused by your product, because the persona has no contact with it. Reliability assessments of persona-conditioned models find that the conditioning changes output in ways that need checking rather than assuming, which is the right posture even before you reach an attribute the conditioning cannot reach at all.⁶

Transaction questions fail for a simpler reason. The answer already exists in your billing system and is a fact rather than an opinion. Asking a panel to estimate it substitutes a guess for a record you own.

Memory questions fail because the model has nothing to recall. A stated reason for churn produced by a persona is a plausible reason for churn in general, which is a hypothesis worth testing and not a finding about your cohort.

Evidence questions fail on the record rather than the accuracy. Where an output has to stand up in a substantiation file or a regulator's review, what is being examined is whether people were asked, and a generated panel has nothing to file even when it happens to be right.

A fifth case sits between the two sides. Where you have a small number of real customer completes and want a wider read, the answer is an estimation problem with a known structure rather than a substitution, covered in combining synthetic and human respondents.

Can you build a synthetic panel from your CRM data?

You can use first-party data to specify the draw, which is worth doing. You cannot use it to transfer experience. Feeding customer records into a generator returns a panel shaped like your base and still answering from population priors, and it puts personal data back into a workflow that was clean without it.

Several vendors now sell personas built from a client's own analytics or CRM. The useful half of that offer is real: your first-party data tells you what your customers look like, and a composition you measured beats a composition you guessed. Take that part.

The part that does not follow is the inference. A profile assembled from records describes the people in the records. It does not give the generated persona their history. The persona still answers from what the model knows about that demographic profile in general, which is the same place it answered from before you uploaded anything. The upload changed the specification, not the source of the answers.

There is a cost on the other side too. A panel grounded in aggregated public statistics processes no personal data, which is the whole basis of the privacy argument for synthetic personas. Routing your customer file into persona generation gives that back and buys composition you could have specified from an aggregate summary instead.

The practical version is short. Pull the demographic and firmographic summary of your base, specify the draw against it, and keep individual records out of the pipeline entirely.

How should a synthetic panel sit around a customer study?

Upstream of it and beside it, not in place of it. Use it to pretest the questionnaire before it reaches paying customers, to generate hypotheses worth spending real completes on, to answer the prospect side of a decision, and to cover markets where you have no customer list yet.

Customer completes are the scarcest input most insight teams have. You can only ask a customer so often before response rates fall and goodwill goes with them. That scarcity is the argument for the panel, and it is an argument about sequencing rather than replacement.

Five steps cover most of it.

First, pretest the instrument. Run the questionnaire on a panel drawn to your base's composition and look for questions that produce confusion, order effects or flat distributions. Fix those before the survey reaches a customer.

Second, generate and rank hypotheses. Put the open question to the panel, collect the reasons it produces, and treat the output as a list of things to test rather than a result. Diagnostic work on LLM consumer panels is built on exactly this posture of checking before trusting.⁷

Third, split the study. Send the experience questions to your customers and the category and prospect questions to the panel. Report them as two sources, which is what AAPOR's 2026 task force report on responsible AI integration in survey research and the ESOMAR and GRBN guideline on online sample quality both point toward: disclosure of provenance rather than a blended number.⁸⁹

Fourth, measure your own offset. Once you have both, compare the panel's answers against your customer answers on the questions both could take. That gap is your category's offset, and it is more useful than any published benchmark.

Fifth, extend to markets without a base. Where you have no customers yet, the look-alike panel is not a substitute for anything, because there is nothing to substitute for. This is where a census-grounded draw does its cleanest work, and it is also where market sizing on a synthetic panel sets out which inputs still have to come from elsewhere.

PersonaHive draws panels grounded in national census data in nine countries, each calibrated to its own statistical office and validated against real national surveys, so the composition of a look-alike draw is documented rather than asserted. The free tier includes 250 credits and no card. The smallest useful first step: take the last customer survey you fielded, re-run only its category questions on a panel drawn to your base's demographic summary, and put the two sets of answers side by side. The questions where they agree are the ones you can stop spending customer goodwill on.

What else do teams ask about synthetic customer research?

Five questions recur once the boundary is clear: whether a tight persona brief can simulate a customer, whether synthetic NPS means anything, what to tell a stakeholder who was promised customer personas, whether the panel can model churn, and how to describe the sample in a methods note. Short answers follow.

**Can a detailed persona brief make a panel behave like our customers?** Detail helps up to a point and then stops. Conditioning carries roughly one to two effective dimensions, so a brief stacking six customer-specific constraints returns fluent text built on one or two of them.

**Is a synthetic Net Promoter Score worth anything?** Not as a score. NPS measures people who have used a product, and a look-alike panel has not. The panel can pretest the wording and the follow-up probe. The number has to come from customers.

**A stakeholder was promised personas of our customers. What do we tell them?** That they are getting a panel matched to the customer base's composition, which answers category and prospect questions, and that experience questions are staying with the customer survey. The distinction survives a methods review; the looser claim does not.

**Can the panel model churn?** It can produce reasons people leave categories like yours, which is useful for building the churn survey. It cannot tell you why your cohort left, because that sits in an event it has no access to.

**How should the sample be described in the report?** As a census-grounded panel drawn to match the demographic composition of the customer base, with the country, census vintage, draw specification, unweighted cell counts, model version and run date recorded. Do not call it a customer sample. It is not one, and the label is the part a reviewer will catch.

Sources

• Subjects Included in the Survey (American Community Survey) — U.S. Census Bureau
• Large language models simulate intersectional synthetic identities with a budget of one to two dimensions — arXiv
• Synthetic Replacements for Human Survey Data? The Perils of Large Language Models — James Bisbee and colleagues, Political Analysis
• When Synthetic Users Fail: A Cross-Domain Benchmark of LLM-Simulated Human Survey Responses — arXiv
• Plausible but Not Valid: A Psychometric Audit of LLMs as Synthetic Survey Respondents — arXiv
• Assessing the Reliability of Persona-Conditioned LLMs as Synthetic Survey Respondents — arXiv
• When Can You Trust Your Synthetic Users? Diagnostics and Corrections for LLM Consumer Panels — arXiv
• Responsible AI Integration in Survey Research (Task Force Report) — AAPOR
• ESOMAR/GRBN Guideline on Online Sample Quality — ESOMAR and GRBN