City-Level Synthetic Personas: Where Geography Stops

Methodology · 10 min read

TL;DR: City-level synthetic personas ask a census-grounded panel for something national grounding was not built to deliver, and two separate floors set the limit. The first is publication: Eurostat's NUTS 2024 classification carries 1,165 NUTS 3 regions, but below that only total population and area are collected, and national offices perturb or swap small-area census counts before release. The second floor is the model. A 2026 study of 21 million simulated response distributions found one feature explains a two-feature persona's answers better than the additive combination in 75 to 82 percent of subgroups, and a third feature adds almost nothing. Region is usually the third. The model floor binds first, and it is the one to test.

Can a synthetic panel represent a specific city?

Not the way a national panel represents a country. Two separate floors decide how local a census-grounded panel can go: what your statistics office publishes at that geography, and how many persona dimensions the model will actually honor. The second floor sits higher than the first, and it is the one vendors rarely raise.

A brief arrives that says Manchester, or Bavaria, or the Midwest. The audience builder takes the word. A panel comes back. Nothing in the output tells you whether the geography did any work.

That silence is the problem. City-level synthetic personas fail quietly rather than loudly: a model asked to answer as a Manchester shopper will produce a fluent Manchester-sounding answer whether or not Manchester changed anything in the distribution underneath.

So the useful question is not whether a platform accepts a place name. It is which floor you hit first, and how you would know.

How far down does a national census actually go?

Further than most briefs need, and in worse condition than most briefs assume. Eurostat's NUTS 2024 classification carries 1,165 NUTS 3 regions, and below that sit municipalities for which Eurostat collects only total population and total area. Depth of detail does not survive the descent. Cross-tabulations collapse first.

Start with what exists. Eurostat maintains NUTS, the classification that lets EU statistics be reported by region. The NUTS 2024 version, valid from 1 January 2024, defines 92 regions at NUTS 1, 244 at NUTS 2 and 1,165 at NUTS 3.¹ A NUTS 2027 version takes effect on 1 January 2027, so regional boundaries are not fixed furniture either.¹

Below NUTS 3 sit the local administrative units, the EU's municipalities and communes. Eurostat calls them the building blocks of NUTS and says that where available it receives each unit's total population and total area through the annual LAU lists.² Population and area. Not income distribution, not household composition, not the attributes a persona needs.

The condition the numbers arrive in matters as much as their existence. The UK Office for National Statistics applies targeted record swapping and the cell key method to Census 2021 outputs, and states that a typical dataset has around 14 percent of cell counts perturbed by a small amount, with small counts more likely to have been perturbed.³ Record swapping changes the geography recorded for between 7 and 10 percent of households.³

The United States has a different floor for a different reason. Detailed local demographics come from the American Community Survey, which is a sample rather than a complete count, published with margins of error at a 90 percent confidence level.⁴ The Census Bureau tells researchers to be careful drawing conclusions from small differences between small-area estimates, and suggests aggregating tracts or block groups to get estimates that are more reliable.⁴

Local census data is not useless. It is a different kind of input: a total you can read, not a cell you can condition a persona on.

Why does the model drop the geography even when the data exists?

Because a persona prompt has a budget. A 2026 study of 21 million simulated response distributions found that one feature explains a two-feature persona's answers better than the additive combination in 75 to 82 percent of subgroups, and that a third feature adds almost nothing. Region is usually the third.

This is the finding that should change how you read a regional brief. Rennard and Xypolopoulos, in a preprint posted on 24 August 2026, simulated responses across 15 waves of Pew's American Trends Panel using eight models from five organizations.⁵ Given a persona with two identity features, a single one of them explained its answers better than the additive combination in 75 to 82 percent of subgroups: 78.4 percent for GPT-4o-mini, 79.1 percent for GPT-5.5, 77.8 percent for Claude Sonnet 5.⁵ The collapse, they write, survives every prompting strategy they test.⁵

Two details matter more than the headline. Real people move the other way: the paper reports subgroup opinion growing 2.5 times more distinctive as identities intersect.⁵ The gap between model and population widens exactly where a regional brief wants detail. And the models systematically discard race and religion, which the paper identifies as the strongest real drivers of opinion.⁵ A persona does not drop the feature you would least miss.

The paper did not test geography as one of its dimensions, so treat the application as an inference rather than a measurement: when a brief already fixes age, sex and income, a region tag enters as the third or fourth feature, which is the territory the study measured.

A second 2026 paper points the same way. Taday Morocho and colleagues, in work in the Companion Proceedings of the ACM Web Conference 2026, found that persona prompting does not yield a clear aggregate improvement in survey alignment and in many cases significantly degrades performance, with the largest shifts concentrated in small-n strata.⁶ A city is a small stratum.

Which of the two floors binds first?

The model floor, in nearly every brief. Statistics offices publish usable regional totals well below the level at which a persona stops carrying its region. So the question worth testing is not whether the census covers your city. It is whether the panel answers differently when you change the city and change nothing else.

Vendors argue about data coverage because data coverage is the floor they control. Adding a country, licensing a tabulation, extending a classification: all of that is procurement. The conditioning limit is a property of the model, and no amount of data buying moves it.

There is a measurement reason to distrust a national validation result here too. Tigre and Souto argue that aggregate validation metrics conceal systematic failures, and report subgroup error balloons of 10 to 30 percentage points.⁷ They cite a result of 90.2 percent aggregate correlation alongside 67 percent segment accuracy against a 50 percent chance baseline.⁷ A panel that matched a national benchmark has told you nothing about its Lombardy cell. If you intend to publish regional readings, the validation study you run has to contain a regional comparison.

The practical test is cheap. Hold every other persona attribute fixed, change only the region, and look at whether the distribution moves and whether it moves in a direction the market would predict. If two regions you know to differ return the same distribution, the geography is decoration.

Which regional readings from a synthetic panel still hold?

A rank order inside one region holds best, because every option carried the same geography tag. A contrast between two large and genuinely distinct regions holds next, provided region is the only dimension you varied. Local incidence, local market share and region-by-segment cross-tabulations do not hold at all.

Sort the outputs by how much of the geographic claim each one rests on.

A ranking inside one named region rests on it least. Every option in the set was shown to the same panel with the same region tag, so whatever that tag did applies across the set. The ordering survives, which is the same reason message testing travels well on synthetic panels.

A contrast between two regions rests on it most, and only pays out when the regions are large, well documented and culturally distinct, and when nothing else about the two panels differs. Three small regions compared against each other is a cross-tabulation wearing a map.

A local number, meaning incidence, penetration or share in a named place, does not come from a synthetic panel at all. It is the same boundary as customer-base research, where no statistical office publishes who buys from you. Rarity and place are different constraints, so an audience that is thin on the ground everywhere is a separate problem from one that is concentrated in a place.

How do you run a regional study on a synthetic panel?

Six steps. Check what your statistics office publishes at the named geography, make region the only dimension that varies, run a region-swap control before the real study, keep the attribute list short, read rank orders rather than levels, and send any local number to a source that counts people.

Step one: look up the level. Before the brief is agreed, find out whether the geography in it is a published statistical region or a marketing convenience. NUTS 3 is published. A sales territory is not.¹

Step two: cut the attribute list. Every dimension you add competes with region for the budget the model actually spends.⁵ If the study is about place, make place the second attribute, not the fifth.

Step three: run the swap control first. Two regions, identical questions, identical attributes, nothing else moved. If the distributions are indistinguishable, rewrite the study as a national one. This costs a few hundred credits and saves a deck.

Step four: fix the run conditions. Same model, same date, same prompt structure across every region, recorded in the methods note. A regional comparison made across two model versions is not a comparison.

Step five: read ordering, not level. Report that region A ranked the second concept highest and region B ranked the first. Do not report that region A scored 6.2 and region B 5.4.

Step six: route the local numbers out. Incidence, share and penetration go to measurement data or a local sample.

PersonaHive grounds each panel in its own national census across nine countries: the United States, Germany, France, Austria, the Czech Republic, Hungary, Romania, Denmark and Finland. Personas are specified from aggregated public statistics and validated against real surveys, which is what makes the national layer trustworthy and the sub-national layer worth testing rather than assuming. Why national grounding beats one global panel sets out the first half of that argument.

The smallest useful next step is the swap control itself. New accounts get free credits without a card, and one credit covers one persona answering one question, so two regions and a ten-question battery is a small run. Do that before you accept a brief with a city in it.

Frequently asked questions about regional synthetic panels

Five questions recur once the two floors are visible: whether more census data fixes the problem, whether naming a famous city helps, how large a region needs to be, whether weighting rescues a regional cell, and what to tell a client who asked for city-level work.

Would licensing more granular census data fix this? No, and that is the counterintuitive part. The publication floor is real, but the model floor sits above it. Better local data gives you a better specification for a panel that still spends its budget on one or two dimensions.⁵

Does naming a well-known city work better than naming a small one? It changes the answer, which is not the same as improving it. A model has read far more about Paris than about Clermont-Ferrand, so a Paris tag pulls in associations from text rather than from the population.

How large does a region have to be before a contrast is worth reading? No threshold is published, so treat it as an empirical question per study. The region should be one your statistics office reports on in its own right, and the swap control should show movement in a direction you can defend.

Can weighting fix a regional cell? Weighting corrects composition, not conditioning. If the model never honored the region, raking the file to regional totals makes the margins right and leaves the answers where they were.

What do you tell a stakeholder who asked for city-level work? That the panel can rank options for them this week at national or major-region level, and that the city number they want is a measurement purchase, not a research design.

Sources

• NUTS: Nomenclature of territorial units for statistics — European Commission, Eurostat
• Local administrative units (LAU) — European Commission, Eurostat
• Protecting personal data in Census 2021 results — Office for National Statistics, Office for National Statistics
• Understanding and Using American Community Survey Data: Sampling Error in the ACS — United States Census Bureau, United States Census Bureau
• Large language models simulate intersectional synthetic identities with a budget of one to two dimensions — Rennard and Xypolopoulos, arXiv
• Assessing the Reliability of Persona-Conditioned LLMs as Synthetic Survey Respondents — Taday Morocho et al., arXiv, Companion Proceedings of the ACM Web Conference 2026
• When Can You Trust Your Synthetic Users? Diagnostics and Corrections for LLM Consumer Panels — Tigre and Souto, arXiv