From official statistics to a tested audience

Prefield runs research with simulated audiences. Each market has its own audience, built from that country’s official statistics and interviewed in its own language. Where a country publishes a recurring national benchmark survey, we test the audience against it and correct the measured gaps on the benchmark’s questions. We keep three kinds of evidence apart: fits to results already published, tests on later results, and forecasts sealed before the results exist.

This paper explains the method, the checks at each stage and the limits that remain. It is written for research, insight and product teams who want to understand what stands behind a study before relying on one. It describes methods and safeguards rather than system internals, and it reports no accuracy figures: benchmark evidence is assessed separately for each market, and sealed forecasts are registered on the track record.

The method at a glance

Every market passes through the same six stages. The details differ by country because the data, language and benchmarks differ.

1. Source the data, licenses first

   Identify the country’s population microdata, official population totals, benchmark surveys and release calendars. Record each source’s terms before using its data.

   In plain terms: we know where every number comes from and what we are allowed to do with it.

2. Build the population from real records

   Reweight national survey records so the population matches official totals by age, sex, geography and other characteristics, without inventing combinations the survey never observed.

   In plain terms: if one adult in eight in a region is over 65, one in eight of that region’s simulated adults is too.

3. Write the questions in the market’s language

   Author each market’s interview framing natively, in the language and register its people expect, rather than translating it from English.

   In plain terms: a respondent in Quezon City is asked in formal Filipino; one in Monterrey, in Mexican Spanish.

4. Interview each simulated respondent individually

   Every respondent answers from their own profile. Shares, intervals and verdicts are then calculated from the answers.

   In plain terms: one profile, one set of answers, and arithmetic done in code rather than estimated in prose.

5. Calibrate against a national benchmark

   Ask the audience a benchmark survey’s own questions, for the benchmark’s population and period, then measure and correct the gaps on those questions.

   In plain terms: we check the audience against a survey whose real answers we know, and adjust where it is off.

6. Test forward and show the limits

   Seal forecasts before the official figures exist and score them when they are released. State each market’s evidence stage, sources and limits wherever its results appear.

   In plain terms: being right about the past is not enough; we also make the call before the answer is known.

Principles

Five commitments shape the decisions described in this paper.

Start from the population, not a prompt

A simulated audience can only be as representative as the population behind it. We begin with official statistics and real survey records rather than with personas invented to fit a brief.

Local by construction

Sources, language, corrections and evidence are established separately for each market. Nothing fitted in one country is carried into another: response styles differ enough between countries to reverse the direction of a correction.

Measure before claiming

A claim is only as strong as the evidence behind it. We distinguish fitted results, later-period tests and sealed forecasts, and we describe each for what it is.

Missing means missing

When no source can supply a characteristic, a profile leaves it unstated. We do not fill gaps with default values or borrow them from another country.

Simulated means simulated

The respondents are simulated, and their answers are an early research input. They help decide what to investigate with real customers; they do not replace that research.

Two kinds of audience

A study draws its respondents in one of two ways, and the report says which.

Population panels

A population panel is drawn at random from a market’s full simulated population, or from the part of it that matches the audience filters the market supports, such as age, sex, region or employment. Filters select from the same national population, so any combination of them remains consistently weighted. Everything this paper says about population construction, calibration and evidence applies to population panels, within the scope stated for each market.

Constructed audiences

Some audiences cannot be isolated from population statistics: creative-agency owners, for example, or hospital procurement leads. For these, Prefield can build a constructed audience of archetypes spread across the attributes that define the role. Reports label constructed audiences as not representative samples. They carry no census weighting or benchmark calibration and are best read as a map of hypotheses to test with real people.

Sourcing and licensing data

Before we build a market, we check whether the country publishes the evidence a credible audience needs. The check covers five kinds of source.

Licenses before data

We record each source’s terms of use before using its data, and those terms decide what we do with it. Where a license requires it, source records stay on our servers and are never redistributed. The acknowledgements a license requires travel with every report. When a benchmark’s terms do not permit republication, we use it for calibration without reproducing its published values in reports or on this site. A source whose terms for commercial use are unresolved stays out of the product.

Reproduce before use

Before a source feeds a population or a calibration, we check that our processing reproduces figures its publisher has released. Two examples from current markets:

• In Mexico, weighted average household income computed from the ENIGH 2024 files matches INEGI’s published figure to the peso. Our reconstruction of the Consumer Confidence Index from ENCO’s respondent-level files matches the official series within 0.03 index points in every regular survey month since January 2019.
• In the Philippines, average annual family income computed from the 2023 Family Income and Expenditure Survey files reproduces the Philippine Statistics Authority’s published national figure within one peso.

Where a check like this is automated, a new release that fails it stops the build until the difference is explained.

Building the population

Real records as the skeleton

Where a statistics office publishes person-level survey files, each simulated profile starts from a real survey record: a consistent bundle of age, sex, household, education, work and income observed together. We never assemble a profile by drawing each characteristic separately, which would create combinations that do not occur, such as a teenager with a doctorate. The trade-off is that survey files are samples, often a year or two old, so their totals rarely match today’s population.

We therefore reweight the records. Each record carries a survey weight, and we adjust those weights until the weighted population matches current official totals, while moving the original weights as little as the totals allow. This is raking, or iterative proportional fitting, a long-established survey method. It corrects the totals while preserving the relationships between characteristics that the survey observed.

Each weighted population is checked before use:

• Margin fit. Every calibrated total is compared with its target, category by category.
• Effective sample size. Weighting can hit its targets by leaning heavily on a few records, which quietly reduces the information in the population. We monitor effective sample size and the spread of weights, and treat a large drop as a sign of too many or too fine constraints.
• Matching universes. A target must describe the same population as the records. Published US totals include people living in group quarters, such as dormitories and care homes, so that group is removed from the targets before weighting records that cover households only.

The finished population for each market contains 100,000 profiles, sampled from the weighted records. Study panels are drawn from it at random.

Adding what the core survey does not measure

No single survey records everything that shapes an answer. We add characteristics in three ways, in decreasing order of strength, and record which way each one arrived.

1. Calibrated to an official total. The characteristic has a published population count and enters the weighting itself, so its population share is correct by construction.
2. Joined from the same household. The characteristic belongs to the same respondent or household in a linked file, as household income and composition do in the US and Mexican surveys. No modeling is involved.
3. Matched from another survey. The characteristic exists only in a different survey. Each profile receives the answers of a donor respondent who shares key characteristics, such as age, sex and education. Values and attitudes come from the World Values Survey in every calibrated market. In the United States, party identification comes from the General Social Survey and urban or rural setting from the Consumer Expenditure Survey; in the Philippines, household economics come from the Family Income and Expenditure Survey.

Matched characteristics are modeled, and we treat them that way. They reproduce the donor survey’s distribution and its differences between the groups used for matching, but not relationships with characteristics outside the match. We check that matched values reproduce the donor survey’s own figures and record how closely each profile could be matched.

Missing stays missing

If a market has no source for a characteristic, profiles leave it unstated. A profile never receives a default value, and we never import a value from another country’s data. The Philippine population, for example, has no in-market source for a measure of money worries, so that field is carried as unknown rather than guessed.

Geography

Each market uses the geography its statistics office publishes: 9 census divisions covering the 50 US states and the District of Columbia; the Philippines’ 18 regions; Mexico’s 32 states, grouped into the four macro-regions of Banco de México’s regional reporting; and Saudi Arabia’s 13 administrative regions. Region is available as an audience filter in every market, but a filter is not a calibrated claim. Published regional differences in Mexican consumer confidence, for example, are not reproduced by the audience, so no region-level calibrated claim is made there.

When a country publishes no person-level data

Some countries do not release person-level survey files; Saudi Arabia is one. Its population is constructed from the General Authority for Statistics’ complete 2022 census tables and 2025 population estimates, with employment assigned from the Labour Force Survey using published rates for groups defined by citizenship, sex and age. The result is 100,000 profiles of Saudi and non-Saudi residents aged 16 and over across 13 regions. Characteristics these tables do not link to individuals, including income, education and occupation, are left unspecified rather than invented. A constructed population is labeled exploratory until evidence specific to it exists.

Diversity has a ceiling

A population built from survey records resamples a finite set of real respondents, so its diversity is bounded by the number of distinct source respondents, not by the number of profiles. Reports on these populations state how many distinct source respondents stand behind a panel, and results for small groups are shown as counts rather than percentages.

Speaking the market’s language

People answer differently depending on the language, formality and phrasing of a question, and response habits, such as how readily people choose the ends of a scale, differ between countries. Survey methodologists treat translation as part of questionnaire design rather than a clerical step. Each market’s interview framing is therefore written natively in that market’s language, not machine-translated from English. Where the right register is uncertain, we test it.

• Philippines. We compared formal Filipino, Taglish and English interviews against the Philippine sample of the World Values Survey before selecting formal Filipino.
• Mexico. Mexican Spanish, in the formal usted register.
• United States. English.
• Saudi Arabia. Arabic, in contemporary Modern Standard Arabic, or English, chosen for each study. We do not infer an individual resident’s preferred language.

Reports are written in English for every market.

Anchored in time

Many benchmark questions ask about change: compared with a year ago, or over the next twelve months. A respondent who does not know the date cannot answer those as a person would. In the United States and Mexico, interviews therefore include a short, dated note of facts, such as the date and who governs. The note contains no economic statistics and never the benchmark results the audience will be compared with.

Interviewing the panel

Responses are generated by AI from each respondent’s profile and the study’s questionnaire. Every respondent in a panel receives the same questionnaire and instructions; only the profile differs, and each respondent is interviewed individually.

• Numbers come from answers. Closed questions are answered using the question’s own options. Shares, top-two-box scores, intervals and hypothesis verdicts are calculated from those answers in code. Depending on the scoring method stated in the report, a result counts the options respondents chose or averages the probability each respondent assigned to each option.
• Thresholds come first. In hypothesis studies, the supporting answers and the support threshold are set before the panel responds, so a verdict cannot be fitted to the result.
• Quotations stay attached to their source. Open-ended answers are linked to the profile that gave them, so a quotation can be read alongside who is speaking.
• Diversity is checked. Simulated panels can homogenize, with many different profiles giving the same answer. We measure the spread of answers to closed questions and treat very low spread as a warning sign, not a finding.
• Runs are reproducible. Each study records its run identifier and random seed, so the same population and audience filters select the same respondents.

Intervals in a report describe variation within the simulation. They do not measure how accurately it predicts real people, and a narrow interval is not evidence of accuracy.

Calibrating against benchmarks

A demographically accurate population can still answer inaccurately. Calibration asks whether the audience’s answers match what a country’s people actually said, on questions where the real answers are known.

What counts as a benchmark

A benchmark is a recurring survey, run by a central bank, statistics office or university, that publishes results from a documented method on a known schedule. Today’s benchmarks measure consumer sentiment and expectations.

The two US benchmarks are treated as separate domains and never pooled, because their questions and subgroup definitions differ.

How a market is fitted

1. Match the benchmark’s population and period. The audience is drawn from the population the benchmark covers and compared with the wave for the same period. ENCO samples Mexico’s 32 principal cities, for instance, so Mexican calibration uses residents of localities with 100,000 or more people and measures how the rest of the country differs.
2. Ask the benchmark’s questions. The audience answers the benchmark’s own questions, with their wording and answer options, in the market’s interview language.
3. Score it the benchmark’s way. Answers are summarized with the benchmark’s own measures, such as net balances and the published index formulas.
4. Fit corrections. Corrections are estimated for each benchmark question and kept as simple as the evidence supports. They include subgroup terms only where the benchmark publishes the matching results, such as sentiment by political party in the United States.
5. Check subgroup transfer. After a national correction, we compare the benchmark’s published differences between groups, by age, sex, education, income, employment or region, with the audience’s. This shows which patterns the audience reproduces.

How corrections are used

• Only on the benchmark’s own questions. A correction applies only when a study asks a benchmark question verbatim, with the same wording and answer options, on a compatible population panel. Every other question is reported exactly as the audience answered it. When a correction cannot apply, the question stays uncorrected and the reason is recorded.
• Never ported. Corrections are specific to a market, its interview language and the period they were fitted on. They are never reused in another market.
• Revisited as benchmarks publish. The size of a gap can change when economic conditions change, so corrections are refitted as new waves are published. Replacing a correction in production is a reviewed decision, recorded with the evidence behind it.

What calibration typically finds

Calibration has shown consistent patterns across markets. We describe them because they shape how results should be read.

• Order transfers better than size. Audiences usually reproduce which groups are more or less positive, but the gaps between groups are often smaller than in the benchmark. Some patterns do not transfer at all, which is why Mexico carries no region-level calibrated claim.
• Levels can lean either way. An audience can be too optimistic in one country and too pessimistic in another. The direction and size of the gap belong to the market, which is why corrections are never shared between markets.
• Current conditions are hard. On questions that depend on recent economic news, such as interest rates or unemployment, unguided answers can point in the wrong direction. Corrections on those questions are larger and less stable over time.

For measured examples from Mexico, the Philippines and the United States, read Same model, opposite mistakes.

What calibration does not do

Calibration evidence covers the benchmark’s questions, population and period. It does not validate questions about a specific product, brand, price or purchase intention: no public benchmark measures those, so their levels are extrapolated rather than calibrated. Studies of those questions remain directional. They are most useful for comparing alternatives and planning research with real customers.

Keeping fit and forecast apart

A correction fitted to a published wave closes the gap on that wave by construction. That demonstrates alignment, not predictive skill. We therefore separate three kinds of evidence and describe each market by the strongest kind it has.

The evaluation protocol

1. Before fitting, specify the target population, the unit being compared, how questions map to the benchmark, the scoring rules and the main error measure.
2. Fit on training waves, and choose settings using separate validation waves.
3. Freeze the configuration and predictions before any test result is read, and bind them together with content fingerprints.
4. Score without changing anything. Report the uncorrected result, the corrected result and simple baselines, including cases where correction made the result worse.
5. Treat each result as evidence only for the questions, population and period it covers.

The baselines matter. A forecast earns little credit for landing near the official figure if simply repeating the last published value would have landed nearer. We compare against carrying the last published value forward and against the same period a year earlier.

Sealed forecasts

A sealed forecast names its target, the baselines and its pass rule before the benchmark’s fieldwork begins. The forecast file is fingerprinted and timestamped by an independent timestamp authority, and a registration that contains the fingerprint, but not the forecast values, is published on the track record. The values are revealed with the score. Nothing is refitted, rerun or swapped after the outcome is known, and a miss stays on the record.

Mexico’s first sealed forecast shows how this works. It covers the October 2026 Consumer Confidence Index, together with its five components and fifteen question balances. It passes only if its headline error is within 2 index points and smaller than the error of carrying forward the last published value. The 2-point tolerance was set in advance from the index’s history of month-to-month movements; it is an error tolerance, not a confidence interval.

Why forward tests matter for AI audiences

AI models learn from large amounts of published text, which can include the very results an audience is tested against. A fit to published data cannot rule out that the model has seen the answer; a forecast sealed before the answer exists can. For the same reason, benchmark figures are kept out of interviews and out of any context the audience sees.

What a report shows

Findings are only useful alongside the evidence that produced them. Before a study runs, the studio shows the market’s status, Calibrated or Exploratory · uncalibrated, and its coverage limits where they apply. A completed report keeps the findings connected to:

• Audience. Who was in the panel, which filters were applied and, for populations built from survey records, how many distinct source respondents stand behind it.
• Questionnaire. Every question and answer option the audience saw.
• Responses. Distributions, counts and simulated verbatims, each linked to the profile that gave it.
• Method and limits. What the report can and cannot establish, and how its figures were calculated.
• Data sources. The acknowledgements and restrictions that apply to the underlying data.
• Reproducibility. The run identifier and random seed.

Constructed audiences are labeled as such, and their reports make no calibration claim.

Reading a report responsibly

Keep the brief, audience, questionnaire, response basis and limitations together when you share findings. Preserve the source notices, and label quotations as simulated. Read differences between groups for direction rather than size, and check the count behind every cut before quoting a percentage.

Scope and limitations

Every research method has limits. These matter most when you read a Prefield study.

• Simulated, not observed. Responses are generated for simulated respondents. They are not observations of real people, and stated intentions are not behavior. Do not read a response share as demand, sales or conversion.
• Calibration has a scope. Current benchmarks cover consumer sentiment and expectations. Questions about products, brands, prices and purchase intentions have no external benchmark; their levels are extrapolated.
• Group differences are compressed. Differences between subgroups tend to be smaller than in reality, and some do not transfer at all.
• Coverage differs by market. Sources describe specific populations. Mexico’s benchmark covers principal cities; the Philippine benchmark surveys households, while profiles describe individuals; US profiles cover adults living in households rather than institutions; and Saudi profiles are constructed from tables rather than observed records.
• Matched characteristics describe groups. Values, party identification and some household economics are matched from other surveys. They are reliable for groups but not certain for any individual profile.
• Source data have their own errors. Official surveys carry sampling error, non-response and measurement error. Weighting corrects totals, not reporting errors, and a population is only as current as its sources.
• Knowledge and timing. AI models know more about some places and periods than others, and they can miss turning points. Dated interviews and refreshed corrections reduce this risk; they do not remove it.
• Wording matters. As in human surveys, small changes to wording, answer options or context can move results. Compare alternatives under identical conditions.
• Not a substitute for customers. Use results to refine ideas and questions, compare alternatives and plan fieldwork. Test consequential decisions with real customers and independent evidence.

Privacy

• No simulated respondent corresponds to a real person. Profiles are built from de-identified statistical files and published tables.
• We do not attempt to re-identify anyone, and we do not link source records with other data to identify people.
• Where a license restricts it, source records stay on our servers and are never included in reports, exports or shared links.
• Results for very small groups are shown as counts rather than percentages.

How Prefield handles account and study data is described in the privacy notice.

Reproducibility and provenance

• Scripted builds. Each market population is rebuilt by script from recorded source files. A build records source versions, retrieval dates and file fingerprints, and writes a quality report covering margin fit, effective sample size, matching depth and coverage.
• Recorded runs. Each study records the configuration it ran on, including the population, interview language and questionnaire, together with its run identifier and random seed.
• Fixed evaluations. Predictions, instruments, configurations and baselines are bound together by content fingerprints before scoring, and a scored evaluation cannot be overwritten.
• Reviewed changes. Changes to a market’s population, corrections or evidence stage are recorded with the evidence that supports them. Earlier versions are preserved.

Markets at a glance

Status, evidence and sources change as markets develop. This table records the position on the date shown; the studio shows each market’s current status.

Glossary

Benchmark survey

A recurring survey with a documented method and published results, used to calibrate and test a market’s audience.

Calibration

Comparing the audience’s answers with a benchmark’s published results and fitting corrections for the benchmark’s questions.

Constructed audience

Archetypes built along the attributes of a role that population statistics cannot isolate. Not a representative sample and never calibrated.

Correction

An adjustment fitted for a specific benchmark question in a specific market. It applies only when that question is asked verbatim on a compatible panel.

Effective sample size

The number of equally weighted records that would carry the same information as a weighted set. It falls when a few records carry large weights.

Out-of-time holdout

A later benchmark wave, unused in fitting, on which a frozen configuration is scored.

Population panel

Respondents drawn at random from a market’s full simulated population, or from the part that matches the audience filters.

Raking

Also called iterative proportional fitting: adjusting survey weights until weighted totals match official totals, while staying as close as possible to the original weights.

Register

The language and level of formality used in interviews, such as formal Filipino or Mexican Spanish in the usted form.

Retrodictive fit

Calibration against results that were already published when the fit was made.

Sealed forecast

A forecast frozen, fingerprinted and independently timestamped before the benchmark’s fieldwork begins, then scored against a pass rule set in advance.

Statistical matching

Adding characteristics from a donor survey by pairing each profile with a donor respondent who shares key characteristics.

Universe

The population a source describes, such as adults in private households or residents of principal cities.

Data sources and notices

These are the sources behind each market today. Each market’s full source notice is reproduced below it, and the applicable notice also appears in every report.

United States

U.S. Census Bureau

American Community Survey public-use microdata, 2024 one-year person and household files; 2024 one-year tables; Population Estimates, Vintage 2025, for independent checks

U.S. Bureau of Labor Statistics

Consumer Expenditure Survey, 2024

NORC at the University of Chicago

General Social Survey, 2024

World Values Survey Association

World Values Survey, Wave 7, United States, 2017

University of Michigan

Surveys of Consumers, used as an internal calibration reference; published values are not reproduced

Federal Reserve Bank of New York

Survey of Consumer Expectations

Philippines

Philippine Statistics Authority

Labor Force Survey, October 2025; Family Income and Expenditure Survey, 2023; 2024 Census of Population

Bangko Sentral ng Pilipinas

Consumer Expectations Survey

World Values Survey Association

World Values Survey, Wave 7, Philippines, 2019

Mexico

INEGI

National Household Income and Expenditure Survey (ENIGH), 2024; National Consumer Confidence Survey (ENCO)

CONAPO

Population projections for Mexico and its states, 2020–2070

Banco de México

Economic Information System, consumer confidence series (CR354), for historical checks

AMAI

Socioeconomic level rule, 2022

World Values Survey Association

World Values Survey, Wave 7, Mexico, 2018

Saudi Arabia

General Authority for Statistics (GASTAT)

Census 2022 complete-data exports; Population Estimates, 2025; Labour Force Survey

KAPSARC and the Royal Commission for Riyadh City

Independent control tables

References

Selected methodological literature behind the practices described in this paper.

1. Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J. R., Rytting, C., & Wingate, D. (2023). Out of one, many: Using language models to simulate human samples. Political Analysis, 31(3), 337–351.
2. Baumgartner, H., & Steenkamp, J.-B. E. M. (2001). Response styles in marketing research: A cross-national investigation. Journal of Marketing Research, 38(2), 143–156.
3. Bisbee, J., Clinton, J. D., Dorff, C., Kenkel, B., & Larson, J. M. (2024). Synthetic replacements for human survey data? The perils of large language models. Political Analysis, 32(4), 401–416.
4. Deming, W. E., & Stephan, F. F. (1940). On a least squares adjustment of a sampled frequency table when the expected marginal totals are known. The Annals of Mathematical Statistics, 11(4), 427–444.
5. D’Orazio, M., Di Zio, M., & Scanu, M. (2006). Statistical Matching: Theory and Practice. Wiley.
6. Groves, R. M. (2006). Nonresponse rates and nonresponse bias in household surveys. Public Opinion Quarterly, 70(5), 646–675.
7. Harkness, J. A., van de Vijver, F. J. R., & Mohler, P. Ph. (Eds.). (2003). Cross-Cultural Survey Methods. Wiley.
8. Hyndman, R. J., & Athanasopoulos, G. (2021). Forecasting: Principles and Practice (3rd ed.). OTexts.
9. Ireland, C. T., & Kullback, S. (1968). Contingency tables with given marginals. Biometrika, 55(1), 179–188.
10. Kish, L. (1965). Survey Sampling. Wiley.
11. Nosek, B. A., Ebersole, C. R., DeHaven, A. C., & Mellor, D. T. (2018). The preregistration revolution. Proceedings of the National Academy of Sciences, 115(11), 2600–2606.
12. Santurkar, S., Durmus, E., Ladhak, F., Lee, C., Liang, P., & Hashimoto, T. (2023). Whose opinions do language models reflect? Proceedings of the 40th International Conference on Machine Learning, PMLR 202.
13. Tashman, L. J. (2000). Out-of-sample tests of forecasting accuracy: An analysis and review. International Journal of Forecasting, 16(4), 437–450.
14. Tourangeau, R., Rips, L. J., & Rasinski, K. (2000). The Psychology of Survey Response. Cambridge University Press.
15. Wang, A., Morgenstern, J., & Dickerson, J. P. (2025). Large language models that replace human participants can harmfully misportray and flatten identity groups. Nature Machine Intelligence, 7, 400–411.

Version 1.0, 30 September 2026: first public edition. We revise this paper when the method changes. Market status is dated where it appears.