Usage and Attitudes: What a Synthetic Panel Recalls

Methodology · 11 min read

TL;DR: A usage and attitude study asks two kinds of question in one questionnaire. One asks what you did: which brands you bought, how often, how much you spent, where. The other asks what you think: what matters, what annoys you, what you would want instead. A census-grounded panel can carry the second and not the first. A persona has no purchase record, so a behaviour question returns a plausible population estimate written in the first person. Work published in October 2026 found that behavioural history predicts a simulated response better than any description of the person, and a census-grounded panel is a description by construction. Split the questionnaire, source the behaviour block from measured data, and run the attitude block early and often.

Can a synthetic panel run a usage and attitude study?

Half of one. A usage and attitude study measures behaviour and attitude inside a single questionnaire, and only the attitude half survives. A persona has no purchase history, so every behaviour question returns a population estimate phrased as a personal report. Run the attitude block on the panel and source the behaviour block from measured data.

A usage and attitude study, called a U&A, is what consumer teams field when they want to understand a category rather than a concept. It asks what you buy, how often, in what sizes, from which shops, alongside what you value and what you would change.

Those two halves look alike on the page. They are different measurements, and a synthetic panel handles them differently.

The attitude half asks for a judgement, and a judgement is produced at the moment of asking. The behaviour half asks for a report, and a report needs something to report on.

What follows is which outputs fall on which side, why grounding cannot move the line, and where the behaviour numbers should come from.

What does a persona return when you ask what it bought?

A prior, not a report. The model holds no record of this persona's purchases, so it produces the most probable answer for someone carrying those attributes. Ask a thousand personas and you get a distribution shaped by the training data and the attribute list, carrying no information about your category.

Ask a persona how many times it bought ground coffee last month and you get a number. Three. It is plausible, it is stable across reruns, and nothing produced it except the question.

The mechanism is now measured directly. A paper published in October 2026 compared two ways of conditioning a synthetic persona: a description of the person, and that person's own behavioural history. The behavioural history won.¹ Descriptions of who someone is predict their answers less well than records of what they have done.

A census-grounded panel is a description. That is its design, and for attitude work it is enough. For a behaviour question it is the weaker input, and it is the only one available. Cross-domain benchmarking of simulated survey responses points the same way, with answers holding up unevenly across question types rather than failing everywhere at once.² The Nuremberg Institute for Market Decisions frames the field as promise, progress and limits rather than replacement.³

The practical tell is variance. Human frequency data is lumpy, with zeros, heavy users and long tails. Synthetic frequency data is smooth and central, because an average is what the model was asked to produce.

Is a human usage and attitude study any better at this?

Better, and still not good. Survey reports of behaviour have been checked against administrative records for decades and come back wrong in documented directions. People under-report spending, misplace events in time and forget routine purchases. The difference is that a person misremembers something that happened to them.

This is the part most criticism of synthetic research skips, which makes that criticism easy to dismiss.

When survey answers are linked to administrative records, misreporting turns out to be large and systematic rather than random noise. Meyer and Mittag checked reports of programme participation across three major surveys against the records themselves and found error patterns that move estimates, not just widen them.⁴

Expenditure is the clearest case. The U.S. Bureau of Labor Statistics publishes a standing concordance between the Consumer Expenditure Survey and Personal Consumption Expenditures so the differences between the two can be examined rather than assumed away.⁵ Recall length is a design variable for the same reason: how far back you ask changes the answer.⁶

So the comparison is not accurate against inaccurate. It is a biased measurement of a real event against a well-formed guess about a population. That decides what you can do next. Recall bias has corrections: shorter reference periods, aided recall, validation against records. A synthetic behaviour answer has no record behind it to correct toward.

Which usage and attitude outputs survive on a synthetic panel?

Sort them by whether the question asks for a record or a judgement. Penetration, frequency, volume, spend and brand repertoire all ask for a record, and none survives. Importance, annoyance, need-state language and reactions to a described change ask for a judgement, and those hold as rank orders.

Most U&A questionnaires can be split in an afternoon. Read each question and ask whether answering it correctly would need access to something that happened.

The table applies that split to the outputs a U&A deck contains. The right-hand column is the part teams skip: a question that fails on the panel does not disappear, it moves to a different source.

Brand repertoire fails in a direction rather than at random, because a persona already knows the famous brands and will over-report them. Occasion language sits in the middle: the words a persona uses for a usage moment are worth harvesting as hypotheses, the share of occasions it reports is not.

Why does census grounding not fix a behaviour question?

Because a census counts people, not purchases. National statistics publish age, income, household size, education and region, which is what PersonaHive uses to build each country panel. No statistical office publishes how often a 34-year-old in Lyon buys oat milk. The attribute that would answer your question was never collected.

Census grounding does real work. It makes the sample look like the country rather than like the internet, and it is why a panel can be specified by market at all. PersonaHive builds each panel from that country's own national census across nine countries, from aggregated public statistics, and validates against real surveys.

What it cannot do is add a column no statistical office collects. Public statistics describe who people are. Commercial measurement describes what they buy, and the two live in different files for a reason.

This is why PersonaHive does not describe its personas as grounded in consumer behaviour. A vendor who does is claiming a data source, so ask which one, at what coverage, and how a national panel was rebuilt from it.

The second-order version matters more. Take a penetration figure from a panel, use it as the base of a market model, and the model now rests on an assumption rather than a measurement, which is the failure described in what breaks when you size a market on a synthetic panel.

How do you split a usage and attitude questionnaire?

Six steps. Classify every question as record or judgement, source the record block from measured data, run the judgement block on the panel, read ranks rather than levels, keep a human anchor on any number quoted as a level, and record which source produced which figure.

1. Classify every question. If answering it correctly needs access to something that happened, it is a record question. Everything else is a judgement question. Borderline items go in the record pile.

2. Source the record block first. Published national statistics, then measured commercial data, then your own transaction files, which for your own buyers beat both.

3. Run the judgement block on the panel. On PersonaHive a survey question costs one credit per persona, so a twenty-item attitude battery is cheap enough to field at every stage of a project.

4. Read ranks, not levels. Which attribute leads, which statement divides the sample, which option wins against which.

5. Anchor anything that leaves the team as a level. A rank survives a methods review on its own. A percentage needs human respondents behind it, for the reason set out in why a synthetic intent score is not a forecast.

6. Write the sources into the methods note, line by line. A deck that mixes measured penetration with synthetic attitudes is defensible. One that hides which is which is not.

The smallest useful next step is step 1, on a study you have already run. Mark every question in your last U&A questionnaire as record or judgement, then count how many charts in that deck came from the record pile. That count tells you what a synthetic panel would and would not have given you, before you spend anything.

When does a usage and attitude study still need people?

Four cases. Any figure that will size a market or set a budget, any question about a specific past episode, any category where small or regional brands decide the outcome, and any attitude number that will be quoted as a level rather than a position in a ranking.

The first case is the common one. U&A findings end up as the base of a planning model, and a base has to be a measurement. If the number will be multiplied by anything, it does not come from a panel.

The second is episodic. Questions about one remembered occasion, service interaction or shopping trip have no synthetic equivalent, which is the gap that breaks experience metrics like NPS on a synthetic panel.

The third is brand coverage. Categories run by regional, private-label or recently launched brands are where a model's brand knowledge is thinnest. If the decision turns on a challenger brand, the panel will understate it.

The fourth is the number that leaves the room. Once a figure reaches a board pack it stops carrying its method with it. Work on imputed synthetic samples makes the point generally: a well-formed number is not a measured one, and the distinction disappears fastest in summary.⁷

Frequently asked questions about usage and attitude studies on synthetic panels

Short answers to five questions that come up when a U&A questionnaire meets a synthetic panel: whether to ask about the market instead of the persona, what plausible frequency data proves, what behavioural grounding claims are worth, whether a synthetic pass can design the human study, and what a written rationale adds.

**Can I ask the persona to estimate the category instead of reporting its own behaviour?** That turns a report into a judgement, which helps the wording and not the number. You are now asking a model what it believes about a market, which restates its training data. If published statistics exist, use those.

**My synthetic frequency numbers look plausible. Is that not evidence?** Plausibility is the one thing a language model guarantees. Check variance and shape, not level. Synthetic frequency runs smooth and central where human frequency is lumpy.

**A vendor says its personas are built from behavioural data. Does that solve it?** It might, and it is a claim about a data source, so treat it as one. Ask which dataset, what it covers, and how it maps to a national population. Behavioural history is a stronger conditioning signal than a description.¹ One company's web analytics is not a national behavioural record.

**Can I use a synthetic pass to design the human U&A?** Yes, and this is the strongest use of the two together. Field the full questionnaire on the panel to find dead items, ambiguous wording and batteries that produce no separation, then take the cleaned instrument to people.

**Does forced rationale help with behaviour questions?** It helps you see the problem rather than fix it. A written justification beside each answer exposes when a persona is reasoning from a general population pattern rather than anything specific to itself.

Sources

• Behavioral History Outperforms Descriptions of the Person for LLM Synthetic Personas — Khashayar Pourtaheri and Ahmad Zareei, arXiv
• When Synthetic Users Fail: A Cross-Domain Benchmark of LLM-Simulated Human Survey Responses — arXiv
• Leaving Insight to Digital Twins? Promise, Progress and Limits of Synthetic Respondents — NIM, Nuremberg Institute for Market Decisions
• What Leads to Measurement Errors? Evidence from Reports of Program Participation in Three Surveys — Bruce D. Meyer and Nikolas Mittag, National Bureau of Economic Research
• Comparing expenditures from the Consumer Expenditure Survey with the Personal Consumption Expenditures: Results of the CE/PCE Concordance — U.S. Bureau of Labor Statistics
• Optimal recall length in survey design — Journal of Health Economics
• The Risks of Using LLM Imputation of Survey Data to Produce Synthetic Samples — Verasight