Questions and answers
How Studio builds people, where their answers come from, how far to trust them, and how plans work.
What Studio is
What is Voxari Studio?
A place to build a synthetic population (your employees, a county, a customer base, a fictional city) and ask it questions. Studio creates every person one by one, with the traits and proportions you describe, then asks each of them your questions and shows how the answers split, which groups differ, and why.
It is for exploring and pre-testing: finding which groups to worry about, which wording backfires, and which questions are worth a real survey. It does not replace one.
Who is it for?
People who need a feel for how a group might react before acting: HR and internal communications testing a policy change, researchers and students exploring a question, teams drafting messages or surveys, writers and game designers building a world.
How it works
How does Studio build a population?
You start from a template (a workplace, city or school staff, patients or physicians, business buyers, an association's members, a homeowners' association, a congregation, a university's alumni, a nonprofit's donors, an online game's players, a streaming service's subscribers, an online course's learners, gig workers, volunteers, factory or restaurant staff, nurses, family caregivers, a health center's patients, NHS staff or a GP practice's patients, care workers in Europe, high school students, early educators, adult learners, a fictional city, a world from the books and films such as Middle-earth or Tatooine, or a real place) and adjust it: the traits people have, the share of each value, how traits depend on each other (tenure on age, role on department), and the subgroups people belong to. Studio then creates every person so the population matches those proportions exactly, within rounding. The same seed always gives the same people.
Where you do not link two traits, Studio treats them as unrelated. If you never say that income and union membership go together, they will not.
Templates of real places, all fifty US states by county and the District of Columbia (on Team: Alabama, Alaska, Arizona, Arkansas, California, Colorado, Connecticut, Delaware, Florida, Georgia, Hawaii, Idaho, Illinois, Indiana, Iowa, Kansas, Kentucky, Louisiana, Maine, Maryland, Massachusetts, Michigan, Minnesota, Mississippi, Missouri, Montana, Nebraska, Nevada, New Hampshire, New Jersey, New Mexico, New York, North Carolina, North Dakota, Ohio, Oklahoma, Oregon, Pennsylvania, Rhode Island, South Carolina, South Dakota, Tennessee, Texas, Utah, Vermont, Virginia, Washington, West Virginia, Wisconsin and Wyoming), twenty-six Brazilian states by município and the Distrito Federal (on Team: São Paulo, Minas Gerais, Rio de Janeiro, Bahia, Paraná, Rio Grande do Sul, Pernambuco, Ceará, Pará, Santa Catarina, Goiás, Maranhão, Amazonas, Espírito Santo, Paraíba, Rio Grande do Norte, Alagoas, Piauí, Mato Grosso, Mato Grosso do Sul, Sergipe, Rondônia, Tocantins, Acre, Amapá and Roraima) and the United States and Brazil by state, the United Kingdom by region and the European Union by country and each of its 27 member states by region (on Pro; these from Eurostat's tables, since no person records are public), work differently: each person is a real combination of traits taken from the Census's anonymised person records, weighted so that every county or state matches its published totals. Each comes with an error report saying how close every place lands.
Where do party lean and 2024 turnout come from, in Georgia and the US?
They are modelled, not recorded, and labelled "(modelled)" wherever they appear. No Census record says how anyone votes. Each person gets a lean (Democratic, Republican, or neither) and whether they voted in 2024, drawn from what Cooperative Election Study 2024 respondents like them (race, age, education and sex, in the same Census region) said and did, with turnout checked against the voter file. Each Georgia county, or each state in the US template, is then calibrated so its people's turnout and Harris-Trump margin match its 2024 result: certified by Georgia's Secretary of State for counties, the Federal Election Commission's official results for states. Non-citizens cannot have voted.
So the totals for each county or state are right by construction, but any one person's lean is a likely guess, not their stated party. Demographics alone place them poorly (each error report shows by how much), so within a county or state the lean follows its result more than the person. Use it to include or compare groups, never to present results as how Democrats or Republicans think: these are not polls of real people.
Where do the answers come from?
From Jev, the AI model behind Voxari. For every person, Jev reads a short description of them (their place, each trait in plain words, what you wrote about the population, and the study's briefing) and judges how that person would most likely answer, as a probability for each option.
Nothing in Studio is a rulebook about people, and the people do not talk to each other. Everything that makes a nurse answer differently from a driver comes from what the model learned about how people in such situations tend to think.
Why does the briefing matter so much?
The model only knows what it is told. Ask about changing the health plan without saying what the current plan is, and most people come out undecided, because they would want details first. The briefing is where you give everyone the facts they would have in real life. The question coach points out when one is missing.
What are scenarios?
A scenario runs time forward for a population. You write an event (what happened, who it reaches) and everyone it reaches is asked how it leaves them and what it does to their trust in those responsible: two responses a person. Letting a month pass is free: word spreads within groups (a department, a county) and reactions drift toward the group's.
Run a study on a scenario and each person carries what they lived through into their answers. Fork a scenario to run two futures from the same point, then compare the runs side by side. The memory is a short summary of the person's own two answers, so it adds judgement, not invented detail.
How far to trust it
How accurate is it?
Building the people is exact: the population has the proportions you set. The answers are a model's judgement, and their accuracy is uneven.
What it does best is direction and contrast: which groups lean which way, and which wording lands worse. Research on this method has found that language models can reproduce many differences between groups on public-opinion questions.
What it does worst is exact levels. Simulated answers tend to come out too tidy: less variety within a group than real people show, closer to stereotypes, and sensitive to wording. Treat a figure like "37% oppose" as a sketch, and "drivers oppose far more than office staff" as the finding.
We measure it. Asked 30 questions with published results (Pew, Gallup, AP-NORC, KFF, the General Social Survey, the Cooperative Election Study and Quaest) of the United States, Georgia and Brazil templates, Studio's shares were on average about 21 points away from the real ones overall, and about 23 within groups by age, education, race or party (the share of people who would have to change answer for the two to match; a second real survey would be 1 to 2 points away). A few landed within 5 points (abortion, school vaccine rules, marijuana); questions about how strongly people feel, and the Brazilian questions, landed furthest. Answers lean too much to the middle of a scale: few simulated people pick "extremely proud" or "strongly favor".
How can I make results more trustworthy?
Write a briefing with the facts people would know, above all what any law, programme or local term in the question is, and the year if timing matters: the model knows little about recent or local things (on a Brazilian poll about ending the 6x1 work schedule, a two-sentence briefing moved the result from 35 points off to 9). Use the question coach before running. Add rewordings of a question: if the answer changes with the wording, it is fragile. Look at Jev's confidence and at how the undecided lean. Prefer contrasts between groups over single percentages.
Above all, compare with reality: if you have real results for one question (an engagement survey, a published poll), add them as a benchmark. When Studio lands close where you can check it, you can trust it more where you cannot.
What should I not use it for?
Not as a survey: results must never be presented as the views of real people or attributed to a real pollster. Not for decisions about any individual (hiring, firing, credit, health care). Not as a forecast of an election. Every page and report says the answers are synthetic, and that notice stays.
Reading results
What do the ± figures mean, and how wrong can a result be?
There are two kinds of error, and Studio shows them apart. The ± beside each share, net lean and average level is a 95% sampling interval: how much the figure could move if the same population were drawn again. It uses the people's effective number after weights, so a small or unevenly weighted group gets a wide interval. Rows under 30 people, or with an interval wider than ±15 points, are greyed out and never become a finding.
The second kind is the model's own error, and it is much larger. Measured on 44 published questions with known results, four answers in five landed within about ±22 points of the real survey's share (±25 within groups); on scales, within ±23, and the average level within ±15 points of 100, because the model avoids the ends of a scale. This error is not chance: more people do not shrink it. Read the sampling interval as "is this difference more than noise in the simulation?" and the model error as "how far might real people be?".
What is a conjoint study, and how far can I trust it?
A conjoint question shows each person pairs of options described by a few attributes (pay, days in the office, shift) and asks which they would choose, several times, with the levels drawn at random for each person. From the choices Studio works out how much each level moves the chance of being chosen (the AMCE, as published conjoints report it), each attribute's importance, and part-worths that a simulator turns into shares of preference for options you build. It comes with Team, as a beta: each choice is one response, so 1,000 people making five choices use 5,000. In the report, pick a breakdown (role, region, a segment) to see each group's own estimates side by side, and read every chart for any one group.
We checked it against two published conjoints of US adults with their replication data: which immigrants to admit (Hainmueller, Hopkins and Yamamoto) and which income tax plan to prefer (Ballard-Rosa, Martin and Scheve). Studio's effects pointed the same way as the real ones for 27 of 31 clear effects and ranked the attributes much as people did, but came out about half as large (5 points off on average) and missed some effects entirely: prior trips to the United States, and support for higher taxes on the top bracket. Asking the invented people to weigh what they think is fair for the country, not only their own stake, did not change this. Read which attributes matter and which way; treat the sizes as a sketch.
What is a MaxDiff study, and how far can I trust it?
A MaxDiff question shows each person small sets of items from a list (4 to 20 items, 3 to 5 at a time) and asks which matters most and which least, several times, with the sets drawn for each person. Studio counts how often each item was picked as the most and as the least when it was shown, and fits a best-worst model whose scores, out of 100, give each item's share of preference. It comes with Team, as a beta: each pick is one response, so 1,000 people answering six sets use 12,000. In the report, pick a breakdown (role, region, a segment) to see each group's own estimates side by side, and read every chart for any one group.
We checked it against two published MaxDiff studies of US adults. On 15 national issues (the 2035 Initiative, June 2026) Studio's order came out close to the real one: a rank correlation of 0.86, the same top three (cost of living, the economy, health care) and 0.76 to 0.92 within each party, though more extreme at both ends. On 20 health problems ranked for research funding (Cook and others, September 2020) it did not: it put rare diseases (muscular dystrophy, multiple sclerosis, spinal cord injuries) far lower and smoking and being overweight far higher than people did, and landed further off than giving every item the same score, with or without a briefing giving the date (which pushed COVID-19 and related items too high). Read the order among familiar items, treat the scores as a sketch, and check a specialist list against real results before relying on it.
What does "of the undecided, 31% lean support" mean?
Among people whose most likely answer was undecided, how many were nearer to supporting, nearer to opposing, or truly torn. It tells you which way a group might break once it knows more.
What are "what drives the answers" and the robustness check?
What drives the answers ranks the traits that best explain the differences in an answer, so you can see whether role, age or site matters most. The robustness check compares the answers to rewordings of the same question: a result that holds across wordings is sturdier than one that moves.
What is a benchmark?
Real results you add for one question, overall or by group, from a survey you ran or a poll you trust. Studio shows how far its answers land from them, in points, with a verdict: close, rough or far.
What does "Check with Claude" do?
On a finished run, Claude, a larger AI model from Anthropic, answers the same questions for 50 of the same people, described to it exactly as Jev read them. Studio shows Claude's answers under Jev's, how many points apart they are, and the groups where they differ most. Jev's own answers never change.
Adjusted with Claude (beta): with a check, each question in the report opens with Jev's shares moved toward Claude's. Among the 50 people, Studio sees how far Claude's answers are from Jev's in each group of the question's strongest driver (age, role, region), and moves everyone in that group by that much; a group with fewer than 5 of the 50 moves by everyone's average. On 17 held-out questions with published results this landed about 8 points from the real result on the older set (Jev 21) and 18 on the questions asked after Claude's training data (Jev 34), close to Claude alone and at no extra cost, since it reuses the check's answers. Jev's own answers follow; hypotheses, drivers, segments and downloads keep using them, and the Excel methodology sheet lists the adjusted shares.
Why: measured against 30 published surveys, Claude's answers landed much closer to the real results than Jev's (about 8 points off against 21). That is not only because it may have seen those polls: on 7 questions fielded from July 2026 on, after its training data (Pew on supplements, the economy and data centers; Quaest on online betting, weight-loss pens and food prices), Claude was closer on 6, about 16 points off against 34. Where the two models agree, the result depends less on the model; where they are far apart, treat the figure with care and compare with real results if you can. Each check costs about US$0.15 to US$0.35 of Claude and counts against the plan's checks with Claude, apart from coach checks: 2 a month on Free, asked by hand; on Pro (15 a month), Team (60) and Enterprise every finished run is checked by itself while they last, and after that you can still ask. A run is checked once.
How do I test a hypothesis?
On a study's page, under Hypotheses, write what you expect in your own words, even as a question ("do night-shift drivers dislike this more?"). "Check this hypothesis" asks the question coach whether the study can test it, suggests a clearer phrasing and fills in the parts, which you can change: two groups on one question, or one group on two questions; the figure (net lean for support-or-oppose and yes-or-no, the average level for a scale, or the share giving one answer); the groups (a trait's value, an age band, a place, up to three traits combined such as drivers on the night shift, or, for a question asked in several wordings, the wording people saw); and the direction: higher, lower, or different by at least some points. Add them before or after a run; suggested studies come with one or two.
Each is tested as a null hypothesis against yours: H₀ says the two figures are the same, H₁ is what you expect. Every finished run reports a verdict on size (supported when the gap goes the expected way by 5 points or more, contradicted when it goes the other way by 5 or more, not supported in between), both sides' figures, the gap, the sizes and a chart of how people's figures are spread. A contradicted hypothesis is listed under What stands out in the report.
The test: two groups are compared with Welch's t-test, which does not assume they vary equally, with degrees of freedom from the data; against everyone, the group is compared with everyone else. One group on two questions is compared person by person with a paired t-test, since the same people answer both. One group on the same question in two studies of the population (with different briefings, say) is compared with Welch's t-test between this study's run and the other study's latest finished run, the two being separate asks. Each person's figure is an average of probabilities, so the tests compare means; people count by their population weights, with effective sizes. "Higher" and "lower" are one-sided, "differs" two-sided, at the 5% level. With large groups this matches a z-test; the t-test matters for small ones.
Under each hypothesis, "People needed" says how many people its test needs to find its gap (5 points, or the number you set for "differs") at the 5% level with 80% power: in each group, and the population size that gives them from the groups' shares, with whether this population is big enough. The spread of people's figures comes from the latest run; before any run, a cautious figure measured on Studio's own answers. Before you create a population, the builder shows the same for the template's suggested studies.
Pro allows two per study. Team and Enterprise allow five and show the test itself (t, degrees of freedom, p-value, a 95% confidence interval for the gap, the effect size and each side's mean and spread), the p-value adjusted by Holm's method across the run's hypotheses (testing several at once makes a lucky one likelier; the adjustment keeps the chance of any false rejection at 5%) and, after Check with Claude, Claude's figures. Read p-values with care: invented people's answers vary less than real people's, so p-values come out smaller than a real survey's would. How big a gap is matters more.
Data and privacy
Can I use my own data?
Use proportions, not people: "35% of staff are drivers" is fine; a list of employees, patients or customers is not. Studio never needs records of real people, and you must not type or upload them.
Who sees what I write?
The members of your workspace. To produce answers, populations and questions are sent to the AI models that judge them: Jev, and Anthropic's Claude when you use the question coach. Neither is told who you are. The terms have the details.
Plans
What is a response?
One person answering one question. A study of 1,000 people with 3 questions uses 3,000 responses. Each plan includes responses every month, from the 1st (UTC); answers that fail are not counted.
What do the plans include?
Free: one population of up to 1,000 people and 30,000 responses a month. Pro (US$15 a month or US$120 a year): ten populations of up to 10,000, 750,000 responses, rewordings, benchmarks, two hypotheses per study, downloads as CSV or Excel (with a codebook and a methodology sheet) and 15 runs a month checked with Claude by themselves. Team (US$59 a month or US$472 a year): three members, populations of up to 25,000, 3,000,000 responses, five hypotheses per study with significance checks, 60 runs a month checked with Claude, conjoint studies (beta), Parquet downloads and reports with your own logo instead of the Studio mark. Extra responses come in packs of 1,000,000 for US$10 and never expire.
Can I cancel?
At any time, from Plan and usage. The plan stays until the end of the period paid for. Any charge is refunded in full if you ask within 14 days.
Answers are kept as long as the plan says (Free 30 days, Pro a year, Team two). After a move to a plan that keeps them for less time, the previous plan's period still holds for 30 days, so you have time to download them or change back; then the new plan's applies. The headline numbers of every run stay.