Draw a population.
Ask it anything.
We model any demographic slice you can define, then answer your questionnaire the way that slice would answer it. Weeks of fieldwork, returned before you finish reading the question.
Tap any plane to change a variable, then run the wave.
What lands in your inbox.
A topline, a banner, and the frame it was estimated on — the same shape a field house would send you, without the six-week wait.
| Response | Total | 18–34 | 35–54 | 55+ | Owners | Renters |
|---|---|---|---|---|---|---|
| Strongly support | 31 | 41 | 30 | 22 | 24 | 42 |
| Somewhat support | 28 | 29 | 29 | 26 | 27 | 29 |
| Somewhat oppose | 17 | 13 | 18 | 21 | 21 | 12 |
| Strongly oppose | 18 | 11 | 18 | 24 | 23 | 11 |
| Don't know | 6 | 6 | 5 | 7 | 5 | 6 |
| Net support | +24 | +46 | +23 | +3 | +7 | +48 |
Every wave ships as a topline, a banner of your chosen crosstabs, and the stratification frame it was post-stratified on — so another analyst can check the arithmetic. Figures above are illustrative and show the format only.
Three steps, one wave.
Draw the slice
Gender, age band, ethnicity, education, household income, region, urbanicity, party ID, marital status, home ownership, religious attendance, parental status, recalled vote. Combine as many as you like, down to slices a phone bank could never reach in usable numbers.
Field the questionnaire
Paste the wording you would send to a field house. Forced choice, Likert scales, favourability, most-important-problem, open-ended. Question order and framing are preserved, because they move the answer.
Read the crosstabs
A topline, your banner points, and the calibration error measured for that slice. Change one variable and re-field — a wave costs seconds, so you can afford the follow-up you would normally cut.
What a respondent actually costs.
Published figures from the 2020 US election cycle, where a probability panel, a random-digit-dial phone poll and an AI-polling pipeline were run against the same result.
ANES 2020
ABC / Washington Post 2020
from digital traces
Columns one to three, and the AI-polling column, are published figures from Cerina & Duch, Artificially Intelligent Opinion Polling (Royal Society Open Science, 2026; preprint arXiv:2309.06029), Table 4 and §5.1. The same paper puts AI polling between 500 and 2,500 times cheaper per respondent than random-digit dialling, and scores FiveThirtyEight's 2020 state-level forecast at RMSE 0.05 on the same margin — so the cheap pipeline lands within about a point of the aggregator, not ahead of it. The 6% response rate is Pew Research Center's measured rate for its own telephone polls in 2018, down from 36% in 1997. Dondarrion's own calibration figures are not yet published.
We don't read what people post. We watch what they follow.
Posts are the wrong instrument: they are contaminated by automation, and the minority of users who post are not the population. Follows are near-universal, cheap to observe over time, and they move before opinion is spoken aloud.
The estimator is multilevel regression and post-stratification with a bias-correction term for online selection, following Cerina & Duch. The full write-up, with the Toronto 2026 mayoral deployment as the worked example, is on the method page.
- Not a replacement for your benchmark wave.Field one real poll a cycle. We calibrate to it.
- Not a read on breaking events.A shock needs live respondents until the follow graph catches up.
- Not a person.It reports how a slice tends to answer. It does not know what anyone thinks.
Poll the slice you could never afford to reach.
Private beta. Campaigns, brands and newsrooms first.