About
We built the data first and the simulation second. That is the opposite of how this usually goes.
Mimi started as an infrastructure problem, not a product idea. Years of public behavioral data had been collected and organized before anyone asked what could be simulated with it. Everything on this page follows from that ordering.
The claim we are making
A synthetic audience is easy to build and almost impossible to check. Generate a plausible network, populate it with personas inferred from a demographic profile, and you have something that produces confident answers immediately. Whether those answers correspond to anything is a separate question that the architecture makes very hard to ask.
Our claim is narrower and, we think, more useful: every audience in this platform is an aggregate of real people, grouped by what they actually engaged with rather than by a demographic guess, and because that behavior is recorded over time, a prediction can be scored against an outcome that already happened. That last property is the one that cannot be retrofitted.
What we are not claiming
We are not claiming to model everyone. Audiences are built from people who engage in public, so people who only watch are missing, and audiences inherit that skew. Our population panel is weighted to the census to correct the mix; it does not invent the people who are absent. We are not claiming simulation replaces research; it narrows what you take into research. And we are not claiming one headline accuracy figure: each kind of question is tested on its own, against real outcomes, and the results are published, misses included.
What the product is built against
Aggregates, never individuals
At least 100 people in every audience, enforced when it is built rather than filtered at display. There is no audience of one, and no product surface that simulates a named person.
Nothing sensitive is inferred
Age and gender come from what people say about themselves. Ethnicity, religion, health, sexual orientation and politics are never inferred, and labels describe content, never beliefs.
The range is part of the answer
A share without a range is not a result. Where two options are close or the evidence is thin, the platform says so instead of returning a confident number.
A winner only when it is earned
A winner is named only where that kind of test passed a check on data it had never seen. Everything else is shown as a lean or called too close, never as a winner.
Reproducible by construction
Every study is pinned to a sealed data release, so a result can be re-read later against exactly the evidence it ran on.
Publish the nulls
Results that miss are reported in the same place, at the same prominence, as the ones that succeed. Pass rules are written down before the test is scored.
What the data is, and is not
Audiences are built only from publicly visible activity: what people chose to post, comment on and engage with in public. Identifiers are replaced with salted keys before any audience is built.
Nothing in the platform ingests private messages, purchased contact lists, or data behind a login the account holder did not make public. No individual record reaches an audience brief, a simulation, or an export: every figure a customer sees is a mix across at least 100 people.