See how your audience reacts before you ship.

mimi runs your ad, name, price or post past audiences compiled from real behaviour: what people actually posted, followed and commented. You get the split, the reasons and the range in minutes, not weeks.

Run your first study See the method

Built from real, observed behaviour
Every audience ≥ 100 real accounts
Ranges shown, never a single number
Answers grounded in members' own words

mimi · study run

running · 4s

creative

“Same price as regular. Carbon-neutral from farm to shelf.”

question · Would you pick this over your usual brand?

Men in the US into sports highlights and debates106K members
Women 18–34 in the US into mom-life content and routines46K members
Men in the US into NBA and basketball debates42K members
Men in the US into UFC and MMA fights40K members

result · all audiences

Yes—

Not sure—

No—

real audiences from the live catalogue · example answers

compiled from TikTok Instagram YouTube X Reddit

The difference

Most synthetic audiences are invented. Ours were observed.

The usual way: A model is told to be “a 35-year-old mum in Ohio” and answers like the stereotype of one. The audience is a prompt, so nothing it says can be checked.
Our way: Every audience is compiled from accounts that exist, grouped by what they actually engage with, and carries the comments its members wrote. The model answers from that evidence.
Why it matters: An audience built from recorded behaviour can be scored against outcomes that already happened. One built from a prompt cannot be scored at all.

Generateduniform degree

Nodes placed on a grid and wired at random. Everyone has the same number of connections, so there are no communities for an opinion to travel through, and nothing to check a prediction against.

Observedrecorded follows

Follows we recorded. Dense clusters, a heavy tail of high-reach accounts, and a handful of bridges between communities: the structure that decides whether a message spreads or stalls.

Method

Four stages, each with rules it cannot skip.

Full method

stage 1 / 4
Observe
Public posts, follows and comments are recorded across platforms and resolved to the people behind them.
- public activity only
- English-language comments only
stage 2 / 4
Compile
People are grouped into audiences by what they engage with, then described and labelled from that evidence.
- ≥ 100 accounts per audience
- ≥ 500 per US panel cell
- labels must pass a grounding check
- under-18s excluded by default
stage 3 / 4
Simulate
Each audience reads your creative and answers your questions from its own recorded comments, many times over.
- every version simulated on its own
- answers grounded in verbatim quotes
stage 4 / 4
Report
You get each answer's share with its range, how every audience split, and the reasons in their own words.
- ranges shown, never a single number
- overlapping ranges reported as a tie

Specimen

One audience, exactly as the catalogue holds it.

Not a mock-up. This is a live record from the catalogue, read when this page loaded. Every audience you can ask carries the same fields, and the ones that fall short of a rule are flagged on the audience itself rather than hidden.

Browse the catalogue

aud-v3-tiktok-7765

Men 18–34 in the US into car modification and repair

members: 59,594
platform: TikTok
gender: men
age: 18–34
location: US
verbatim quotes: 320
focus: one subject
interest: car modification and repair

topics its members engage with

Mustang modifications
truck towing and diesel engines
car repair tips
car aesthetics and styling

Output

A distribution you can defend, not a verdict you have to trust.

The split, with its range

Every answer's share and the range it could sit in. Two answers whose ranges overlap are called a tie, not a win.

Who disagreed, and by how much

Each audience's own split, and the gap between the two furthest apart, so a divided audience is never averaged into a calm one.

Why, in their words

The reasons each audience gave, drawn from its members' own comments. Then ask them a follow-up, and they answer from the same evidence.

figures above are illustrative

What we will not claim

A simulation is worth exactly what its validation is worth.

Accuracy figures are easy to produce and hard to earn. Ours are reported against simple baselines on held-out outcomes, with the misses shown as plainly as the hits.

Read the research

Every audience is an aggregate of real accounts. It is never one person, and never a stand-in for one.
Results are model estimates of how an audience is likely to respond, not observed responses. Every report says so.
Where an audience is thin, mixed or unverified, the audience itself carries the flag.
A study informs what you test with real people. It does not replace that test.

Test the reaction before you commit to the decision.

Bring a message, a name, a price or a campaign. Get the split, the reasons and the range, in the time it takes to brief an agency.

Run a study Talk to us