Platform
Four layers between raw behavior and a defensible answer.
Nothing in Mimi is generated from a demographic assumption. Each layer below only ever groups, describes or checks what real people already did. That restraint is what makes the output auditable.
Real people, grouped by what they do
Public behavior, grouped into audiences of real people. The counts below are read live from the catalog.
Audiences you can test
Every one of them listed in the app, ready to test.
People observed
Grouped into audiences by what they engage with, not by who a demographic model says they are.
People in every audience
Enforced when an audience is built, not filtered at display. There is no audience of one.
Kinds of audience
Interests, brands and media, individual creators' audiences, and a US population panel. Each is described below.
Audiences, labeled only with what is true
Audiences come from co-engagement: people who engage with the same creators share an audience. Each one carries a brief that describes it, and a label that says nothing its members would not recognize.
Interest audiences
5,731
People who share an interest, found by what they engage with. Gender or age appear in a label only when most members share them.
Brand & media audiences
1,391
The audience of a brand, organization or media account, as a group you can test on its own.
Creators' audiences
11,739
The audience of an individual creator. Found only by searching their exact @handle, never browsed.
US population panel
171
Age, gender and life-stage cells weighted to the US census, for questions about adults in general.
Behavior, not tags
Grouping by hashtag clusters tagging habits. Grouping by who engages with whom clusters real audiences. Near-duplicate audiences are merged, so two labels never mean the same people.
Traits people state themselves
Age and gender are learned from what people say about themselves, calibrated, and shown only as an audience's mix. Ethnicity, religion, health, sexual orientation and politics are never inferred.
Every label fact-checked
A second model checks each label against members' own words it did not see when the label was written. Only labels that pass are offered. Sensitive labels are reviewed and describe content, never beliefs.
Checked against what members say
Among members who state their gender, 97% of those in a “Women …” audience are women (74% across all audiences), and 79% of those in a “Men …” audience are men. Stated age ranges hold for 76–82%.
What an audience brief contains
# Women 25–34 in the US into home cleaning routines AGGREGATE of thousands of real people. Not a person. ## Who they are Age and gender mix, with the model's stated accuracy Themes they engage with, and what they look at ## How they write Voice lines from members' own comments, stratified across topics, scrubbed of names, handles and links ## Known limits People who only watch are not represented
Stated limits
Each brief names its own gaps. Audiences are built from people who engage publicly, so people who only watch are missing, and a thin audience says so on the audience.
Simulation
Each audience answers from its own brief: who its members are, what they engage with and how they write. Nothing is added from a demographic stereotype.
The audience states its split
Each audience says how its members would divide across the options, instead of one answer sampled many times. On fresh survey questions this landed about three times closer to real splits.
Reactions can spread
Where audiences are linked by what they engage with, reactions pass between them in rounds, so a message that only lands with one hub behaves differently from one that lands broadly.
A winner only when it is earned
Comparing versions, Mimi names a winner only where that kind of test passed a held-out check on real outcomes. Otherwise it shows a lean, a shortlist or too close to call, and images and video say there is no reliable call.
Follow-ups in their own voice
Ask an audience why. Answers are written the way its members write, measured from their own comments. They explain a result; they never change it.
A written recommendation
What to do, how sure the result is, what is working and what is not, with each finding linked to the question it came from.
Reproducible by construction
Every study is pinned to a sealed data release. Re-reading a study later reads the same evidence it ran on.
Validation
A simulation is worth exactly what its validation is worth. Every pass rule is written down before the held-out data is scored, and the misses are reported next to the wins.
87%
right when it named a winner on news-headline A/B tests
1,500 held-out headline A/B tests from the Upworthy Research Archive (2013 to 2015), each with a clear winner. Mimi named a winner on 44% of them; the rest it called too close. Scored in October 2026 on a simpler setup than a full study. Other kinds of copy are not scored on their own yet, and no winner is called for images or video.
94%
similar to real US opinion splits
Nearly 800 survey questions from waves never used to build or tune it, compared with how real respondents divided.
0.86
AUC at spotting backlash
Ads that drew a regulator ruling, a boycott or an apology, against ads that drew none, on a held-out half. Good, but simple signals do well on these sets too, so we treat it as not yet proven.
87%
accurate on engagement with known channels
Whether an audience engages with a channel, forecast from recorded co-engagement and checked on nearly 1,800 fresh cases.
Held out means untouched
Choices are made on development data. The held-out set is scored once, against a rule written before any call was made on it. Locked sets stay locked until a method has earned them.
What we do not beat yet
Predicting one person's answer: 63% against the 76% at which people repeat their own answers. Engagement with brand-new content from a description alone: 77%, below where we want it. Both are reported here until they move.
Ranges, always
Every share comes with its range. Where two options' ranges overlap, the report calls them too close rather than ranking noise.
Build quality is published
How coherent an audience is, whether its label passed the check, and how much evidence stands behind it are shown on the audience itself, including when the number is unflattering.