What an audience is, and what it will never say about a person

When Planetarium names an audience, it is describing a real group of people, boiled down to a summary. It is not a character, it is not a named person, and nothing in it can be traced back to someone. This post explains how an audience gets built, how we check the labels, and the lines we will not cross.

Grouped by who engages with whom

The easy way to make an audience is to pick a hashtag and gather everyone who used it. We do not do that. A hashtag mostly measures tagging habits: people who tag things the same way, which is not the same as people who care about the same things.

Instead we look at who engages with whom. People who comment on the same creators, over time, tend to belong together, whatever they happen to tag. We group on that behavior, then merge near-duplicates so two labels never secretly mean the same people.

What a brief contains

Each audience is described by a short brief, written in plain terms. It holds the mix of ages and genders and how reliable that mix is, the themes the audience engages with, and a sample of how its members write, taken from their own comments and scrubbed of names, handles and links. It opens with a plain reminder of what it is: an aggregate of a stated number of real people, never fewer than 100, not a person.

When you test a message, the audience answers from its brief. Nothing is added from a demographic stereotype. If the brief does not say it, the audience does not know it.

How we check a label

Naming a group is where a lot of audience tools go wrong. A label sounds right and nobody checks. We check in two ways.

First, a second AI model reads each label against members’ own words that it did not see when the label was written. If the label does not hold up, it is not offered. Labels on sensitive subjects are reviewed, and they describe content, never beliefs.

Second, we compare what we say about an audience with what its members say about themselves.

What we never infer

Age and gender come from what people say about themselves, calibrated, and appear only as an audience’s mix. Ethnicity, religion, health, sexual orientation and politics are never inferred. That is a rule about what the system will compute, not a filter on what it displays.

The raw material is public activity only: what people chose to post, comment on and engage with in public. Identifiers are replaced with salted keys before any audience is built. The platform does not take in private messages, bought contact lists, or anything behind a login the account holder did not make public.

The floor of 100

Every audience has at least 100 people in it, and that is enforced when the audience is built, not tidied up at display time. So there is no audience of one, and no screen anywhere in the product that simulates a named person. Every figure you see is a mix across at least 100 people.

As our privacy policy says, account-level records are access-controlled, kept only so an audience’s construction can be audited and exclusion requests honored. What the product and its evaluation read are the aggregates.

The gaps we state on the audience itself

An honest audience description includes what it is missing, so each brief carries its own limits. The biggest is simple: audiences are built from people who engage in public, and people who only watch leave no trace. They are missing, and the audiences inherit that skew. Our population panel is weighted to match the census, which corrects the overall mix, but it cannot invent the people who are absent.

There is a second, more specific gap. Some of our earlier audiences were built from the connections between commenters and the creators they follow, and the comment text itself was not kept. They know what their members watch and follow, but not how they talk, so they answer from topics and mix alone. We say so, and we do not describe them as backed by member voices. Where a brief is thin, it says it is thin.

We also tested whether audiences caricature themselves. Asked to write about something neutral, making coffee at home, the simulated audience used its own identity words in 36 of 59 cases, 61%, against 20.6% on average in the real posts of those groups: about 3 times as often. That check ran on 13 August 2026 on 59 creator-based audiences, and it is a flaw we track, not one we have solved.

What this adds up to

A good audience is not the one with the most flattering description. It is the one whose description you can check: grouped by behavior, labeled by evidence, built from public activity, large enough to hide any individual, and honest about who is missing. If you ever find a place where we fall short of that, tell us.