How Planetarium works

What Planetarium is

Planetarium simulates an audience. Its cohorts are groups of real people, clustered from what they watch and how they comment on public social platforms. When you run a study, model agents answer as the members of each cohort, grounded in what those members actually say and the way they say it. You get a read cohort by cohort: who leans in, who pushes back, why, and how sure the result is.

You work with Planetarium in three places: the console (this app), the REST API for your own code, and the MCP server for Claude, Cursor and other assistants that speak MCP. All three run the same studies on the same cohorts.

How a study works

1. Goal. Say what you want to learn: read one creative, compare versions, find objections, test a concept or run a survey.
2. Creative. Paste the copy, script or idea, or attach images, video and links. Planetarium reads attached media once and describes it to the audience.
3. Audience. Choose cohorts yourself, or describe who should answer and let Planetarium pick a balanced panel.
4. Questions. Up to ten: pick-one, open answers, 0 to 100 scores and rankings.
5. Research (optional). Planetarium reads current web context about the subject before the audience answers.
6. Run. Each cohort is sampled, objections are checked, and a second round lets cohorts see one another's reactions.
7. Report. Shares with ranges, splits by cohort and version, objections, quotes and a written recommendation.

Start in the console from New study. Smart Run drafts the whole study from a sentence ("test this headline with moms who meal-prep"); Advanced takes you through each step yourself.

What do you want to test?

Where surveys are on, Run a survey asks questions with no creative at all.

Audiences and cohorts

A cohort is a group of real people with something in common, and its label says who they are, for example "Women 25-34 in the US into skincare". Every label states the gender, the age range, the place and the interest where the data supports it, and never invents what the members did not reveal.

Kinds of cohort

• Interest cohorts: people grouped by what they watch and talk about.
• Creator audiences: the audience of one brand, business or media account.
• US panel: general-population cells by age and gender, weighted to the US census. Planetarium uses these when you describe an audience by demographics alone.
• Audience: private cohorts built from your own brand's followers, visible only to your account.

Choosing them

Search the catalog under Cohorts, or describe who you want in plain words ("dads in their 30s into home espresso"). Planetarium reads the description into gender, ages, place and topics, drops cohorts whose label contradicts it, and ranks the rest with a one-line reason for each match. In a study, Describe the audience has Planetarium choose three to five cohorts with deliberate contrast, so the panel is not five copies of the same room, and it warns when the brief asks for someone no cohort covers.

Only cohorts with a real label and one clear subject can be added to a new study, and a study holds up to twelve. You can also add audience requirements (up to eight lines, such as "has bought a subscription box in the last year") that every answer must respect.

Website tests

A website test sends people from an audience to a live site and records what each one does there: what they read, what they tap, where they leave and why. Open it from Website test in the console.

1. Site. Enter a public web address. If the part worth testing sits behind a sign-in, you can sign in yourself in the test's own browser before people arrive.
2. Audience. Search the catalog for who should visit, or let Planetarium suggest audiences likely to care about the site. A fit check says how well the audience you chose matches who the site is for.
3. Who they are to the site. Customers by default, or investors, job candidates, partners, press, or people you describe in your own words.
4. Task and device. Give people something to do, such as finding a product or picking a plan, or let them look around. Choose desktop or phone, and how many people visit: up to 100, 20 by default.

The result says how each visit ended (tapped a call to action, might come back, not for them, or gave up), where people left, what put them off and what works. Watch any visit as a replay with a heat map of where people looked and tapped, read each person's reasons, and copy ready-made prompts for fixing the site. A test can be shared as a read-only link, and the API and MCP server can start and read tests too.

Brand audiences

Under Audience, connect a brand's social accounts and Planetarium rebuilds its audience from the people who actually comment on them. It looks up what else those people engage with and splits them into cohorts. Building one runs once and takes a few minutes; the result is private to your account.

Open a built audience to see its cohorts and how they differ, choose which to model, and put the question you want to test to them. Each cohort answers from what its own members say. Open Audience from the console sidebar to build and manage your cohorts.

Questions

At least one question needs answer options, because open answers alone have nothing to tabulate. In a survey every question needs options. Write questions the way a respondent would read them, and keep options distinct.

Running a study

Depth sets how many responses each cohort gives: Quick (6), Standard (12), Extended (24) or Max (40). Deeper runs give tighter ranges and take longer. Each cohort gives the number of responses its depth sets; there is no adaptive option.

A run moves through these stages, shown live on the study page:

1. Researching context (only when research is on)
2. Preparing cohort evidence: each cohort's own comments and vocabulary
3. Collecting open responses and checking potential objections
4. Sampling the cohorts, and retrying any incomplete answers
5. Evaluating social influence: a second round where cohorts see one another's reactions (the first round is kept as a control)
6. Saving results, then writing the analysis

Most runs finish in about a minute; with web research, several minutes. You can leave the page: the run continues and the study shows its result when it is done. To stop a run, use Stop run on its progress card: it ends before its next paid step, the responses so far are discarded, and its credit hold is returned. A run still going after three hours is stopped and refunded on its own. A run that fails says why, and you can run it again. When fewer than half of a run's planned answers came back, its report says so, and it is charged only for the answers it got.

Reading results

• Shares and ranges. Each option's share of answers comes with a range. When the ranges shown for the leading option and the runner-up do not touch, the lead is clear (the API calls this rangesSeparate); when they overlap, treat it as too close to call.
• By cohort. The same question split by cohort, which is usually where the story is: who loves it, who is indifferent, who is put off.
• By version. For comparisons, every question per version, and a side-by-side test. Each cohort sees all the versions together under neutral letters, in a balanced rotation of orders, and splits 100 points by what its members would do at first glance and by which they would call better. The verdict comes from what they would do (which they would call better, for a logo or name change). A version is ahead only when its margin reaches the cutoff for its kind and the margin's 80% interval stays above zero; that interval shows how far the cohorts agree, not error against real outcomes. A winner is named only when that kind of test has passed an accuracy check on real outcomes. Otherwise the report shows a lean, a shortlist or too close to call, and image and video comparisons say there is no reliable call.
• Objections. Each cohort's strongest objection, how upset they are, its severity (none, mild, serious), its kind (harms a group, objects to who is included, objects to the stance, or taste) and what they would likely do: nothing, post about it, stop buying or complain to a regulator.
• Quotes and themes. Representative answers in each cohort's own voice, and the themes in the open answers.
• AI analysis. A recommendation with its confidence, the findings behind it, the risks and what to test next.

Chat with your study

Open Chat on a finished study and ask about its results in your own words: "which cohort liked Version B least, and why?", "what do the people who said no have in common?", "what should I change first?". The answer leads with one bold line, cites the questions (Q2) and versions it rests on, and offers questions to ask next.

• What it can answer. Anything in this run's recorded results: every cohort's answers, the shares and ranges, objections, quotes and the AI analysis.
• What it cannot. It runs nothing new and asks the cohorts nothing, so it cannot say how they would react to something the run did not test. For that, ask the cohorts a follow-up, or run a new version. When the results do not cover a question, it says so.
• Charts. When an answer compares numbers it may draw one chart: bars for one number per cohort, a stacked bar for how each cohort splits across the options, versions side by side with the winner, or a heat map of cohorts by questions.
• Model estimates. Its numbers come from the simulated run, and the answer is written by a model reading them. Treat it like the rest of the report: a strong prior for what to test, not a measurement.
• Cost and limits. Each answer costs a few credits, shown under the chat box, and is charged once it is written; a failed answer is free. Each account can ask 60 questions an hour, counting questions asked through the API and MCP server.

Code and assistants can ask the same questions with POST /studies/{id}/ask or the ask_study tool.

Follow-ups and sharing

• Ask a follow-up. Put a new question to the study's cohorts after the run; each answers in its own voice.
• Group discussion. Let up to six cohorts discuss the creative with one another and read the transcript.
• Share. Make a read-only link to the report, or to a single question, and turn it off at any time. Export the report as a PDF.

Limits

What results mean