Find the segments that behave differently
Discover which groups in your audience react in opposite directions to the same decision — and which of your existing segments are really one group wearing two labels.
MiroFish approaches segmentation behaviourally: instead of grouping people by attributes, it simulates how different groups react to a specific decision and reports where those reactions genuinely diverge. You describe your audience and the change you are considering, and the engine models the cohorts as agents, surfacing which segments split, which move together, and which distinction you are maintaining that makes no behavioural difference. The output is a reaction-based segmentation with estimated confidence, useful for deciding who to message differently. It is a hypothesis generator for your real data, not a replacement for analysing it.
How MiroFish simulates it
Behavioural segmentation runs the pipeline against a decision rather than a population, because segments only exist relative to something:
- 1
Describe the audience and the decision
Both matter. "Our users" is not a segmentable input; "our users, facing a move from perpetual licence to subscription" is, because segments are defined by differing reactions to something specific.
- 2
Cohort modeling with distinct priors
Each described group becomes agents with their own incentives, switching costs, and sensitivities. Where you have not described a difference, the simulation will not invent one — vague inputs produce vague segments.
- 3
Divergence detection across rounds
Agents react over several rounds and the simulation tracks where reactions separate and where they converge. Convergence is as informative as divergence: two segments that behave identically do not need separate messaging.
- 4
Reaction-based segment report
Which cohorts diverge and on what dimension, which of your existing segments collapse into one, the segment most at risk from the decision, and the messaging angle each genuinely distinct group needs.
A worked example
A design tool with 40,000 users plans an AI feature. Existing segments: hobbyists, freelancers, in-house teams, agencies. Which actually differ?
What you give it
- Audience
- 40k users, 4 declared segments
- Decision
- Ship a generative AI feature
- Known tension
- Anti-AI sentiment in design
- Channels
- In-app, X, design communities
What the report estimates
- Genuine divergence
- Ideological, not by customer type
- Collapse into one
- Freelancers + in-house teams
- Highest-risk cohort
- Vocal craft-identity minority
- Implication
- Segment by stance, not by plan tier
Illustrative output. Segment hypotheses need validating against your real usage and revenue data before you act on them.
What it can’t do (honest limits)
- —It has no access to your actual usage, revenue, or cohort-retention data — it reasons from what you describe, so it generates hypotheses rather than findings.
- —It will not discover a segment you have given it no reason to suspect. Naming an unexpected group is often what produces the useful result.
- —Segment sizes are not estimated. It can tell you a cohort behaves differently, not how many people are in it.
- —Behavioural segments from a simulation should be validated against real data before they drive spend, pricing, or roadmap.
Questions people ask
Can AI help with audience segmentation?
It can produce behavioural segment hypotheses quickly: given an audience and a specific decision, MiroFish simulates how described cohorts react and reports where those reactions genuinely diverge. That is different from clustering your CRM data, which tells you who people are rather than how they will respond. Use the simulation to generate hypotheses and your own data to confirm them.
What is behavioural segmentation versus demographic segmentation?
Demographic segmentation groups by attributes — role, company size, plan tier. Behavioural segmentation groups by how people respond to something. The distinction matters because attribute-based segments frequently behave identically, meaning you maintain separate messaging for groups that did not need it, while a real behavioural fault line runs invisibly across all of them.
How do I know if my segments are real?
Real segments react differently to the same decision. If two of your declared segments produce near-identical reactions across several simulated scenarios, they are probably one segment with two labels — and the report will say so. That finding usually saves more effort than discovering a new segment does.
Can it estimate how large each segment is?
No, and it should not pretend to. The simulation reasons about behaviour, not population counts, and any size figure it produced would be invented. Segment sizing has to come from your own data; the simulation tells you which cuts of that data are worth making.
Which segment is most likely to churn after a change?
The report identifies the cohort most at risk from the specific decision you describe, with reasoning — typically the group whose switching cost is lowest relative to how strongly they object. It is a directional estimate, so pair it with your real retention data before acting.
Can I segment by attitude rather than firmographics?
Yes, and it is often the more useful cut. Attitudinal splits — early adopters versus sceptics, craft identity versus efficiency focus — frequently predict reaction better than company size or job title, and they cut across conventional segments in ways firmographic analysis cannot see.
How is this different from a survey?
A survey asks people what they think; this estimates how groups react to each other over time, which is where opinion actually forms. Neither replaces the other. The simulation is faster and cheaper for generating hypotheses; a survey gives you real answers from real people to test them.
How many segments should I describe in the prompt?
Three to five, named with the distinction you believe separates them. Fewer gives the simulation nothing to differentiate; more tends to produce overlapping cohorts that dilute the output. If you are unsure a segment is real, include it — being told it collapses into another is a useful answer.
Go deeper on the method: read the full guide on the blog →
Run it on your own scenario
No install, no API keys. Monthly plans from $2.99 — results in about 30 seconds, cancel anytime.
Start predicting