Subscriber Twins
Ask a segment of your own subscribers what they would do — before you ship the change.
A twin is a model of a group of your subscribers, built from what that group actually did: when they subscribed, whether they trialled first, whether they renewed, when they left. You then interview it — "would you renew at £9.99?", "what would have stopped you cancelling?" — and it answers as that group would, based on their behaviour.
Available on Pro and Business.
What a twin is, and is not
It is a prediction built from your own subscription events, aggregated across many people.
It is not a real customer, a person, or a guarantee. It is not a survey — nobody was asked. And nothing is trained on your data: the twin retrieves the aggregate facts about its group, and the AI answers in character from those facts. Your data is never used to train a model, here or anywhere else in SubSovereign.
Why it models a group, never a person
Every fact a twin knows is a proportion or an average across its cohort — "68% of this group cancelled within 14 days of the first renewal charge". No individual subscriber's record is ever put in front of the AI.
That is a deliberate design decision, and it has a hard floor: a segment with fewer than 50 subscribers is refused. An "average" over four people is those four people wearing a hat, and someone could work backwards from a handful of percentages to an individual. If you see the refusal, it means the segment is too small to be genuinely aggregate — not that something went wrong.
The four segments
| Segment | Who it models | Good for |
|---|---|---|
| Cancelled or lapsed | Everyone whose subscription ended | Why they left, what would have kept them |
| Trialled, never subscribed | People who started a trial and never paid | What stops people at the paywall |
| Converted from trial | Subscribers who came through a trial | What won them over, what they expect next |
| Current subscribers | People paying you right now | How a change would land with your base |
Anyone who trialled and later converted is excluded from "trialled, never subscribed" — they are a customer, not a lapsed trial.
Using it
- Open Subscriber Twins in the dashboard and choose an app.
- Pick a segment. The twin is built from that segment's history — usually a few seconds.
- Click Interview and ask a question in plain English.
- Read the answer, then open "What this answer was built from" to see the aggregate facts behind it.
Ask about decisions, not facts. "Would you pay £2 more for offline downloads?" is a good question. "How many people cancelled in March?" is not — that is analytics, and the dashboard already answers it exactly rather than approximately.
Why the answers are blunt
Twins are instructed to disagree with you. They push back on price, distrust marketing language, and give lukewarm or negative answers when the group's behaviour suggests they would. That is deliberate: a twin that agrees with everything predicts nothing, and an encouraging answer you cannot trust is worse than no answer.
If a twin tells you your price rise will cost you subscribers, that is the feature working.
How to read an answer
- The cohort size sits next to every twin's name. A prediction from 3,000 subscribers deserves more weight than one from 60.
- The facts behind the answer are always available under the answer itself. If the reasoning rests on one weak statistic, you will be able to see that.
- Every answer is marked as AI-generated, because it is. Treat it as a well-informed opinion from someone who has read your data — not as a measurement.
Where the AI runs
Interviews are processed by Mistral in France through Merlyn, our own AI layer — the same EU-resident path as every other AI feature in SubSovereign. No AI request goes to a provider in the United States. If your workspace has AI switched off, Twins is off with it.
Limits today
- Twins are built from a snapshot. Rebuild one to pick up recent behaviour.
- The four segments above are what is available; custom segments are not yet supported.
- Twins model segments only. Twins of individual, consented subscribers are a future capability and will require explicit consent from those subscribers before anything is built.