How Litmus works
From your problem statement to a structured report — in plain language.
1
You tell us what you’re testing
A problem statement, an idea (or design images), and a target audience. The problem statement is the highest-signal input — people post about problems on Reddit and X, not about solutions, so this is what drives our search later.
2
We figure out who your users are
A model parses your audience description into structured axes — geography, age range, profession, life stage. These become the diversity constraints for the personas we generate.
3
We scan Reddit and X for real voices
Using your problem statement, idea, and audience axes, we find ~6–12 relevant subreddits and ~6–10 X search queries. We pull the last 30 days of posts, drop noise (low-engagement, deleted, off-topic), and keep ~30–50 genuine snippets that match your problem space.
This is the “real voices” part. Every quote we pull keeps its source URL so you can verify the provenance later.
This is the “real voices” part. Every quote we pull keeps its source URL so you can verify the provenance later.
4
We generate the synthetic personas
A model creates 10 diverse personas spanning your audience axes. Critically, each persona gets 2–3 verbatim quotes from real Reddit and X postsbaked in as their authentic voice samples. We also force a 33/33/33 split across sympathetic / neutral / contrarian stances so the test doesn’t skew positive.
5
We interview each persona in character
All 10 personas run in parallel. Each is given their full profile (including the grounded voice samples) and the interview script. They answer in character — speaking with the voice of the real people they’re grounded in, not generic AI-speak.
6
We synthesize the report
A model aggregates all 10 interview transcripts into a structured report: headline finding, executive summary, where it works, where it fails, themes with verbatim quotes, sentiment distribution, stance breakdown, and concrete next steps.
Two things to know
- 10 synthetic personas is a qualitative read, not a statistical one. That’s sufficient for finding themes and surfacing concerns (Nielsen’s 5-user rule × 2). It is not enough for “X% would pay” claims.
- Quotes in the report are from the simulated interviews— not directly copy-pasted from Reddit or X. Each persona’s voice was grounded in real posts, but their interview answers are generated in character. Click any persona in the report to see their original grounding source URLs.