Last weekend I found research showing you can run accurate surveys using synthetic audiences built from real persona data.
Wild concept. I had to try it.
I built a synthetic audience engine to optimize subject lines for this newsletter. Based on survey data from you and 10,000 other readers.
I backtested it against previous emails. The accuracy? 85%.
Here's how it works:
You enter a subject line + preview text. Exactly what shows up in someone's inbox.
7 synthetic personas evaluate it. Each one represents a segment from our actual survey data: freelancers, corporate innovators, productivity seekers. They have different preferences for subject line styles.
Each persona answers 4 questions:
Would you open this email?
How relevant is this to you right now?
Do you instantly understand what it's about?
Does this feel clickbaity or spammy?
We convert responses to scores. Instead of asking for numerical ratings (which always drift to the middle), I use Semantic Similarity Rating. The system compares their written response to benchmark statements like "Must open now" vs "I'd ignore this" and calculates proximity.
You get a SubjectLine Index score. This composite score combines:
Open Intent (main driver)
Relevance boost
Clarity boost
Authenticity boost
Test variants side-by-side. Run A/B/C tests before you send.
Here's one of the persona definitions:
{
name: 'General AI-Curious Professional',
personaType: 'General AI-Curious Professional',
snapshot: 'roles: employed, professional, founder, ceo; industries: technology, other, education, academia,
healthcare, pharma; goals: General Learning, Automate Workflows, Build AI Business',
aiMaturity: 3.0,
aiUsage: 'weekly',
subjectLineStyle: 'short (<42 chars); angles: starter guides, safe defaults, avoid pitfalls; tones: specific,
credible, no-hype; formats: numbers upfront, template/verb-led, before→after; avoid: ALL CAPS, !!!, vague, clickbait',
examples: [
'5‑min fix for [task]',
'starter guides in 3 steps',
'From 60 mins → 10: the workflow we use'
],
ssrWeights: {
OpenIntent: 0.35,
Relevance: 0.3,
Clarity: 0.2,
SpamRiskPenalty: 0.15
}
}This same approach works across your entire marketing stack:
Test product names before you commit to branding. Build personas from customer interviews, then run 20 variations through them to see what actually resonates.
Validate messaging angles for landing pages. Does your "save time" angle hit harder than your "increase revenue" angle? Your synthetic audience will tell you which pain point connects.
Optimize ad copy variations. Facebook ads, LinkedIn sponsored content, Google search ads. Test headlines and body copy against segment-specific personas before you spend a dollar on impressions.
Pressure-test positioning statements. That tagline you've been workshopping for three weeks? Run it through personas built from your actual customers to see if it lands or falls flat.
Content topic selection. Before you write that 2,000-word guide, ask your synthetic audience which topics they'd actually engage with. Save yourself from producing content nobody asked for.
The core insight is the same across all of these: you're not asking AI to guess what works. You're encoding real human preferences into AI personas, then using them to stress-test your ideas before they go live.
I'm still exploring how to plug this into bigger workflows.
Example: We're launching our consulting business and planning to let our AI assistant (Fred) run the company LinkedIn autonomously. He'll mine for content ideas, publish, measure, optimize.
We're building him a synthetic audience panel based on our ICP. He'll test content before publishing.
The early results are promising enough that I'm curious where else this applies.
Hit reply if you want a more detailed behind-the-scenes of this build.
Until next week,
Martin
PS: Go connect with me on LinkedIn and say hi! It’s always fun to chat with readers
PS2: Want to kickstart you business’ AI transformation? Fill out this interest form and my consultancy will be in touch asap for a free discovery call.