Voice fidelity is the moat
A demonstrable, measurable answer to "does this AI actually sound like me?" — and why the answer is the only durable advantage in this category.
The default thesis for any AI content product is: train on your data, sound like you, save you time. Customers nod, sign up, churn three months later. Why?
Because "sound like you" is a feeling, not a number. The product feels close at week one, drifts at week three, and by week six the founder is editing every draft so heavily it isn't saving them time anymore. There's no instrument on the dashboard that tells them whether the voice is improving, holding, or degrading.
We think the durable advantage in this category — maybe the only durable advantage — is making voice fidelity a measurable, falsifiable, dashboard-visible number. Not "how confident is the model in its own output." That's marketing. The actual question is:
In a blind A/B comparison against a generic LLM, would a forensic reader pick our draft as more like the founder's authentic writing?
If the answer is yes, the product has a moat. If the answer is no, the product is a wrapper.
What we built
A loop that runs nightly without human intervention:
- Gold set. Each founder's published, authored writing — verified at ingest time, never touched by the profile builder. The ground truth.
- Brief synthesis. Reverse-engineer what brief would have produced each gold sample.
- A/B generation. Run the synthesized brief through (a) Replacer with the founder's profile and (b) a generic LLM with no profile.
- Blind judge. A separate model picks which draft is more like the gold across five voice dimensions, in randomized order, with rationale.
- Score. % of judgments where the profile-trained draft wins.
That number lives on the dashboard. When a profile rebuild regresses fidelity, we auto-roll back to the previous version and log the incident — no human needed for routine drift.
What we learned testing it on ourselves
The first live test surfaced a hard truth about where voice training data should come from.
A founder's Drive folder is dominated by artifacts — agendas, CSVs, handoff docs, third-party letters, AI-generated text the founder saved for reference. After we built an authorship verifier and ran it on the test corpus, 58% of the candidate gold samples got quarantined. Most "business" content in the Drive was non-authored work product. The remaining authored content skewed personal — letters, family memorials, vulnerable first-person reflections.
The implication: voice gold sets shouldn't come from generic file repositories. They should come from published outputs — LinkedIn posts the founder wrote, blog posts under their byline, op-eds, conference transcripts. That's where in-context business-voice authored writing lives. It's also where the founder has already proven they can produce something publication-ready.
This is why the first onboarding step is exporting your LinkedIn archive. Not because we want your data. Because that's where the signal actually is.
The honest number
We don't claim "indistinguishable from human." We claim "demonstrably closer to your published voice than a generic LLM, by a wide margin, with the math on the dashboard." That's a number that holds up.