Welcome to the Draper desk
Why we built an AI copywriter trained on ads that actually performed — and what we'll be writing about here.
Most copywriting AIs are general-purpose models with a system prompt taped to the front. They've read about ad copy. They haven't been shaped by it.
Draper is different. The model was fine-tuned on thousands of real paid ads — ads that ran on Meta, TikTok, X, Reddit, and Pinterest — paired with how long they kept running and how much engagement they earned. The pattern of what worked is in the weights, not in a prompt.
This blog is where we'll write about that — what we're learning from a corpus of winning ads, how we think about evaluation, and the craft of paid-social copy in 2026.
What you'll find here
- Pattern notes. Things we've noticed in the corpus — recurring shapes, hooks, failure modes — that you can use even if you never touch Draper.
- Eval write-ups. Blind comparisons of Draper output against frontier general-purpose models, judged by working media buyers.
- Product updates. New platforms, new training runs, new tools shipped.
If any of that sounds useful, get started free — or just check back. We'll be here.