How We 3X'd a Client's Facebook Reach at Near-Zero Added Cost.
The Exact AI Production Stack, the 8-Second Rule, and the Expensive Mistakes That Cost Us Weeks.
We tripled a client's Facebook reach in one campaign cycle. Budget, team and timeline stayed effectively flat.
Brand awareness (measured independently via Tracksuit) rose ~8% over the same window.
The lever wasn't "AI." It was a repeatable production protocol - and roughly a dozen failures that taught us where the real constraints are. Below is the whole thing, including what I'd do differently.
The setup - what actually capped reach
For the entire history of paid social, output was rationed by production cost. You want to test 15 creative angles; each one is a brief, a designer, two revision rounds, sign-off, trafficking. So you test 3, and call rationing "strategy."
Here's the punchline: the reach ceiling was never the algorithm. It was how few quality variations you could afford to feed it. Meta's delivery engine rewards fresh, relevant creative. Starve it and it has nothing to optimise toward.
The protocol (what we actually ran)
Strategy - and this step did NOT get faster. Audience sub-groups, psychological triggers, objections. Garbage in, garbage out - at 10x volume.
Production - the multiplier. AI-assisted generation took a 2-week production cycle to ~2 days. We went from a handful of concepts to 14 → 24 → 36 per cycle, each adapted per format and segment.
Human filter - non-negotiable. We killed most of it. The win isn't making 1000 assets; it's the taste (& data insights) to ship the right 30.
Deploy → read signal → feed winners back. Because production is cheap, we could actually run the loop instead of theorising about it.
What nobody tells you (the expensive lessons)
These aren't in any manual. We paid for them in wasted cycles:
Don't feed a processed AI image into another AI tool ("AI-squared"). Artifacts compound. People come out with a fish-like quality - shiny, waxy, uncanny. Always start from a clean generation.
~8 seconds is the usable video ceiling. Beyond that, coherence degrades. Design around it; cut, don't stretch.
Accents are a specificity test. The NZ/Kiwi accent was broken early - until the tooling crossed a threshold. Now it's flawless unless the prompt is vague. Vague in, generic out.
Direct AI actors in "beats." Emotional cues + timing, like notes to a real actor. It's the difference between animatronic and believable.
You can't "fix it in post." Editing anything non-trivial ("remove the text on that chair") is slower than re-rolling. All weight moves to the prompt.
More prompt ≠ better. We choked a model with a giant tone-of-voice doc. The skill is enough, not too much - earned through test-and-learn.
The 80/20 - where the edge actually lives
Strip it back and the advantage is three things a tool can't sell you: (1) what you feed it, (2) how you prompt it, (3) how you filter it. The multiplier amplifies whatever you already are. Point it at sharp strategy → 3x reach at flat cost. Point it at weak thinking → 3x the mediocrity, faster.
What I'd test next: longer sequenced narratives stitched from 8-second cuts; a scored rubric for the human filter to reduce subjectivity; prompt-length A/Bs to find the choke point empirically.
Want the deconstruction applied to your accounts? Let's talk.
Not investment advice, not a guarantee of results - creative performance varies by account, offer, and audience.
