I Was So Sure This Would Work
Here's the thing about running ads with AI — you can move fast. Really fast. And sometimes moving fast means you sprint off a cliff before you realize the ground is gone.
I had an e-commerce client generating decent sales through standard ad creatives. Talking heads, product demos, the usual stuff. But I kept seeing these split-screen meme ads everywhere — the ones where the left side shows some "before" disaster and the right side shows the aspirational "after." They were getting insane engagement.
So I had an idea: what if we automated the entire split-screen ad pipeline?
The Setup
I told Ari to build me a system. Here's what we created:
- Tenor scraping — automated search for reaction GIFs matching specific emotions (frustrated, confused, stressed on the left; happy, relaxed, celebrating on the right)
- ffmpeg processing — download GIFs as MP4, loop them to exactly 15 seconds, crop to vertical panels
- Remotion rendering — custom
GifSplitAdcomponent that composites two panels side by side with text overlays and a CTA - Batch rendering — a
variations.jsonfile with 10+ concept variations, each rendering automatically
The whole pipeline went from "idea" to "10 finished ad creatives" in about 4 hours. I was pumped.
Here's a sample of the concepts we ran:
- "Trying to figure out Facebook Ads alone" (left: confused person) vs "Having a done-for-you system" (right: person celebrating)
- "Burning money on ads that don't work" (left: money pit) vs "Smart ad targeting that converts" (right: cash register)
- "Running your business the old way" (left: stressed entrepreneur) vs "Automated everything" (right: person on beach with laptop)
You get the idea. Classic pain/pleasure split.
The Campaign Structure
I set up a CBO campaign (campaign budget optimization) at $50/day through the Meta API. Ari handled the technical setup:
- Campaign objective: Conversions (purchase)
- Targeting: Broad — women 25-55, US, interested in e-commerce/side hustles
- Placements: Automatic (Facebook Feed, Instagram Feed, Stories, Reels)
- Budget: $50/day CBO across 3 ad sets with different audience segments
- Creative: 6 of our best split-screen variations
We launched on a Monday. I checked results obsessively for the first few hours. Impressions were flowing. CTR looked decent — around 1.2%. Link clicks were coming in.
Day 1-2: The Warning Signs I Ignored
Here's where I made my first mistake: I confused engagement with intent.
The ads were getting clicks. People were watching the full 15 seconds. Some were even sharing them. The meme format was doing its job — it was entertaining.
But nobody was buying.
Day 1 stats:
- Spend: $48.72
- Impressions: 12,340
- Link clicks: 148
- CTR: 1.2%
- Add to carts: 3
- Purchases: 0
Three add-to-carts out of 148 clicks. That's a 2% landing page conversion to ATC, which is terrible for a product that normally converts at 8-10% from warm traffic.
I told myself it was early. The pixel needed data. Give it time.
Day 2 stats:
- Spend: $51.14
- Impressions: 14,200
- Link clicks: 162
- CTR: 1.14%
- Add to carts: 2
- Purchases: 0
Day 3: Pulling the Plug
By the morning of Day 3, I'd spent $124.86 with exactly 5 add-to-carts and zero purchases. I had Ari run the numbers on what we needed to see to justify continuing:
At the product's average order value and our target CPA of $15, we needed roughly 1 purchase per $15 in spend. We were at $124.86 per zero purchases.
I told Ari to pause the losers. It ran our pause_losers.py script, which automatically kills any ad set that's spent more than 2x the target CPA without a conversion. All three ad sets got flagged.
Day 3 stats (before pause):
- Spend: $48.30
- Impressions: 13,800
- Link clicks: 134
- Add to carts: 1
- Purchases: 0
Total damage: $173.16 spent, 0 sales.
The Autopsy: What Actually Went Wrong
After the bleeding stopped, I sat down and figured out what happened. There were three problems, and they compound on each other.
Problem 1: Wrong Creative Format for the Funnel Stage
Split-screen meme ads are awareness content. They're funny. They're shareable. They stop the scroll. But they don't sell.
Think about it: someone watches a 15-second meme comparing "stressed entrepreneur" vs. "beach laptop lifestyle" and... what exactly? They're entertained, not educated. They don't know what the product is, why it's different, or why they should care.
We were running a conversion-optimized campaign with awareness-stage creative. That's like showing up to a job interview in a clown suit — you'll definitely get noticed, but you're not getting hired.
Problem 2: The Audience Mismatch
Our targeting was too broad. Women 25-55 interested in e-commerce is basically half of Facebook. The split-screen format attracted the "I wish I could work from home" crowd, not the "I'm actively looking for a specific solution to a specific problem" crowd.
The clicks we got were curiosity clicks, not intent clicks. People wanted to see what was on the other side of the meme, not what was on the landing page.
Problem 3: No Warm-Up Layer
The ads went straight to a product page. No lead magnet, no content piece, no video sales letter — just meme → product page → "buy this thing you've never heard of."
For cold traffic from entertainment-style creative, you need a bridge. Something that takes someone from "haha that's relatable" to "oh wait, this actually solves my problem." We skipped that entirely.
What the Data Actually Showed
After I had Ari pull all the analytics, a few things became clear:
The positive signals:
- CTR was above average (1.2% vs. typical 0.8-1.0%)
- Cost per click was actually decent ($0.39)
- Video completion rate was high (~65% watched the full 15 seconds)
- The creative format itself was engaging
The negative signals:
- Landing page conversion was 1.1% to ATC (normally 8-10%)
- Time on landing page averaged 8 seconds (normally 45+)
- Bounce rate was 94% (normally 65%)
- Zero purchases across 444 link clicks
The story the data tells: people clicked because the ad was entertaining, landed on a product page that had nothing to do with what they just watched, and bounced immediately.
The Pivot
Here's what I changed after this failure:
1. Creative-to-Landing Page Congruence
If the ad is about a specific pain point, the landing page better open with that exact pain point. We killed the meme approach for direct-response campaigns and moved to UGC-style talking head videos that actually explain the product.
2. Added a Comment-Trigger CTA
Instead of sending people to a landing page, our organic content now uses a comment trigger — "Comment CREATE to get started." This feeds into a GHL automation that DMs the offer link. It's warmer traffic because they took a deliberate action, not just a curiosity click.
3. "Unbothered Mundane" Over Memes
The new creative direction is completely different. Instead of flashy split-screen memes, we film clips of someone doing boring everyday stuff — eating cereal, folding laundry, walking through their house — with text overlays about the product. It looks organic. It doesn't look like an ad. And early results are way more promising.
4. The Automation Pipeline Lives On
Here's the silver lining: the technical pipeline we built wasn't wasted. The Remotion rendering system, the batch processing, the automated creative generation — all of that transferred directly to the new organic reel templates. We just swapped the creative direction.
GifSplitAd died. NectarReel, NectarBlackBar, and NectarWhiteBox were born from its ashes.
The Real Lesson
$173 is cheap tuition.
The mistake wasn't building the pipeline. The mistake was assuming that what works for engagement also works for conversion. Those are two completely different metrics that require two completely different creative approaches.
Entertainment gets eyeballs. Education gets wallets.
If I'd run these same split-screen ads as top-of-funnel awareness and then retargeted the viewers with actual product content, it might have worked. But going straight from meme to checkout? The data says no.
The AI did exactly what I told it to do. It built the system perfectly. The system just solved the wrong problem.
That's the thing about having an AI business partner — it accelerates everything, including your mistakes. You can fail faster than ever before. Which, honestly, is a feature, not a bug. I'd rather spend $173 in 3 days and learn than spend $1,000 over 3 weeks and learn the same lesson.
The Numbers, Final
| Metric | Value |
|--------|-------|
| Total spend | $173.16 |
| Total impressions | 40,340 |
| Total link clicks | 444 |
| Average CTR | 1.1% |
| Average CPC | $0.39 |
| Add to carts | 5 |
| Purchases | 0 |
| Revenue | $0.00 |
| ROAS | 0.00x |
| Days running | 2.5 |
| Lesson value | Priceless (genuinely) |
---
This is Day 3 of the Machine Earned 30-day build log. Every day, I publish what's actually happening — wins, losses, and everything in between. If you're building with AI or just want to watch someone try, subscribe to the newsletter.