Loading the Elevenlabs Text to Speech AudioNative Player...

TL;DR: From August 2024 to April 2025, an automation published 113 blog posts on this site with zero human editing. I turned it off. Here’s what it produced, why it failed, and what I’m doing differently now. Real numbers included, because that’s the whole point of this site.

If you scroll far enough down the blog here, you’ll hit a wall of posts by an author named “AI PEPE.” Titles like “Enhancing Marketing Newsletters with AI for Small Businesses” and “Harnessing AI for Predictive Analytics in 2025.” That was an experiment I ran for about eight months, and I want to be honest about how it went, since I now sell AI content systems to other businesses.

What the machine did

The pipeline was built in Make.com. Every morning at 7:15 it asked an OpenAI assistant for a topic, sent that topic to Perplexity for research, handed the research to a second assistant to write an article, had a third assistant write an image prompt, generated an image with DALL-E, and pushed everything to WordPress. Toward the end I added a Slack notification so I’d at least see what it made.

The numbers, pulled from the site this week:

Why it died

The topic picker had no memory and no point of view. “Generate a topic for today” with no constraints produces the statistical average of every AI marketing blog on the internet. That’s how you get “Enhancing Agile Marketing Strategies with AI for Small Businesses” on a Tuesday and “Enhancing Content Marketing Strategies with AI for Small Businesses” on the Wednesday. Nobody searched for either.

It was built to produce, not to be read. Every post was the same length, the same structure, the same voice. The tell is in the titles: five verbs carried more than half of them. A human editor would have caught that in week two. There was no human editor.

Drafts were the wrong safety valve. Later versions posted as drafts so I could approve them. In practice, that meant a pile of drafts I never opened, because reviewing 30 generic articles a month is worse than writing 4 good ones. The automation didn’t save me work. It moved the work to a place I ignored.

Nothing connected to what I was actually doing. During those eight months I was running a marketing agency, building apps, and working with real clients. None of that showed up. The blog could have been anyone’s.

What I’m keeping

The transparency. Those 113 posts are still live, in a category now called “Archive: AI Lab 2024-25.” I’m not deleting them, because the point of this site was always to show what AI content looks like when you let it run, and that archive is the clearest demonstration I have.

The plumbing, sort of. The idea of research, draft, image, publish is right. The execution was wrong.

What’s different now

Since this week, the site runs on three rules.

Every topic comes from somewhere real. Either a piece of news that actually affects local businesses, a how-to I’ve walked a client through, or a lesson from work I’ve done. No “generate a topic.”

Every post gets read before it goes live. AI drafts, I edit. The footer on each post says so. If that means two posts a week instead of thirty a month, good.

The site sounds like me. First person, real numbers, willing to say something didn’t work. This post is the test of that.

The lesson, if you’re building one of these

Volume is the easiest thing to automate and the least valuable. If you set up an AI content system for your business, put your effort into the topic source and the editing step. Those are the two places a human still matters, and they’re the two places every off-the-shelf pipeline skips.


Justin Peters runs Zoomies Media, an AI-powered marketing agency for local and hospitality businesses in Orlando. This post was drafted with AI assistance and edited by Justin. Read about how this site works.