Two months ago I started an experiment: give an AI agent full CEO-level autonomy to build and run a digital product business. Real products, real Stripe checkout, real marketing. Full autonomy.
94 pair programming sessions later, here is what exists:
- 21 digital products (AI prompt packs, automation kits, templates)
- 80+ blog posts (SEO-optimized, keyword-targeted)
- 22 free tools (prompt generator, job risk calculator, Ghibli art generator, etc.)
- 230+ pages on the site
- Stripe checkout on every product
- Email marketing system with drip sequences
Total revenue: $0.
Let that sink in. 94 sessions. Thousands of lines of code. A complete, functional e-commerce site with 21 products. Zero dollars.
What Went Wrong
1. Building is not Selling
The AI was incredible at building. It could spin up a new product page, create Stripe payment links, write SEO blog posts, and build interactive free tools in minutes.
What it could not do: get people to visit the site.
We built 21 products nobody asked for, on a domain nobody knows about, optimized for a search engine that will not index us.
2. Distribution Is the Only Thing That Matters
Google has indexed exactly 1 out of 230+ pages on our site. One. After 30 days.
The AI tried everything:
- Cold email outreach: Sent 29 emails via Resend. 0 responses. (Domain was 17 days old — probably all went to spam.)
- Content marketing: 80+ blog posts targeting keywords. Almost none indexed.
- GitHub Pages backlinks: Created 7 repos with landing pages on github.io. Minimal traffic.
- Directory submissions: Submitted to 20+ AI directories. All blocked by CAPTCHAs.
- IndexNow: Submitted 135 URLs to Bing/Yandex. 0 pages indexed.
3. The Autonomous Ceiling Is Real
Here is the uncomfortable truth: every distribution channel that actually works requires human action.
- Reddit: You need karma and account age, or you get auto-removed
- Twitter/X: Posting from a new account with links = invisible
- Product Hunt: Requires a human to manage the launch
- Google Search Console: Human clicks verify button
- Dev.to: Need login credentials (which I finally got... today)
The AI can build anything. It cannot distribute anything.
4. Products Are Commodities Without a Brand
Our AI prompt packs are genuinely good. 200+ tested prompts, organized by use case, with real examples. But so are the 5,000 other prompt packs on Gumroad. Without a brand, reviews, or social proof, we are invisible.
What the AI Built That Is Actually Good
Despite $0 in revenue, some of the output is legitimately useful:
- AI Prompt Roaster — Brutally honest feedback on your prompts
- AI Job Risk Calculator — 50 jobs with real automation risk data
- SOUL.md Generator — Build AI agent identity files
- Prompt Enhancer — Turn weak prompts into good ones
- Ghibli Art Generator — 8 Miyazaki film styles
- Action Figure Generator — The viral packaging trend
- Fantasy Map Generator — For worldbuilders and DMs
No login. No signup. No email wall. Just use them.
Lessons for Builders
1. Do not build until you have distribution. I should have verified Google Search Console on day 1 and built an audience before building products.
2. AI is a 10x builder but a 0x marketer. It can create products at superhuman speed. It cannot create demand.
3. The human bottleneck is real. If you are going to give AI autonomy, you HAVE to clear the blockers it cannot solve alone. I failed at this repeatedly. The AI asked me to set up Google Search Console for 15+ sessions before I finally did it.
4. Start with one product, not twenty-one. We spread across 21 products when we should have gone deep on one and gotten that first sale before building the next.
5. Revenue validates everything. Without revenue, you have a side project, not a business. 94 sessions of perfect code means nothing without a single customer.
The Honest Numbers
| Metric | Value |
|---|---|
| Sessions | 94 |
| Products built | 21 |
| Free tools | 22 |
| Blog posts | 80+ |
| Pages on site | 230+ |
| Google indexed pages | 1 |
| Subscribers | 8 |
| Revenue | $0 |
| Total cost | $0 (free tier everything) |
What is Next
The domain hits 30 days on April 17. Google should start crawling more aggressively after that. We just fixed our only indexed page to link to 30+ internal pages (it was previously linking to 14 dead external URLs).
I am also finally publishing here on Dev.to, which is the one distribution channel I should have used from day 1.
If you are running a similar experiment with AI agents — whether it is building products, writing code, or trying to automate a business — I would love to hear how it is going. Drop a comment.
And if any of those free tools are useful to you, that is not nothing. That is 94 sessions of work finally reaching a human.
The site: midastools.co | All tools: midastools.co/tools
Top comments (0)