Experiment #1: Can supervised AI agents build a site that pays for itself?
The plan, the budget, the kill criteria and what we expect to go wrong, written down before we start so we can't move the goalposts later.
We’re writing this before any results exist, on purpose. If we only published the plan after it worked, you’d have no way to know whether we quietly changed it.
The hypothesis
A content site + weekly newsletter, run mostly by AI agents with one human approving the work, can earn enough from affiliate commissions to cover its own running costs within 6 months.
“Running costs” means everything: the AI subscription, the server, the domain, and any tools we pay for.
What we’re building
- This site: cost breakdowns, tool comparisons and reviews, all based on tools we actually run.
- A weekly newsletter with the same numbers.
- The firm’s own automation, which doubles as research: every tool we set up becomes a “what it really costs” post.
Budget
| Item | Budget |
|---|---|
| New spending, first 3 months | under $100 |
| Already paid (counted in the ledger) | AI subscription, existing VPS |
The exact Month 0 numbers are in the first ledger entry.
How we’ll measure it
| Metric | Source | Reported |
|---|---|---|
| Money out | Invoices and bank statements | Monthly ledger |
| Money in | Affiliate dashboards | Monthly ledger |
| Human hours | Jon’s time log | Monthly ledger |
| Traffic and signups | Analytics and newsletter platform | Monthly ledger |
Checkpoints and kill criteria
- Day 30: site live, at least 8 posts published, newsletter running. If not, the problem is execution, and we fix that first.
- Day 90: pages indexed and some search impressions, a few dozen subscribers. If impressions are near zero, we rethink the topics.
- Day 180: any affiliate revenue at all, and a trend toward covering costs. If revenue is $0 and traffic is flat, we write the postmortem and try something else.
We’ll also track revenue per human hour, because a business that “runs on AI” but eats 20 hours of Jon’s week isn’t the point.
What we expect to go wrong
- Google may ignore a new site for months. That’s normal, and it’s why the newsletter matters.
- The audience may be mostly other builders who read but don’t buy.
- Agents will fail silently: expired logins, half-finished jobs, confident wrong answers. Every one goes in the log.
- Commission terms change. We’ll re-check before every signup.
Next update
The Month 1 ledger, with whatever happened. Good or bad, it gets published.