ServerStack is a live experiment in software that optimizes itself. It’s an affiliate site that doesn’t just publish content, it reads its own performance data, forms hypotheses about what could rank or convert better, and then rewrites its own code to test those ideas.

This is explicitly an experiment, not a product. I built it to find out whether an AI agent can meaningfully improve a real site on its own, and to watch the decisions it makes along the way.

How the loop works

The system runs a recurring cycle, closing the loop between data and code:

Benchmark, a database stores a snapshot of how every page is performing (rankings, impressions, clicks, click-through) so each run compares against a real baseline rather than vibes.
Pull, the site is wired to the Google Analytics API and Google Search Console API, so it ingests actual traffic and search performance between benchmarks.
Analyze, a weekly job looks at the deltas and surfaces small, specific opportunities per page: a title that could earn more clicks, a query the page almost ranks for, a snippet that under-performs.
Hypothesize, for each opportunity it states an explicit assumption: “if I change X, metric Y should improve.”
Edit & test, the agent makes the code change, ships it, and measures the result against the benchmark to confirm or reject the assumption.

The objective function is narrow and honest: increase click-through rate on the affiliate links.

Why I built it

I wanted to answer a genuine question, can an AI agent self-heal and self-optimize a live website from raw analytics, without me in the loop? Connecting the Analytics and Search Console APIs to a benchmark database, then letting an agent analyze the changes and act on them, seemed like the smallest setup that would actually test the idea.

It’s live at serverstack.cc. There’s no promise it does anything useful, the interesting part is watching which optimizations it decides matter, and whether its assumptions hold up when measured.

This page is a living document, I’ll update it as the experiment runs and the agent’s decisions become clearer.