SEO Autopilot

Crawls a site weekly, fixes what it finds, and pushes the change. Monthly, it does keyword research and schedules a post.

An AI agent allowed to commit needs a blast radius, not a review queue.

The only AI agent in the fleet with write access to something public. Each week it crawls a site I run, runs Lighthouse, reads Search Console, fixes what it finds and pushes, which auto-deploys. Monthly it does keyword research and schedules a new bilingual post.

How it is built

Autonomy is bounded by what it may touch, not by approval. It can edit metadata, headings, alt text, internal links, structured data and page copy. It cannot touch layout, dependencies, or configuration. The guardrail is the file list, which is checkable, rather than a judgement call about whether a change is “safe”.

Every change is logged with a before and after, so a bad week is reversible without reading diffs.

What it got wrong

It pushed changes to files a human had edited by hand. Some files in that project are maintained through a visual editor, and overwriting them silently discarded work. The fix was a pre-flight check that refuses to push if certain paths have changed since the last run, the AI agent now has to prove it is not clobbering someone.

The lesson

“Human in the loop” is usually theatre, nobody reads the hundredth diff. Constraining what an AI agent can reach works, because it keeps working on the day nobody is paying attention.