What happened when we let an agent rewrite our metadata
A three-week experiment on this very site: same rankings, 32% more clicks from search — and a complete log of every change the agent made along the way.
Metadata is the least glamorous part of a website and one of the highest-leverage. Titles and descriptions are the copy Google actually shows — yet on most sites they were written once, years ago, and never touched again. This site was no different, so we handed the problem to the SEO agent.
The brief was narrow on purpose: rewrite titles and meta descriptions across the blog, one batch per week, and change nothing else. Rankings were stable going in, which made click-through rate the clean thing to measure.
The rules the agent worked under
Nothing on this site ships without a permission check. For this experiment the agent could:
- Draft new titles and descriptions for any published post
- Ship changes to pages ranking below position 3 without review
- Queue changes to top-3 pages for human approval first
- Roll back any change whose CTR dropped after 7 days
What changed
The agent's rewrites had a pattern humans rarely sustain: they were specific. Vague descriptions became claims with numbers; generic titles picked up the exact phrasing people were already searching. It also killed every truncated description — small, boring, and worth clicks.
Same rankings. Same content. 32% more clicks — from copy nobody had looked at in a year.
The part that matters
The result isn't that an AI can write a meta description — of course it can. It's that the whole loop ran without a project: measure, rewrite, ship, watch, roll back. The agent did in three weeks what sits in most teams' backlogs for three quarters.
Every change from this experiment is in the site's audit log, attributed and reversible. That's the standard for everything the team does here — and what AgenticWebTeam runs for client sites.
This article is live in the demo. Open the AWT console and try one of these on it: