The problem was not fixing SEO issues. It was building a system that proves each fix is reversible, asks a human before writing, and confirms the public result independently.
Operating principle
Observe → propose → approve → write → verify.
V1 writes deterministic values. AI-generated SEO copy is a later recommendation layer, not a direct write authority.
A business with a WordPress site needed ongoing SEO health monitoring. Missing meta descriptions, absent Open Graph tags, missing structured data—the fixes are straightforward. The real question was whether a fix actually took effect on the public page.
Most tools tell you what is wrong; some fix it. Very few prove the fix took effect, let you undo it cleanly, and keep reporting when the site is healthy. Silence is the wrong signal for a monitoring product.
I wanted a managed SEO service that behaves like a trustworthy mechanic: explain the issue, wait for approval, use a narrow and reversible path, check the public result, and report even when everything is fine.
The system is a chain of six specialized workflows coordinated by one orchestrator. The orchestrator owns the schedule; the workers are reusable and independently testable.
The audit separates “audit failed” from “zero findings,” so a healthy site remains a valid product state. Every supported write captures a before-value and generates an undo token. The final worker sends reports on every run, including clean ones.
The interesting challenges were decisions about what the system should and should not be allowed to trust.
An Open Graph fix was approved and written successfully. WordPress returned success. The public page still showed the tags missing, so verification marked the fix regressed rather than verified. A CMS confirmation is not proof; the visitor-facing HTML is.
The first orchestrator stopped after an audit with zero issues, sending no report. I changed it so clean runs report “zero issues found” and show the current health state.
Every WordPress write captures the before-value and creates an undo token before the change is applied. The worst case for a fix is “put it back how it was.”
The engine detected missing structured data, emailed a proposal, received approval, wrote through the narrow bridge, waited for cache clearance, and independently verified the public page. The ledger records each state and timestamp.
The engine is currently in supervised pilot with manual trigger runs. Production scheduling, additional sites, and broader check coverage are the next layer—not the current claim.