AI does the reading.
People do the deciding.
Nothing reaches your site because software decided it should. Here is exactly where a human sits in the loop — the same commitments that ship inside every proposal we send.
Five commitments
Each one is a property of how the platform is built, not a policy we ask you to trust us on.
No write access to your site by default
No CMS logins, no FTP, no deploy keys. Our analysis only reads your public pages — the same pages any visitor loads. In this mode we deliver recommendations and briefs; your team implements them.
One finding becomes one task, and one task becomes one brief
No AI is ever given a broad instruction like “improve this site.” Each issue is scoped to a single, specific change with its own written brief: what to change, why, and how to tell it worked. Small units are reviewable units — that is the entire point.
A person reviews the brief before anything happens
The brief is the gate. It is written to be read by a human first, and it is approved or rejected before any work begins.
We verify by re-measuring, not by asking
After a change, the next scan re-checks the underlying signal. If it did not actually land, the task reopens itself. Nothing is marked done on an AI’s say-so.
Regressions are watched for
Your site is monitored between visits. Uptime, certificates, and a drop in measured score raise an alert and a task automatically.
When you do want us to implement
Advisory-only is the default. Where you would rather we carry out the work than hand it to your team, you grant scoped access that you control and can revoke at any time. Even then, three things hold:
- the change must already have been reviewed and approved as a specific task — authorization to act is never authorization to decide;
- work is scoped to that one approved task, not open-ended access to your site;
- every change records who approved it, who applied it, and when.
On sites we host, AI-drafted content is written to a draft layer only. The live site does not change until a person presses Publish, and any publish can be rolled back.
Why we put this in writing
“We use AI” now describes most of the industry, so it tells you nothing about whether anyone is checking the output. The questions worth asking a vendor are narrower: what access does the AI have, what is the smallest unit of work it is given, who approves that unit, and how does anyone know the change actually landed. This page answers those four for us. Ask them of anyone else you are evaluating.