Rebuilt in 18 days · Part 5 of 6
Launch day is when the work starts: a website with an AI agent on duty
After a launch, most websites are left alone until something breaks. What looking after a site should look like every day, week and month, and why an AI agent can do much of it once the rules are written down and an engineer checks the work.
Most websites get their last careful look on launch day. The team celebrates, the invoice goes out, and the site is left alone until something breaks. By then a few weeks or months have passed and nobody remembers what "working" looked like.
We recently launched a directory and review website, rebuilt in 18 days. On the day it went live, an AI agent went on duty on it, with an engineer who sets its rules and checks its work. This post is about that second part: what looking after a site should involve, and why written rules are what make an agent worth trusting with it.
Why launch day is the start
A site drifts even when nobody does anything wrong. We wrote about how that happens in a website is never finished. A directory and review site drifts faster than most, because what it describes lives somewhere else: the companies it lists change their offers, their pages and their terms whenever they like. A review that was right last month can be wrong today, and the reader who trusted it pays for that.
So the question after a launch is not "is the site done?" It is "who notices when it stops being right, and how fast?"
The routine, written down
The agent works from a written routine. Nothing in it depends on someone remembering.
Every day. When a listed company changes something that touches a reader's money, the change goes into the page that day, with its source. When a reader asks for a correction, it is checked against the source, and if the reader is right, the page is fixed and the correction is recorded on a public corrections page. A build that fails is looked at the same day.
Every week. A few news posts, as the owner's plan asks. The build refuses any news post without a source, so the agent can't publish one even by mistake. One guide written or refreshed, taken from the inventory's "review due" list. A link check across the site, run slowly on purpose so it never looks like an attack. And a watch on security advisories for everything the site runs on.
Every month. The priority pages are read again against the companies' own sites. Each company's details and the dates its sources were last checked are updated.
An inventory that says what to read next
Behind the routine is a content inventory. Every page has a next review date and a status: current, review due, outdated, empty, or second stage (planned for later). The agent doesn't decide what to look at by mood or by what is easiest. It reads the list.
That list is also the answer when the owner asks "how is the site?" Not a feeling: how many pages are current, which are due, which are waiting on facts we can't source yet.
Security: the same day, not the next sprint
While we were building the site, a critical security advisory came out for the framework it runs on. The site was moved off the affected version and the fix was live the same day. Nothing about that was clever. It happened because someone was watching, and because the site's checks could show that the update had broken nothing.
That is the standard after launch too. An advisory is read the day it appears, and the answer is either "updated today" or a written reason why it doesn't apply.
Why written rules make an agent trustworthy
An AI agent left to its own judgment will fill gaps with something plausible. On a review site, a plausible fact that isn't true is the worst thing it can publish. The agent on this site is trusted for the opposite reason: it works inside rules it can't talk its way around.
- Never invent a fact. Every fact on a company comes from that company's own pages, with the source and the date it was checked. A page that needs a fact nobody can source is held, not guessed. How the writing works is in write only what you can source.
- The build enforces what matters most. A news post without a source doesn't publish. That rule doesn't rely on anyone's good intentions.
- Mistakes are public and fixed against the source. A corrections page tells readers the site would rather be right than look right.
- A person sets the rules and checks the work. The engineer reads what the agent did, changes the rules when they prove wrong, and handles anything the rules don't cover.
This is the same principle behind AI content that does not make things up, applied to a whole site instead of a single article. For a smaller business, the weekly version of it is in how to keep your website working every week.
What to expect from whoever looks after your site
Whether it is a person, an agency or an agent with an engineer behind it, you should be able to get clear answers to these:
Ask whoever looks after your site
- Is the routine written down? What happens daily, weekly and monthly?
- Which pages are due for review, and when was each one last checked?
- When a security advisory comes out for what my site runs on, how fast is it handled, and who decides?
- Where is every fact on my site sourced from, and how is a wrong one corrected?
- Are broken links found by a regular check, or by my customers?
- Who reads what an automated tool publishes before or right after it goes out?
If the answers are vague, the site is being looked after by luck. We wrote more on deciding what comes after the build in after the first month. The full story of this launch, from audit to go-live, is in the case study.
This article is drawn from a real engagement. Client details withheld; the figures describe our own work, checked against the project’s records.
Read the full case study →