Sitearc

What breaks after an AI-built site launches

Seven failure classes seen in the first month. Each has a detector, a filled check and the fix that closes it.

Answers
Know the failure classes to check for in the first month.
Included checklist
A filled first-month check for one site, class by class
Written for
Solo developers who publish with an agent

The failures are boring

An agent-built site rarely fails because the code is exotic. It fails in the same places every other site fails, one week later.

The classes, with a detector

Class Detector
Metadata written once Every page has a unique title and a written description
Sitemap lists a host you no longer serve Fetch the sitemap, open five URLs
Canonical points at a preview host Read the canonical on three pages
Forms post nowhere readable Submit one response and find where it landed
Analytics installed twice Count the tag in one built HTML file
Images never resized Open the home page on a phone, slowly
A page that only works with JavaScript Read the page source, not the render

A filled first month

Meridian Landscaping, launched 2026-09-18, checked on 2026-10-16.

1. metadata        61 pages, 4 with the same description from the template
                   fix: write four descriptions
2. sitemap         131 urls, all on the live host
                   nothing to fix
3. canonical       three pages checked, all on the live host
                   nothing to fix
4. forms           one test response stored, receipt recorded
                   nothing to fix
5. analytics       tag counted once in the built html
                   nothing to fix
6. images          home page 2.1s on a slow profile, one image 1.8MB
                   fix: resize, then re-measure
7. javascript      every page renders with javascript disabled
                   nothing to fix

total: three fixes, all under an hour

Why they cluster in week one

Launch day proves the site renders. It does not prove anything downstream, because nothing downstream has run yet.

Impressions, enquiries and crawls all start after launch. So failures in those paths appear in the first month, not before.

Acceptance: the check is done when

  • every class has a recorded result, including the passing ones
  • each failure has a fix and an owner
  • the fixes are re-checked after they ship
  • the date of the check is recorded

Troubleshooting

Symptom Likely cause What to do
Every page has the same description The template hardcodes it Fix the template, then re-check
The sitemap points at a preview host The build used an environment value Set the production origin
Two tags on one page A layout tag and a plugin tag Remove one, then count again
A form posts nowhere It was built for email only Move to an endpoint that stores first
The site is slow on a phone Full-size images Resize, then re-measure

Limits

These are the classes seen on sites this team operates. Your site may fail somewhere else, and the detectors still find it.

No detector replaces a backup you have restored once.

Where this came from