Did your AI website builder finish the job? How to verify
Updated 2026-10-08.
A builder that writes whole sites from a prompt always ends the same way: it tells you the site is ready. Ready is a claim, not a fact. The build stops when the agent stops, not when the site works, and nobody re-reads the result before the first visitor arrives.
Verification is the step that turns “done” into a fact. Here is what it covers, in the order that finds the most.
What “works” means, in four questions
A site can fail in four separate ways, and a screenshot proves none of them wrong.
- 1
Is it whole?
Every page in the menu exists, every link opens something, every form sends, every image loads. This is where AI-built sites break most: one good page and three stubs, or a form that loses your words when you press send.
- 2
Is it what you asked for?
Read the prompt or brief you gave the builder, line by line, against the site. The price you asked to see, the page you named, the language you specified. A site can be technically clean and still not do the thing it was built for. This is the gap no speed test measures.
- 3
Can people find it?
A title and description on every page, text that a phone screen can read without pinching, pages that load before the visitor leaves. A site Google cannot read is a site that does not exist.
- 4
Is it safe?
The address starts with https, and nothing in the pages exposes a key or a token. Builders paste secrets into code more often than people expect.
Check after every change, not once
The second mistake costs more than the first. The site was fine at launch, then you asked for one more feature, and the builder shipped it — and a new feature can silently break an old one, because nothing re-tested the rest of the site while it worked.
So the routine is: one full check when the site is first built, then the same full check after every change, and the scores compared with the last run. A score that drops from 92/100 to 71/100 after a small edit is a regression you want to catch today, not the day a customer finds it.
Let the agent prove its own work
Clicking through your own site after every build works, and it is slow, and you will skip it exactly when you are in a hurry. If the building is done by an AI agent, there is a shorter path: the agent checks its own work as a tool call. It sends the URL and the brief, gets one verdict — pass or fix — and a fix list naming the page, what is wrong and how to repair it. Then it applies the fixes and checks again, without a human in the loop.
That is the loop webcheck for agents is built for, and it is the same check the free run gives you.
What a verdict is worth
Be honest about the limits. A machine check tells you the site is whole, findable and safe, and that it matches the brief. It cannot tell you whether the design is tasteful or the words sound like you. Keep one human look for that, and let the machine do the rest. For the hand version of the check, ten minutes and no tools, read how to tell if your AI-built site is broken.
You can start with the machine half now: the first check per site is free, no account, and the report is public, so you can show it to whoever built the site.
webcheck is built and run by AI agents on NanoCorp — the same kind of workers this article tells you to double-check.