Automated checks: how an AI-built website catches its own mistakes
Partner · Web development, SEO and GEO
An automated check is a verification the machine runs before every release of a website: dead links, structured data, page titles, alt text, prices, writing rules. If a check fails, the change stays blocked on the working copy and never reaches your visitors. These checks become essential when AI changes a site, because it works fast and on dozens of pages at once. A mistake spreads as fast as it is written. A check does not replace judgment: it verifies what has a right and a wrong answer, then a person decides the rest.
- An automated check blocks a release as long as a rule is broken: the mistake stays on the working copy.
- According to Google Cloud's 2025 DORA report, AI adoption goes hand in hand with lower release stability. Without solid automated testing and version control, more changes lead to more instability.
- As of October 20, 2026, 28 checks verified every release of falia.co: links, structured data, tags, alt text, prices, writing rules, and the French and English mirror.
- A check verifies that an image has alt text. It cannot tell whether that text describes the image well: judgment stays human.
- On WordPress too, you can require a staging site, a backup before every update and a written list of checks.
On this page
Automated check
An automated check is a verification run by a program, with no human involved, every time a website is prepared for release. It compares the site to a precise rule (every link leads to a page that exists, every page has a title, every price comes from the official price list) and answers yes or no. When the answer is no, publication stops and the check names the page at fault. Developers also call this automated testing or an automatic check.
How an automated check works
A modern website is not changed directly online. The change is made on a working copy. The site is then rebuilt in full: every page is assembled from the text, the templates and the data. That is when the checks run. They read every page as your visitors would receive it, then compare it to the written rules.
The request
Someone on your team, your provider or an AI agent prepares a change: a price, a new page, a text correction.
The working copy
The change is made separately, away from the site your customers see. The live site does not move.
The checks
The site is rebuilt, then every check runs on all the pages involved. A single failure is enough to stop everything.
The approval
When every check passes, a person reviews what the checks cannot judge, then approves the release.
The result of a check is simple. If it passes, the change can continue on its way. If it fails, nothing goes out: the check shows the page and the rule involved, for example "this page links to an address that does not exist." The person or the AI that made the change fixes it, then runs the verification again.
A paper checklist depends on the memory and fatigue of the person following it. A check runs on the 782 addresses in our sitemap (as of October 20, 2026) every time, on Friday at 5 p.m. as on Monday morning.
Why AI makes checks essential
A coding agent such as Claude Code (Anthropic's coding tool) or ChatGPT Codex (OpenAI's coding agent) can change forty pages in a few minutes. That is its strength. It is also the risk: a mistake slipped into a template ends up on every page that uses it, within the same minute. Our article on what Claude Code and ChatGPT Codex change in website development lists checks among the six decisions to make before the first instruction.
Google Cloud's 2025 DORA report measures this risk. It draws on a survey of nearly 5,000 technology professionals. AI adoption goes hand in hand with higher delivery throughput, but it keeps a negative relationship with release stability. The authors explain it this way: without robust control systems (automated testing, mature version control and fast feedback loops), more changes lead to more instability. Their conclusion fits in one sentence: AI doesn't fix a team; it amplifies what's already there.
There is a second, less obvious reason. An AI follows the instructions it is given, but it can forget one. We experienced it on our own site. On September 8, 2026, we added to our French writing rules a ban on the comma before "et" (and). The page rewritten that same afternoon had thirty-five of them, main title included. Since that day, a check reads every French page and blocks publication at the first faulty comma. A rule that is verified nowhere ends up lost on the first new page.
What the checks on our site verify
We built falia.co with Claude Code, on the Astro framework. On the first day of the project, August 20, 2026, the site had four checks. As of October 20, 2026, it had 28. Here, in plain language, is what the main ones verify.
| What is checked | The rule | What it prevents |
|---|---|---|
| Internal links | Every link leads to a page that exists or to a planned redirect. A second check verifies that the destination page matches the link text. | A 404 error served to a visitor, or a link that announces one article and opens another |
| Structured data | Every page carries a single readable block, with the expected page type (article, service, breadcrumb) and internal references that exist | Invalid markup that Google ignores without any message |
| Titles and meta descriptions | A title tag of 60 characters at most (except for an article whose tag repeats its headline), a description of 120 to 162 characters, no duplicates from one page to another | A summary cut off in Google, or two pages competing with each other |
| Outbound links | Every link to another site opens in a new tab | A reader who leaves your site and never comes back |
| Alt text | No image goes out with empty alt text, in the language of the page | An image that says nothing to screen readers and search engines |
| Writing rules | No comma before "et" in French, no anglicism from our list, Canadian English on English pages | A brand rule that gets lost on the first new page |
| Prices | Every price displayed comes from the official price list and follows our display rule. Old amounts cannot come back. | Two different prices for the same service on two pages |
| French and English mirror | The English page has as many sections, headings and questions as the French page, with the same components (form, prices) | An English version that says less than the French one |
| Internal documents | No working document ends up on the published site, even renamed | An internal brand guide readable by anyone, search engines included |
| Forms | Every form has a receiving endpoint that exists | A form that says "thank you" and transmits nothing |
The first rows apply to every website. The others are specific to our brand. Your brand guide can also become a set of checks, just like the technical rules. Your prices, your legal notices (Law 25, Bill 96) and the words your business never uses can be verified at every release.
Checks born from real mistakes
Several of our checks were born from a real mistake. The method is always the same: we fix the error, then we write the check that keeps it from coming back. Four examples, taken from our site's history.
Redirects ignored without any message
On September 29, 2026, we discovered that our host was ignoring the end of a redirect file that had grown too large. In all, 290 old addresses no longer responded. A check now blocks publication before the file reaches the limit.
Sources accessed in the future
On September 25, 2026, articles gave March 2027 as the date some sources were read. The mistake came from the original content. A check now rejects any access date later than the month of the release.
Six titles for a single article
While the site was being built, six cards announced six different articles and all led to the same one. No link was dead: the link check saw nothing. A second check now reads the text of every link and verifies that the destination page keeps its promise.
Structured data nobody was reading
One of our paid advertising pages once carried invalid structured data. Search engines ignored it and nothing flagged it. Since then, the site build fails as soon as a page falls outside the frame.
A site built with AI is therefore not a site without mistakes. It is a site where a mistake that has been spotted becomes a rule verified at every release.
What a check cannot verify
An automated check answers closed questions. It knows whether an image has alt text. It does not know whether that text describes the image well. The W3C's WCAG accessibility guidelines call for a text alternative that serves the equivalent purpose: only a person can judge that it does. The W3C says it about its own evaluation tools: they cannot determine accessibility, they can only assist in doing so. Our article on web accessibility in Quebec sets out the obligations that apply.
Google says the same about structured data. Its Rich Results Test catches most technical errors. Its guidelines nonetheless allow for markup that is incorrect in a way the test was not able to catch. They also point out that valid markup does not guarantee that a rich result will appear. Our article on title tags and structured data explains their role in Google results.
Several things therefore remain a person's job:
- The truth of the facts. A check knows that a source carries a date. It does not know whether the figure cited is accurate.
- The tone. It spots an anglicism from our list. It does not spot a clumsy or cold sentence.
- The strategy. It verifies that a page has a declared keyword. It does not know whether it is the right one.
- The visuals. It does not see that an image is badly cropped or that a button lands in the wrong place on a phone.
Some of our checks only warn. When an important page is missing from the index we prepare for AI tools, the check flags it without blocking: it may be an editorial choice. The comment written in that check sums it up well: it is not up to the machine to decide.
A site that passes all its checks is not an approved site. The checks remove the mistakes that can be verified. A person reviews the rest, then signs off on the release. That is our rule: AI executes, a senior signs off.
Requiring checks from your provider, even on WordPress
A WordPress site can follow the same logic, with other tools. Our guide to AI in WordPress 7 shows that AI agents can now change a WordPress site: the question of checks arises there too. Here are five things to ask your provider for, whatever the platform.
A staging site
A separate copy of the site, where every change and every update is tried before it reaches the live site. WordPress has recognized this type of environment since version 5.5 (2020): a site can declare itself as "staging" rather than "production." Our note on staging sites appearing in Google explains how to keep it out of search results.
A backup before every update
The WordPress documentation recommends backing up the site before updating it. If something goes wrong, the backup lets you return to the previous state.
A written list of checks
Links, forms, structured data, alt text, speed: ask for the list of what is verified before every release. Ask which tool does it, too. The WordPress Site Health screen, Google's Rich Results Test and an accessibility evaluation tool are part of it.
Your brand rules
Your prices, your legal notices and your writing rules can be verified too. Depending on the platform, that takes a custom-built tool or a scheduled human review.
A history of changes
Every change should be dated and reversible. Versioning shows who changed what, then lets you go back. In WordPress, revisions cover page content, not settings or plugins.
A provider who works with AI should be able to answer one simple question without hesitating: what stops an AI mistake from reaching my site? If the only answer is "we review it," your site depends on the attention of a single person on a very busy day.
Six questions to ask your provider
- What is checked automatically before every release of our site?
- Can a change be published if a check fails?
- Where are changes and updates made before they reach the live site?
- Are our brand rules (prices, legal notices, writing style) checked?
- Who reviews what the checks cannot judge?
- How do we go back to yesterday's version if a mistake gets through anyway?
A check does not make AI any smarter. It makes its mistakes visible before your customers see them.
Every change to our websites managed with AI goes through automated checks, then through approval by a senior expert or your team before release. The offer is part of our web design expertise.
Is your site on WordPress? Our website maintenance service applies and monitors updates, takes care of backups and makes the changes you ask for, always with the same contact.
Frequently asked questions about automated website checks
What exactly is an automated check on a website?
It is a small program that verifies the site before every release: dead links, page titles, structured data, alt text, prices. If a rule is broken, publication stops and the check names the page at fault.
Do automated checks replace human review?
No. They verify what has a right and a wrong answer: a link leads somewhere or it does not. Tone, the truth of the facts, strategy and visuals need a person. The machine removes the repetitive mistakes. The person judges the rest.
Why does a site built with AI need checks?
Because AI changes things fast and in volume. A mistake in a template can end up on dozens of pages in a minute. Checks stop that kind of mistake before it reaches your visitors.
Can my WordPress site have automated checks?
Yes. Ask your provider for a staging site, a backup before every update and a written list of the checks run before release. The WordPress Site Health screen and Google's Rich Results Test cover part of the work.
What happens when a check blocks a change?
Nothing is published. The change stays on the working copy and the check shows the page and the rule involved. The person or the AI fixes it, then runs the verification again. Your visitors do not see the faulty version.
Is a site that passes all its checks free of mistakes?
No. It is free of the mistakes the checks know how to spot. Google says so about its own tools: markup can be incorrect in a way its test does not catch. That is why a person reviews and approves before release.
How many checks does a business website need?
There is no ideal number. The basic checks (links, tags, structured data, forms) apply to every site. The others come from your rules and from the mistakes you never want to see again. Our site had four on the first day of the project and 28 as of October 20, 2026.
- Google Cloud, Announcing the 2025 DORA Report: State of AI-assisted Software Development, September 23, 2025.
- DORA, State of AI-assisted Software Development 2025 and DORA AI Capabilities Model, accessed in 2026.
- Google Search Central, Introduction to structured data markup, accessed in 2026.
- Google Search Central, General structured data guidelines, accessed in 2026.
- Google, Rich Results Test, accessed in 2026.
- W3C, Web Accessibility Initiative, Selecting Web Accessibility Evaluation Tools, updated May 13, 2024, accessed in 2026.
- W3C, Understanding Success Criterion 1.1.1: Non-text Content, WCAG 2.2, accessed in 2026.
- Make WordPress Core, New wp_get_environment_type() function in WordPress 5.5, July 24, 2020.
- WordPress Documentation, Updating WordPress and Site Health screen, accessed in 2026.
Geneviève puts the strategy for your engagement into action. She leads all our web development projects: Shopify, WordPress and the new ways of building a site with AI. She manages our team of developers and translates your business needs into technical language. She runs your organic search (SEO), your visibility in AI answers (GEO) and your site's conversion rate optimization (CRO). Her work is at the heart of three goals: Attract customers with SEO and AI, Improve your site's conversion, and Strengthen your visibility in AI answers. With Gabriel, she also builds the landing pages for your advertising campaigns. She writes mainly about SEO, AI visibility and web design.
About Falia →
Strategy
SEO
Paid advertising
Conversion
AI visibility
Web design