AI website check: using an LLM with real scan results

An AI website check works when the model gets real scanner data. How to have AI explain a report, write the fix brief and not fall for hallucinations.

SEO · Published: 5 October 2026 · 6 min read · Марк Лебедев

"Check my site for errors" is one of the most common things people type into a chat assistant, and one of the least useful. A language model cannot measure your LCP, see your response headers or click your checkout button. Pasted a URL, it either fetches one raw HTML file or simply guesses from what it knows about sites like yours. An AI website check becomes genuinely useful only when the work is split properly: a scanner measures, the model explains and drafts the fix, and a second scan verifies.

What a model cannot see on its own

Even with browsing switched on, a chat assistant typically gets one HTML response, without rendering, without the console, without the timing of a real page load. It does not know:

• what the server sends in its headers and redirects;
• which JavaScript errors appear after the page has loaded;
• how long the largest element takes to paint, or how much the layout shifts;
• whether the form actually submits, or what the certificate expiry date is.

Without data, the model fills the gap with plausible text. "Your site may have slow-loading images" is true of almost any site — and tells you nothing.

Split the work: the scanner measures, AI explains

The sensible pipeline has three steps. First a tool collects facts: status codes, headers, Core Web Vitals, console errors, HTML validation messages. Then a model works on those facts: it explains a finding in plain language, groups related ones, and drafts a patch for your stack. Finally the same tool checks the result. The model never becomes the source of truth about your site — the measurement does.

Prompts that work with a report

Give the model the finding, not the whole site. Good requests look like this:

• "Here is a finding from my site audit. Explain in two paragraphs what it means for visitors and for search, without jargon."
• "These are the HTTP response headers of my home page. Write an nginx config block that adds the missing security headers. Do not add HSTS preload."
• "This JavaScript error appears on 14 pages. Here is the stack trace and the file. Suggest the most likely cause and a minimal fix."

The narrower the input, the less room for invention. For console errors in particular, read our guide to console errors first — half the messages there come from third-party scripts you cannot change.

A fix brief for an AI coding assistant

Coding assistants in the editor work best from a structured task: what is wrong, where, how severe, how to tell it is done. That is exactly the shape of a good audit brief. Paste it with a short instruction — "Fix these one at a time, starting from the top; show me the diff for each and explain anything that touches server configuration." — and review every change like a colleague's pull request.

Where AI gets it wrong most often

• Content-Security-Policy. Models happily write a strict policy that blocks your analytics, payment widget and inline scripts. Roll it out as Report-Only first; see CSP.
• HSTS preload. Suggested casually, it is hard to undo. Read HSTS preload before accepting it.
• robots.txt and canonical. One wrong line removes whole sections from search. Never accept these without understanding every directive.
• Outdated advice. FID instead of INP, deprecated headers like X-XSS-Protection, meta keywords "for SEO".

And never paste secrets, tokens or customer data into a prompt — a report and a config fragment are enough.

Trust, but re-scan

Code that looks right is not the same as a fixed site. The only acceptance criterion that matters: a fresh scan no longer reports the problem, and nothing new appeared in exchange. Our article on reading audit results explains how to compare two runs; if you want to know how ready your site is for AI search itself, that is a different topic — see the AI website audit.

How to do it with SiteScan

1. Run a free scan. Pages are loaded in a real headless browser; forms are exercised by AI agents, as described in our piece on form testing.
2. Use the help centre for the meaning of each row before asking a model — it gives the context the model lacks.
3. Click "Copy fix brief". You get plain Markdown: critical, high and medium findings, sorted by severity and reach, with check ids, affected pages and acceptance criteria. It pastes cleanly into a task tracker or an AI assistant.
4. After deploying, scan again and look at "What changed since the last scan": new, fixed, still there.