AI website detector: what one can actually prove
An AI website detector answers one of two different questions, and it matters which. The first is how the site was built: which tool generated the code. The second is who wrote the words. The first is checkable, because tools leave traces in what they ship. The second is guesswork, because prose carries no fingerprint.
This page is about the first, and about what you should expect from any tool that claims to answer either.
The two approaches, and what each is worth
Code fingerprints. The detector fetches the page server-side and looks for what the builder left: a generator meta tag, a script host like gpteng.co, a hosting domain, a marker inside the JavaScript bundle. When one of these matches, it names the tool, and you can verify it yourself in the page source. This is evidence, not inference.
Style and text heuristics. The detector reads the copy and the layout and scores how "AI-ish" they feel: stock phrases, three even sections, a certain rhythm. This can be suggestive, but a careful human writes that way too, and a language model asked to judge will give a different answer on different days for the same page. Treat any single percentage produced this way as an opinion.
A tool can combine both. What matters is whether it tells you which one produced the verdict, and shows you the proof.
What a detector cannot know
If someone generated a site with Cursor, Claude or ChatGPT, edited the code, and deployed it from their own repository, there is usually nothing left to find. No tool can honestly call that AI-built from the outside. Any detector that still returns a confident number is scoring style, not evidence.
The same is true for a project exported off its builder: a Bolt or v0 app hosted elsewhere often ships no marker at all. The correct answer there is "no AI builder detected", which is different from "written by a human".
How to check a site yourself, with no tool
Open the page source with Ctrl+U and search, in this order: generator for a meta tag naming the builder; the script and image hosts for a name you do not recognise; the address itself for a builder subdomain; and the main file under /assets/ for the builder name or a key like base44_access_token.
Four searches, two minutes, and no tool needed. Everything a detector worth using does on top of that is speed, coverage of tools you have not heard of, and telling you what the finding means for your site.
What to ask of the result
Which rule matched, where it matched, and what the matched text was. If a tool cannot show you that, you cannot check it, and you are trusting a number. That is why every verdict here comes with the evidence table, and why the whole rule set is published on the method page.
Worth knowing: the most useful thing a scan of an AI-built site finds is usually not the builder. It is an API key shipped to the browser in the JavaScript bundle, which on many generated sites is enough to read the database behind it.
Questions
Can an AI website detector be 100% accurate?
For the question "which tool built this", it can be conclusive when the tool left a marker, because the marker is checkable. For "was this written by AI", no: there is no reliable test, and tools that claim one produce false positives on ordinary writing.
What is the difference between an AI website detector and an AI content detector?
A website detector looks at how the site was built, in the code. A content detector guesses whether the text was generated. They answer different questions and have very different reliability.
Which builders can be detected?
Lovable, v0, Bolt, Replit, Base44 and Durable among the AI builders, and Framer, Webflow, Wix and Squarespace among the visual ones, each with its own markers. The full list, with the rules and their weights, is on the builders pages.
Is it free?
The verdict and the full evidence are free and need no account. The report of what to fix needs a free account.