This AI website detector fetches the live page and weighs more than seventy observable signals across ten groups: the build platform, hosting headers, leftover AI artefacts, markup and code style, accessibility care, metadata quality, imagery sourcing, business trust signals, writing style and page structure. It recognises the fingerprints of AI website builders such as Durable, 10Web, B12, Lovable, Bolt, v0, Butternut, Mixo, Zyro and Jimdo Dolphin, and separates them from visual builders like Wix, Squarespace, Webflow and Elementor and from genuine developer stacks such as Next.js, Laravel, Astro and custom WordPress themes.

The score is a weighted probability estimate from observable evidence, not a certainty. When a site declares an AI builder in its generator tag or ships an AI builder badge, the verdict is close to definitive. When the evidence is purely stylistic — phrasing, typography, layout patterns — it is an informed indication. A carefully edited AI-built site can read as human, and a template-heavy agency site can read as AI. That is why every result shows its confidence figure and the full evidence ledger rather than a bare number.

The scale runs from 0, meaning entirely human, to 100, meaning entirely AI. Below 25 the site reads as hand-built. From 25 to 41 it is likely hand-built. From 42 to 51 the evidence is genuinely inconclusive and the tool says so instead of guessing. From 52 to 64 the site looks human-built with AI-assisted copy or sections. From 65 to 79 it is likely AI-generated, and above 80 the evidence is strong.

Yes, and that case is common. The writing-style group looks for LLM stock phrasing, unusual em-dash density, the "not just X, but Y" construction, rule-of-three list phrasing, unnaturally uniform sentence rhythm and low vocabulary diversity. A hand-coded site with AI-written copy typically lands in the AI-assisted band: human platform signals pull the score down while the language signals push it up.

Code-generating models reach for the same small set of typefaces on almost every brief. Monospace faces such as JetBrains Mono, IBM Plex Mono, Space Mono and Geist Mono applied to body or interface text — rather than to code samples — are a reliable tell, as are stock pairings built from Inter, Manrope, Space Grotesk or Plus Jakarta Sans. Monospace used properly for code blocks is scored as neutral context, and a licensed foundry typeface or a self-hosted webfont counts in the other direction, because someone chose it deliberately.

No. Only publicly reachable pages can be analysed, and the scanner reads the server-rendered HTML rather than executing JavaScript. A single-page application that renders everything client-side will return fewer content signals, so the confidence figure drops accordingly. Internal, local and private network addresses are blocked outright for security reasons.

Google judges content by helpfulness and quality rather than by how it was produced. There is no penalty for using AI as a tool. What does get penalised is thin, unoriginal, mass-produced content built to game rankings — which happens to be what unedited AI output usually looks like. If a scan flags your own site, treat the evidence list as an editing checklist rather than a verdict on your search rankings.

Work down the evidence list. Replace boilerplate section headings with copy specific to the business, cut the stock LLM phrasing, and add first-person detail, real dates and named people. Swap placeholder and stock imagery for real photography, write meaningful alt text, publish working contact details, and fix dead placeholder links. Choose a typeface deliberately instead of keeping the default, and set proper titles, meta descriptions and structured data. Each fix removes a signal and moves the score.