What "AI-Built" Actually Means: The Six Signals We Check
"AI-built" isn't one thing. A page can be generated end-to-end by a hosted builder, hand-coded by a developer using an AI coding agent for assistance, written entirely by a human on a CMS, or somewhere in between. Rather than guess at a single label, our detector scores six independent categories of evidence and shows you exactly what fired — so you can judge the confidence yourself instead of trusting a black-box number.
The six categories, briefly
Builder fingerprints look for the specific CDN domains, class prefixes, and generator tags that design-tool platforms leave in their output. Coding-tool traces look for an explicit claim of attribution from full-stack app generators. LLM API references catch pages whose own client-side code calls an LLM provider directly. Hosting signals note infrastructure common to AI deployment platforms, weighted lightly since plenty of hand-built sites use the same hosts. Human/CMS counter-signals actively lower the score when we see WordPress, Drupal, or similar markers. And boilerplate heuristics flag structural patterns typical of one-shot, unedited AI output.
A false positive we found on our own site
Shortly after launch, we ran our own detector against our own homepage — the page that lists every tool we check for, complete with links to each one's real homepage — and it came back a maximum-confidence result, naming a tool we had obviously never used. The cause was simple in hindsight: several of our coding-tool rules matched on a platform's bare marketing domain rather than an actual claim of attribution. Our own catalog page linked out to those same domains to describe the tools, and the detector read an outbound reference link as if it were a confession.
The fix was to require real attribution language — an actual claim of having been built with a named tool — next to the tool's name, instead of accepting a bare domain anywhere in the page. A domain by itself is not evidence; a page recommending, reviewing, or simply mentioning a tool by name should never score the same as a page that platform actually generated. We added regression tests asserting exactly that scenario returns zero signals, specifically so this doesn't regress quietly later.
Why this matters more than the score itself
A detector that inflates its own confidence on easy, common patterns — like a hyperlink — isn't just wrong occasionally; it's systematically wrong in a predictable direction, which is worse than being occasionally wrong in random ones. We'd rather under-claim on a genuinely AI-built site that hid its tracks well than over-claim on a hand-built site that happened to mention the right words. That trade-off is also why every signal we do fire on shows its exact matched evidence on the results page, rather than asking you to trust a single number.
Read the full breakdown of every signal category on the methodology page, or try the detector yourself.