A tool, and an argument
Most AI-made websites fail the same five ways.
Not because they're ugly - they're usually clean. Because they're identical: the same safe font, the same indigo gradient, the same centered hero over the same three-icon grid. The tell is sameness. Here's a working tool to catch it - the auditor below scores the cluster, not any single flaw, and it's honest about which "tells" are folklore.
Audit a site
Open the site you're vetting and tick every default it uses untouched. The verdict updates live. The rule is the one the evidence actually supports: 3 or more of the five core defaults, stacked, reads as AI-made.
Supporting signals (they add weight, but don't decide it)
Don't score these — the research refuted them
Analyse the copy
Paste a chunk of the site's writing. This flags the vocabulary and rhythm tells client-side - buzzwords, hedging, cliched openers, em-dash density, and sentence-length uniformity. Treat it as a weak hint: modern models defeat these, and humans use "cutting-edge" too. The strongest content tell isn't a word - it's the absence of proof.
The signal catalog
The full research report as a reference. Confidence is stated per signal; the five core cluster signals are marked.
Make a site stop reading as AI-made
The inverse of the tells, in priority order. You don't have to remove every signal - just de-stack until fewer than three or four remain.