Nine detection methods.
Each one explained.

Observed bigrams in the page text are compared against a 50-entry English frequency model. High KL-divergence means the word-pair distribution is unusual. That can happen in AI output, domain jargon, or non-native writing.

Measures information density per character. AI-generated text tends to cluster around a narrower entropy band than human writing at the same reading level.

Standard deviation of sentence lengths. Natural writing varies between short sentences and longer explanations. AI output can be unusually consistent.

Unique word ratio (type-token ratio). AI-generated content reuses key terms at higher rates than human writing at equivalent length.

Elevated em-dash density is treated as one weak additive signal. Writing style varies, so this indicator is never used as a verdict by itself.

Checks 39 promotional phrase patterns and 14 filler bigrams that occur frequently in synthetic marketing copy. Phrase matches are treated as supporting evidence, not proof.

AI text frequently targets a specific reading level (grade 9–12) and overshoots on sentence complexity. Grade level outside natural variation is a weak but additive signal.

URL paths containing listicle patterns (/best-X-for-Y/), AI tool names, or affiliate structures are analyzed separately from text. Useful for classifying programmatic content farms.

Lovable, Bolt.new, v0.dev, Framer, Webflow, Wix, and 10 other AI-assisted builder signatures detected from HTML class names, script URLs, and asset paths.