Sitefire published a method to detect AI-generated commercial web content. It uses structural patterns instead of word-level analysis. The method was tested on 2,250 human blog posts and 11,250 AI mirrors. A 214-feature tool achieved 98.0 macro-F1 on held-out companies. Performance stayed at 98.1 even when AI posts were reworded. The tool attributed 79.3% of AI posts to the correct model. Human posts showed rare structural configurations. The approach replicates earlier fiction-focused findings. Pipeline, instrument, prompts, code and artifacts were released. The paper was revised to correct counts and add an AI disclosure.
What changed
Word-level detectors failed under rewording and could not attribute text to a model.
What it unlocks
Detecting AI-generated commercial web content from structural patterns alone.
- 2,250 human blog posts
- 11,250 AI mirrors
- 214-feature instrument
- 98.0 macro-F1
Sources