Developer fixes AI content penalty by detecting structural duplicates with n-gram math
A developer running multiple B2B websites noticed that AI-generated comparison articles were indexed normally but ranked poorly, averaging position 76 across five domains despite generating around 50,000 impressions over three months. On closer inspection, hundreds of articles shared identical sentence-level patterns, passing plagiarism checkers as unique while remaining structurally near-identical. Using n-gram analysis and Jaccard similarity scoring, the developer built a detector that measured overlap between sliding word-window sets across article pairs, revealing several hundred paragraph pairs with a perfect similarity score of 1.00. The investigation found that long recurring text blocks dominated each article's structural fingerprint, and simple variant rotation only created a fixed number of distinct skeletons repeated across the entire content batch. The developer subsequently rebuilt the content generation pipeline to ensure every article pair scored measurably distinct under the same mathematical test.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in