AI Writing Patterns Research Data - Geostar

Why This Data Matters for SEO & GEO

AI is now a primary discovery channel. Algorithms are starting to filter AI-generated content, which gives authentic writing a real ranking advantage.

For GEO, knowing how AI models write helps brands differentiate their content, improve citation rates, and maintain the authenticity signals that both algorithms and readers are looking for.

Model Identification

97%

Accuracy identifying which AI wrote text

Detection Bias

61.3%

Non-native English wrongly flagged as AI

Word Surge

1000%

'underscore' increase (2022→2025)

Structural Tell

25-27

Avg AI sentence length (humans vary widely)

Executive Summary

Detection tools can identify machine-generated text with 97% accuracy by analyzing word choice and sentence structure. ChatGPT, Claude, Gemini, Grok, and DeepSeek each leave distinct fingerprints in the content they produce.

The numbers tell the story: 54% of LinkedIn long-form posts, 20-30% of academic papers published in 2024, and 9.1% of US news articles now include machine-generated text—often without disclosure.

Specific words exploded in frequency after November 2022. "Delve" surged 654% in academic papers—46% of all historical uses appeared in just 15 months. "Underscore" jumped from 3% to 30% prevalence in papers (2022-2025).

Detection bias is a real problem: 61.3% of non-native English speakers get falsely flagged due to similar low-perplexity writing patterns. This creates systemic issues for global content creators and academic authors.

Universal AI Tells

Structural patterns that reveal AI writing (AI vs Human)

Low burstiness (uniform rhythm) vs human variation | Register leveling across all genres

Top AI Signature Words

Most reliable vocabulary tells (% surge since ChatGPT launch)

“delve” + “underscore” co-appeared in 98.8% of flagged papers

AI Content by Domain

Where AI-generated content is most prevalent (2024)

54% of LinkedIn long-form posts | 60-85k papers in 2023

Key Research Findings

LinkedIn AI Content

54%

Long-form posts AI-generated (2024)

Academic Papers

60-85k

Papers with LLM influence in 2023 (1-2% of total)

PubMed Delve

46%

Of all historical 'delve' uses appeared in 15 months

Co-occurrence

98.8%

'delve' + 'underscore' appear together in flagged papers