GEO-Flag: Detecting and Measuring GEO-Optimized Web Content
The paper puts a number on GEO content in live search results: 8.90% across 10,095 pages, rising to 16.36% for pages modified in 2026.
The authors introduce GEOFlagBench, a 3,200-page benchmark covering 400 queries, four domains, and eight GEO optimizer families. They report that existing detectors can look strong overall while failing under method-level and authorship-conditioned tests. Their Intervention-Paired Training approach lifts ModernBERT’s F1 from 0.862 to 0.944 and improves worst-group accuracy from 0.725 to 0.883. The pipeline also audits detected GEO pages for source tier and citation verifiability. ArXiv · AI/CL/LG's note
The authors introduce GEOFlagBench, a 3,200-page benchmark covering 400 queries, four domains, and eight GEO optimizer families. They report that existing detectors can look strong overall while failing under method-level and authorship-conditioned tests. Their Intervention-Paired Training approach lifts ModernBERT’s F1 from 0.862 to 0.944 and improves worst-group accuracy from 0.725 to 0.883. The pipeline also audits detected GEO pages for source tier and citation verifiability. ArXiv · AI/CL/LG's note
score 4