Scientific research › Kim et al., SAGEO Arena, 2026

SAGEO Arena: The Study Showing Most GEO Tricks Backfire — and Structure Saves Them

SAGEO Arena (Kim, Jeong, Kim, Lee & Lee — arXiv:2602.12187) built a realistic end-to-end environment for testing GEO and found that many existing optimization approaches are impractical under real conditions and often degrade retrieval and reranking performance — while structural information (the layer most optimizers throw away: titles, headings, markup) helps mitigate those losses. The uncomfortable conclusion: rewriting your body text to “please the AI” can make you less visible, not more.

What the study did differently

Earlier GEO benchmarks evaluated content changes in isolation and often stripped real web documents down to plain text. SAGEO Arena keeps the document whole — structure included — and evaluates visibility end-to-end: retrieval, reranking, and the generated answer. That matters because a real AI search pipeline has stages, and a tactic that helps at one stage can hurt at another.

The three findings that should change how you edit pages

Finding	Practical translation
Many GEO approaches degrade retrieval and reranking under realistic conditions	Aggressive body-text rewriting is a risk, not a free lunch. Test before you scale it.
Structural information mitigates those losses	Titles, heading hierarchy, markup and metadata are optimization surface — not decoration.
Effective optimization must be tailored to each pipeline stage	One uniform “AI-optimize everything” pass is the wrong model. Retrieval, reranking and generation each need their own levers.

Why this converges with the other research

Zhang, He & Yao (2026) found the same shape from the measurement side: Q&A formatting alone does not improve absorption, while modular structure and evidence density do. And Aggarwal et al. (KDD 2024) showed evidence-rich content wins. Put the three together and the doctrine is clear: structure decides whether you are retrievable; evidence decides whether you are absorbed. Skip either layer and the other cannot save you.

What this means for your business

Before touching body copy, fix the structural layer: title, meta, H1–H3 hierarchy, schema markup, llms.txt.
Treat every body-text “AI optimization” as an experiment with a rollback path — the study shows it can backfire.
Audit per pipeline stage: are you failing at retrieval (never fetched), reranking (fetched but buried) or generation (read but not used)? The fix is different each time.

This is precisely why CapstonAI’s tooling starts at the structural layer — the WordPress plugin and Shopify app fix titles, metas, structured data and llms.txt with human approval — and why the platform diagnoses which stage loses you before recommending a fix.

Frequently asked questions

What is SAGEO Arena?

A realistic evaluation environment for search-augmented GEO built by Kim et al. (Yonsei University, arXiv:2602.12187). It tests optimization tactics end-to-end — retrieval, reranking and generation — on documents that keep their real structure.

Does it say GEO does not work?

No. It says naive GEO often backfires under realistic conditions, and that structural information and stage-specific strategies are what make optimization hold up.

What counts as “structural information”?

The document layer beyond body text: titles, meta descriptions, heading hierarchy, markup and structured data — the elements many optimization pipelines discard.

What should I change first on my site?

The structural layer of your highest-value pages: correct heading hierarchy, precise titles and metas, schema markup. Then measure per engine before and after — not just rankings.

Find out which pipeline stage is losing you — free AI visibility audit →