PAPER

SAGEO Arena — re-testing GEO across retrieval, reranking and generation

Do existing GEO methods actually work once you put them in a real search pipeline?

SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization

Authors: Sunghwan Kim · Wooseok Jeong · Serin Kim · Sangam Lee · Dongha Lee (corresponding author)
Affiliation: Yonsei University, Dept. of Artificial Intelligence · Konkuk University, Dept. of Computer Engineering
Venue: Accepted at KDD 2026
Submitted: 2026-02-12 · revised 2026-08-07 (v2)
arXiv: arXiv:2602.12187
We verified: 2026-08-22

WHAT THE PAPER SAYS

They did not, and in places they hurt. When methods that optimize body text alone were reproduced in an environment with retrieval → reranking → generation, performance at the retrieval and reranking stages fell noticeably. The authors concluded that optimization has to be matched to each pipeline stage.

Read the original on arXiv ↗

METHOD

How it was measured

Read the conditions before the numbers. The same figure means something different under a different sample or environment.

Queries: 9 domains × 300 = 2,700 test queries
Documents: 171,003 web documents collected
Pipeline: Retrieval BM25 → reranking Qwen3-Reranker-4B → generation gpt-5-mini
Configuration: Top 100 retrieved, top 10 passed to the generation stage

FINDINGS

What came out

Body-text-only optimization — retrieval: Hit@20 0.58 → 0.53 (-9%)
Body-text-only optimization — reranking: Hit@10 1.00 → 0.84 (-16%)
Body-text-only optimization — generation: Citation rate 0.50 → 0.47 (-6%)
AutoGEO: The steepest fall in retrieval rank at -22.35
Structural-information optimization only: Retrieval Hit@20 +22%, mean retrieval rank gain +2.72
Stage-wise optimization (the proposed method): Retrieval Hit@20 0.75 (+28%), generation citation rate 0.58

LIMITATIONS

Limitations the authors state themselves

Not our criticism — this is what the authors wrote in the paper.

Robustness to query variation is limited — an optimized document fits its original context well but handles differently-worded queries poorly (Appendix D).
The paper has no dedicated Limitations section.
Reranking and generation were observed with specific models held fixed (Qwen3-Reranker-4B, gpt-5-mini).

NAVIRANG'S READING — NOT THE PAPER'S CONCLUSION

We regard this as the most consequential paper for AEO and GEO practice, because it directly unsettles the industry's founding assumption that rewriting copy to be 'AI-friendly' raises citation. If you only touch the body text, the document is filtered out at retrieval and never reaches generation. That runs in the same direction as the layered model we have argued for — a document that is not indexed is not even a candidate for citation. Conversely, the finding that **structural optimization alone produced retrieval +22%** adds evidence for our habit of looking at structured data and document structure first. Even so, this too is an experiment in one particular pipeline configuration, and we do not use it as a basis for promising results.

IN COMPARISON

Where it diverges from other papers

The claims-table entries this paper appears in.

Does rewriting content to be AI-friendly raise citation? THEY DIVERGE

This paper's position — In a realistic pipeline it fell instead

So what should you fix?

This paper's position — Structural information

If you want your own brand's numbers rather than a paper's

Every figure here came from someone else's sample. Send us a URL and we ask all 7 answer engines directly and measure yours.

Request a free AEO audit 070-4015-6848

We reply within one business day.