The practical brief
The finding
In the paper’s benchmark, optimizing body text alone reduced average visibility at every stage of the tested generative-search pipeline, while optimizing structural information alone improved retrieval and slightly improved citation rate but still struggled at reranking. [c1] [c2] [c3] [c4] [c5] [c6] [c8]
Why it matters
If your team is editing prose to sound more “AI-friendly” without equal attention to titles, headings, meta descriptions, and schema-like structure, you may make pages less likely to reach the answer engine at all. In this study, retrieval and reranking acted as gatekeepers before citation could happen.
| Visibility stage | Baseline avg. | Body-text-only optimization avg. |
|---|---|---|
| Retrieval visibility (H@20) | 0.58 H@20 | 0.53 H@20 (-9% vs. baseline) [c4] |
| Reranking visibility (H@10) | 1.00 H@10 | 0.84 H@10 (-16% vs. baseline) [c4] |
| Generation visibility (citation rate) | 0.50 citation rate | 0.47 citation rate (-6% vs. baseline) [c4] |
What to try
BLURSOR’s practical interpretation
Our interpretation: treat AI-answer optimization as a pipeline problem, not a copywriting-only task. Prioritize cleaner structural fields that mirror key entities and terms already supported by the page, then test body edits cautiously for whether they preserve topical focus and answer prominence.
Study boundary
The results come from the authors’ benchmark, not a live commercial engine. It uses a specific stack and excludes proprietary ranking and user-behavior signals, and the paper reports point estimates without confidence intervals or significance tests.
Before rewriting for AI, ask what could disappear
A practical decision for a business owner is whether to spend the next content sprint rewriting page copy for AI answers. This study suggests that, in a realistic test environment, that decision carries downside risk if it focuses on body copy alone.
The reason is simple: answer engines do not start at the answer. They first retrieve documents, then rerank them, then generate a response. If a page loses position in the first two steps, it may never be available to the model for citation, no matter how polished the prose becomes.
- The benchmark evaluated 2,700 queries across 9 domains on a corpus of 171,003 web documents.
- Its default pipeline used BM25 retrieval, Qwen3-Reranker-4B for reranking, and GPT-5-mini for generation.
- Only the top-10 reranked candidates reached the generator.
What the researchers actually tested
The paper introduces a benchmark called SAGEO Arena for evaluating “search-augmented generative engine optimization,” or optimization across the whole AI-search pipeline rather than only inside a model’s final context window.
Crucially, the authors preserved both body text and structural information from real web pages. Structural information here means fields such as title, meta description, headings, and schema/JSON-LD markup. They then compared what happened when optimization targeted body text only, structural information only, or both.
- Visibility at retrieval was measured with H@20: the share of queries where the target document appeared in the top 20.
- Visibility at reranking was measured with H@10: the share of queries where the target document stayed in the top 10 reranked candidates.
- Visibility at generation was measured as citation rate: the share of queries where the target document was cited in the final response.
The main business takeaway is a gatekeeping problem
Across the default system, body-text-only optimization made results worse on average at all three stages. Retrieval H@20 fell from 0.58 to 0.53, reranking H@10 fell from 1.00 to 0.84, and citation rate fell from 0.50 to 0.47. That means common rewriting strategies often reduced visibility instead of improving it in this benchmark.
Structural-information-only optimization told a different story. Retrieval H@20 rose from 0.58 to 0.71, and citation rate edged up from 0.50 to 0.52, though reranking still dropped from 1.00 to 0.83. So the tested benefit was strongest in getting pages surfaced earlier, not in solving every downstream ranking problem.
The reranking step remained a bottleneck. In the experiments, 5.8% of target documents slipped from reranking rank 10 to 11. Because only the top 10 advanced to generation, that one-position loss removed them completely from possible citation.
- Body-text-only average change: retrieval -9%, reranking -16%, citation -6%.
- Structural-information-only average change: retrieval +22%, reranking -17%, citation +2%.
- Small rank losses can matter more than modest copy gains because of hard cutoffs between stages.
What we think a cautious operator should do
Our interpretation is not that body copy never matters. The paper also indicates that body text and structural fields play different roles: structural information helps pages get surfaced, while body text supplies much of what the generator later cites. But the study warns against assuming that stylistic rewrites alone are a safe path to AI visibility.
A realistic operating approach is to start with page structure before aggressive prose rewrites. Make titles, headings, meta descriptions, and schema-like fields clearly reflect the page’s core entity, claim, and relevant terms already supported by the content. Then treat body edits as controlled experiments, checking whether they keep the main answer early, specific, and tightly aligned to likely query intent.
The authors’ own stage-aware method performed best among their tested combined approaches, reaching 0.75 retrieval H@20 and 0.58 citation rate, but even it did not restore reranking to baseline levels. So the lesson is not “use this exact prompt,” but “optimize by stage and validate end to end.”
- Audit important pages for weak or generic titles, headings, and descriptions before commissioning wholesale rewrites.
- When revising body text, protect common query vocabulary and avoid expanding scope just to sound more expert.
- Judge success across retrieval, ranking, and citation together; a citation-only test can miss where visibility is being lost.
What this study cannot tell you
This is strong benchmark evidence, not live-platform proof. The environment was designed to be realistic, but the authors state that it does not reproduce proprietary commercial ranking signals or user behavior signals.
There are also measurement limits. The paper reports point estimates without confidence intervals or significance tests, and its target documents were sampled from pages that were already visible enough to reach generation at baseline. So you should not read the exact percentages as guaranteed live impact for your site.
- Useful for direction: copy-only AI optimization can backfire in an end-to-end pipeline.
- Not established here: the exact lift or loss you will see in Google, Bing, ChatGPT, or Perplexity.
- Not established here: that any specific optimization prompt is a proven production tactic.
Evidence [c8]
The source and its limits
SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization
Paper-wide extraction of marketer-relevant methods, findings, and limitations from the full body of “SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization,” focusing on end-to-end visibility, structural vs. body-text optimization, stage bottlenecks, robustness, and practical optimization guidance for AI-search discoverability.
- The paper reports point estimates in tables and percentages, but does not provide confidence intervals, standard errors, or significance tests for the main visibility results.
- Main effects are measured in the authors’ benchmark pipeline, not directly on live commercial systems.
- Target documents are sampled only from documents that already reached the generation stage at baseline, which narrows evaluation to initially visible documents.
- The default generator, reranker, and retriever are specific model choices; robustness checks vary components, but the benchmark still cannot capture proprietary production signals or real user behavior.
Evidence behind this briefing
[c1] Method/sample scope: the benchmark evaluates optimization on 2,700 test queries drawn as 300 queries from each of 9 domains, using a final corpus of 171,003 unique web documents. In the default test system, retrieval uses BM25, reranking uses Qwen3-Reranker-4B, and generation uses GPT-5-mini. This defines the denominator and system behind later visibility results.
§3.1 Corpus Construction · Read the study
[c2] Method/sample scope: after crawling and filtering, the final corpus size was 171,003 unique web documents, and each query had 63 candidate documents on average. This matters for marketers because the benchmark evaluates optimization in a large, multi-domain web setting rather than on fixed candidates.
§3.1 Corpus Construction · Read the study
[c3] Method/test system: the default evaluation pipeline used BM25 retrieval, Qwen3-Reranker-4B reranking, and GPT-5-mini generation, with retrieval over top-100 passages and generation from top-10 reranked candidates. This is the comparator system for the main optimization results.
§4 Experimental Setup, Pipeline Configuration · Read the study
[c4] Causal benchmark result: compared with the pre-optimization baseline, body-text-only optimization reduced average visibility across all stages in the default system. In Table 2, the average retrieval H@20 fell from 0.58 to 0.53 (-9%), reranking H@10 fell from 1.00 to 0.84 (-16%), and generation citation rate fell from 0.50 to 0.47 (-6%). No uncertainty intervals are reported.
§5.1 Main Results; Table 2, Avg. row under Body Text only · Read the study
[c5] Causal benchmark result: optimizing structural information alone improved retrieval and slightly improved generation versus the same baseline, while reranking still declined. In Table 2 averages, retrieval H@20 increased from 0.58 to 0.71 (+22%) with +2.72 average retrieval rank gain, generation citation rate increased from 0.50 to 0.52 (+2%), and reranking H@10 decreased from 1.00 to 0.83 (-17%). No uncertainty intervals are reported.
§5.1 Main Results; Table 2, Avg. row under Structural Information only · Read the study
[c6] Causal benchmark result: reranking remained a bottleneck even when optimization helped retrieval. In the experiment, 5.8% of target documents dropped from reranking rank 10 to 11, which is a hard cutoff because only top-10 reranked candidates reach the generator in SAGEO Arena. For marketers, small ranking losses can fully remove AI-answer visibility.
§5.1 Main Results, Reranking as a Persistent Bottleneck · Read the study
[c7] Causal benchmark result: among evaluated ‘Both’ methods in Table 3, the authors’ StageAware approach delivered the strongest overall visibility. Against the baseline of 0.58 retrieval H@20 and 0.50 citation rate, StageAware reached 0.75 retrieval H@20 (+28%), 0.80 reranking H@10, and 0.58 citation rate, with +4.86 retrieval ΔRank and +1.01 generation ΔRank. No uncertainty intervals are reported.
Table 3 · Read the study
[c8] Author-acknowledged limitation: the environment is realistic but does not fully reproduce commercial generative search because proprietary ranking signals and user behavior signals are absent. This limits direct generalization of measured effects to live platforms used by marketers.
Appendix D Limitations · Read the study