A page can become easier for a language model to cite and simultaneously become less likely to reach that language model at all. That is the central warning from SAGEO Arena, a new end-to-end benchmark that tests generative engine optimization across the full search pipeline rather than assuming the optimized document has already been selected for generation.
The SAGEO Arena paper, accepted at KDD 2026 and updated on August 7, evaluates optimization strategies across 2,700 queries and 171,003 unique web documents. Instead of measuring only whether an LLM prefers a rewritten passage once it is placed in context, the framework follows a target page through retrieval, reranking and final answer generation.
That change in experimental design produces a much less comfortable result for simplistic GEO playbooks. Body-text rewrites that are intended to make content more persuasive, fluent or citation-friendly can deliver little benefit at the generation stage while degrading performance earlier in the pipeline. If a rewrite causes the page to fall out of the retrieved or reranked candidate set, its theoretical citability becomes irrelevant: the generator never sees it.
The study also points toward a more integrated view of SEO and GEO. Structural elements including the title, meta description, headings and JSON-LD schema can help preserve or improve visibility in upstream search stages, while the body remains the primary material available for answer generation and citation. The practical objective is therefore not “write for the LLM” in isolation. It is to remain competitive at every gate between the corpus and the generated answer.
Most GEO benchmarks start too late in the pipeline
The authors identify a fundamental limitation in much of the earlier GEO literature. Existing experiments often begin with a predetermined set of candidate documents and ask whether rewriting one of them makes it more likely to be cited in the generated response. That is useful for studying generation behavior, but it abstracts away the search process that determines whether the document becomes a candidate in the first place.
Real search-augmented systems generally contain several stages. A retriever searches a large corpus and produces an initial candidate set. A reranker applies a more computationally expensive relevance model to narrow or reorder those candidates. Only a small number of surviving documents are then passed to a generative model that synthesizes the answer and decides which sources to cite.
A content optimization can affect each stage differently. Adding persuasive language might make a passage attractive to the generator but reduce lexical alignment with the original query. Expanding a page with additional topics can make the prose richer while diluting the signals that helped the retriever identify the page as relevant. A benchmark that injects the optimized document directly into the final context cannot observe either failure.
SAGEO Arena was built specifically to restore those missing stages. The authors call the broader setting Search-Augmented Generative Engine Optimization, or SAGEO, because the optimization target combines traditional search visibility with generative visibility rather than treating them as independent problems.
The benchmark uses 2,700 queries and 171,003 real web documents
The researchers assembled 2,700 queries by sampling 300 queries from each of nine established information-retrieval datasets. The domains include general web search, question answering, multi-hop reasoning, biomedical information, community questions, finance, debate, ecommerce and academic research.
For each query, the researchers collected up to 100 search results through the Google Custom Search API and crawled the resulting pages. After filtering malformed content, the final corpus contained 171,003 unique web documents, with an average of 63 candidate documents associated with each query.
The use of Google Custom Search helped construct a realistic corpus of pages already associated with real queries, but it is important not to confuse corpus construction with evaluation of Google's production ranking system. SAGEO Arena then runs its own controlled retrieval, reranking and generation pipeline over the collected documents. The benchmark is designed for reproducibility and stage-level experimentation, not for reverse-engineering Google's current proprietary search stack.
SAGEO Arena preserves the parts of a web page that text-only benchmarks throw away
Another major design difference is structural information. Many LLM benchmarks reduce a web page to plain body text. SAGEO Arena retains fields that search systems can use separately: document title, meta description, headings and JSON-LD structured data, alongside the page's body passages.
This matters because those fields are not interchangeable. A title provides a compact topical signal. A meta description summarizes the page. Headings expose its information hierarchy. Structured data can make entities and page attributes explicit. Flattening all of them into one block of text erases the distinction between signals that may influence early retrieval and the prose that a generator eventually quotes.
The benchmark therefore lets the researchers ask a more realistic question: if a publisher changes the body, the metadata or both, does the target page maintain visibility at each stage of the pipeline?
Body-only GEO optimization often made retrieval worse
The paper evaluates ten optimization strategies adapted from prior GEO research, including transformations intended to make content more authoritative, fluent, statistically supported or otherwise attractive for generative answers. When these strategies were applied only to the body text, the average effect was not an end-to-end visibility gain.
Cross-checks of the paper's reported results show average top-20 retrieval presence falling from approximately 0.58 to 0.53 under body-only optimization, a decline of about 9%. After reranking, average top-10 presence fell from 1.00 to roughly 0.84, while the final citation rate moved from about 0.50 to 0.47.
The exact magnitude varies by strategy, but the direction is the important finding. Optimizing a page for generation can damage the upstream probability that the page will be available for generation. Some more aggressive automated rewrites produced even larger retrieval-rank losses.
This resolves an apparent contradiction in GEO research. A technique can look successful when evaluated on a fixed candidate set because the experiment asks only whether the generator prefers the modified text. The same technique can fail in an end-to-end system because the modified document no longer survives retrieval or reranking.
Citability is conditional on being retrieved
The easiest way to understand the problem is probabilistically. A page's final citation visibility depends on several conditional events: it must be retrieved from the corpus, survive reranking, enter the model's context and then be selected or cited by the generator.
Improving only the final probability is not enough if an earlier probability falls by more. A rewrite that increases the chance of citation from 50% to 60% after the document reaches the LLM can still reduce total visibility if it cuts the probability of reaching the LLM in half.
This sounds obvious when stated explicitly, yet many practical GEO recommendations implicitly assume retrieval is fixed. They focus on sentence length, quotations, statistics, answer formatting or stylistic authority without testing whether those changes alter search relevance. SAGEO Arena shows why that assumption can produce misleading optimization advice.
Why a rewrite can lose retrieval relevance
The study's pipeline illustrates one mechanism particularly clearly. Lexical retrieval systems reward overlap between query terms and document signals. A rewrite that paraphrases the original language, broadens the topic or replaces specific terms with more polished alternatives can change that overlap even if a human reader considers the new version equally accurate.
Reranking introduces a different objective. A cross-encoder can evaluate semantic relevance more deeply than a lexical retriever, but it still rewards focused alignment with the user's intent. Additional persuasive or tangential material may make a page sound more comprehensive while making the core answer less concentrated.
The generator then has another preference set again. Once a page is in context, well-structured, explicit and self-contained passages may be easier to synthesize or cite. Effective optimization therefore has to respect the different jobs performed by retrieval, reranking and generation instead of assuming one textual style maximizes all three.
Titles, metadata, headings and schema can protect the upstream stages
SAGEO Arena's structural experiments provide the constructive side of the result. Optimizing structural information rather than rewriting only the body improved retrieval performance in the benchmark. Technical summaries cross-checked against the updated paper report the retrieval hit rate rising from about 0.58 to 0.71, with an average retrieval-rank gain of 2.72 positions.
The structural fields include title, meta description, headings and JSON-LD. These elements give the search stages compact relevance signals without requiring the publisher to distort the body text that the generator will eventually use as evidence.
This does not mean adding schema automatically boosts visibility in Google, ChatGPT or every production answer engine. SAGEO Arena is an experimental pipeline with specific retrieval and reranking models. The result is narrower: in this controlled environment, preserving and optimizing web-page structure helped mitigate the retrieval losses associated with generation-focused rewriting.
That is nevertheless a useful principle for real-world GEO work because it aligns with a long-standing SEO lesson. Content does not exist as an undifferentiated text blob. Pages expose meaning through multiple layers, and optimization that coordinates those layers is less likely to sacrifice one stage for another.
The authors propose stage-aware optimization instead of one universal rewrite
The paper's response to the problem is a stage-aware strategy that tailors optimization to the objectives of retrieval, reranking and generation. Rather than applying one transformation and hoping it works everywhere, the method considers what each stage needs from the document.
Reported benchmark results show this approach improving top-20 retrieval presence to roughly 0.75 and final citation rate to around 0.58. Those numbers are specific to the SAGEO Arena setup, but they demonstrate the central idea: end-to-end performance improves when optimization explicitly protects upstream search visibility instead of maximizing generation behavior alone.
For practitioners, the concept is more valuable than the benchmark score. A stage-aware content workflow asks separate questions. Can the page still be discovered for the intended query? Does it remain highly relevant after more sophisticated ranking? Once selected, does it contain clear passages the model can use? Does its structure make the topic and entities unambiguous?
A rewrite should be judged against all of those questions before it is considered a GEO improvement.
The benchmark also challenges the idea that GEO replaces SEO
The study's framing is explicitly joint: search-oriented optimization and generation-centric optimization are two parts of the same visibility problem. That is a useful correction to narratives that treat GEO as a replacement for SEO.
If an AI system retrieves information from a search index or another ranked corpus, traditional discoverability remains upstream of generation. Titles, headings, topical focus and structured information can still determine whether the content enters the candidate set. The generative layer adds another optimization target; it does not erase the earlier ones.
This also explains why publishers should be cautious about agencies selling “LLM-friendly rewrites” as a standalone service. A before-and-after test that places both versions directly in a model prompt measures conditional preference, not end-to-end visibility. Without retrieval and reranking, it cannot reveal whether the optimized version would ever reach that prompt naturally.
Reproducibility is one of SAGEO Arena's main contributions
SAGEO Arena is valuable not because it perfectly reproduces any commercial search engine but because researchers can inspect and rerun the pipeline. The authors provide an experimental environment in which optimization strategies can be evaluated on the same queries and documents across multiple stages.
The paper's robustness analysis also tests alternative components rather than relying on one exact stack. Technical documentation around the work reports experiments across multiple retrieval approaches, rerankers and generators. The goal is to determine whether the stage-level conclusions survive reasonable changes in system architecture.
This is an important direction for GEO research. Production systems such as Google Search, ChatGPT, Gemini and Perplexity change continuously and expose limited information about their internal retrieval stacks. A reproducible benchmark cannot tell marketers exactly how those proprietary systems rank a page tomorrow, but it can test whether an optimization principle survives under controlled conditions.
What SAGEO Arena does not prove
The strongest mistake would be to convert the benchmark into a claim about Google's or OpenAI's live algorithms. SAGEO Arena does not demonstrate that editing a page body will reduce Google rankings by 9%, nor that adding structured data will increase ChatGPT citations by 22%. Those figures belong to the benchmark configuration and its evaluation protocol.
The corpus was assembled from web search results, but the experimental retrieval and reranking stages are not Google's proprietary ranking system. Likewise, using a generative model in the evaluation pipeline does not reproduce the full production retrieval, safety, ranking and citation logic of ChatGPT.
The paper also evaluates a finite collection of optimization strategies across nine benchmark domains. Real websites have additional constraints involving links, freshness, authority, crawlability, JavaScript, localization, personalization and many other signals that the environment does not attempt to reproduce completely.
The correct takeaway is methodological rather than algorithmic: a GEO technique should be evaluated end to end because a gain at generation can be cancelled by a loss before generation.
GEO audits should measure stage failure, not just final mentions
This suggests a more useful diagnostic framework for practitioners. When a page stops appearing in AI answers, the first question should not automatically be “how do we make the prose more citable?” The failure may have happened much earlier.
If the page is not being retrieved, work on topical relevance, discoverability, metadata, structure and the surrounding authority signals. If it is retrieved but loses during reranking, examine whether it answers the intended query directly and competitively. If it reaches the model context but is not cited, then generation-focused improvements such as clearer claims, evidence, concise passages and explicit sourcing become more relevant.
Production AI products rarely expose this full funnel directly, so marketers may need proxies. Traditional rankings, indexation data, referral patterns, citation monitoring and controlled prompt tests can each reveal a different part of the system. The key is not to mistake the last observable stage for the entire pipeline.
Do not sacrifice a page's retrieval identity for LLM-friendly prose
The practical content lesson is conservative. Publishers should preserve the terms, entities and structural signals that make a page unambiguously relevant while improving the clarity of passages a model may eventually use. Optimization should add machine-usable precision without rewriting away the page's retrieval identity.
Titles should remain specific. Meta descriptions should accurately summarize the page. Headings should expose the information architecture. Structured data should describe what is actually present rather than serving as decorative markup. The body should answer the query clearly and provide evidence, but it should not be transformed merely to imitate a supposed “AI writing style.”
Most importantly, every significant rewrite should be monitored for downstream search effects. If a GEO experiment improves citation rate in a controlled prompt but the page loses rankings, impressions or retrieval visibility, the optimization has not succeeded end to end.
The next generation of GEO research has to include the search system
SAGEO Arena's most important contribution is not another optimization tactic. It is a change in what counts as a valid test. A document cannot be considered optimized for a search-augmented generative engine if the experiment begins after search has already selected it.
The benchmark puts retrieval and reranking back into the evaluation loop, using 2,700 queries and more than 171,000 real web documents while retaining structural information that simpler text benchmarks discard. Under those conditions, many body-only GEO rewrites become far less impressive. Small generation-stage benefits can coexist with larger upstream losses, leaving the page less visible overall.
For SEO and GEO teams, that is a useful warning against optimizing the last mile while ignoring the road that leads to it. A page has to be retrievable before it can be reranked, and it has to survive reranking before it can be cited. The best generative optimization is therefore not the rewrite that makes an LLM like a passage most when it is handed the page. It is the change that improves—or at least preserves—the page's probability of surviving the entire journey from search corpus to generated answer.