Your GEO rewrite can lose the retrieval

Retrieval loss is what happens when a page rewritten to win an AI answer becomes less likely to be pulled into that answer's candidate set in the first place, so the rewrite optimizes a stage the page never reaches.

That failure mode is no longer hypothetical. A benchmark published in February 2026 and presented at KDD this month tested the standard GEO playbook against a full generative search pipeline instead of a pre-selected set of documents, and most of the playbook came out negative.

#The number everyone quotes assumed the page was already there

The tactics that define practical GEO come from one paper: Aggarwal et al.'s GEO: Generative Engine Optimization, which reported that its methods "boost visibility by up to 40% in generative engine responses" and named citation inclusion, quotation addition, and statistics addition as the strongest levers. Three years later those three tactics are the backbone of nearly every GEO checklist in circulation.

The constraint attached to that result is the part that rarely travels with it. The experiment measured what happens to a document that is already inside the engine's context window. It compared rewritten sources against original sources within a fixed candidate set, so it answered a narrow question: given that an engine is already reading your page, how do you get more of it into the answer?

A critical survey of 45 GEO studies published in July 2026 made this the central complaint about the field's evidence base. Its reading is that experiments strongly establish that a document already placed in an engine's context can change how it is cited, establish far less often that the page will be retrieved at all, and almost never establish a durable effect on clicks or conversions. GEO, it argues, is not one ranking task but a pipeline: search activation, crawling, indexing, retrieval, reranking, context allocation, citation, then absorption.

Optimizing one stage of a pipeline tells you nothing about the others unless you measure the others.

#What happens when you measure the whole pipeline

SAGEO Arena is the first environment built to do that. Kim et al. assembled 171,003 web documents across nine domains with their structural markup intact, paired them with 2,700 queries, and ran ten optimization strategies through retrieval, reranking, and generation as separate measured stages. The strategies tested were the familiar ones: Authoritative, Cite Sources, Fluency, Quotation, Easy Language, Statistics, Technical Terms, Unique Words, an All-in-One combination, and the automated AutoGEO method.

Preserving the structural markup matters more than it sounds. Earlier GEO benchmarks stripped schema and metadata out of their documents, which quietly removed the signals real search systems weigh most heavily before generation happens.

The results for body-text rewriting were consistently bad upstream. In the survey's summary of the same experiment, optimizing only the body of a page reduced average top-20 presence by roughly 9 percent and cut top-10 presence after reranking by 16 percent. Individual strategies fared worse: the paper records a 14% hit-rate drop at retrieval for Technical Terms, and AutoGEO, the most aggressive method, lost 36% at retrieval and 42% at reranking. Every strategy tested produced a negative effect at the reranking stage.

The page got better at being quoted and worse at being found. Net of both, it lost.

#Why the stages disagree

Retrieval and generation are not scoring the same thing, and the tactics that flatter one distort the other.

Stage	What it rewards	What GEO rewrites do to it
Retrieval	Lexical and semantic match to the query, structural signals	Padding with quotations, statistics, and jargon dilutes topical focus
Reranking	Relevance density, direct correspondence to intent	Added authority language reads as off-topic filler
Generation	Quotability, attributable claims, specificity	Genuinely improves, which is what the original study measured

A retriever scoring your page against a query does not care that you added an authoritative tone. It sees a document whose text now contains more material unrelated to the query, which means a lower similarity score. Technical Terms is the clearest case: substituting jargon for plain phrasing moves the page away from the vocabulary users actually search with.

#Structural fields were the exception

The same benchmark found one category of change that helped rather than hurt. Optimizing structural information alone, meaning metadata and schema markup rather than prose, improved average retrieval hit rate by 22% and raised average rank by 2.72 positions, according to the SAGEO Arena results. Optimizing body text alone produced no significant retrieval gain at all.

Doing both at once was worse than doing structure alone. The body-text modifications partially cancelled the structural gains.

The authors' stage-aware method, which applies different optimizations at different stages instead of one rewrite everywhere, reached a 28% retrieval hit-rate improvement while holding the rerank change to essentially flat at -0.08, per the same paper. That is the shape of the finding: the same effort, allocated by stage, flips from negative to positive.

#What to change on Monday

Stop treating "make the page more quotable" as a complete strategy. It is one stage of four, and it is the only stage the classic evidence covers.

Audit structure before prose. Schema markup, titles, headings, and metadata are the fields the retrieval layer reads, and they are where this benchmark found real gains. They are also the cheaper class of fix, because correcting them requires no editorial rewrite and no stakeholder sign-off on tone.

Do not bulk-apply quotation and statistics injection across a site. Adding a supporting statistic to a paragraph that needed one is good editing. Adding statistics to hit a density target is the exact intervention that lost 16% of top-10 presence after reranking.

Measure retrieval separately from citation. If your reporting only counts answers you appeared in, a rewrite that halves your retrieval rate and doubles your quote rate looks flat, and you will never see which half moved.

#The honest limits of this

One benchmark is not a verdict. SAGEO Arena is a reconstructed pipeline, not ChatGPT or Perplexity, and its retrievers and rerankers are its own choices. Commercial engines run different models, different index freshness, and different context budgets, and the 45-study survey is equally sceptical of anyone claiming a stable cross-platform effect in either direction.

What the work does establish is narrower and still useful: the standard GEO evidence was collected under an assumption that does not hold in production, and when you remove that assumption the sign of the effect can flip. Before you roll a rewrite across a content library on the strength of the 40% figure, check whether you are optimizing the stage your pages are actually failing at.