Yes, if body text is the only thing you optimize. A study from Yonsei University and Konkuk University, accepted at KDD 2026, tested the eight most widely used generative-engine optimization strategies in the literature against a complete search pipeline and measured what happens at each stage. The result: optimizing body text alone degrades visibility across all stages. Average Hit Rate fell 9%, and the most aggressive automated strategy sank it 36%, with a drop of 22.35 positions in retrieval rank.
The reason is that an AI answer is not produced in a single step. Before a model writes a word, a retriever has to pull your page out of millions of documents, and a reranker has to push it into the handful the model will actually read. The tricks that make text sound citable — loading it with statistics, quotations, technical terminology, unusual words — act on the last stage and hurt the first two. A document that is never retrieved is not competing to be cited: it does not exist.
The same study measured what corrects that drop, and that is the finding that matters to anyone who owns a site.
Because the stages before generation do not evaluate the same thing the model does. The retriever and the reranker look for a match with the query; the generator looks for material to build an answer with. When you rewrite text to please the generator, you change the signals the two earlier stages depend on.
The authors tested eight strategies drawn from prior literature — authoritative tone, citing sources, fluency, quotations, easy language, statistics, technical terms and unique words — plus a combination of all of them and a rule-based automated strategy. None escaped the reranking bottleneck: all of them showed degradation at that stage, regardless of optimization scope. In concrete terms, 5.8% of target documents dropped from rank 10 to rank 11 — the line between making it into the set the AI actually reads and not making it.
It does not contradict it: it completes it, and explains why earlier results were so optimistic. Previous benchmarks evaluated optimization strategies over a set of candidate documents chosen in advance — that is, assuming your page had already been retrieved and reranked. With that condition guaranteed, improving text for the generator does improve citation, and that is how the 40%-plus gains circulating since 2024 were measured.
What makes this study different is that it assumes nothing: it measures the full chain, from retrieval over a large corpus to the final answer. And that is where the cost the lab could not see shows up. The citation gain is still real; what was discovered is that reaching the stage where that gain applies is a separate problem, and that optimizing text alone makes it worse.
For anyone who owns a site, the practical reading is simple: the optimization advice in circulation is not false, it is incomplete. It was measured at one stage of the process and interpreted as if it held for the whole process.
The parts of a document a search system reads as a signal separate from the body text. The study lists them: the title, the meta description, the headings and the JSON-LD markup. This is what is commonly called structured markup, and it is exactly what earlier AI-optimization benchmarks discarded when evaluating — which is why they reached conclusions that do not hold outside the lab.
Extending optimization to those components improved retrieval by 22% in Hit Rate and 2.72 positions on average. The authors' explanation is straightforward: structural information is inherently designed to be dense with query-relevant terms, which increases the lexical overlap that retrievers prioritize.
No, and this is the part worth not skipping. The same study analyzed where the citations in the final answer come from and found the opposite: body text is what gets cited.
The authors attribute this to information density: body text provides richer evidence for addressing a user query than the concise, keyword-oriented nature of structural information.
Hence the operational conclusion, and the sentence that organizes the whole topic: structural markup makes you findable; body text makes you citable. They are not interchangeable, and optimizing one at the expense of the other leaves you out at one of the two stages. That is why the authors conclude that effective optimization requires tailoring the work to each stage of the pipeline separately.
In practice, the structural half is the neglected one. In Nostos's own study of 39 Argentine e-commerce sites, markup is the weakest point across the board.
23% of the sites analyzed have no structured markup at all. Not incomplete markup: none. At the retrieval stage, those sites compete without the signal the Yonsei and Konkuk study identifies as the highest-contributing one.
62% scored below 30 out of 100 on content structure, and the overall average was 35/100, with a maximum of 62 and no site in the top tier.
It is the opposite of the paper's scenario: not sites that over-optimized their text, but sites that never touched the half that decides whether they get found.
Work the two halves separately, knowing what each one does:
Get your GEO score across eight academically-backed criteria, including your technical markup and your content structure.
Run the analysis at nostosgeo.app