A researcher named Olivier Martinez spent months tracing the most repeated statistic in GEO, generative engine optimization, the practice of optimizing content for LLM-driven answers, back to its original source. He found that the number does not mean what vendors have been selling it to mean. His review of 45 studies, posted to arXiv on July 15 and reported by PPC Land on July 20, concludes that no technique in the literature delivers a dependable effect, holding across platforms, on either discoverability or the traffic that follows from it. Martinez did not run new experiments. He graded the existing evidence, and the grade for the industry’s headline claim is low.
The number traces to a 2024 paper by Aggarwal and colleagues at the ACM SIGKDD conference, which introduced Position-Adjusted Word Count, a metric that weights how much of a generated answer’s text can be traced to a given source, discounted by where that text lands in the response. Under one optimization strategy, the metric rose from 19.3 to 27.2, a jump of roughly 41 percent in relative terms. That gain was measured in a test setup where five documents were supplied to the generator before measurement even started. The paper never claimed a 40 percent lift in clicks or in the odds a page gets retrieved in the first place. Martinez places the broader “40 percent visibility” claim in his lowest confidence tier.
The more useful finding sits deeper in the paper, and it should worry any team currently paying for body-copy rewrites. In an end-to-end test called SAGEO Arena, run across 171,003 documents and 2,700 queries, rewriting only a page’s body reduced its average presence in the top 20 results by about 9 percent, cut its top-10 share after reranking by roughly 16 percent, and lowered its final citation rate by 6 percent.
The cause becomes clear as soon as you separate retrieval from citation. A rewrite can raise the odds that a document gets cited once a system has already pulled it into context. That same rewrite can lower the odds the document gets pulled into context in the first place. Multiply a higher citation rate by a lower retrieval rate and the combined effect can turn negative. No study that starts with a document preloaded into the model’s context window would ever catch this failure, because the loss already happened before the test began.
Two variables survived the review with real support. How closely a document answers the query is the lever that reproduced most consistently across the 45 studies, and where a passage sits within the retrieved context ranks second. A large factorial study cited in the survey, which ran 252,000 trials spanning six large language models, identified the same two factors as the primary drivers of first citation. Verifiable, dated, and properly attributed evidence carries a moderate edge on top of that, but only when the figures are checkable. Martinez notes that a fabricated statistic can increase how often a passage gets reused even as it degrades the accuracy of the answer. The lesson is not “add numbers.” It is add numbers a reader could verify.
Generic heuristics fared worse across the corpus. A separate benchmark in the survey, C-SEO Bench, evaluated optimization approaches over 54 method-domain pairings and found only three produced a significant gain, none of them in question answering. Keyword stuffing carried over from classic search actively lowered the position-weighted metric in several of the reviewed tests.
For a team currently paying a GEO vendor, this changes the diligence questions. Ask which pipeline stage a reported gain was measured in: retrieval, reranking, or citation after a document was already retrieved. Ask whether the test environment supplied the document to the model before measuring the outcome, because that setup cannot detect the failure SAGEO Arena found. Ask for the retrieval rate alongside the citation rate. A vendor reporting only the second number may be describing a page that performs beautifully once selected and gets selected less often.
PPC Land reported on the survey, authored by Olivier Martinez and posted to arXiv on July 15, 2026, in a story published July 20, 2026.
