GEO survey: foundational gains hold only in controlled conditions
The field's foundational gains are real inside the lab. Outside it, nine pipeline stages stand between your content and a citation.
Key takeaways
- GEO's foundational gains are valid only when a source is already present in the retrieval context, not in organic conditions.
- Search activation (whether the model retrieves at all) is the most important pipeline stage and the one most GEO research ignores.
- Inconsistent metrics across 45 studies make vendor claims about citation-rate improvements impossible to compare reliably.
- Multilaterals and policy bodies should focus GEO effort on citation and prominence stages, where evidence is strongest.
- Financial services firms lose citation value earlier, at crawling and indexing, making those the priority gates to fix.
The foundational GEO paper's most-cited finding rests on a condition its citations rarely mention: the source must already be present in the retrieval context. Remove that assumption and the gains evaporate. A critical survey published on arXiv, reviewing 45 studies from November 2023 to July 2026, makes this case systematically, and the implication for any organisation investing in generative engine optimisation is harder than the field's boosters have admitted.
The survey's central argument is structural. GEO is not a ranking task with a single lever. It is, as the authors put it, a stochastic, partially observable pipeline spanning at least nine distinct stages: search activation, crawling and indexing, retrieval, reranking and context allocation, citation, prominence, factual absorption, fidelity, and user behaviour. Most published research intervenes at one stage and measures outcomes at one stage. The gains are real within those walls. Outside them, the evidence is thin.
The stage most research skips
Search activation is the first and most consequential gate. A generative engine must decide to trigger a retrieval step at all. If the query is handled from parametric memory, no external source gets cited, regardless of how well-optimised it is. The survey notes this stage is almost entirely absent from the experimental designs of existing GEO studies. The foundational paper, whose "widely cited gains" the survey describes as "valid within its experimental setting but conditional on a source already being present in a fixed context," never had to solve this problem: sources were pre-loaded into the context window.
For a bank publishing a policy brief, a UN agency releasing a technical report, or an industrial group staking out a position on supply-chain standards, this matters directly. Their content competes not just against other sources at retrieval time, but against the model's own parametric knowledge. If ChatGPT or Perplexity already has a confident internal answer, crawling and indexing are irrelevant. Optimisation work that ignores search activation is optimising for conditions that may not obtain.
The terminology problem compounds this. The survey finds that across 45 studies, metrics, vocabulary, and evidence standards are heterogeneous enough to make cross-study comparison unreliable. "Citation rate," "prominence," and "factual absorption" are not consistently defined. A study reporting a 30 percent improvement in citation frequency may be measuring something entirely different from a study using the same phrase. Senior marketers commissioning GEO audits or buying GEO tooling have no stable common currency with which to evaluate the claims vendors make.
Who the findings actually reward
The survey's implicit winner is the organisation that already has strong organic discoverability and institutional authority, because those are the conditions under which the foundational gains are most likely to transfer from controlled experiment to live system. A source that search engines index reliably, that retrieval systems surface consistently, and that models have encountered often enough to treat as credible, is a source positioned to benefit from prompt-level and content-level optimisation. Everything else is optimising the last mile while ignoring the first eight.