Generative-engine visibility needs more than a ranking score
A deep-dive on Olivier Martinez's new critical survey of 45 GEO studies, separating retrieval, citation, fidelity, and business outcomes and showing why the popular 40% claim does not generalize.
A widely repeated claim in generative-engine optimization is narrower than its marketing shorthand suggests. In the foundational GEO benchmark, adding quotations raised a position-adjusted word-count score from 19.3 to 27.2, about a 41% relative increase. But the source had already been placed inside a five-document context. That result measures how a retrieved document changes an answer; it does not show that the page becomes easier to discover, earns more clicks, or drives more conversions.1
Olivier Martinez's new critical survey of 45 GEO studies makes that distinction the organizing principle. Its central judgment is simple: generative-engine visibility is a chain of conditional events, and most published experiments observe only one link at a time. The paper was submitted to arXiv on July 15, 2026, in information retrieval, and covers work published from November 2023 through July 2026.2
The claim gets narrower under scrutiny
The original GEO result is not invalid. It used 10,000 queries, supplied the top five Google results to GPT-3.5-turbo, generated five answers per query at temperature 0.7, and compared nine source-rewriting strategies. Quotation addition produced the highest position-adjusted word-count score in the setup summarized by Martinez: 27.2 versus 19.3 for the baseline. Statistics addition reached 25.2; fluency reached 24.7; keyword stuffing fell to 17.7.1
The denominator changes the interpretation. Because every source was already in the context, the experiment held retrieval fixed. The measured question was: once the system can see this document, does a rewrite change how much of the answer is attributed to it? The unmeasured questions were upstream and downstream: will the system retrieve the page in the first place, will it keep the page after reranking, and will a reader click or act on the answer?
Martinez calls the popular "up to 40%" formulation conditional evidence, not a general visibility law. The survey does not argue that document changes are useless. It argues that a citation-share gain should not be promoted into a claim about organic discoverability or business impact without measuring those outcomes directly.1
A pipeline, not a single rank
The survey models GEO as a stochastic, partially observable pipeline:
Activation -> crawling and indexing -> retrieval -> reranking and context allocation -> generation and citation -> absorption and fidelity -> attention, click, and conversion
Each stage answers a different question. A page can be indexed but not retrieved, retrieved but pushed out during reranking, cited but not used in the answer, or used accurately without generating a click. The survey's warning is aimed at the common practice of treating one visible output, usually a citation, as a proxy for the whole chain.1
| Stage | What to measure | What it does not prove |
|---|---|---|
| Activation | Whether search or retrieval was triggered | That a particular source was visible |
| Crawling and indexing | Whether the content was available to the engine | That it entered a candidate set |
| Retrieval | URL or domain presence, recall@k | That the source shaped the answer |
| Reranking and context | Rank, top-k inclusion, allocated tokens | That the source was cited accurately |
| Generation and citation | Mention, citation rate, first-citation position | That the source materially contributed |
| Absorption and fidelity | Claim support, entailment, source contribution | That users noticed or acted |
| Behavior | Clicks, referrals, conversions, revenue | Causality without a control or baseline trend |
This decomposition also explains why two studies can appear to disagree while measuring different estimands. A controlled context experiment can find a real effect on answer composition. A live audit can find no stable change in organic retrieval. Those findings are compatible if the first study conditions on retrieval and the second includes it.1
Seven coordinates for source visibility
To keep those stages separate, Martinez proposes a visibility vector for a source,
V_s = (D_s, K_s, C_s, P_s, H_s, F_s, B_s):D_s: discoverability, or the probability that the source is retrieved.K_s: exposure inside the context, including rank, top-k inclusion, and token allocation.C_s: the probability of a mention or citation.P_s: observable prominence, such as position, repetition, or attributed share.H_s: absorption, meaning the source's contribution to the answer's facts, language, or structure.F_s: fidelity, or whether the answer's claims are actually supported by the cited source.B_s: behavioral or economic outcomes such as clicks, referrals, or conversions.1
The vector is useful because it blocks an easy but misleading shortcut: adding all seven coordinates into a single score. A scalar is defensible only when its weights correspond to an explicit objective. A citation can be accurate but economically irrelevant. A source can be retrieved often but contribute little to the final answer. A page can be prominent while the answer misrepresents it. These are different product and research outcomes, not noisy measurements of one hidden rank.
The survey expands the same point into nine metric families: activation, retrieval, mention, citation, prominence, coverage, absorption, fidelity, and behavior. Coverage matters because an answer may cite many domains while omitting important subquestions. Fidelity matters because a visible citation can still fail to support the sentence beside it. Behavior matters because even a well-supported citation is not evidence of a click.1
Where the methods agree, and where they break
The reviewed evidence is not uniformly negative. Relevance and context position recur as the most reliable determinants of citation. Martinez cites a 2026 study with 252,000 trials across 18 factors and six language models that identifies relevance and position as primary determinants of the first citation. That result has a practical implication: a well-structured page cannot help if the engine never retrieves it, and a retrieved page can still lose when the context budget or reranking stage favors another source.1
The weak point is the fixed recipe. In the survey's account of C-SEO Bench, roughly 1,900 queries and 16,360 documents produced only three significantly positive method-domain combinations out of 54 in the main experiment; none was positive on the question-answering task. In an e-commerce benchmark, 10 of 15 initial heuristics were neutral or negative. The reported pattern is not that optimization never works. It is that generic advice transfers poorly across domains, tasks, and engines.1
One result makes the conditional-versus-total distinction especially concrete. The survey reports that SAGEO Arena, using 171,003 documents and 2,700 queries, found that body-only optimization reduced top-20 presence by about 9%, top-10 presence after reranking by 16%, and final citation by 6%. A rewrite that looks helpful after the document is placed in context can damage the earlier retrieval stages. That is a systems interaction, not merely a weak effect size.1
Measurement itself is unstable. The survey reports daily source-level Jaccard scores around 0.34-0.42 in one audit, with similar overlap for repetitions within 24 hours. Another study found URL-level Jaccard similarities of 0.11-0.18 across a large query set. The survey also notes a more basic failure mode: if search is not activated, an experiment that analyzes only cited answers silently drops the hardest part of the visibility problem. In one configuration, 57.8% of ChatGPT repetitions did not activate web search.1
The practical consequence is that a single run is a poor unit of evidence. Query wording, engine, date, locale, account state, search activation, context order, and competing sources all change the observation. A reported average without its denominator, repetition policy, and failed-search cases can sound precise while describing only a selected slice of the pipeline.
How the field changed from 2023 to 2026
Martinez's timeline shows a field moving through four questions:
- 2023-2024: Can a source change its share of an answer? Early work established controlled interventions and metrics for answer-level visibility.
- 2024-2025: Can retrieved content manipulate a recommendation? The literature began treating the same source-selection channel as an attack surface, separating helpful optimization from adversarial behavior.
- 2025: Do heuristics survive competition and domain changes? Counterevidence, multi-actor interference, and cross-engine variation weakened claims built from one benchmark or one platform.
- 2026: How do we measure the full pipeline? Newer work moves upstream toward retrieval and downstream toward absorption, fidelity, traffic, and governance.1
The sequence matters because each step exposes a variable that the previous step had treated as fixed. First the context was fixed and the answer was measured. Then the retrieval process became part of the question. Then competition and engine variance made the treatment non-isolated. The current problem is not finding a universal rewrite. It is building measurements that survive those extra variables.
The minimum defensible GEO study
The paper's most useful practical contribution is its minimum protocol for research that others can inspect and repeat.
- Define the estimand before collecting results. State whether the target is a conditional effect after retrieval, a total effect across the pipeline, an observational association on a commercial surface, or a production business outcome. Name the primary metric, denominator, exclusion rule, and smallest meaningful difference.
- Vary the conditions that can change retrieval. Use multiple named engines and search modes, multiple intents and domains, three to five paraphrases per information need, closely spaced repetitions, and more than one time window. Keep an untreated baseline and add a length-matched placebo when possible.
- Record activation and failure. A no-search answer is part of the result. Log whether search was triggered before analyzing retrieval or citation. Randomize or counterbalance context order when the study controls the context, and model competing sources when several actors can adopt the intervention.
- Treat the data as clustered. Queries, sources, engines, dates, and repeated runs are not independent observations. Report absolute and relative effects, intervals, and distributions rather than a mean alone. A hierarchical model or clustered bootstrap is more defensible than pretending every answer is an independent trial.
- Preserve the audit trail. Keep prompts, raw responses, citations, search status, timestamps, locale, account type, user agent, and URL-canonicalization rules. Use a stratified human sample to check citation attribution, claim support, factual preservation after rewriting, and errors from automated judges.1
This protocol is demanding because the system is variable. That is the point. A small, carefully defined study can say something useful about a narrow stage. A larger but under-specified study can produce a bigger number without producing a clearer claim.
What builders should ask before buying a GEO claim
The survey gives practitioners a way to interrogate a vendor report, internal dashboard, or content experiment:
- Was the source retrieved organically, or was it injected into the context? These are different experiments.
- Was search activated in every trial? If not, how were no-search outputs counted?
- Which outcome moved? Mention, citation, prominence, absorption, fidelity, click, or conversion?
- How stable was the result across engines, paraphrases, dates, and repetitions?
- What happened to competing sources and upstream retrieval after the rewrite?
- Can the team show raw outputs and verify that the citation supports the generated claim?
The answers determine what the result is worth. A measured citation lift may be valuable for a narrow answer-composition goal. It becomes a much stronger business claim only after retrieval, fidelity, and user behavior are measured separately.
Bottom line
The survey's contribution is not a better list of phrases to add to a page. It is a boundary around what current GEO evidence can honestly support. The strongest finding is that a document already in context can influence an answer's rank, citation, or use. The weaker claims concern durable organic discovery, cross-platform transfer, and downstream revenue, where the reviewed evidence remains unstable or absent.
For researchers, the paper supplies a better estimand and a reproducible measurement checklist. For builders, it supplies a practical warning: do not buy a single visibility score when the system can fail at indexing, retrieval, reranking, attribution, fidelity, or user response. Measure the stage that matters, and report the stages you did not observe.
References
- 1
- 2arXiv abstract page for the survey
arxiv.org
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.