The evidence behind generative engine optimisation — the fast-growing practice of tuning web content so that AI assistants cite it — is far thinner than the sales pitch around it suggests. A critical survey of 45 GEO studies, posted to the preprint server arXiv on 15 July by the researcher Olivier Martinez, concludes that no reviewed technique shows a stable, cross-platform effect on whether a page is discovered organically or earns lasting traffic from AI-generated answers.

The timing matters. B2B marketing teams are increasingly pitched GEO services, and much of that pitch rests on a single headline figure: that the right optimisation can raise visibility by up to 40 per cent. The survey does not claim GEO never works. It finds that the version being sold — reliable, durable, engine-agnostic visibility gains — mostly is not supported by the evidence, and that some of the most common rewrites can make a page harder for an AI to retrieve, not easier.

What does the survey actually conclude?

Martinez reviews 45 studies published between November 2023 and July 2026 and models GEO not as a single ranking task but as a multi-stage pipeline: search activation, crawling, indexing, retrieval, reranking, context allocation, citation, prominence, factual absorption and, finally, user behaviour. The central point is that an intervention can improve one stage while quietly harming another, so a gain measured at one point in the chain says little about the outcome a marketer actually wants.

Across that corpus, one thing is reasonably well supported: a document that has already been retrieved and placed in a model’s context can causally change whether and how it is cited. What the evidence does not establish is the step before it — that optimisation reliably gets a page retrieved organically in the first place — or the step after it, that any gain persists across different engines and over time. In the survey’s own terms, no reviewed technique shows “a stable, longitudinal, cross-platform causal effect” on organic discoverability or downstream behaviour.

Two caveats sit around the finding. It is a preprint and a critical scoping review by a single author, not a peer-reviewed systematic review, and the survey says so plainly. And the field it maps is young: many of the 45 studies are themselves preprints, using different metrics and definitions, on systems that change from one month to the next.

Where does the ‘40% visibility’ figure come from?

The 40 per cent claim traces to the foundational 2024 paper that coined the term GEO, by Aggarwal and colleagues, which reported that GEO-specific strategies could boost visibility by up to 40 per cent in generative-engine responses. The survey does not dispute the arithmetic; it reframes what the number measures. That result came from a closed test in which a source was already supplied to the model, and “visibility” meant the position-weighted share of citations that source received within a fixed context — not whether a page is found in the first place, and not clicks or traffic.

Pulled out of that setting, the survey argues, the figure cannot carry the weight the market puts on it: it establishes neither organic discoverability nor durable traffic effects. The paper also documents how uneven the underlying effects are. Under one “cite sources” strategy in the original work, the fifth-ranked source gained about 115 per cent in prominence while the top-ranked source lost roughly 30 per cent — the same tactic helping one page and hurting another, depending on where it started.

More striking for anyone about to rewrite a page is the evidence that optimisation can backfire. One large study cited in the survey (Kim and colleagues, 2026), run across 171,003 documents and 2,700 queries, found that optimising the body text alone reduced a page’s presence in the top 10 after reranking by about 16 per cent, its top-20 presence by around 9 per cent, and its final citation rate by 6 per cent. Tuning the words on the page, in other words, pushed those pages down the shortlist the model draws from.

What should marketing teams take from it?

The survey is clearest on which levers replicate and which do not. Topical relevance and a page’s position in the retrieval context are the most reproducible advantages; generic heuristics — stuffing in quotations, converting every page into a question-and-answer format — transfer poorly between engines, and any edge erodes as more competitors adopt the same tactic. The practical implication is unglamorous but steadying: credibility, technically fetchable pages, clear structure and genuinely relevant content do more reliable work than any proprietary “GEO formula”.

Measurement is the other warning. The survey highlights how easily an AI-referral uplift can be misread: in one analysed case, referrals from ChatGPT to a site rose 5.7-fold after an optimisation, but untreated pages on the same site had already risen 3.5-fold as the platform itself grew. A controlled estimate put the genuine additional lift at 1.82 times, and a placebo test left even that short of statistical significance — “suggestive rather than causal”, as the paper puts it. The survey’s recommended alternative is procedural rather than magical: repeated measurements, paraphrased queries, proper controls, human validation and an eye on what rival optimisers are doing at the same time.

The survey is a preprint, has not yet been peer-reviewed, and no GEO vendor has publicly responded to it. Its most useful legacy for a marketing manager may be a single question to put to any AI-visibility pitch: which stage of the pipeline does the promised number describe, and does it survive more than one run.