The Critical Survey of 45 GEO Studies (arXiv 2607.14035), Digested

The paper is Optimizing Visibility in Generative Engines: A Critical Survey of GEO (2023-2026), posted to arXiv as 2607.14035, version 1, on 15 July 2026. It reviews 45 published studies on generative engine optimization: the practice of changing content so AI answer engines cite or mention it.

The corpus covers the research most GEO advice traces back to. That includes the original GEO paper, C-SEO Bench, and a 252,000-trial factorial study on what gets cited in competitive answer engines.

The author matters for how you weigh it. The survey is single-author and unaffiliated. It is not peer reviewed. It contains no primary data of its own.

A critical survey reads existing studies and asks what their designs can actually support. It does not run new tests. So the survey's strength depends on two things: which studies it included, and how carefully it judged each one's design.

That makes it useful in a specific way. It can tell you where the published evidence is thin, where studies agree, and where marketing claims outrun the data. It cannot tell you anything no study has measured.

The pattern it surfaces is about where measurements happened. Nearly every published GEO effect size, including the original GEO paper, C-SEO Bench and the 252,000-trial factorial, was measured on documents already placed inside the model's context, with no live search index involved. The original GEO paper, for example, used an engine its authors built with sources supplied in context.

The central sentence, quoted exactly:

"Within this corpus, the evidence is narrow: already-retrieved content can causally alter its citation or use, but no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior."

A second passage names what does replicate:

"Reviewed work indicates that topical relevance and context position are the most reproducible levers, generic heuristics transfer poorly, competition can erode individual gains, and citation-oriented rewrites can impair retrieval."

In plain terms:

• Topical relevance holds up. A page that genuinely fits the question does better once it is in the candidate set.
• Context position holds up, but no publisher can set where their page lands in a model's context window.
• Generic heuristics transfer poorly. A trick that worked in one study tends not to carry to another setup.
• Competition erodes gains. C-SEO Bench found that as more sites adopt the same tactic, overall gains decrease, which its authors describe as "congested and zero-sum".
• Rewrites can backfire. C-SEO Bench reported that "most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking."

The 252,000-trial factorial agrees that topical relevance and list position are the biggest drivers of being cited first. It is Sprinklr-authored and used a synthetic corpus with two candidates per trial, so two interested sources agreeing is corroboration, not replication.

The survey's finding only makes sense if you separate the steps an AI answer goes through. Four are useful:

Almost all GEO experiments live in the last stage. They assume retrieval already succeeded. The survey's point is that a measured effect at the generation stage says nothing about whether your page reaches that stage in a live engine. Our comparison of lab GEO studies and live engines goes through what transfers and what does not.

Be precise about its limits, in both directions.

• It is not proof that content quality is irrelevant. The reviewed nulls are within-context. They do not test whether anything helps a page enter the consideration set.
• It is not new evidence. With no primary data, its conclusions are only as good as the 45 studies it read and its judgement of them.
• It is not peer reviewed. A single unaffiliated author wrote it. Nobody has formally checked its inclusion criteria or readings.
• It does not cover current models. Much of the underlying work uses 2025-era models, and platform behavior changes. Treat its conclusions as a statement about the literature as of July 2026.
• It does not measure access. Server-side rendering, per-bot robots.txt rules and firewall reachability sit outside its scope. Those levers rest on operator documentation, such as OpenAI's crawler page and Vercel's finding that ChatGPT and Claude crawlers fetch JavaScript files but "don't execute them".

If you want the wider context on why results conflict, see why GEO studies disagree.

Use the survey as a filter for every GEO claim you hear.

1. Ask where the effect was measured. If the study placed content directly into the model's context, the result is about citation among already-retrieved pages. It is not evidence about getting found.
2. Discount uplift numbers. The widely repeated "up to 40% more visibility" figure comes from a lab engine with sources pre-loaded. It does not describe live search.
3. Fix access first. These are the levers with a documented mechanism: content in the HTML before JavaScript runs, robots.txt that allows retrieval bots such as OAI-SearchBot, PerplexityBot and Claude-SearchBot, a firewall that does not return 403 to them, and no accidental noindex or nosnippet. Our breakdown of on-page rewrites versus access fixes lays out the evidence side by side.
4. Stay topically relevant. It is the one substantive factor the survey, C-SEO Bench and the factorial agree on. It is not a formatting trick.
5. Measure, then decide. The free checker on this site shows whether AI engines name you for up to three buyer questions. It is a snapshot, not proof of cause.

Nothing here guarantees a citation. What it gives you is a way to spend effort on the gates you control. For the broader picture, read what AI SEO research actually shows.

Does the survey mean GEO is a scam? No. It means the evidence is narrower than the marketing. Rewriting can change how content already in a model's context gets cited, and topical relevance matters. What no reviewed study shows is a stable, cross-platform effect on getting discovered. Advice that promises discovery from rewriting or markup goes beyond what was measured. For a longer answer, see [whether GEO is legit or hype](/questions/is-geo-legit-or-hype).

Why does it matter that the survey is not peer reviewed? Peer review checks whether a paper's inclusion criteria and readings hold up. Without it, you are trusting one unaffiliated author's judgement of 45 studies. That does not make the conclusions wrong, and they line up with C-SEO Bench and the 252,000-trial factorial. It does mean you should treat the survey as a careful summary, not as settled fact or as new data.

What is the difference between citation and discoverability? Discoverability is whether your page gets crawled, indexed and pulled into an AI engine's context for a query. Citation is whether the model uses and credits that page once it is there. Most GEO experiments skip the first step by handing the model the content directly. Their results describe citation behavior, not whether a live engine would ever have found the page.

Which GEO levers does the evidence actually support? Access levers, which work by removing blocks. Put important content in the HTML before JavaScript runs. Allow the retrieval bots you want in robots.txt, remembering GPTBot governs training while OAI-SearchBot governs ChatGPT search. Confirm your CDN or firewall does not block those bots. Check for accidental noindex or nosnippet directives. Beyond that, stay topically relevant. None of this guarantees a citation.