45 GEO studies from 2023 to 2026: what actually matters for AI search visibility?

Study
Paper	Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023–2026)
Author	Olivier Martinez
Published	15 July 2026, arXiv:2607.14035 (Information Retrieval)
Scope	45 studies published November 2023 – July 2026, plus related RAG and evaluation research
Analysis	Navirang AEO Research

This article is Navirang’s reading of a preprint posted to arXiv in 2026 (not yet peer reviewed). We mark the paper’s claims and our practical interpretation separately, and recommend reading the original.

What this study examined

The survey places the 45 GEO studies that appeared rapidly after the first GEO paper in 2023 side by side and reviews them critically. The author’s starting concern is that terminology, metrics and standards of evidence differ from study to study — the same word “visibility” means citation frequency in one paper and share of the answer in another.

The core conclusions in seven lines:

• GEO is not a single “raise the ranking” problem
• AI search visibility has to be read as a multi-stage pipeline
• The most reproducible factors are topical relevance between question and document and position within the context
• Increasing citation for an already-retrieved document and making a document discoverable in the first place are different problems
• Performance cannot be judged from one ChatGPT test — results differ run to run
• Measurement needs repeated runs, paraphrases, several engines and several dates
• No single technique has yet been demonstrated to raise organic discoverability across platforms over a sustained period

GEO is not one ranking problem

The core model the paper offers is to view the stages a brand passes through before appearing in an AI answer as a probabilistic, only partially observable pipeline. The paper’s stages:

Stage	The question
① Search activation	Does the engine run a web search for this question at all?
② Crawling / indexing	Is our document collected and indexed?
③ Retrieval	Is our document retrieved as a candidate for the question?
④ Reranking / context allocation	Among the candidates, is it actually supplied to the model, and in what position?
⑤ Citation	Among the supplied documents, is it chosen as the answer’s source?
⑥ Prominence	In what position and at what weight does it appear in the answer?
⑦ Absorption	Is the document’s information absorbed into the answer’s content?
⑧ Fidelity	Is the brand and service information expressed accurately?
⑨ User action	Does the answer lead to a visit or an enquiry?

Why this division matters in practice is simple: the prescription differs entirely according to which stage you are stuck at. Polishing citation phrasing for a site blocked at stage ② is the wrong prescription in the wrong order.

Discoverability and citation are different

The most important distinction for a practitioner to take from this paper.

• Discoverability — before AI generates an answer, does our page enter the retrieval and collection candidate set? (pipeline ①–③)
• Citation — among the retrieved material, does the generated answer choose our page as a source? (pipeline ④–⑤)

Making AI cite an already-retrieved document more often, and making that document retrievable in the first place, are different optimization problems. A good deal of GEO advice in the market does not distinguish them, bundling both under the single phrase “AI visibility”. That produces the misunderstanding in the next section.

“GEO raised AI visibility 40%” may be only half right

The roughly 40% visibility improvement reported by the first GEO paper (Aggarwal et al., 2023) is the most widely quoted figure in this field. This survey does not deny it — it attaches the precise conditions.

In the paper’s terms, the improvement is valid within that experimental environment but depends on the condition that documents had already been retrieved and supplied to the model as fixed context. And that result does not demonstrate organic discoverability or sustained traffic effects.

In practitioner language: there is evidence that “for a document already in the candidate set, polishing it can raise the probability of citation”. There is no evidence that “the probability of being discovered across the web rises by 40%.” Citation optimization and discoverability optimization must not be treated as the same thing.

Factors confirmed repeatedly across the studies

The factors the survey found reproducible across 45 papers. The left two columns carry the paper’s terms; the right column carries Navirang’s practical reading.

Factor	Meaning in the research	AEO practical reading
Topical relevance	The most reproducible lever	One page has to answer one question clearly
Context position	Position within the supplied context affects citation	Competition continues after retrieval — the direct answer has to be near the front
Limits of generic heuristics	Generic heuristics transfer poorly when the environment changes	There is no all-purpose GEO checklist
Source competition	Competition erodes an individual document’s gain	One good article does not guarantee continued citation
Run-to-run variability	Substantial variability confirmed in audits of commercial tools	Do not judge from one search screen
Low source overlap between engines	Referenced documents differ by engine for the same question	ChatGPT, Gemini and Perplexity have to be measured separately
Fidelity gaps	Cases where the mention occurs but the information is inaccurate persist	Track accuracy of content, not just whether a mention occurred

Translating the paper’s concepts into our measurement

The paper proposes treating visibility not as a single score but as a vector separating discoverability, citation, absorption and economic outcome. Below is the six-lens version Navirang reconstructed for practical measurement, adding the pipeline’s prominence and fidelity stages to that vector — the paper does not present this form.

1. Discoverability — does our document enter the retrieval candidate set for the question?
2. Citation — is it actually chosen as the answer’s source?
3. Prominence — at what position and weight does it appear in the answer (first-appearance position, recommendation inclusion)?
4. Absorption — do our document’s facts get reflected in the answer’s content?
5. Fidelity — is the company and service information expressed accurately?
6. Business outcome — does it lead to actual visits and enquiries?

How those six lenses get measured as concrete metrics (mention rate, citation rate, recommendation appearance, share) is set out in how to measure AI visibility, and the measurement unit we applied to ourselves is the baseline of 40 questions × 7 answer engines = 280 cells in the eight-week improvement series.

Why one test is not enough to judge by

Users looking for the same information phrase the question differently. “Tell me about AEO specialist firms”, “recommend a Korean AEO agency”, “which agency is good for ChatGPT visibility?”, “tell me about GEO specialists”, “which company is good at answer engine optimization” — all paraphrases of the same intent, and when the phrasing changes, the answer and the brands that appear can change too.

So the measurement protocol this survey proposes has repeated runs, paraphrases, a control group and human verification. As a measurement unit:

Engine × Query × Paraphrase × Run × Date

For instance, 20 core questions in several phrasings each, across several engines, repeated on several dates (the specific counts are a matter of your own resources; the paper does not prescribe numbers). “Always first” is not an achievable goal in a probabilistic system, so the KPI should be a frequency metric such as the proportion of repeated measurements in which the brand was among the candidates — for example “candidate inclusion rate across 100 repeated runs of non-brand AEO queries” (an illustration of measurement design, not an industry-standard metric).

Why citation-friendly prose should not be trusted blindly

The part of this survey most likely to be uncomfortable for practitioners. An observation is reported that content rewritten to be citation-oriented can harm the retrieval stage.

That is, an article AI finds easy to cite and an article a retrieval system finds easy to surface are not always the same. An approach that only polishes citation phrasing looks at one stage of the pipeline. So AEO has to be approached not as content rewriting alone but as a compound problem covering crawlability, topical relevance, retrieval, citation and external trust signals together. It is also why our audit checks indexing and accessibility before content.

Navirang’s conclusion

Navirang does not regard AEO as “how to write sentences AI likes”. We approach it as a search optimization problem that has to be observed stage by stage, from search activation through citation and fidelity, with the same question set measured repeatedly. This paper adds academic weight to that approach — and it is equally a basis for demanding measurement conditions from every claim an agency like ours makes. Please apply the same standard to our figures too.

If you want to know which stage of the pipeline your brand is stuck at, a free AEO audit will tell you.

References

• Martinez, Olivier. “Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023–2026).” arXiv:2607.14035 (2026). DOI: 10.48550/arXiv.2607.14035
• Aggarwal, Pranjal, et al. “GEO: Generative Engine Optimization.” arXiv:2311.09735 (2023) — the first GEO paper, the survey’s starting point