Awesome GEO

Generative engine optimization papers and what they found

By · Updated

Start with three papers on generative engine optimization (GEO): “GEO: Generative Engine Optimization” for the founding idea, “Measuring Google AI Overviews” for a measurement study, and “Critical Survey of GEO (2023–2026)” for a critical review. This page is written by Dmitry (Dee) Kargaev, who maintains Awesome GEO and wrote two of the papers on its list.

The Awesome GEO reading list holds 32 academic papers on generative engine optimization and AI search, checked on 11 November 2026. Of the 32, 17 appear here, chosen because their finding carries a number or a clear result. Groups follow what the papers study, and entries inside a group keep the list’s order. The full list is in the deeflect/awesome-geo repository, where anyone can propose a paper with a pull request.

Benchmarks and methods

Three papers cover how GEO is tested and scored: the foundational benchmark paper and two on what a visibility score captures.

GEO: Generative Engine Optimization

Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, 2024, KDD. The foundational GEO paper introduces the GEO-BENCH benchmark and reports up to 40% visibility gains from adding statistics and citations in its benchmark setting. The evidence is experiments on that benchmark.

Measuring GEO Visibility

  1. The paper argues that the prompt corpus and weights behind a GEO visibility score define an “answer market” that may not match real demand. The evidence is an argument about how such scores are built.

Scoring With the Engine

  1. In a single-author audit of 15 commercial prompts, 589 citations and 4 engines, 84.9% of engine pairs shared no cited URL. The audit also concludes that engine-free page scores estimate something narrower than end-to-end visibility.

How engines choose and cite sources

Six papers study which sources generative engines retrieve and cite, each reporting its own measurements.

Mahe Chen, Xiaoxuan Wang, Kaiwen Chen and Nick Koudas, 2025. Chen and co-authors find that AI search has systematic bias toward earned media over brand-owned content, and that engine-specific strategies are needed.

Measuring Google AI Overviews

Haofei Xu, Umar Iqbal and Jacob M. Montgomery, 2026, IMC. Xu and co-authors report that AI Overviews appeared on 13.7% of 55,393 queries (64.7% of question-form queries); about 30% of cited domains were not in first-page results and 11.0% of claims were unsupported.

What Drives AI Citations?

Mahe Chen and Nick Koudas, 2026. Across 158,099 answers with 468,037 citations, the authors report that the engine’s own retrieval explains far more variation than page properties (32% vs. 2.8%).

Understanding ChatGPT Citations

Yunxuan Jiang, Shiqi Zhang, Anqi Xing, Xiangyu Chang and Xiaokui Xiao, 2026. Jiang and co-authors trace query rewriting, retrieval and citation across ChatGPT 5.2–5.5 settings, and find that pages retrieved more often are cited more often.

LIVE-GEO

  1. LIVE-GEO tests five GEO strategies on live webpages across ChatGPT, Gemini and Grok. In its results, engines agreed on the top product for only 48.8–66.3% of brands.

CITECHOICE

  1. This preprint tests 113 same-call document pairs in agentic search. It finds that structured rendering adds about 0.5 citations per answer to the target without raising total citations, so it shifts credit rather than admitting more sources.

Effects on publishers and the web

Two papers look at clicks to publishers and at AI-generated sources among the citations.

AI in Search Reduces Publisher Referrals

Stephanie T. Wang, Jeffrey Gleason, Yakov Bart, Christo Wilson and Danaé Metaxa, 2026. This preregistered field experiment on Google Search (N=1,100) found that removing AI Overviews and AI Mode raised clicks to publishers. An AI Mode-only experience lowered clicks and user trust.

Synthetic Sources?

Allaham and Diakopoulos, 2026. They audited 712 real queries on politics, health and environment across ChatGPT, Copilot, Gemini and Perplexity. About 16% of the cited sources were AI-generated.

Risks and critiques

Three papers cover manipulation of conversational search, defenses against it and a test of llms.txt-style manifests.

Ranking Manipulation for Conversational Search Engines

Samuel Pfrommer, Yatong Bai, Tanmay Gautam and Somayeh Sojoudi, 2024, EMNLP. Pfrommer and co-authors show how conversational search ranking can be manipulated, which the list calls a useful safety caveat for GEO tactics.

2026, IEEE FLLM. The paper finds no isolated evidence that llms.txt-style manifests improve AI retrieval. It also evaluates their security risks.

Counter-GEO-Bench

2026, EMNLP. This benchmark of defenses against information-distorting GEO reports that off-the-shelf defenses cut attack success by at most 5.7% relative.

Frameworks and syntheses

One paper here reviews 45 studies. The other two are SSRN working papers by this page’s author, and their findings are provisional.

Critical Survey of GEO (2023–2026)

Olivier Martinez, 2026. The survey reviews 45 studies: topical relevance and context position are the most reproducible levers, and the foundational paper’s gains hold only when a source is already in the model’s context. The evidence is a literature review.

Dmitry Kargaev, 2026, SSRN working paper. Kargaev compares traditional SEO and GEO ranking-factor evidence and argues for a reweighting model rather than SEO replacement. It is an exploratory comparative synthesis, and its findings are provisional.

SEEN: A Four-Layer Framework for Generative Engine Optimization

Dmitry Kargaev, 2026, SSRN working paper. Kargaev groups GEO work into four layers: Structure, Evidence, Entity and Notability. The paper is a framework with a validation agenda for future empirical work, and it does not claim that the framework guarantees citations, recommendations, rankings or traffic.

All entries at a glance

EntryYearWhat it shows
GEO: Generative Engine Optimization2024Up to 40% visibility gains in its benchmark setting
Measuring GEO Visibility2026Prompt corpus and weights behind a score define an “answer market”
Scoring With the Engine202684.9% of engine pairs shared no cited URL
How to Dominate AI Search2025Bias toward earned media over brand-owned content
Measuring Google AI Overviews2026AI Overviews on 13.7% of 55,393 queries
What Drives AI Citations?2026Engine retrieval explains 32% of variation, page properties 2.8%
Understanding ChatGPT Citations2026Pages retrieved more often are cited more often
LIVE-GEO2026Engines agreed on the top product for 48.8–66.3% of brands
CITECHOICE2026Structured rendering adds about 0.5 citations per answer to the target
AI in Search Reduces Publisher Referrals2026Removing AI Overviews and AI Mode raised clicks to publishers
Synthetic Sources?2026About 16% of cited sources were AI-generated
Ranking Manipulation for Conversational Search Engines2024Conversational search ranking can be manipulated
Manifest-Assisted Retrieval in AI Search2026No isolated evidence that llms.txt-style manifests improve retrieval
Counter-GEO-Bench2026Off-the-shelf defenses cut attack success by at most 5.7% relative
Critical Survey of GEO (2023–2026)2026Topical relevance and context position are the most reproducible levers
The SEO-to-GEO Gap2026Provisional: a reweighting model rather than SEO replacement
SEEN: A Four-Layer Framework for GEO2026Four layers: Structure, Evidence, Entity and Notability