AI SEO Research Papers

The paper discusses the development, training, risks, safety evaluations, and mitigations of OpenAI's Deep Research model, focusing on its browsing capabilities, cybersecurity risks, and autonomy.

The paper introduces Agentic Reasoning, a framework that improves LLM reasoning by integrating web search, coding agents, structured memory (Mind Map), and much more.

The paper models ranking manipulation in LLM search as a repeated Prisoner’s Dilemma, analyzing how attack costs, success rates, and defenses shape adversarial behavior and content provider cooperation.

The paper introduces GASLITE, an attack that injects adversarial passages into embedding-based search to manipulate rankings. It outperforms baselines, demonstrating vulnerabilities in dense retrieval models with minimal poisoning effort.

The paper examines how poisoning only 0.1% of pre-training data can make harmful behaviors persist in large language models, even after alignment. It demonstrates attacks affecting model beliefs, prompt security, and safe outputs.

The paper explores how prompt injections manipulate rankings in conversational search engines, demonstrating an attack that boosts product rankings across LLMs and transfers to real-world systems like Perplexity.ai, highlighting fairness concerns.

CONFLICTBANK is a benchmark evaluating knowledge conflicts in LLMs, using 7M claim-evidence pairs and 553k QA pairs to study misinformation, temporal, and semantic conflicts across four model families.

This paper explores how language models judge evidence for controversial questions using the CoNFLICTINGQA dataset, finding they prioritize text relevance over stylistic features humans value.

This paper introduces Preference Manipulation Attacks, showing how crafted content can bias LLMs in search engines and plugins to favor attackers, creating a prisoner's dilemma that degrades results.

The paper introduces Generative Engine Optimization (GEO) to boost content visibility in generative engine responses by up to 40%, using a flexible framework and GEO-BENCH for evaluation.