Blog

GEO field notes

Research and consulting notes on Generative Engine Optimization, citation selection, and reward-model alignment — published from Firestore. New posts rotate across English, 繁體中文, Español, Français, and Deutsch.

No 繁體中文 posts yet — showing all languages until the next rotated publish.

• 2026年7月6日

  The Retrieval Bottleneck in Generative Search

  Retrieval-Augmented Generation (RAG) systems and generative search engines rely on a critical assumption: that the retrieval pipeline can consistently surface the most relevant, context-rich documents to feed into the generator's context window . If the retriever fails to capture the precise information required, the d

• 2026年7月6日

  The Mechanics of Attention Allocation in Context-Rich Prompts

  In Retrieval-Augmented Generation (RAG) architectures, the fidelity of source attribution and factual recall is fundamentally constrained by how retrieved context is structured before injection into the Large Language Model (LLM) context window. While significant engineering effort is directed toward vector database op

• 2026年7月6日

  The Architecture of Positional Bias in Long-Context Language Models

  Long-context Large Language Models (LLMs) suffer from a structural degradation in retrieval accuracy known as the "lost-in-the-middle" (LitM) phenomenon. When processing extensive context windows (ranging from 32k to over 128k tokens), autoregressive transformers exhibit a pronounced U-shaped attention curve. These mod

• 2026年7月6日

  Resolving the Semantic Collision Problem in LLM Retrieval

  Retrieval-Augmented Generation (RAG) systems frequently fail when processing queries containing ambiguous brand names, proprietary product lines, or corporate entities that share lexical tokens with common nouns or other enterprises. In high-dimensional vector spaces, dense embeddings project semantically related conce

• 2026年7月6日

  Mathematical Foundations of Semantic Delta and Shannon Entropy in LLM Retrieval

  The transition from lexical search indexing to generative synthesis requires a fundamental shift in how digital content is engineered. In modern search architectures, retrieval-augmented generation (RAG) pipelines and direct-to-answer generative engines do not merely match keywords; they compress, synthesize, and filte

• 2026年7月5日

  Adversarial RAG Engineering: Defending Generative Search Citations against Semantic Manipulation

  The transition from traditional lexical search engines to Retrieval-Augmented Generation (RAG) systems has fundamentally altered the mechanics of digital visibility. In generative search architectures—such as Google’s Search Generative Experience (SGE), Perplexity AI, and Bing Copilot—the primary objective of Generativ

• 2026年7月5日

  Generative Engine Optimization: Reward-Model Alignment as the Nexus of Content Visibility

  The landscape of digital information retrieval is undergoing its most significant structural transition since the inception of the commercial web. For nearly three decades, Search Engine Optimization (SEO) operated on a retrieval-based paradigm. Search engines functioned as indexers, matching user queries to a ranked l