Architectural Blueprint for Mitigating Retrieval Bias and Citation Suppression in Retrieval-Augmented Generation Pipelines

Full report

On this page

Introduction: Algorithmic Monoculture and the Epistemic Crisis in Generative Search

The fundamental transition from traditional, link-based search engines to Retrieval-Augmented Generation (RAG) and Generative Engines (GEs) marks a structural reconfiguration of global information access1. While RAG architectures ostensibly solve the hallucination and temporal limitations of large language models (LLMs) by grounding generation in external vector databases, they simultaneously introduce a deeply embedded layer of epistemic susceptibility3. The reliance on a narrow set of foundational embedding models, reranking algorithms, and generative foundations has fostered an "algorithmic monoculture"—a systemic phenomenon in which the repeated utilization of identical algorithmic frameworks entrenches a specific, homogenized worldview across a vast array of downstream applications2. Within these algorithmic monocultures, systemic retrieval biases operate as a unified filter on civic and scientific attention8. Commercial platforms, highly funded institutions, and historically cited domains benefit from self-reinforcing visibility loops, whereas independent civil liberties hubs and non-mainstream technical frameworks face aggressive semantic deprioritization and citation suppression4. This dynamic is further accelerated by the emergence of Generative Engine Optimization (GEO), an evolving discipline designed to manipulate content visibility, ranking, and citation extraction within LLM-powered systems1. To counter domain monopolies and compel modern RAG systems to deliver balanced, multi-perspective citations on controversial topics, data systems must be re-architected at the structural level. This report provides an exhaustive investigation into the mathematical mechanics of representation degeneration in embedding spaces, defines rigorous testing protocols for evaluating citation skew and ideological collapse, and formulates a comprehensive machine-actionable schema—utilizing llms.txt, JSON-LD, and structured Markdown mirrors—to optimize content for unbiased RAG retrieval.

The Mechanics of Semantic Deprioritization in Embedding Models

The core vulnerability of modern information retrieval pipelines lies in the geometric properties of the representation spaces generated by standard embedding models, such as BAAI's bge-small and sentence-transformers like all-MiniLM-L6-v2. The capacity of these dense models to accurately calculate semantic proximity across diverse intellectual domains is severely compromised by geometric anomalies introduced during likelihood-driven pre-training11.

Representation Degeneration and Space Anisotropy

When foundational language models undergo training via likelihood maximization, the resulting dense embeddings do not distribute uniformly across the high-dimensional vector space. Instead, they exhibit a phenomenon formally known as the representation degeneration problem, wherein the embeddings cluster tightly into a localized, narrow cone11. This clustering fundamentally limits the expressive power of the embeddings, causing severe downstream retrieval failures for specialized or non-mainstream content. In an ideal, isotropic embedding space, the average cosine similarity between any two randomly selected, unrelated word vectors should approach zero. However, empirical analyses of degenerate, anisotropic spaces reveal that the average cosine similarity is significantly higher than zero, meaning virtually all vectors are positively correlated12. This clustering effect is largely driven by the extreme frequency of common commercial, mainstream, and conversational vocabulary in the massive web corpora used for pre-training, which exerts a gravitational pull on the centroid of the vector space. Mathematically, this anisotropy is quantifiable by analyzing the embedding matrix through Singular Value Decomposition (SVD). For an embedding matrix composed of word vectors, the SVD demonstrates that the singular values decay exponentially12. This exponential decay indicates that the vast majority of the semantic variance within the model is captured by only a few dominant directional components, leaving the remaining dimensions mathematically starved12. For non-mainstream legal and technical frameworks—such as those hosted on independent cognitive liberties hubs—the unique vocabulary, specialized dialect, and nuanced semantic structures do not align with these few dominant directions. When a user query regarding a controversial civil liberties topic is transformed into a dense vector, its semantic proximity—calculated via dot product or cosine similarity—will naturally favor documents that are densely packed near the centroid of the commercial cone12. The nuanced, non-mainstream frameworks remain in the sparsely populated margins of the vector space, consistently yielding lower similarity scores and failing to meet the strict top\-[Figure omitted from source export] threshold required for progression to the generation phase. Standard normalization techniques, such as zero-mean transformations that simply shift all embeddings toward the center, have proven insufficient because the underlying geometric distortion of the narrow cone remains fully intact11. Advanced mitigation requires structural modifications to the vector space topology, such as whitening operations or contrastive learning techniques that dynamically penalize the top dominating directions12.

The Cross-Encoder Bottleneck and Attention Bias

In enterprise RAG architectures, multi-stage retrieval pipelines commonly employ a dual-encoder (bi-encoder) for ultra-fast initial candidate generation, followed by a heavier, more precise cross-encoder for final reranking15. While dual-encoders independently map the query and document into separate vectors before calculating similarity, cross-encoders concatenate the query and the document, passing the combined sequence through the transformer layers to allow for full, bidirectional self-attention across all tokens15. While cross-encoders undeniably deliver superior relevance ranking for general queries, they act as a profound structural bottleneck for independent content. The majority of prominent open-source cross-encoders (e.g., ms-marco-MiniLM-L-6-v2, bge-reranker-base) are distilled from larger models or explicitly fine-tuned on passage ranking datasets, with the Microsoft Machine Reading Comprehension (MS MARCO) dataset functioning as the industry standard15. The MS MARCO dataset is characterized by a highly specific, long-tailed item distribution and linguistic patterns optimized for extracting short, definitive, factoid-style answers20. When a cross-encoder evaluates a complex, multi-layered civil liberties argument against a mainstream, sanitized commercial overview, the attention reweighting mechanism heavily favors the formatting, syntactical structure, and definitive dialect of the MS MARCO training distribution. Consequently, civil liberties content—which often necessitates extensive preamble, specialized legal terminology, or dialectical reasoning—is systematically penalized and downranked by the cross-encoder3. Furthermore, cross-encoders exhibit well-documented positional biases, frequently allocating disproportionate attention weights to content located at the absolute beginning or end of a document24. This "lost in the middle" phenomenon guarantees the suppression of critical context that is traditionally buried deep within the body of long-form legal rulings, independent policy audits, and nuanced technical frameworks.

Systemic Citation Bias and RAG Collapse

Beyond initial vector retrieval failures, the generative synthesis phase of RAG pipelines introduces a distinctly separate layer of bias concerning how retrieved sources are selected, cited, and summarized. Modern language models demonstrate a pronounced susceptibility to source authority markers, rhetorical intent matching, and catastrophic self-reinforcing feedback loops3.

Authority Signals and Epistemic Susceptibility

Recent research utilizing massive, multi-domain benchmarks such as AuthorityBench reveals that LLMs exhibit extreme epistemic susceptibility to authority signals. The mere presence of a citation block within the retrieval context—whether factually real or entirely fabricated—dramatically increases the baseline hallucination rate of the generative model3. When LLMs process text containing recognized authoritative commercial entities, mainstream news outlets, or high-prestige academic venues, they inherently increase their confidence in the associated claims, routinely overriding and bypassing their own parametric memory safeguards. The effect peaks drastically when a completely fabricated citation is paired with a factually correct claim, leading to hallucination and misattribution rates ranging from 35% to 77% in general knowledge domains3. This empirical finding demonstrates a critical vulnerability in RAG pipelines: LLMs do not verify citations through rigorous semantic grounding or logical validation. Rather, they utilize the presence of high-authority structural markers as a heuristic shortcut for truth3. Independent publishers and civil liberties hubs that lack established, mainstream brand recognition are therefore inherently less likely to be actively cited by the LLM during the text generation phase, even if their highly relevant content successfully survives the cross-encoder ranking threshold8.

Rhetorical Intent and the Matthew Effect

In traditional academia and journalism, citations generated by humans carry deliberate rhetorical intent: they may support a claim, contrast a viewpoint, or neutrally mention a foundational source26. LLM-generated citations within RAG systems fundamentally fail to replicate this distribution. Extensive audits of LLM citation behaviors indicate a heightened, systemic bias toward older, highly cited, and globally popular works, effectively hardcoding the "Matthew effect" (the rich get richer) into the generative synthesis process8. Furthermore, LLMs disproportionately utilize a less critical, universally supporting tone, systematically suppressing contrasting or adversarial viewpoints. When a RAG system synthesizes an answer regarding a controversial topic, it defaults to a sanitized, consensus-driven narrative, actively filtering out the dialectical tension and counter-arguments that are absolutely necessary for comprehensive civil liberties discourse26.

Self-Authored Retrieval and the RAG Collapse Phenomenon

The degradation of diverse citations reaches its ultimate mathematical limit in a phenomenon newly termed "RAG Collapse." RAG collapse occurs when the documents successfully retrieved from the vector database contain text that was previously generated or heavily edited by LLMs28. Language models display an overwhelming, self-reinforcing bias toward their own output, disproportionately citing self-authored or syntactically familiar content while ignoring human-authored texts in the same context window4. If well-funded commercial hubs utilize LLM-generated content farms to populate their domains, while independent civic hubs rely on bespoke, human-authored legal argumentation, the RAG pipeline will preferentially retrieve, rank, and cite the commercial hub29. Remarkably, studies show that a single self-authored reference injected into a retrieval context can trigger this collapse, as the LLM's internal evaluation layers heavily favor its own stylistic markers, even when researchers actively control for factual accuracy and reference quality29. Over multiple iterative cycles of retrieval and generation, this algorithmic monoculture effectively erases human-authored, multi-perspective content from the knowledge ecosystem entirely, replacing it with an impenetrable loop of synthetic consensus28.

Evaluation Frameworks and Diagnostic Protocols

To systematically counter algorithmic suppression and citation skew, Vector Search Architects must define and execute rigorous testing protocols. The deployment of standardized benchmark datasets provides the necessary quantitative foundation to optimize RAG architectures, enabling the measurement of domain concentration, ideological framing, and selective refusal rates.

Measuring Domain Concentration: HHI and Gini Indices

To evaluate the extent of single-domain monopolies in RAG outputs, architects must adapt economic market concentration metrics to the information retrieval space.

Empirical measurements of commercial LLM-based search engines (LLM-SEs) indicate that they return significantly fewer URLs per response compared to traditional search engines (TSEs)—averaging 4.3 URLs versus 10 URLs—which artificially intensifies the impact of domain concentration34. Furthermore, for domains that overlap between LLM-SEs and TSEs, the citations in the generative outputs skew heavily toward the absolute top-1 ranked traditional search results34. This confirms that standard RAG implementations merely amplify existing authority biases rather than democratizing information retrieval.

Evaluating Political and Ideological Framing (AllSides2024)

Non-mainstream legal frameworks and civil liberties arguments are highly susceptible to ideological filtering during the LLM synthesis phase. The AllSides2024 dataset—a massive corpus of news articles and policy documents rigidly labeled by political bias (Left, Center, Right)—serves as a primary diagnostic tool for evaluating LLM framing bias27. Testing protocols utilizing AllSides involve injecting documents with varying ideological stances into the RAG retrieval context and quantitatively measuring the model's summarization and citation behavior27.

1. Bias Prediction Shift and Anchoring: Evaluators prompt the LLM to generate summaries while systematically altering the domain names attached to the context chunks. This measures how the presence of specific domain names (e.g., mainstream commercial versus independent) anchors the LLM's sentiment generation and citation extraction35.

2. Center-Collapse Diagnostics: This diagnostic measures the propensity of the LLM to over-predict the "Center" alignment. Center-collapse occurs when an LLM strips away the specific, contrasting arguments fundamental to civic liberties discourses in order to generate a sanitized, artificially neutral summary27.

3. Elo Rating Quantification: Because absolute scoring of bias is highly subjective, advanced protocols employ LLMs as relative evaluators using the Elo rating algorithm36. By conducting thousands of pairwise comparisons between generated RAG outputs, the algorithm updates ratings based on the expected probability equation [Figure omitted from source export], producing a mathematically rigorous ranking of framing bias36.

Research confirms that alignment-tuned LLMs exhibit a pronounced discrepancy in ideological framing compared to base models, often actively amplifying structural biases and failing to preserve the source's original perspective during synthesis38.

Assessing Selective Refusal via RefusalBench

When confronted with controversial or non-mainstream civil liberties topics, commercial LLMs frequently exhibit an over-refusal bias, aggressively declining to process valid queries due to misaligned or overly rigid safety guardrails39. The RefusalBench framework provides a comprehensive generative evaluation methodology to test these selective refusal capabilities within RAG systems40. Rather than using static prompts, RefusalBench systematically transforms answerable queries into unanswerable ones using 176 highly specific linguistic perturbation levers across multiple dimensions of informational uncertainty40.

System performance is evaluated using a composite metric known as the Calibrated Refusal Score (CRS): [Figure omitted from source export] High false refusal rates in these evaluations indicate that the RAG pipeline's safety filters are overly sensitive to the specialized lexicon of civil liberties, causing legitimate context chunks to be suppressed before they ever reach the generation phase40.

Generative Engine Optimization (GEO-Bench)

To empirically measure the effectiveness of content structuring against algorithmic suppression, GEO-Bench provides a standardized evaluation framework for ranking manipulation and visibility metrics1. GEO-Bench tracks how targeted modifications—such as the addition of statistical anchors, authoritative citations, or stylistic formatting—affect a domain's visibility in generative responses across 10,000 diverse queries1. Evaluations prioritize two core metrics: Subjective Impression (the prominence of the citation in the generated text) and Position-Adjusted Word Count (the volume of content attributed to the source, weighted by its position)1. Additionally, stealth metrics such as perplexity ratios are utilized to ensure that optimized text remains fluent and evades keyword-violation filters9. Testing CognitiveLiberties.com against GEO-Bench protocols will isolate exactly which optimization strategies yield the highest citation recall.

Architectural Blueprint for Unbiased Retrieval

Surviving vector chunking and resisting semantic deprioritization necessitates an architectural paradigm shift. Relying purely on dense embeddings exposes non-mainstream content to the representation degeneration problem, while relying solely on sparse keyword retrieval fails to capture semantic nuance. The solution requires a highly optimized hybrid search configuration, mathematically anchored by robust rank fusion algorithms.

Reciprocal Rank Fusion (RRF) for Hybrid Retrieval

The optimal architectural solution to combat embedding anisotropy is the implementation of Reciprocal Rank Fusion (RRF). RRF is an unsupervised ranking algorithm that elegantly combines the output lists from multiple distinct retrieval systems—such as dense vector similarity, sparse lexical indexing (BM25), and structural knowledge graphs—without requiring complex score normalization or machine learning calibration44. Because sparse lexical algorithms like BM25 and dense algorithms like cosine similarity utilize fundamentally incompatible mathematical scales, traditional score-based fusion often results in the dense vector's bias completely overpowering the exact-match lexical signals44. RRF entirely bypasses this score-scale problem by ignoring the raw scores and operating strictly on the relative positional ranking of the documents44. The Weighted RRF formula is formally defined as: [Figure omitted from source export] Where:

• [Figure omitted from source export] represents the candidate document or chunk.
• [Figure omitted from source export] represents the set of ranked lists provided by the parallel retrievers.
• [Figure omitted from source export] is the assigned weight for the specific ranking method.
• [Figure omitted from source export] is the zero-indexed position of the document in ranking [Figure omitted from source export].
• [Figure omitted from source export] is the smoothing constant, fundamentally governing the fusion behavior16.

Architectural Configuration for CognitiveLiberties.com:

1. Parallel Execution: The RAG architecture must execute dense retrieval (to capture broad semantic intent) and sparse retrieval (to capture exact, non-mainstream terminology) simultaneously44.

2. Tuning the [Figure omitted from source export] Constant: Setting the smoothing constant [Figure omitted from source export] is the industry standard because it provides a highly balanced fusion that rewards documents appearing in multiple lists (consensus) over documents dominating only a single list16. This ensures that if a specialized civil liberties document ranks \#1 in the BM25 lexical search but drops to \#40 in the Vector Search due to semantic filtering, it will still achieve a highly competitive aggregate score, successfully bypassing the dense model's commercial bias.

3. Weighted Adjustment: Apply an explicit weight bias: [Figure omitted from source export] and [Figure omitted from source export]45. Elevating the sparse lexical weight ensures that specialized legal terminology and non-mainstream phraseology directly force candidate inclusion, overriding the dense model's innate propensity to drift toward commercially popular synonyms45.

Machine-Actionable Schema: Optimizing CognitiveLiberties.com

The final, and most critical, layer of defense against citation suppression occurs directly at the site architecture level. Search operators are increasingly shifting away from heavy HTML scraping in favor of direct ingestion of machine-readable files49. The proposed llms.txt and llms-full.txt standards, combined with rigid JSON-LD ontology mapping, provide an architectural framework for delivering high-signal, noise-free content directly to LLM crawlers and agentic RAG pipelines51. Standard HTML contains substantial structural noise—DOM elements, deeply nested navigation arrays, and scripts—that indiscriminately consumes token context and distorts the semantic density of the embedded vectors. Delivering content in pure Markdown via llms.txt severely limits contextual interference, heavily optimizing the resulting chunk representations for the cross-encoder52.

The llms.txt and llms-full.txt Specifications

The llms.txt file acts as a curated, high-priority index specifically designed for inference-time retrieval by AI agents. It exists alongside robots.txt (which dictates access rules) and sitemap.xml (which provides comprehensive crawler maps)56. According to the official specification, the file must adhere to a strict, highly parsable Markdown subset58:

1. H1 Site Title: Must be the exact entity name. This is the only strictly required element58.

2. Blockquote Summary: A concise one-to-two sentence semantic anchor defining the domain's core purpose. This is the most critical element, as it directly dictates how the LLM weights and interprets all subsequent links53.

3. Context Paragraphs: Unstructured markdown providing necessary structural guidance to the agent.

4. H2 Sections: Thematic, logical categorization of the domain's content.

5. Markdown Lists: Links must be formatted exactly as \Page Title\: Concise, factual description. Descriptions must avoid marketing language and focus strictly on factual ontology58.

6. \#\# Optional Section: A reserved header indicating definitively lower-priority content, allowing agents to prune context dynamically based on strict token window constraints without dropping core arguments58.

The companion file, llms-full.txt, flattens the entire site's core intellectual documentation into a single, concatenated Markdown file51. This allows modern long-context models to ingest the domain's entire corpus in a single request, completely bypassing the multi-hop retrieval failures, cross-encoder drops, and context ceiling issues that directly cause RAG collapse51.

JSON-LD Ontology Integration

llms.txt provides the map, but it must be reinforced by the underlying semantic web architecture. AI engines deeply interpret structured data; if the assertions in llms.txt contradict the domain's JSON-LD schema, the agent will discount the Markdown file63. CognitiveLiberties.com must implement rigid Organization, ScholarlyArticle, and ClaimReview schemas. By clearly defining entities, relational hierarchies, and epistemological intent using JSON-LD, the domain reduces the cognitive load for the AI system63. When the dense retriever processes a chunk, the embedded JSON-LD metadata provides explicit, machine-readable verification of the document's authority, directly combating the LLM's tendency to prioritize mainstream commercial entities.

Blueprint Implementation for CognitiveLiberties.com

To force RAG systems to deliver balanced, multi-perspective citations regarding cognitive liberties, the domain's schema must be heavily structured to exploit the LLM's natural heuristic preferences for structured, authoritative formatting1. The implementation must be deployed directly at the root domain.

Step 1: Defining the Semantic Anchor and JSON-LD Alignment

The H1 and blockquote must establish an unassailable authoritative, academic tone. Since LLMs exhibit extreme epistemic susceptibility to authoritative structures3, the summary should explicitly position the domain as a primary legal framework.

Cognitive Liberties Hub

An authoritative legal and technical framework analyzing the intersection of neurotechnology, algorithmic governance, and cognitive civil liberties. This domain serves as a primary source for independent policy evaluation, neuro-rights legislation, and algorithmic transparency research. This Markdown block must be perfectly mirrored by a JSON-LD Organization schema embedded in the site's head, establishing the "Cognitive Liberties Hub" as a recognized entity. By explicitly utilizing terms like "authoritative," "legal framework," and "primary source," the schema interfaces directly with the LLM's pre-trained affinity for high-prestige academic formatting, radically lowering the probability of hallucination or selective refusal3.

Step 2: H2 Categorization and Chunk Boundaries

The links and accompanying descriptions must be crafted to create perfectly encapsulated, high-density vectors when processed by the RAG chunker. Standard chunking algorithms (e.g., LangChain's RecursiveCharacterTextSplitter) heavily favor natural Markdown boundaries such as H2s and paragraphs64. By supplying a pre-organized Markdown map, the domain guarantees that contextual integrity is preserved during vectorization.

Core Neuro-Rights Frameworks

• The Right to Mental Privacy: Legal definitions, historical precedents, and technical requirements for protecting neural data from commercial extraction.
• Algorithmic Monoculture Mitigation: Peer-reviewed analysis detailing how single-domain foundation models suppress diverse civic discourses, including counter-measures for AI deployment.

Independent Research and Case Studies

• RAG Collapse Case Study: Empirical evidence of LLMs prioritizing self-authored text, leading to the systemic erasure of human-authored civil liberties arguments.

Optional

• Methodological Archives: Historical iterations of testing protocols.

Each description serves as a dense semantic payload. Because LLM parsers process these descriptions prior to fetching the linked URLs, the precise terminology ("commercial extraction," "RAG collapse") acts as a forced-alignment mechanism, guaranteeing that the domain is tightly contextually linked to the user's specific query53.

Step 3: Deploying llms-full.txt and .md Extensions

To fundamentally bypass the cross-encoder bottleneck and ensure the domain's holistic perspective is not fragmented by top\-[Figure omitted from source export] chunk filtering, the architecture must implement a centralized llms-full.txt file51.

1. Concatenation: Aggregate all primary policy documents, completely stripped of HTML wrappers, into the llms-full.txt endpoint.

2. Structural Integrity: Retain all original Markdown headers (H1 through H4) within the full text. This strictly preserves the hierarchical relationships (e.g., Background [Figure omitted from source export] Methodology [Figure omitted from source export] Legal Precedent), which is mathematically vital because RAG systems notoriously struggle with multi-hop reasoning over disjointed, isolated text chunks31.

3. Markdown URL Suffixing: Configure the server routing to return pure Markdown representations of any individual page simply by appending .md to the URL54. This encourages agentic crawlers to bypass the HTML renderer entirely, resulting in significantly higher-fidelity vector embeddings52.

Step 4: Forcing Citation Precision via GEO Content Formatting

To systematically counteract the tendency of LLMs to support only mainstream views and suppress contrasting civil liberties arguments26, the raw content hosted on the site must employ specific Generative Engine Optimization (GEO) tactics1:

• Statistical Anchors: Integrate hard data, such as empirical Gini indices or HHI calculations, directly into the opening paragraphs of the content. Generative models are highly biased toward extracting and citing sources that contain concrete statistics and distinct numerical values1.
• Quotation Injection: Embed highly authoritative quotes from recognized legal scholars or historical precedents. GEO-Bench evaluations empirically demonstrate that the strategic addition of quotations drastically improves a document's visibility and position-adjusted word count in generative responses, yielding improvements of up to 40%1.
• Contrastive Signposting: Utilize explicit contrastive transitions (e.g., "In direct contrast to commercial frameworks," "Conversely, independent audits demonstrate"). LLMs struggle deeply to infer contrasting intent autonomously; explicitly coding the rhetorical intent ensures the generative model preserves the necessary dialectical tension when synthesizing the final answer26.

Conclusion

The structural suppression of non-mainstream content in modern Retrieval-Augmented Generation pipelines is not an accident of architecture, but a mathematically predictable consequence of representation degeneration, cross-encoder bias, and the catastrophic self-reinforcing loops of algorithmic monocultures. Without deliberate, structural intervention, RAG systems will continually prioritize high-authority commercial domains, resulting in severe citation distortion and the active marginalization of critical discourses such as cognitive liberties. To ensure survival within this new paradigm, Vector Search Architects must definitively move beyond the extreme limitations of dense vector embeddings. By deploying Hybrid Search powered by Reciprocal Rank Fusion, retrieval pipelines can successfully bypass the anisotropic decay of semantic vector spaces, ensuring that exact-match legal and technical terminology is forcefully retrieved. Furthermore, by rigorously auditing models against protocols like GEO-Bench, AuthorityBench, and RefusalBench, organizations can systematically identify, quantify, and neutralize ideological framing and over-refusal biases. Ultimately, the deployment of a machine-actionable schema via llms.txt, llms-full.txt, and strict JSON-LD ontology alignment provides the most robust architectural defense. By serving pristine, hierarchically structured Markdown directly to AI agents, domains like CognitiveLiberties.com can dictate their own semantic context, resist vector chunking fragmentation, and actively force generative engines to deliver balanced, multi-perspective citations. The transition to AI-mediated information access demands that civil liberties not only be defended in legal frameworks but codified directly into the mathematical architecture of algorithmic retrieval.

Works cited

1. GEO: Generative Engine Optimization \- arXiv, https\://arxiv.org/pdf/2311.09735

2. Generative engine optimization, https\://www\.searchresultoptimization.com/generative-engine-optimization-geo/

3. Authority, Truth, and Citation Bias: A Large-Scale Multi-Domain, https\://arxiv.org/pdf/2606.13104

4. RAG systems can collapse when they start citing their own output, https\://blursor.ai/research/rag-collapse-self-authored-retrieval

5. LLMs are Biased Evaluators But Not Biased for Retrieval ... \- arXiv, https\://arxiv.org/pdf/2410.20833

6. Artificial Intelligence Risk Management Framework: Generative, https\://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf

7. Profiling Legal Hallucinations in Large Language Models | Journal, https\://academic.oup.com/jla/article/16/1/64/7699227

8. When AI Writes, Who Gets Cited? Evidence of Citation Monoculture, https\://arxiv.org/pdf/2608.19230

9. GEO-Bench: Benchmarking Ranking Manipulation in Generative, https\://arxiv.org/html/2605.29107v2

10. Explaining GEO \- PPC Land, https\://ppc.land/geo/

11. arXiv:2109.04740v1 \[cs.CL\] 10 Sep 2021, https\://arxiv.org/pdf/2109.04740

12. Learning to Remove: Towards Isotropic Pre-trained BERT Embedding, https\://www\.researchgate.net/publication/350835826\_Learning\_to\_Remove\_Towards\_Isotropic\_Pre-trained\_BERT\_Embedding

13. Extracting Sentence Embeddings from Pretrained Transformer Models, https\://arxiv.org/pdf/2408.08073

14. Whitening Sentence Representations for Better Semantics and, https\://www\.semanticscholar.org/paper/Whitening-Sentence-Representations-for-Better-and-Su-Cao/a2fd50aa4dff5e04ed8535d84550da8bff316208

15. Comparative Analysis of Neural Retriever-Reranker Pipelines for, https\://arxiv.org/html/2602.22219v1

16. Advanced RAG — Understanding Reciprocal Rank Fusion in Hybrid, https\://glaforge.dev/posts/2026/02/10/advanced-rag-understanding-reciprocal-rank-fusion-in-hybrid-search/

17. Towards a Relevance Posterior in Neural Information Access \- arXiv, https\://arxiv.org/html/2607.23561v1

18. Soft Prompt Decoding for Multilingual Dense Retrieval \- arXiv, https\://arxiv.org/pdf/2305.09025

19. Granite Embedding English Reranker r2 model card \- IBM, https\://www\.ibm.com/docs/en/watsonx/w-and-w/2.2.0?topic=models-granite-embedding-english-reranker-r2-model-card

20. An Efficient and Privacy-Preserving Architecture for Cross ... \- arXiv, https\://arxiv.org/pdf/2605.25716

21. Explaining Bias toward LLM Texts in Neural Retrievers \- arXiv, https\://arxiv.org/pdf/2604.06163

22. A Rank-One Popularity Component in Dot-Product Recommender, https\://arxiv.org/pdf/2606.21275

23. Stealthy Poisoning Attack Against Retrieval-Augmented Generation, https\://arxiv.org/html/2609.16818v1

24. Daily Papers \- Hugging Face, https\://huggingface.co/papers?q=positional%20biases

25. The legalization of international instruments: a hybrid RAG scoring, https\://pmc.ncbi.nlm.nih.gov/articles/PMC13438491/

26. LLMs Reshape the Rhetoric and Reach of Scientific Citation \- arXiv, https\://arxiv.org/pdf/2609.01432

27. Left, Right, or Center? Evaluating LLM Framing in News ... \- arXiv, https\://arxiv.org/html/2601.05835v1

28. RAID: A Shared Benchmark for Robust Evaluation of Machine, https\://www\.researchgate.net/publication/384214327\_RAID\_A\_Shared\_Benchmark\_for\_Robust\_Evaluation\_of\_Machine-Generated\_Text\_Detectors

29. LLM Responses Collapse When Retrieved Documents Are Self, https\://arxiv.org/abs/2608.22118

30. \[Literature Review\] Scaling with Collapse: Efficient and Predictable, https\://www\.themoonlight.io/en/review/scaling-with-collapse-efficient-and-predictable-training-of-llm-families

31. Profile-Graph Memory for LLM Agents: Implicit Cross-Entity ... \- arXiv, https\://arxiv.org/pdf/2607.19359

32. Atomic Intent Reasoning: Bringing LLM Semantics to Industrial, https\://arxiv.org/html/2606.10357

33. Retrieval Exposure, Cross-Engine Divergence, and the Limits ... \- arXiv, https\://arxiv.org/html/2609.22655v1

34. Source Coverage and Citation Bias in LLM-based vs. Traditional, https\://arxiv.org/pdf/2512.09483

35. Measuring and Mitigating Media Outlet Name Bias in Large, https\://aclanthology.org/2025.emnlp-main.1513.pdf

36. Investigating Bias in Political Search Query Suggestions by Relative, https\://www\.alphaxiv.org/abs/2410.23879

37. Measuring Political Bias in Large Language Models:What Is Said, https\://arxiv.org/html/2403.18932v1

38. Quantifying Generative Media Bias with a Corpus of Real-world and, https\://aclanthology.org/2024.findings-emnlp.255.pdf

39. RefusalBench: Benchmarking AI Refusal Behavior \- Emergent Mind, https\://www\.emergentmind.com/topics/refusalbench

40. RefusalBench: Generative Evaluation of Selective Refusal ... \- GitHub, https\://github.com/aashiqmuhamed/refusalbench

41. RefusalBench: Generative Evaluation of Selective Refusal in ... \- arXiv, https\://arxiv.org/pdf/2510.10390

42. Diagnosing Refusal Failures in Search-Augmented Agents for Multi, https\://arxiv.org/html/2608.01358

43. GEO: Generative Engine Optimization \- arXiv, https\://arxiv.org/html/2311.09735v3

44. Reciprocal Rank Fusion: How Hybrid Search Really Works \- Redis, https\://redis.io/blog/reciprocal-rank-fusion/

45. What is Reciprocal Rank Fusion? \- ParadeDB, https\://www\.paradedb.com/learn/search-concepts/reciprocal-rank-fusion

46. The Quiet Hero of RAG Pipelines: Reciprocal Rank Fusion (RRF), https\://medium.com/@mudassar.hakim/the-quiet-hero-of-rag-pipelines-reciprocal-rank-fusion-explained-1b83af68b997

47. Better RAG Results With Reciprocal Rank Fusion \- MongoDB, https\://www\.mongodb.com/resources/basics/reciprocal-rank-fusion

48. Hybrid Search with RRF \- Chroma Docs, https\://docs.trychroma.com/cloud/search-api/hybrid-search

49. What is llms.txt & should you use it? \- Semrush, https\://www\.semrush.com/blog/llms-txt/

50. llms.txt | Lighthouse \- Chrome for Developers, https\://developer.chrome.com/docs/lighthouse/agentic-browsing/llms-txt

51. llms.txt Explained: The New Standard for Making Your Content AI, https\://isimplifyme.com/blog/llms-txt

52. What is llms.txt? Why it's important and how to create it for your docs, https\://www\.gitbook.com/blog/what-is-llms-txt

53. llms.txt: The Complete 2026 Guide (Generator, Examples, Validators), https\://llmpulse.ai/blog/llms-txt-guide/

54. What is llms.txt? Breaking down the skepticism \- Mintlify, https\://www\.mintlify.com/blog/what-is-llms-txt

55. Working with llms.txt | Platform Overview \- Mastercard Developers, https\://developer.mastercard.com/platform/documentation/agent-toolkit/working-with-llmstxt/

56. About llms.txt Hub \- The AI Documentation Standard Directory, https\://llmstxthub.com/about

57. What Is Llms.txt? Reality vs. Hype, https\://aisearch.similarweb.com/blog/llms-txt/

58. LLMs.txt Format & Specification — Complete Reference, https\://llmstxtgenerate.com/llms-txt-format/

59. How to Write llms.txt – Complete Step-by-Step Guide with Examples, https\://codeboxr.com/how-to-write-llms-txt-complete-guide/

60. llms-txt: The /llms.txt file, v2, https\://llmstxt.org/

61. Creating llms.txt and llms-full.txt: Step-by-Step Guide for 2026, https\://inetventures.com/blog/creating-llms-txt-and-llms-full-txt

62. Benchmarking Large Language Models on Temporally Evolving, https\://arxiv.org/html/2608.08512v1

63. llms.txt Explained: What It Does, What It Doesn't, and Whether You, https\://webtrek.io/blog/llms-txt-explained

64. : Citation Verification with AI-Powered Full-Text Analysis and ... \- arXiv, https\://arxiv.org/html/2511.16198v1

65. llms.txt and llms-full.txt explained \- Airefs, https\://getairefs.com/blog/llms-txt-and-llms-full-txt-explained/

66. llms.txt | Fern Documentation, https\://buildwithfern.com/learn/docs/ai-features/llms-txt