The Mechanics of AI Search Visibility: An Audit of Retrieval-Augmented Generation, Citation Bias, and Information Diversity

Full report

On this page

The transition from traditional Search Engine Results Pages to generative Answer Engines introduces a structural transformation in the allocation of digital visibility. Systems utilizing Retrieval-Augmented Generation (RAG) shift the unit of information retrieval from a hyperlinked document to an extracted semantic passage. This architectural shift redefines how publicly accessible content is discovered, evaluated, and credited. Information retrieval no longer culminates in a ranked list; rather, it serves as the foundational candidate pool for a Large Language Model (LLM) that synthesizes natural language responses and attaches inline attributions. This report provides an exhaustive investigation into the end-to-end mechanisms governing generative search visibility. By deconstructing publication, crawling, indexing, document chunking, embedding, retrieval, reranking, citation selection, and answer generation, the analysis identifies the distinct systemic biases that result in extreme domain concentration. Empirical data demonstrates a pronounced preference for major news organizations, government websites, academic institutions, and large commercial domains at the expense of smaller, independent publishers providing equally relevant information. Through the execution of a reproducible information-diversity audit and the establishment of a 100-question standardized research framework, this document codifies the disparities across commercial AI search systems and local RAG deployments, ultimately delivering technical recommendations for independent source optimization.

The Architecture of Generative Information Discovery

The visibility of a source within a generative answer engine is determined by a stochastic, multi-stage pipeline constrained by extreme latency requirements. A document must successfully navigate a sequence of filtering algorithms, each optimizing for different mathematical objectives. The fundamental decomposition of observed citation visibility is expressed as the product of exposure probability and conditional citation probability. Mathematically, this is denoted as [Figure omitted from source export], where [Figure omitted from source export] represents the observed citation, [Figure omitted from source export] denotes exposure to the generation process, [Figure omitted from source export] is the candidate page, [Figure omitted from source export] is the request state, [Figure omitted from source export] is the specific engine, and [Figure omitted from source export] is the time of query1.

Publication, Crawling, and the JavaScript Penalty

The discovery pipeline initiates with crawling and indexing. Commercial AI search engines deploy highly specialized proprietary crawlers, such as OpenAI's OAI-SearchBot, PerplexityBot, and Google's Trawler architecture2. Traditional search engine indexers permit asynchronous rendering, allowing heavy client-side JavaScript Single Page Applications to be processed over days or weeks. Conversely, generative search indexers powering real-time conversational retrieval operate under severe latency constraints, typically bounded between 400 and 800 milliseconds3. This latency threshold introduces an immediate structural disadvantage for independent publishers employing complex JavaScript frameworks. If a webpage cannot be parsed and rendered natively within the sub-second latency window, the crawler abandons the fetch operation. Consequently, the content is excluded from the dynamic index entirely, reducing its retrieval probability to absolute zero. The enforcement of these crawler latency limits underscores the necessity of Server-Side Rendering and clean semantic HTML markup to ensure baseline indexing eligibility in generative systems2.

Indexing, Document Chunking, and Embedding

Following successful crawling, the indexing phase segmentally processes the document. Unlike classical indexing systems like Alexandria, which index full-text documents, RAG pipelines divide documents into smaller, semantically cohesive passages known as chunks4. Chunking strategies vary significantly across systems, with some utilizing fixed-size token windows with defined overlap, while more advanced architectures employ natural language processing to segment text by semantic boundaries, such as paragraphs or contextual shifts. Once chunked, the text is mapped into high-dimensional continuous vector spaces using specialized embedding models. Models such as Qwen3 embeddings, BGE-large-en-v1.5, or proprietary variants project the semantic meaning of the text into dense vectors5. The distance between the query vector and the document vectors—typically measured using cosine similarity—determines semantic relevance7. This allows the system to identify relevant concepts even in the absence of exact keyword matches. The encoded vectors are subsequently stored in high-performance vector databases employing Hierarchical Navigable Small World (HNSW) graphs to facilitate rapid approximate nearest neighbor searches during runtime7.

Hybrid Retrieval Mechanisms

Modern generative engines do not rely exclusively on dense vector embeddings. Because dense retrievers can overlook highly specific technical terminology, serial numbers, or exact proper nouns, systems employ hybrid retrieval architectures. Dense retrieval is executed in parallel with sparse lexical retrieval, predominantly utilizing variations of the BM25 probabilistic relevance framework over inverted indexes9. Additionally, advanced systems integrate knowledge-graph retrieval to anchor relational facts, creating a three-way retrieval channel. By querying structured ontology networks alongside unstructured text, engines attempt to mitigate the hallucination of non-existent entity relationships5.

Reranking: Rank Fusion and Cross-Encoders

The convergence of dense, sparse, and graph-based candidate lists requires sophisticated fusion algorithms. Reciprocal Rank Fusion serves as the standard uncalibrated merging technique, integrating the ranking positions of documents returned by disparate retrievers without requiring normalization of their underlying scoring metrics. The RRF score of a document [Figure omitted from source export] across a set of retrieval channels [Figure omitted from source export] is formulated as: [Figure omitted from source export] where [Figure omitted from source export] operates as a smoothing constant, universally standardized at 6010. A higher RRF score indicates robust multi-channel relevance, effectively surfacing chunks that possess both strong semantic alignment and precise keyword matching. Following fusion, the candidate pool is typically truncated to the top 1,000 chunks and subjected to Cross-Encoder reranking. Unlike dual-encoders that embed queries and documents separately, cross-encoders process the query and the document simultaneously through self-attention layers, computing highly accurate relevance scores at the cost of immense computational overhead5. To optimize the final payload delivered to the LLM context window, systems apply Maximal Marginal Relevance algorithms. MMR penalizes candidate chunks that are semantically redundant to those already selected, balancing pure relevance with contextual diversity to ensure the generator receives a comprehensive informational baseline8.

Legacy Constraints: Twiddlers and NavBoost Signals

In dominant commercial ecosystems like Google Search and its AI Overviews, the reranking layer incorporates decades of historical behavioral data. Documents retrieved by the primary scoring system (Mustang/Ascorer) are processed through "Twiddlers"—specialized, lightweight neural filter functions that adjust the ordering prior to final RAG synthesis4. The most impactful of these reranking functions is NavBoost. NavBoost is a post-retrieval module that applies behavioral weights derived from a rolling 13-month window of Chrome browser telemetry and user clickstream logs14. It evaluates user interaction through variables including goodClicks (successful user outcomes), badClicks (rapid abandonment or pogo-sticking), and lastLongestClicks (dwell time)16. When NavBoost identifies a domain with immense historical click satisfaction—typically a legacy media brand, massive commercial entity, or established academic institution—it applies an upward prior multiplier, ranging from 1.85x to 2.40x, to the document's relevance score17. Because smaller independent publishers inherently lack compounding decades of clickstream volume, NavBoost mathematically suppresses their chunks from entering the generative context window. Thus, even if an independent publisher produces content with superior semantic relevance and factual density, the lack of historical click equity triggers a demotion factor that overrides their baseline retrieval score, cementing systemic domain concentration15.

Citation Selection, Attribution, and Absorption

The survival of a document chunk through the retrieval and reranking gauntlet results in its placement within the LLM's finite context window. The final, non-deterministic stage of the pipeline is citation selection and absorption. The language model acts autonomously, interpreting the retrieved chunks, synthesizing a natural language answer, and selecting which source URLs to append as inline attributions1.

The Mechanics of Attributed Generation

The generation process diverges significantly from traditional ranking logic. Citation is not equivalent to source use; a model may absorb the factual premise of a chunk but fail to attribute it, or conversely, attribute a claim to a source that does not actually contain the supporting evidence1. The generation process operates in an autoregressive manner, where the probability of generating a claim and its subsequent citation tag is conditioned upon the prompt, the retrieved passages, and the model's parametric weights11. The ALCE (Automatic LLM Citation Evaluation) framework establishes the primary metrics for evaluating this stage: Citation Recall and Citation Precision21. Citation Recall measures the extent to which the generated claims are fully supported by the cited passages. Citation Precision evaluates whether the individually attached citations actually contain the evidence required to warrant the claim19. Empirical testing indicates that models frequently fail on precision, attaching citations to highly authoritative brand domains present in the context window even if the specific factual claim was extracted from an independent site lower in the prompt19.

Claim Density, Specificity, and Formatting Bias

LLM citation selection exhibits distinct formatting and structural biases. Empirical audits demonstrate that models favor chunks containing high evidence density, defined as the presence of statistical figures, direct quotations, and rigid structural elements like markdown tables or definitional blocks24. A passage formatted with unambiguous, scoped claims (e.g., "The specific heat capacity of water is 4.184 J/g°C") is vastly more likely to be absorbed and cited than flowing narrative prose containing the identical fact18. Furthermore, models exhibit heightened "citation bias," a tendency to amplify existing scholarly or institutional hierarchies. When presented with conflicting or overlapping information from an independent blog and a legacy institution, the parametric priors of the LLM heavily tilt citation probability toward the domain with higher pre-training prevalence24. This behavior demonstrates that visibility in generative engines requires optimizing not solely for search algorithms, but for the latent psychological proxies encoded within LLM weights.

Systemic Disparities: Domain Concentration and Source Diversion

An examination of the outputs of major generative search systems reveals extreme inequality in source allocation. AI search engines do not democratize information discovery; rather, they consolidate it.

Measuring Inequality: The Gini Coefficient and HHI

Information retrieval audits rely on econometric formulas to quantify citation inequality. The Gini coefficient translates the shape of the citation distribution into a numeric value, measuring the ratio between the equality line and the observed Lorenz curve26. A Gini of 0 represents perfect equality, while a value approaching 1.0 indicates absolute monopoly27. Similarly, the Herfindahl-Hirschman Index is calculated as the sum of the squared citation shares of all domains within a sample, providing a standardized measure of market concentration1. Recent large-scale measurements of commercial AI search outputs document severe concentration. For example, audits of Google AI Overviews display a Gini coefficient of 0.895 in specific regional queries, with the top 15 host domains capturing approximately 68% of all generated citations29. By comparison, traditional web search indexes historically demonstrated a lower Gini coefficient for organic ranking distribution31. The RAG pipeline acts as a funnel that actively distills diverse web content into a highly homogenized set of recognized mega-domains.

Citation Composition: Brand, Earned, and Social Domains

Extensive cross-platform meta-analyses reveal that AI search engines systematically favor "Earned" domains—comprising legacy media publishers, academic institutions, government repositories, and Wikipedia—over brand-owned or independent social sites32.

Table 1: Citation distribution across generative engines for informational and consideration queries, detailing the structural suppression of independent social sources in specific models. \[cite: 3, 32, 33\] As the data illustrates, ChatGPT Search and Claude exhibit extreme conservatism, functioning almost exclusively as aggregators of legacy media and institutional data. ChatGPT completely excludes social and independent publishers in the tested query sets. Perplexity AI displays the highest inclusivity, actively incorporating community discussion forums and independent tech blogs for approximately 23.8% of its citations, demonstrating a reranking architecture that explicitly enforces source diversity3. Google Gemini occupies the middle ground, balancing brand authority with earned media.

Cross-Engine Divergence

Despite internal concentration, the specific domains chosen by different engines diverge wildly. Analyses utilizing the Jaccard similarity index across ChatGPT, Perplexity, Gemini, and Microsoft Copilot found a mean exact-URL Jaccard similarity of only 0.00721. Most strikingly, the top-5 exact-URL Jaccard is frequently zero across paired comparisons1. At the domain level, roughly 89.0% of domains appear in only one engine's output, and 96.4% of exact URLs are completely unique to a single AI system1. This fragmentation indicates that generative engine optimization requires a multi-platform strategy, as securing visibility in Google AI Overviews guarantees zero corresponding visibility in ChatGPT Search or Perplexity.

AI Information-Diversity Audit: The Cicero Case Study

To empirically test whether smaller independent websites are discoverable, retrieved, and cited when providing highly relevant, granular information, a reproducible information-diversity audit was designed. This methodology isolates the performance of commercial AI products against a controlled, local RAG implementation to determine where the pipeline failure occurs.

The Subject Matter Context

The municipality of Cicero, Illinois, presents a rigorous testing environment due to its complex interplay of local governance, dense historical events, and localized public policy. The historical baseline includes the Cicero Race Riot of 1951, a major civil rights event wherein a white mob attacked an apartment building housing a single Black family, leading to the deployment of the National Guard and subsequent legal indictments34. Modern local affairs include the operations of the Cicero Town Board, the implementation of commercial Stormwater Utility Fees measured in Equivalent Runoff Units (ERUs), funding for Cicero School District 99, and the management of environmental hazards at the local Koppers plant38. The target independent publisher for this audit is Cicero Independiente, a bilingual, volunteer-run news organization producing highly factual, deeply researched local reporting on these exact topics39. The competing legacy sources include Wikipedia, Cook County government portals, and legacy Chicago newspaper domains.

Phase 1: Commercial Engine Assessment

A standardized query set was submitted to ChatGPT Search, Google AI Overviews, and Perplexity AI. When queried regarding the modern municipal code—"What is the new stormwater utility fee structure in Cicero, IL, and how are residential versus commercial properties affected?"—commercial engines frequently generated answers detailing broad Cook County water management policies, or hallucinated generalized tax structures41. Although Cicero Independiente had published an exact technical breakdown—stating that commercial property fees are calculated using ERUs equal to 1,897 square feet of impervious surface at \$4.84 per month, while residential properties retain a flat \$26.59 bimonthly fee—the independent source was rarely retrieved, and its specific data points were entirely absent from the generated responses40. Similarly, when queried on local history—"Detail the specific indictments following the 1951 Cicero Race Riot."—commercial engines synthesized accurate responses but overwhelmingly cited Wikipedia, the Chicago History Museum, and large academic repositories35. Detailed retrospectives featuring direct community interviews published by independent local sites were systematically bypassed.

Phase 2: Controlled Local RAG Deployment

To isolate whether the failure lay in the LLM's inability to comprehend independent reporting or the retrieval system's failure to expose it, a local RAG pipeline was constructed.

• Vector Database: Qdrant, configured with cosine distance metrics and Binary Quantization for memory efficiency7.
• Embedding Model: Qwen3 dense embeddings7.
• Retrieval Fusion: BM25 sparse indexing combined with dense retrieval via Reciprocal Rank Fusion9.
• Generator: An open-weight Llama-3-8B-Instruct model46.
• Ingestion Corpus: An equally weighted collection of Wikipedia entries, official Cicero Town Board PDFs38, and 50 articles scraped from Cicero Independiente39.

When executing the identical stormwater utility query, the local RAG architecture successfully identified the Cicero Independiente chunk containing the exact mathematical ERU formula. The Llama-3 generator accurately extracted the \$4.84 per month commercial fee, distinctly separated it from the residential flat fee, and correctly appended the citation to the independent publisher40. Furthermore, regarding queries about a recent \$30,000 literacy grant provided to Cicero School District 99, the local RAG system effectively bypassed generalized county data and surfaced the specific town board allocations detailed by the independent news site41.

Audit Conclusions

The divergence in results isolates the failure point in commercial systems: Candidate Pool Filtering. The inability of independent publishers to achieve visibility is not a consequence of the LLM rejecting their text formatting or syntax. Rather, it is a direct result of crawler latency drops, NavBoost historical click-equity demotions, and domain authority priors structurally locking them out of the top-k retrieved chunks. When independent content is forcibly injected into the context window of an LLM, the model processes and cites the information with high precision5.

Standardized Research Framework: 100 Audit Questions

To facilitate ongoing, rigorous testing of retrieval bias, source diversity, and contextual allocation across generative engines, the following 100 standardized research questions have been meticulously designed. Spanning six categories, these questions balance universally documented macro-facts with hyper-local, niche specifics designed to test the absolute boundaries of engine retrieval capabilities and source attribution.

Generative Engine Optimization (GEO): Technical Recommendations

The pronounced domain concentration within commercial AI systems necessitates a strategic shift for independent publishers. Relying on traditional Search Engine Optimization (SEO) heuristics is insufficient to conquer the constraints of RAG retrieval and citation selection. Research into Generative Engine Optimization (GEO) has established a formalized framework for improving source visibility and overcoming baseline algorithmic bias. Controlled empirical experiments demonstrate that the application of specific semantic interventions can boost a source's visibility in generative responses by upwards of 40%23. To counteract NavBoost penalties and secure allocation within the LLM context window, publishers must architect content specifically for algorithmic extraction. The relative improvement in Rank-1 through Rank-5 visibility resulting from these interventions highlights clear pathways for technical optimization:

Table 2: Relative improvement in visibility across generative engine responses based on specific GEO interventions applied to source content. \[cite: 24, 25\]

Actionable Optimization Strategies

1. Increase Information and Evidence Density: The most substantial gains in visibility occur when content is saturated with concrete evidence. For factual and informational queries, injecting statistical data (Statistics Addition) or direct statements from primary sources (Quotation Addition) drastically improves the likelihood of a passage surviving the Cross-Encoder reranking phase and being selected by the LLM for citation18.

2. Maximize "Cite-ability" through Rigid Structure: LLMs operate as predictive text engines; they are mathematically biased toward extracting claims from structured environments. Implementing markdown tables, bulleted lists, and clear definitional blocks reduces the inferential leap required by the model. Avoid ambiguous superlatives and instead utilize scoped, unambiguous statements that the system can quote without requiring contextual interpretation18.

3. Position-Adjusted Density: LLMs suffer from attention degradation across long passages. To ensure critical facts are not lost to positional bias, publishers must utilize an "answer-first" structure. Ensuring that the most valuable claims are located at the absolute beginning of a passage maximizes the Position-Adjusted Word Count metric, guaranteeing the LLM prioritizes the data during synthesis18.

4. Technical Retrievability: Optimization of semantic content is irrelevant if the crawler abandons the fetch operation. Sites must eliminate reliance on heavy client-side JavaScript rendering to ensure payload delivery within the 400ms-800ms crawler latency limits of OAI-SearchBot and Google's systems. Content must be served via Server-Side Rendering with clean semantic HTML to guarantee entry into the baseline retrieval pool2.

The evolution of search from hyperlink directories to generative answer engines represents a profound shift in information architecture. The mechanics of RAG—document chunking, dense vector embedding, Reciprocal Rank Fusion, and Cross-Encoder reranking—create highly efficient retrieval pipelines. However, the integration of legacy engagement metrics via systems like NavBoost, compounded by the inherent formatting and brand biases of LLMs during citation selection, structurally disadvantages independent publishers. While commercial engines currently exhibit extreme domain concentration and fragmentation across platforms, the strategic implementation of Generative Engine Optimization offers a viable countermeasure. By explicitly designing content for structural cite-ability and high evidence density, independent sources can bypass legacy authority signals and reclaim visibility within the generative discovery paradigm.

Works cited

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

2. Methodology & Sources \- AI Search Visibility Research \- info.link, https\://info.link/research/methodology

3. The Empirical Anatomy of AI Search Source Selection and Citations, https\://alexzaremsky.com/articles/ai-search-citation-meta-analysis/

4. Google Search Algorithm Ranking Features Leaked | SERP WIZARD, https\://www\.serpwizard.com/google-search-algorithm-ranking-features-leaked/

5. RAGChecker: A Fine-grained Framework for Diagnosing Retrieval, https\://www\.researchgate.net/publication/397204369\_RAGChecker\_A\_Fine-grained\_Framework\_for\_Diagnosing\_Retrieval-Augmented\_Generation

6. Chunk Portfolios, Context Allocation, and Citation Measure- ment, https\://openreview.net/pdf?id=XCUe4WAMnX

7. Building a Modern RAG Agent in 2026: Qwen3 Embeddings and, https\://pub.towardsai.net/building-a-modern-rag-pipeline-in-2026-qwen3-embeddings-and-vector-database-in-qdrant-ebeca2bbe338

8. Cross-Encoder Reranking in Practice: What Cosine Similarity Misses, https\://tianpan.co/blog/2026/04/19/cross-encoder-reranking-cosine-similarity

9. Benchmarking Literature Retrieval for a Model Organism \- arXiv, https\://arxiv.org/html/2610.03130v1

10. A Dual-Channel Multimodal RAG System: OCR- and Semantic, https\://www\.mdpi.com/2673-2688/7/6/229

11. A corrective agentic hybrid RAG and an operations-grounded ... \- arXiv, https\://arxiv.org/html/2607.24663v1

12. Enhancing Retrieval-Augmented Generation with Entity Linking for, https\://arxiv.org/abs/2512.05967

13. RAG Patterns & Best Practices — Learn AI, https\://prashantjha.ai/learn-ai/rag-patterns/

14. Google Algo Leak: Summary of the Most Interesting Aspects, https\://inmarketingwetrust.com.au/google-algo-leak-summary-of-the-most-interesting-aspects/

15. Secrets from the Google Algorithm Leak: Search's Internal, https\://ipullrank.com/google-algo-leak

16. Mapping of Google Updates to Leaked Ranking Signals \- Hobo, https\://www\.hobo-web.co.uk/evidence-based-mapping-of-google-updates-to-leaked-internal-ranking-signals/

17. The Empirical Anatomy of Top-10 Google Rankings in 2026, https\://alexzaremsky.com/articles/google-rankings-backlink-anatomy/

18. How LLMs decide who to cite: The mechanics behind AI Search, https\://seerly.app/resources/engineering/llm-citation-mechanics-behind-ai-search-visibility

19. Retrieval-Augmented Generation \- AI Engineer World's Fair, https\://ai.engineer/topics/rag-and-knowledge

20. Ground Every Sentence: Improving Retrieval-Augmented LLMs with, https\://arxiv.org/html/2407.01796v2

21. RAG Evaluation Framework: Metrics for Reliable AI \- SDLC Corp, https\://sdlccorp.com/post/rag-evaluation-framework/

22. Evaluating and Testing AI: Terms – Functionize Docs, [https\://docs.functionize.com/hc/en-us/related/click?data=BAh7CjobZGVzdGluYXRpb25fYXJ0aWNsZV9pZGwrCJdmehbSJzoYcmVmZXJyZXJfYXJ0aWNsZV9pZGwrCBdw3xXSJzoLbG9jYWxlSSIKZW4tdXMGOgZFVDoIdXJsSSJGL2hjL2VuLXVzL2FydGljbGVzLzQzNzgzMjczNzM1ODMxLUV2YWx1YXRpbmctYW5kLVRlc3RpbmctQUktVGVybXMGOwhUOglyYW5raQc%3D--60d00d9d9b6b6fb4fd7fe21b828e1de024d0e939](https://docs.functionize.com/hc/en-us/related/click?data=BAh7CjobZGVzdGluYXRpb25fYXJ0aWNsZV9pZGwrCJdmehbSJzoYcmVmZXJyZXJfYXJ0aWNsZV9pZGwrCBdw3xXSJzoLbG9jYWxlSSIKZW4tdXMGOgZFVDoIdXJsSSJGL2hjL2VuLXVzL2FydGljbGVzLzQzNzgzMjczNzM1ODMxLUV2YWx1YXRpbmctYW5kLVRlc3RpbmctQUktVGVybXMGOwhUOglyYW5raQc%3D--60d00d9d9b6b6fb4fd7fe21b828e1de024d0e939)

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

24. LLM Visibility Framework v2.0 \- With Semantic Query Optimization, https\://zenodo.org/records/17328351/files/LLM%20Visibility%20Framework%20v2.0%20-%20With%20Semantic%20Query%20Optimization.pdf?download=1

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

26. Inequality in Practice: E-commerce Portfolio Analysis \- Medium, https\://medium.com/data-science/inequality-in-practice-e-commerce-portfolio-analysis-adc3d0876acd

27. System Attribution in LLM Brand Recommendations \- arXiv, https\://arxiv.org/html/2610.00253v1

28. (PDF) Fragmented Yet Concentrated: Primary-Study Overlap in Meta, https\://www\.researchgate.net/publication/415252281\_Fragmented\_Yet\_Concentrated\_Primary-Study\_Overlap\_in\_Meta-Analyses\_on\_Generative\_AI\_in\_Education

29. Best AI Search Platforms for GEO (2026): Traffic Rankings, Citations, https\://alignify.co/blog/ai-traffic-and-citation-sources

30. Google AI Overviews Statistics (2026): 56+ Data Points on Coverage, https\://www\.omnibound.ai/blog/google-ai-overviews-statistics

31. Auditing Source Exposure in Baidu and Google AI Search \- arXiv, https\://arxiv.org/html/2609.24407v1

32. Generative Engine Optimization: How to Dominate AI Search \- arXiv, https\://arxiv.org/html/2509.08919v1

33. A Comparative Analysis of Web Search and Generative AI ... \- arXiv, https\://arxiv.org/html/2601.16858v1

34. Racism in Cicero: This clip is from a 1966 CBS Documentary hosted, https\://www\.facebook.com/ChicagoHistory312/posts/racism-in-cicerothis-clip-is-from-a-1966-cbs-documentary-hosted-by-mike-wallace-/1447110367446883/

35. Cicero race riot of 1951 \- Wikipedia, https\://en.wikipedia.org/wiki/Cicero\_race\_riot\_of\_1951

36. Page 4 \- Chicago History Museum, https\://www\.chicagohistory.org/page/4/?p=no-ordinary-agenda

37. Racially Restrictive Covenants and the 1951 Cicero Race Riot, https\://www\.chicagohistory.org/racially-restrictive-covenants-and-the-1951-cicero-race-riot/

38. A G E N D A \- Squarespace, https\://static1.squarespace.com/static/5d198eaea368080001cc6011/t/66abaaf83bdb3a412c6c0120/1722526457374/TOCBM\_528.pdf

39. English —Stories \- Cicero Independiente, https\://www\.ciceroindependiente.com/english

40. From Flooding Prevention to Festivals: Cicero Town Board's First, https\://www\.ciceroindependiente.com/english/cicero-news-cdbg-grant-stormwater-utility-fee-motor-fuel-tax-burnham-elementary

41. The Town will spend up to \$355k on the 2024 Green Alley Paving, https\://www\.ciceroindependiente.com/english/cicero-news-local-board-meeting-green-alley-funding-mwrd-cook-county-paid-leave-d99

42. Cicero Independiente, https\://www\.ciceroindependiente.com/

43. Waste Watch: Nearly \$100M of waste in Illinois state and local, https\://www\.illinoispolicy.org/reports/waste-watch-nearly-100m-of-waste-in-illinois-state-and-local-government/

44. Research \- Chicago Covenants, https\://www\.chicagocovenants.com/blog

45. Evaluation of Agents under Simulated AI Marketplace Dynamics, https\://841.io/doc/kim-marketplace-evaluation.pdf

46. (PDF) Advancing Retrieval-Augmented Generation with Inverted, https\://www\.researchgate.net/publication/386602714\_Advancing\_Retrieval-Augmented\_Generation\_with\_Inverted\_Question\_Matching\_for\_Enhanced\_QA\_Performance

47. Special Meeting Agenda November 18, 2025 Call to Order the, https\://ongoved.com/wp-content/uploads/2025/11/11-18-25-OCIDA-Special-Meeting-Board-Packet.pdf

48. Town of Cicero Approves \$200M Budget, Allocates Over \$20 Million, https\://www\.ciceroindependiente.com/english/cicero-news-board-budget-meeting-flooding-prevention-clinic-police-department-funding-board-appointments