PaperlayerView exact revision on arXiv

Source-linked AI summary

Source Coverage and Citation Bias in LLM-based vs. Traditional Search Engines

Peixian Zhang, Qiming Ye, Zifan Peng, Kiran Garimella, Gareth Tyson

arXiv:2512.09483v1cs.CLcs.CY

TL;DR

The paper addresses the lack of systematic evaluation of LLM-SE citation practices and how they compare with traditional search engines. It analyzes results from six LLM-SEs and two TSEs, finding substantial sourcing differences while LLM-SEs do not outperform TSEs on several quality and safety measures.

  • Problem

    There is no systematic evaluation of how LLM-SEs use citations or how their cited results differ from TSEs, despite concerns about transparency and reliability.

    Source: §1 · p. 1

  • Method

    The authors compare user-facing citations from six LLM-SEs and two TSEs across a large-scale dataset and analyze source-selection features.

    Source: §1 · pp. 1–2, §3.1 · p. 2

  • Results

    37% of domains in LLM-SE results are absent from TSE outputs, while LLM-SEs do not outperform TSEs in credibility, political neutrality, and safety metrics.

    Source: §1 · pp. 1–2

  • Takeaways & Limitations

    The authors conclude that future LLM-SEs need more diverse and safer sourcing practices, while source selection can narrow accessible content.

    Source: §7 · p. 9

    The analysis covers user-facing citations but does not measure the intermediate retrieval pipeline effectively.

    Source: §7 · p. 9

Abstract

from arXiv · show

LLM-based Search Engines (LLM-SEs) introduces a new paradigm for information seeking. Unlike Traditional Search Engines (TSEs) (e.g., Google), these systems summarize results, often providing limited citation transparency. The implications of this shift remain largely unexplored, yet raises key questions regarding trust and transparency. In this paper, we present a large-scale empirical study of LLM-SEs, analyzing 55,936 queries and the corresponding search results across six LLM-SEs and two TSEs. We confirm that LLM-SEs cites domain resources with greater diversity than TSEs. Indeed, 37% of domains are unique to LLM-SEs. However, certain risks still persist: LLM-SEs do not outperform TSEs in credibility, political neutrality and safety metrics. Finally, to understand the selection criteria of LLM-SEs, we perform a feature-based analysis to identify key factors influencing source choice. Our findings provide actionable insights for end users, website owners, and developers.

1 Introduction

The paper addresses the lack of systematic evaluation of LLM-SE citations and their transparency and reliability relative to traditional search engines. It compares six LLM-SEs with Google and Bing, finding substantial sourcing differences alongside persistent quality and safety concerns.

1 Introduction

  • The authors identify a gap in systematic evaluation of how LLM-SEs use citations and how their results differ from TSEs.Prior work cited in the introduction reports possible RAG bias, hallucinations, harmful content, and users unaware of errors in responses.Source: §1 · p. 1
  • The study compares Google and Bing with six LLM-SEs across 55,936 queries, 124,287 unique domains, and 1,418,733 unique citation hyperlinks.The authors state that the dataset will be made available for research use.Source: §1 · p. 1
  • 37% of domains in LLM-SE results are absent from TSE outputs, while LLM-SEs do not outperform TSEs on cyber safety.The study also reports that Gemini returns frequent far-left citations and Grok has the lowest partisan neutrality.Source: §1 · pp. 1–2
  • Feature analysis finds that LLM-SE-favored domains tend to have structured hierarchical HTML, easier-to-read text, lower popularity, and more outlinks to reputable sources.The introduction frames these features as factors associated with source selection compared with TSEs.Source: §1 · p. 2

2 Background & Related Work

Traditional search engines rank links, whereas LLM-SEs retrieve and synthesize information into cited answers. Prior work examines user interaction and output risks, but not the sources LLM-SEs cite compared with TSEs.

2 Background & Related Work

  • TSEs crawl and index pages, then rank links using signals including relevance, authority, reputation, and experience.Ranked positions can shape impressions and traffic; prior work also raises concerns about rank and authority bias and echo chambers.Source: §2 · p. 2
  • LLM-SEs use RAG to retrieve documents, synthesize them into summaries, and provide inline citations, while retrieved misinformation and generated malicious links remain concerns.The passage also notes political bias in source ratings and user sensitivity to citation number and type, even when attributions are incorrect.Source: §2 · p. 2
  • Prior comparisons report task-dependent performance, user trust effects from citations, and higher unsafe or malicious link generation for LLM-SEs than TSEs.These studies do not examine which sources LLM-SEs cite or how those sources differ from TSE results.Source: §2 · p. 2

3 Data Collection and Annotation

The study collects user-facing results from six LLM-SEs and two TSEs using a query corpus combining trending keywords, related questions, and controversial questions. It annotates query topics and cited domains using external classification and quality datasets.

3 Data Collection and Annotation

  • The collection captures a snapshot of popular queries and focuses on user-facing citations, although LLM-SEs may process multiple pages before selecting sources.The authors describe Figure 1 as the overall data-collection pipeline.Source: §3 · p. 2

3.1 Search Engines

  • Six LLM-SE platforms—ChatGPT, Gemini, Perplexity, Grok, AI Mode, and Copilot—are compared with Google and Bing as TSE representatives.An LLM-SE is defined here as an LLM system that retrieves web content and provides cited answers paired with summarized text.Source: §3.1 · p. 2

3.2 Search Query Corpus

  • The query corpus combines trending keywords, related natural-language questions, and controversial questions to represent varied search behavior and sensitive topics.The sources include Google Trends and X Trends keywords, recommended related queries from Google Search and AI Mode, and 2,281 controversial queries.Source: §3.2 · pp. 2–3

3.3 Search Query and Response Processing

  • The researchers submit queries through browser-based interactions, use incognito mode, and collect first-page TSE results and LLM-SE citations.ChatGPT memory is disabled to avoid referencing prior interactions; the study also counts repeated URLs only once within a response.Source: §3.3 · p. 3

3.4 Data Annotations

  • Queries are classified into 26 topic categories, while cited domains are annotated for category, popularity, credibility, political bias, and security risks.The annotations use an LLM classifier, Forcepoint ACE, Tranco, Media Bias/Fact Check, and VirusTotal; some domains cannot be matched to these datasets.Source: §3.4 · p. 3

4 Source Citation Behaviors (RQ1)

The section compares the sources cited by LLM-SEs with TSE results, examining citation counts, domain distribution, overlap, and popularity.

4.1 Comparing Citation Patterns

  • LLM-SEs cited a mean of 4.3 URLs and 3.4 domains per response, versus 10.3 URLs and 7.3 domains for TSEs.Grok and Gemini responses contained no cited websites in 82% and 38% of cases, respectively.Source: Figure 2 · p. 4, §4.1 · p. 4
  • All LLM-SEs had lower Gini indices than TSEs, indicating a more uniform domain distribution, although they returned fewer sources overall.Gemini and Grok differed significantly from both Google and Bing; ChatGPT differed significantly from Google only.Source: Figure 3 · p. 4, §4.1 · p. 4

4.2 Disparity in Domain Selection

  • 37% of domains were unique to LLM-SEs, while 38% appeared in both LLM-SE and TSE results.At the response level, overlap between any LLM-SE and TSE pair averaged less than 40%.Source: Figure 4 · p. 4, §4.2 · p. 4
  • ChatGPT more frequently surfaced Reuters and AP News (Google RTD = –0.48, Bing RTD = –0.41), while TSEs favored established broadcasters and opinion sources.The observed news-source divergence could introduce bias, which the paper investigates in §5.1.Source: Figure 5 · p. 5, §4.2 · p. 5

4.3 Measuring Source Popularity

  • For domains shared with TSEs, LLM-SE citations skewed toward top-ranked TSE results: 23.27% were Bing rank-one domains and 14.53% were Google rank-one domains.The rank distribution aggregates overlapping domains across searches.Source: Figure 6 · p. 5, §4.3 · p. 5
  • Except for ChatGPT, LLM-SEs tended to cite less popular domains than TSEs by Tranco rank; Copilot and Gemini averaged ranks over 22,000 higher than both TSEs.Popularity was assessed using Google and Bing rank positions and Tranco ranks; a lower Tranco rank indicates higher popularity.Source: §4.3 · p. 5, Table 2 · p. 6

5 Source Quality (RQ2)

This section compares source reliability and cyber-threat exposure across LLM-SEs and TSEs. It finds that source diversity matters for credibility and political bias, while fewer citations do not ensure safer results.

5 Source Quality (RQ2)

  • Because LLM-SE users receive summaries with only a few embedded references, this section evaluates source quality along reliability and cyber-threat axes.For reliability, the analysis focuses on news and media domains and uses MBFC ratings for credibility and political leaning.Source: §5 · p. 6, §5.1 · p. 6

5.1 Reliability

  • Gemini’s citations rely heavily on Wikipedia: it accounts for 41% of citations, and 62% of responses cite only one MBFC-listed domain, alongside Gemini’s lowest mean and median credibility scores.The paper associates this narrow sourcing with reduced variance and suggests greater source diversity is important for credibility.Source: §5.1 · p. 6
  • All engines have left-leaning mean and median political scores; Grok has the lowest response-level neutrality, with only 16% of responses scoring zero.Gemini’s cited domains include 6.4% in the far- or extreme-left range (< −5).Source: §5.1 · p. 6

5.2 Cyber Threat

  • Malicious-domain sets overlap substantially for the same query: 85% of engine pairs have average Jaccard similarity above 0.5, and 14% exceed 0.8.The authors suggest coordinated mitigation targeting shared malicious sources could reduce harmful-content spread.Source: §5.2 · p. 7
  • More than 50% of queries returning malicious domains fall into four categories: arts and entertainment, people and society, finance, and sports.The authors note that malicious domains can appear even for seemingly innocuous topics.Source: §5.2 · p. 7
  • Despite returning fewer sources, LLM-SEs do not outperform TSEs in cyber safety; most LLM-SE malicious-domain rates fall between Google’s 0.23% and Bing’s 0.42% baselines.Gemini has a higher mean of 1.56 malicious domains per response, 1.5 times Google’s mean.Source: §5.2 · pp. 6–7

6 LLM-SEs vs. TSEs Website Characteristics

The section examines website features associated with source selection by LLM-SEs versus TSEs. LLM-SE-favored domains tend to have more readable content and more structured HTML, while domain-level features also distinguish sources cited exclusively by LLM-SEs.

6.1 Features in HTML

  • LLM-SE-cited pages have more readable content than TSE pages: average grade level is 14.57 versus 18.24, and reading ease is 24.15 versus 12.32.The differences across the six textual and structural features are statistically significant (p < 0.01).Source: §6.1 · p. 7
  • LLM-SE pages show modestly higher HTML tag usage (mean 7 versus 6) and nesting depth (mean 16 versus 15) than TSE pages.The analysis considers readability alongside semantic tags, DOM nesting depth, accessibility attributes, and deprecated or discouraged tag counts.Source: §6.1 · pp. 7–8

6.2 Predicting Unique Inclusion of Domains

  • XGBoost best distinguishes domains cited exclusively by LLM-SEs from those exclusive to TSEs, with weighted F1 0.758, precision 0.668, and recall 0.879 for LLM-SE-unique domains.The reported metrics appear in Table 6.Source: §6.2 · p. 8, Table 6 · p. 8
  • The most influential XGBoost features are Tranco ranking (0.923), number of outlinks (0.799), and .com top-level domain (0.623).Higher numerical Tranco rank means lower popularity; the paper also reports positive influence for subdomain count, .nl domains, and business and economy category.Source: §6.2 · p. 8

6 LLM-SEs vs. TSEs Website Characteristics

  • Overall, LLM-SE-favored domains tend to be less popular, more readable, more structurally hierarchical, and have more outlinks than comparison sources.The analysis uses HTML features and predictive models to examine characteristics distinguishing domains uniquely cited by LLM-SEs from those exclusive to TSEs.Source: §6 · p. 7, §6.2 · p. 8

7 Conclusion and Future work

The conclusion synthesizes the study’s comparison of LLM-SE and TSE sourcing, finding more unique domains but comparable cyber-risk levels alongside concerns about source quality. It also identifies the study’s user-facing citation focus and future research directions.

7 Conclusion and Future work

  • LLM-SEs return 4.3 URLs per response on average versus 10 for TSEs, yet 37% of domains are absent from TSEs.The study assesses sourcing quality across reliability and cyber threats, and concludes that LLM-SE risk levels are comparable to TSEs.Source: §7 · pp. 8–9
  • The study restricts analysis to user-facing citations and does not effectively measure the intermediate retrieval pipeline.Future work could audit full retrieval traces and examine how monetization-driven behaviors influence content delivery.Source: §7 · p. 9

A Ethics

  • The researchers collected publicly accessible information, used rate limiting and other controls to avoid affecting search-engine performance, and obtained an ethics committee waiver.The authors report that their dataset contains no personal identifiers or information, to the best of their knowledge.Source: A Ethics · p. 12

B Bootstrapped Evaluation of the Gini Index

The appendix describes a bootstrap procedure for assessing differences in domain inequality between LLM-SEs and TSEs. It reports that Table 7 contains two-sided bootstrap p-values for Gini-index differences.

B Bootstrapped Evaluation of the Gini Index

  • The analysis resamples queries with replacement for each system, constructs domain-count distributions, and computes Gini-index differences across 1,000 iterations.The observed difference from full domain counts serves as a baseline.Source: B Bootstrapped Evaluation of the Gini… · p. 12
  • A 95% confidence interval is derived from the 2.5th and 97.5th percentiles of the bootstrap distribution of Gini differences.The procedure is described as a non-parametric way to assess robustness to sampling variability.Source: B Bootstrapped Evaluation of the Gini… · p. 12
  • Table 7 reports two-sided bootstrap p-values for Gini-index differences between LLM-SEs and TSEs.Its caption defines significance markers as * p < 0.05, ** p < 0.01, and *** p < 0.001.Source: Table 7 · p. 12

C Data Annotation

The annotation methods draw on four external knowledge bases, with human validation applied only to Forcepoint and VirusTotal; the overlap table reports mean domain overlap between LLM-SEs and TSEs.

C Data Annotation

  • Coverage rates assess how each annotation method retrieves categorization data from Forcepoint, Tranco Ranking, MBFC, and VirusTotal.Source: C Data Annotation · p. 12
  • Tranco and MBFC labels were not validated because Tranco provides popularity scores rather than interpretable labels, while MBFC assessments come from trained human reviewers.Source: C Data Annotation · p. 12
  • Three co-authors independently annotated 100 samples each from Forcepoint and VirusTotal; agreement with original labels reached 92% and 86%, respectively.Labels required agreement from at least two annotators, with unresolved cases settled through group discussion.Source: C Data Annotation · p. 12

D Overlapped Domains with TSEs

  • Table 9 reports mean domain overlap between LLM-SEs and TSEs as the per-response overlap divided by the number of domains retrieved by Google or Bing.Source: D Overlapped Domains with TSEs · p. 12

E Rank Turbulence Divergence (RTD)

The section describes a signed rank-divergence measure for comparing domains across TSE and LLM-SE rankings, then explains a transformation that aligns MBFC factuality scores with a conventional political-score scale.

E Rank Turbulence Divergence (RTD)

  • RTD compares each domain’s TSE and LLM-SE ranks, with positive values indicating greater prominence in TSE results and negative values indicating stronger representation in LLM-SE outputs.Domains missing from one list receive rank max(R)+1 there; α is set to 1/3 to control bias toward higher-ranked tokens.Source: E Rank Turbulence Divergence (RTD) · pp. 12–13
  • Table 9 presents mean domain overlap between LLM-SEs and traditional search engines.Source: Table 9 · p. 13
  • Figure 8 reports Dunn post-hoc comparisons for credibility and political profiles, with all pairwise comparisons statistically significant.Source: Figure 8 · p. 13

F MBFC Factual Score Transformation

G Testing the Independence of the MBFC Factual Score

The authors test whether MBFC factual reporting, credibility, and political leaning vary across search engines, and set a two-vendor threshold for classifying malicious domains.

G Testing the Independence of the MBFC Factual Score

  • Dunn’s post-hoc pairwise comparisons with Bonferroni correction followed the Kruskal–Wallis tests, and Figure 8 reports the credibility and political comparisons.Source: G Testing the Independence of the MBFC… · p. 13
  • Kruskal–Wallis tests found significant differences among search engines for credibility (H=1079054, p<0.001) and political leaning (H=1943828, p<0.001).The reported profiles associate AISEs more often with higher factual reporting scores and credibility, while some show distinct political lean distributions.Source: G Testing the Independence of the MBFC… · p. 13

H Threats Threshold

  • A threshold of at least two vendor detections classifies 0.4% (646) of domains as malicious, balancing false-positive and false-negative risks.A single positive flag maximizes sensitivity for emerging threats but risks false positives due to vendor inconsistencies.Source: H Threats Threshold · p. 13

I Query Categorization

The appendix describes query categorization and its validation, summarizes experiment configurations, and outlines the reported views of susceptible query categories and HTML structure features.

I Query Categorization

  • The categorization uses Google’s taxonomy, accepts labels supported by at least three models, and sends unresolved queries for human annotation.The prompt allows one to three relevant categories and a closest-match subcategory, or Unknown when no category fits.Source: I Query Categorization · p. 14
  • LLM-generated query categories achieved Fleiss’ κ=0.88 against human-verified labels in a 100-instance sample.The authors exclude categories where LLMs fail to reach consensus from this validation; an author annotates those cases.Source: I Query Categorization · p. 14

J Hyperparameter

K Susceptible Query Category

L HTML Structures

  • The HTML-structure features include semantic-tag counts, DOM nesting depth, accessibility attributes, and deprecated or discouraged markup errors.The text treats these as indicators of structure, accessibility, and technical quality.Source: L HTML Structures · pp. 14–15

Related papers

Loading 2512.09483v1…