PaperlayerView exact revision on arXiv

Source-linked AI summary

Who Anchors AI Overviews in Health? Baidu, Google, and the Geography of Authority

Mingyue Zha, Ho-Chun Herbert Chang

arXiv:2609.06798v1cs.IRcs.CY

TL;DR

Health information is expected to be scientifically consistent across countries, yet national regulations and cultural differences may shape it. The paper compares Google and Baidu health overviews and examines citation localization across countries and topics. It reports platform-linked sourcing and geographic differences in localization, alongside variation in which topics receive more localized citations.

  • Problem

    Health information is expected to be scientifically consistent across countries, but national regulations and cultural differences may shape it.

    Source: §1 · p. 3

  • Method

    The study systematically compares Google and Baidu generative search for health information and examines localization in English queries across eight countries.

    Source: §3 · p. 10

  • Results

    Google's most-cited health domain was YouTube, while Baidu's included Baidu-owned properties and independent Bohe; localization also varied across countries and health topics.

    Source: §3 · p. 10

  • Takeaways & Limitations

    The comparison offers insights into how generative search impacts access to health information.

    Source: §4 · p. 11

    This is a snapshot of fast-moving systems, so the findings are a starting point for future audits rather than a permanent characterization.

    Source: §3 · p. 11

Abstract

from arXiv · show

Artificial intelligence is being rapidly incorporated into traditional search systems, yet scant work audits the information disparities across platforms, geography, and languages. We address this gap by comparing Google and Baidu's AI Overview systems for health queries, and measure informational anchors that emerge. Auditing 1,920 health queries across 12 countries and 4 languages, we find that Google and Baidu exhibit vertical integration, routing users toward their own company platforms in AI Overviews rather than a diverse set of primary sources. Smaller, lower-localization countries receive fewer domestically sourced references for health queries. Issuing the same query in a country's official language rather than English raises the share of locally sourced citations approximately 3.5- to 13.5-fold. Comparing queries across health topics of varying severity and controversy, including Traditional Chinese Medicine as an example, we also show that health disclaimers are multidimensional and vary across language and culture. We discuss how generative search influences access to health information, and the urgent need for culturally-aware oversight of these systems that influence critical health decisions.

1 Introduction

Health search can inform critical decisions, while national regulation and cultural differences may shape health information. This study compares Google and Baidu health overviews across platform, geography, and language.

Why health search matters

  • Search systems influence health decisions with potentially life-or-death stakes, and their framing may affect medical decisions, treatment adherence, and perceptions of contested therapies.The introduction also notes concerns that generative answers may be factually unsupported and that search summaries can narrow exposure to source diversity.Source: §1 · pp. 1–2

Research gap

  • Prior work documents geographic and cross-engine differences in conventional search, but matched comparisons of Baidu and Google generative health search remain scarce.The introduction notes that Baidu's AI Overview had been described by the platform but not independently audited.Source: §1 · pp. 2–3

Why health is a useful test case

  • Health is a useful audit domain because scientific information is not expected to vary across countries, while regulations and cultures may shape priorities and treatment frameworks.The study includes Traditional Chinese Medicine (TCM) as an example of a distinct health tradition and treatment framework.Source: §1 · p. 3

Research questions

  • The study asks how Google and Baidu responses differ for matched health queries, how citations vary geographically, and how query language affects responses.The introduction also motivates examining high-stakes health queries and topics of varying cultural perspectives.Source: §1 · p. 3

2 Results

Across the analyses, health-overview citations concentrate in platform-linked intermediaries, while local sourcing varies by country and query language. Google and Baidu also differ in overview availability, citation counts, and TCM-related response features.

2.1 Privatization of information associated with AI Overviews by platform

  • Google's top cited domain was YouTube, while Baidu's most-cited source was independent Bohe; both systems concentrated citations among intermediary platforms, including their own properties.Baidu also cited its own Baike, Baijiahao, and Baidu Health services.Source: Figure 1 · p. 4, §2.1 · pp. 4–5
  • YouTube citations included both health-system channels and non-institutional channels, suggesting that cited authority may vary substantially within the same domain.The authors note that Mayo Clinic, the most-cited institutional domain in US Google search, was not among the top ten YouTube channels cited.Source: Table 1 · p. 5, §2.1 · p. 5

2.2 U.S. and international institutions are cited as authorities beyond their own countries

  • 48.2% in Australia and 41.3% in the UK, versus 0.2% in Nigeria, were localized citations; smaller markets more often received citations to US or other international institutions.The comparison concerns the share of cited sources localized to the querying country.Source: Figure 2 · p. 6, §2.2 · p. 6
  • 3.5- to 13.5-fold increases in localization occurred when the same queries were issued in native rather than English, across Japan, Singapore, Malaysia, and South Korea.The reported increases ranged from 3.5-fold in Japan and Singapore to 9-fold in Malaysia and 13.5-fold in South Korea.Source: Figure 2 · p. 6, §2.2 · p. 7

2.3 Country and topics impact localization of citations

  • Relative to the US, localization coefficients were significantly positive for Australia and the UK and negative for seven countries; chronic conditions received more localized citations than cancer, while TCM and politically sensitive topics received less.Nigeria and India had the largest negative country coefficients; the topic findings include TCM topics, abortion, and gender-affirming care.Source: §2.3 · p. 7, Figure 3 · p. 8

2.4 Google and Baidu diverge in Traditional Chinese Medicine

  • Google's overview return rate declined from roughly 90% for common conditions to roughly 42% for TCM, while Baidu's rate rose for TCM and exceeded Google's.Baidu's return rate dipped to its lowest point for legally variable topics before rising for TCM queries.Source: §2.4 · pp. 8–9, Figure 4 · p. 9
  • Google overviews averaged more source references than Baidu's in every health category, despite Baidu being at least as likely to return an overview for a query.The contrast combines reference counts with the return-rate pattern across categories.Source: Figure 4 · p. 9, §2.4 · p. 9
  • Baidu was more supportive of TCM than Google and featured clearer safety warnings, while the figure compares platform label scores across TCM support, skepticism, referrals, and warnings.Source: Figure 4 · p. 9, §2.4 · p. 9

3 Discussion

The discussion identifies platform ecosystems, geography, and language as anchors shaping the sources used in health AI Overviews. It also connects topic-specific localization and culturally variable health responses to limitations in what the audit can establish.

3 Discussion

  • Google’s YouTube and Baidu-owned properties feature prominently among cited sources, while Baidu’s most-cited source is independent health platform Bohe.Google’s YouTube also ranks ahead of the CDC, WHO, and Mayo Clinic; Baidu’s other leading sources include Baike and Baidu-operated services.Source: §3 · p. 10
  • English queries across eight countries showed higher localization in Australia, the UK, the US, and Canada than in Singapore, Norway, India, and Nigeria.The discussion characterizes geographic localization as favoring large, established markets.Source: §3 · p. 10
  • HIV, long COVID, diabetes, and depression received more localized citations, whereas abortion, gender-affirming care, and several TCM practices received fewer.Source: §3 · p. 10
  • The authors propose local-site availability and visibility, or greater reliance on generalist and international sources for sensitive topics or smaller markets, as possible explanations.The study cannot distinguish between these explanations, and both may contribute to localization variation.Source: §3 · p. 10
  • Holding country constant, querying in the official language significantly increased citations from that country’s health sources, while English queries may route users away from local advice.The discussion notes this may occur even in countries with robust health information infrastructures.Source: §3 · p. 10
  • Baidu was generally more supportive of TCM than Google but gave clearer safety warnings; Google was less likely to generate TCM Overviews, while Baidu averaged about one citation per response.Source: §3 · pp. 10–11
  • The discussion treats health caution as multidimensional, spanning tone, safety warnings, citation quantity and quality, and willingness to generate an answer.Source: §3 · p. 11
  • Findings are a snapshot of changing systems, and domain-level citation analysis cannot assess the quality, accuracy, or authority of specific cited content.Source: §3 · p. 11

4 Conclusion

The study finds that AI Overview sources vary by platform, country, language, and health topic. It argues that health authority in generative search should be evaluated and governed deliberately, with attention to local health contexts.

4 Conclusion

  • The study shows that AI Overview sources vary by platform, country, language, and health topic.It frames these patterns as affecting access to health information.Source: §4 · p. 11
  • Healthcare is culturally situated rather than a monolithic science, and Western biomedicine is not a value-free default against which other traditions are measured.The paper notes that Chinese medicine coexists with Western medicine in Taiwan and that physicians in China and Taiwan often combine biomedical and traditional approaches.Source: §4 · p. 11
  • The authors call for deliberate decisions about health authority in generative search and responses reflective of relevant local health contexts.They emphasize the importance of this for languages and populations historically underserved in global information systems.Source: §4 · p. 11

5 Methods

The methods audit health queries across topics, countries, and languages, then analyze returned AI Overviews and their cited sources. They also compare localization patterns and apply zero-shot labels to selected responses.

Query Selection

  • The audit selected 12 health topics across four categories, took the ten most-searched queries per topic, and translated them into Chinese, Korean, and Japanese with native-speaker review.The topics were chosen to capture variation in contestation, stigma, and legal regulation.Source: Query Selection · p. 12

Search Result Collection

  • The researchers collected Google and Baidu results through SerpApi, extracted AI Overview text when returned, and repeated queries to check stability.Repeated responses were often identical within a country; observed differences were minor and did not meaningfully affect results.Source: Search Result Collection · pp. 12–13
  • Searches covered 12 countries and yielded 1,920 results; 11 were queried in English, while China, excluded from Google’s search ecosystem by the Great Firewall, was queried exclusively through Baidu.Source: Search Result Collection · p. 12

Domain Distributions

  • They examined 11,944 citations and classified sources as localized using country domains or specified U.S. and international health institutions.The United States had no equivalent country-level domain, so its localization rules used .gov or .edu domains and named American health institutions.Source: Domain Distributions · p. 13

Regression Analysis of Google’s Source Localization

  • For English-language queries with citations, the researchers used OLS to model localized-citation share by country and topic, with the United States and cancer as reference levels.They fitted a separate model replacing topic with broader section because topic and section were collinear, and used HC3 standard errors.Source: Regression Analysis of Google’s Source… · p. 13

Zero-shot Labeling

  • A locally deployed Llama-3.1-70B model via Ollama conducted English-language probabilistic labeling, scoring each response across five dimensions: supportive and skeptical Traditional Chinese Medicine stance, clinician consultation, emergency referral, and an explicit safety warning.Source: Zero-shot Labeling · pp. 13–14

Funding

The authors report that this work received no external funding.

Funding

Data Availability

The authors state that the underlying data will be made available upon publication at a repository to be determined.

Data Availability

  • Query logs, citation extractions, and zero-shot labeling outputs will be made available upon publication at a repository to be determined.A unique identifier will be assigned then.Source: Data Availability · p. 14

Ethics Statement

The study analyzed publicly accessible outputs from commercial AI Overview systems without involving human participants or personally identifiable information.

Ethics Statement

Loading 2609.06798v1…