The practical brief
The finding
Across four within-country Google comparisons, local sources accounted for a larger share of health citations in native-language queries than in English queries. [c1] [c4] [c7]
Why it matters
Health businesses assessing discoverability through English-only searches may get an incomplete picture of the local information ecosystem.
| Query country | English queries | Native-language queries |
|---|---|---|
| Japan | 16.6% of Google health citations were local | 57.5% of Google health citations were local [c4] |
| Malaysia | 4.6% of Google health citations were local | 42.1% of Google health citations were local [c4] |
| Singapore | 11.5% of Google health citations were local | 40.2% of Google health citations were local [c4] |
| South Korea | 2.7% of Google health citations were local | 36.4% of Google health citations were local [c4] |
What to try
BLURSOR’s practical interpretation
Our interpretation: compare relevant queries across customer languages and locations before committing to a content or distribution strategy.
Study boundary
This audit measured source selection, not business visibility gains from translation, video publishing or other marketing interventions.
Do not base a local visibility budget on English alone
If your health business serves a multilingual market, an English-only AI search audit could misrepresent which local sources customers encounter. Before deciding that local content is absent—or that international publishers dominate every customer journey—consider whether your monitoring reflects the languages people actually use.
That is the business risk suggested by this study, not a tested marketing prescription. Researchers examined citations in AI Overviews, the synthesized answers search engines display alongside search results. They did not test whether translating a business website improves its chances of being cited.
What the researchers compared
The May 2026 audit collected 1,920 search results across 12 countries and four languages. Researchers submitted 120 health queries per country-language configuration through SerpApi, using country-location settings for Google.
Eleven countries were queried in English. Japan, Malaysia, Singapore and South Korea were also queried in a native language; China was queried only in Chinese through Baidu. Consequently, this was not a design that tested both platforms in every country and language.
The analysis tracked whether an Overview appeared and which domains it cited. “Local” generally meant a source with the querying country's domain ending, such as .jp or .sg. US classification instead used institutional-domain and named-health-institution rules. This is a measure of citation geography, not a certification of local expertise.
Language changed the local citation picture
Within the four Google language comparisons, native-language queries consistently returned higher local citation shares. Japan rose from 16.6% in English to 57.5%; South Korea rose from 2.7% to 36.4%. Malaysia and Singapore showed the same direction of difference. These percentages describe shares of citations, not shares of businesses discovered or customers converted.
Platform source patterns also differed. YouTube led Google's health citations, ahead of major health institutions. Baidu's leading source was the independent health platform Bohe, followed by Baidu-owned properties. The presence of company-owned sources does not, by itself, establish that ownership caused their selection.
For a business owner, the useful distinction is between being surfaced and being credible. A familiar platform domain can host sources with very different credentials; the audit did not assess the medical accuracy of individual cited items.
A realistic interpretation: improve measurement before changing tactics
Our interpretation is to treat language and location as separate monitoring dimensions. The findings support checking whether an English-only view misses local sources; they do not establish that multilingual publishing or creating YouTube videos will earn citations.
- Compare equivalent health queries in the languages your customers use, keeping location settings consistent.
- Record which sources appear and whether an Overview appears; absence of a synthesized answer is different from absence of your business.
- Inspect the cited page or video rather than treating its host domain as proof of authority.
- Use observed gaps to form content hypotheses, then evaluate any changes separately against your business goals.
What the study cannot tell your business
These results are a snapshot of changing systems, not a durable ranking rule. The authors cannot separate the availability of authoritative local sources from algorithms' preference for international sources. Nor does the Google-US versus Baidu-China comparison isolate platform effects from country and language.
The study cannot tell you whether a citation drives visits, trust, appointments or revenue. Its response-stance and warning scores came from model labeling with limited bilingual review, not clinical validation. The practical takeaway is narrower: audit the relevant search environments before interpreting visibility as either market coverage or evidence of credibility.
The source and evidence
Who Anchors AI Overviews in Health? Baidu, Google, and the Geography of Authority
Study limitations and disclosures
- The May 2026 audit is a snapshot. Results may change as search models and retrieval systems are updated, and the study did not test marketing interventions or conversion outcomes.
- Platform comparisons pair Google in the United States with Baidu in China, so platform, geography and language effects cannot be fully separated.
- Localization relies mainly on country-specific domains, with different institutional rules for US sources. Domain-level citation analysis does not establish the accuracy, quality or authority of individual items. The authors cannot distinguish local-source availability from algorithmic preference.
- No uncertainty intervals were supplied for the reported within-country language percentages. Response-stance and warning labels were generated by Llama-3.1-70B in English and checked on selected observations by a bilingual reviewer; these are not clinical assessments.
- Both authors are affiliated with Dartmouth College's Program in Quantitative Social Science. They declare no competing interests and no external funding, while acknowledging support from the Stamps Foundation and Dartmouth College.
- The paper promises release of query logs, citation extractions and labeling outputs upon publication, but specifies neither a repository nor an identifier, limiting immediate reproducibility.
Evidence behind this briefing
[c1] The audit used SerpApi with country-location settings for Google and 120 health queries per country-language configuration. Eleven countries were queried in English, four also in a native language, and China exclusively through Baidu in Chinese, yielding 1,920 results. This is not a fully crossed platform-country-language comparison.
§5 Methods — Search Result Collection · Read the study
[c2] Citation prominence differed by platform: YouTube led Google's health citations, while independent health platform Bohe led Baidu, followed by Baidu-owned properties. These are source-selection findings, not proof that platform ownership caused selection or that the cited content was accurate.
§3 Discussion, first findings paragraph; see also §2.1 · Read the study
[c3] For English-language Google health queries in eight countries, the reported share of domestically localized cited sources ranged from 48.2% in Australia to 0.2% in Nigeria. Localization was inferred primarily from country-specific domains, with separate rules for US sources; this measures citation geography rather than source quality.
§2.2 U.S. and international institutions are cited as authorities beyond their own countries; Figure 2a; §5 Domain Distributions · Read the study
[c4] In the four within-country Google language comparisons, native-language queries had higher local citation shares than English queries: Japan 57.5% versus 16.6%, Malaysia 42.1% versus 4.6%, Singapore 40.2% versus 11.5%, and South Korea 36.4% versus 2.7%. The reported increase was approximately 3.5–13.5-fold; no uncertainty intervals are supplied for these percentages.
§2.2, within-country language comparison; Figure 2b · Read the study
[c5] In the Google-US versus Baidu-China comparison, Google cited more references per health category on average. Google's Overview return rate fell from roughly 90% for common conditions to roughly 42% for Traditional Chinese Medicine. Baidu's higher TCM support and clearer warnings reflect model-labeled response stance, not demonstrated medical accuracy or user preference.
§2.4 Google and Baidu diverge in Traditional Chinese Medicine; Figure 4a–b; §5 Zero-shot Labeling · Read the study
[c6] Response stance, referrals and warnings were scored using locally deployed Llama-3.1-70B through Ollama, with all labeling conducted in English. A bilingual reviewer checked 10 low-score and 10 high-score observations per feature; these labels should not be presented as clinical validation.
§5 Methods — Zero-shot Labeling · Read the study
[c7] The study is a snapshot of changing systems, and its primarily domain-level citation analysis cannot establish the quality, accuracy or authority of individual cited items. The authors also cannot distinguish local-source availability from algorithmic deference as explanations for localization differences.
§3 Discussion — Limitations and Future Work; mechanism ambiguity discussed earlier in §3 · Read the study
[c8] The authors disclose no competing interests and no external funding, while acknowledging Stamps Foundation and Dartmouth College support. The underlying data were not yet assigned a repository or identifier in the supplied paper, limiting immediate reproducibility.
Data Availability; Acknowledgments; Competing Interest; Funding · Read the study