AI Overviews favour own platforms over primary health sources

Vertical integration in AI search has a new, uncomfortable meaning. An arXiv audit of 1,920 health queries across 12 countries finds that both Google and Baidu systematically route users toward their own platforms in AI Overview answers, sidelining the primary health sources that brands, institutions, and governments have spent years cultivating.

The study, conducted across four languages and countries spanning the Global North and South, is the most geographically comprehensive audit of health-query citation behaviour published to date. Its central finding is blunt: neither engine treats its AI Overview as a neutral aggregator. Both treat it as a product-distribution channel.

What vertical integration looks like in practice

When a user poses a health query, the expectation, implicit in the design of AI Overviews, is that the system surfaces the most authoritative answer regardless of its origin. The arXiv data suggests the opposite. Google's AI Overviews disproportionately cite Google-owned or Google-adjacent properties; Baidu routes to its own ecosystem. The primary beneficiaries are not the WHO, national health ministries, academic medical centres, or the peer-reviewed literature. They are the platforms themselves.

This matters because the citation slot in an AI Overview is fast becoming the functional equivalent of the first organic result, except that there are far fewer of them, the methodology is opaque, and users are less likely to click through at all. If a multilateral health institution or a pharmaceutical company's medical-affairs content does not appear in that answer block, it effectively does not exist for a large and growing share of queries.

The language finding compounds the problem. Issuing the same query in a country's official language rather than in English raises the share of locally sourced citations by 3.5 to 13.5 times, depending on the market. That is an enormous multiplier. It means a health query in Mandarin in China yields a citation profile structurally different from the same query posed in English in the same country, with local sources flooding in when the local language is used. For global institutions publishing primarily in English, this is not a niche edge case. It is a systematic exclusion from the citation pool in major non-Anglophone markets.

The geography of authority

Smaller, lower-localisation countries receive fewer domestically sourced references for health queries. The pattern has a predictable shape: the richer and larger the market, the more local sources appear; the smaller the market, the more the AI Overview defaults to globally dominant, platform-adjacent sources. For a regional health ministry in Southeast Asia or West Africa, this means that AI search effectively imports its authority hierarchy from California or Beijing, regardless of what local expertise exists.

The implications for organisations in the multilateral and UN system are direct. Bodies such as UNDRR, CGAP, or the WHO produce primary-source material that, by any editorial standard, should anchor health and risk-related AI answers. The audit suggests that without deliberate effort to publish in local languages and to acquire citations from within regional web ecosystems, that material will be systematically underweighted in the AI Overview layer.