If you sell “get cited in Baidu AI answers,” start with the concentration math. A Politecnico di Milano audit on arXiv — Auditing Source Exposure in Baidu and Google AI Search — compared Baidu and Google AI overviews across 500 MS MARCO intents in English and Simplified Chinese (1,998 query-platform observations). Data were collected 20 June–10 July 2026 from Milan via Playwright on Chrome.
Overview appearance rates alone redraw the map: Baidu Chinese 82.0%, Baidu English only 4.0%, Google Chinese 97.6%, Google English 97.0%. Baidu barely bothers with English AI overviews in this setup; Google fires them constantly in both languages. Host-domain inequality is the knife: Baidu Chinese Gini 0.895 versus Google Chinese 0.500 and Google English 0.462. Top-1 host share is 27.1% on Baidu Chinese against 6.6% / 5.5% on Google Chinese / English. Mean distinct hosts per overview: Baidu CN 15.0; Google CN 18.6; Google EN 16.7.
Cross-engine source overlap is tiny — host-set Jaccard for G-CN/G-EN 0.034, B-CN/G-CN 0.032, B-CN/G-EN 0.008 — even while matched-query answer cosine similarities sit in a middling 0.701–0.813 band. Same questions, different corpora, surprisingly similar answer vibe. Baidu Chinese top hosts in the study include the usual ecosystem gravity wells: baike.baidu.com, baijiahao.baidu.com, zhidao.baidu.com.
For China SEO that is a strategy fork. Winning Baidu AI overview citations looks more like winning Baidu’s owned content surfaces and a short list of dominant hosts than like diversifying across an open web long tail. Google-style “be one of eighteen roughly equal sources” is the wrong mental model. It also explains why Baidu’s new AI Traffic Analysis keyword table will matter: when citations cluster, knowing which queries already name you is leverage.
Caveats: one Milan vantage point, MS MARCO-derived intents, mid-2026 snapshot. Do not treat Gini as a lifelong constant. Do treat “Baidu AI overview ≠ Google AI Overview economics” as settled enough to brief clients this quarter. Combine this audit with Baidu’s evidence-fitness research thread and you get a clear editorial standard: fewer, denser, Baidu-ecosystem-compatible evidence pages beat a spray of thin English blog posts hoping for Chinese AI mercy.
Method note for practitioners quoting these figures: the study translates MS MARCO intents into Simplified Chinese rather than sampling native Baidu query logs, and it observes overviews from a single European IP range. That is still far better than anecdote, but it will under-represent local Baidu verticals and over-represent internationally portable informational queries. Re-run a thin version of the audit on your own money keywords before you rewrite an entire China content strategy around Gini 0.895.