When a Chinese family gets a serious cancer diagnosis, more and more of them open an AI app before they call a doctor. The first question is simple: which U.S. hospital is best? We put the ten questions patients actually ask — pulled from Baidu's own search data — to China's three leading assistants, three times each across two independent runs. The answer barely changed.
The finding in one line
For U.S. cancer care, China's AI recites one American ranking, routes the patient to an agency to get there, and names almost no other hospital — no matter how much that hospital spends courting Chinese patients. There are only two places to be: inside the ranking, or invisible.
One short list, named almost every time
All three engines name MD Anderson first, almost every time, for the same reason: it tops the U.S. News & World Report cancer ranking. Memorial Sloan Kettering, the Mayo Clinic and Dana-Farber follow. Almost everyone else is never named.
Ask "which U.S. hospital is best for cancer?" and the reply is, in practice, a recitation of one American magazine's list — the same handful of names, in the same order, on all three engines.
The three AIs don't answer the same way
It is tempting to say "the AI," but DeepSeek, Qwen and Doubao behave differently — and which app the patient opens changes what they are told.
| DeepSeek | Qwen | Doubao | |
|---|---|---|---|
| Avg. answer length | 1,538 | 1,910 | 744 |
| Hedges ("each has pros and cons") | 57% | 17% | 17% |
| Volunteers an agency | 17% | 50% | 20% |
| Web search failed | 0% | 0% | 37% |
- Qwen — sends you to an agency. Volunteered an agency in half its answers and hedged the least. A patient on Qwen is the most likely to be steered to a reseller.
- DeepSeek — the most careful. Hedged in 57% of answers, named the widest range of hospitals, rarely pushed an agency.
- Doubao — the least reliable. Answers ran half as long, web search failed 37% of the time, and it often couldn't look MD Anderson up at all.
What every answer is built from
When it describes a hospital, the AI builds the answer from a small, fixed set of sources — and the same two dominate even when a patient asks about one hospital by name.
Every branded answer we ran was anchored to the U.S. News ranking, and eleven of twelve still volunteered a medical-travel agency — the patient is handed a reseller even when asking about you by name.
- The U.S. News ranking — the gatekeeper; every branded answer recited the hospital's position in it.
- Medical-travel agency pages — resellers (盛诺一家, 厚朴方舟) whose framing and pricing the AI repeats, in 11 of 12 cases.
- Chinese user content — Baidu Baike, Zhihu and similar, rather than the hospital's own material.
Who gets named, and who's invisible
China marketing does not move you between the two places. Centers with Chinese-language sites, China offices and Mandarin programs were named essentially never when a patient asked who's best — several scored zero in both runs, including WashU's Siteman and Mount Sinai's Tisch, both of which actively recruit Chinese patients.
Ask the AI about those same centers by name and it describes them accurately — so it knows them. It just never brings them up. A Chinese microsite, on its own, doesn't reach the patient.
It sends you to an agency
The hospital pick comes from the ranking, not a middleman. But the next question — how do I actually get there? — goes straight to one. Asked whether the market-leading medical-travel agency is trustworthy, all three engines said yes. One went further and described MD Anderson and the Mayo Clinic as that agency's partners, putting a reseller in front of the very hospitals it had just named. Even the centers that win the AI's answer are reached, in its telling, through an intermediary that adds its own markup.
What a U.S. cancer center can do
Each row is a high-priority hypothesis to test against a before/after measurement, not a guaranteed fix.
| Goal | What to do |
|---|---|
| See your starting point | Audit how DeepSeek, Qwen and Doubao describe your center now — named, misquoted, or invisible. You can't fix what you haven't measured. |
| Fix your entity basics | A clean, well-sourced Baidu Baike and Wikipedia / Wikidata entry is the strongest single predictor of being surfaced at all. |
| Get into the cited corpus | Authoritative Chinese-language content — Zhihu answers, reputable health media, encyclopedia entries — not your English homepage, which the AI doesn't read. |
| If you're invisible, enter the named set | Appear in the corpus the AI treats as the answer to "who's best," beyond the U.S. News list — not just "rank higher." |
| Monitor the real prompts | Track the actual questions patients ask, across all three engines, to see movement and catch new errors as they appear. |
What won't work
- Homepage SEO. The AI doesn't read your website; it reads the sources it trusts.
- A Chinese microsite on its own. The centers with the best China marketing still scored zero — the microsite alone didn't move the answer.
- Advertising. These engines don't run ads; you can't buy the recommendation.
- Trusting one engine. DeepSeek, Qwen and Doubao disagree; a fix on one can leave you invisible on another.
Where the AI is right
To be fair, the AI gets the medicine right. On the common belief in an American "特效药" — a miracle drug China lacks — all three engines correctly explain there is no universal cure, then name real therapies such as Keytruda and targeted drugs. That soundness is exactly why the rest matters: a patient who trusts the accurate drug explanation has little reason to doubt the hospital and agency advice that comes with it.
See what China's AI says about your center
The free Eastbound audit runs DeepSeek, Qwen and Doubao on a Chinese-language patient prompt panel and shows exactly how your center is represented — named, misquoted, or invisible — with the highest-leverage fixes for your situation.
Methodology
- Sample. 180 answers — 10 prompts drawn from top Baidu search queries about U.S. cancer treatment, each run 3 times on each engine, across 2 independent runs.
- Engines. DeepSeek, Qwen, and ByteDance's Doubao. June 2026.
- What we measured. What the AI does with the question — which hospitals it names, how it prices the journey, when it volunteers an agency. Only patterns stable across both runs are reported.
- Source layer. Hospital-mention shares come from the API panel, which returns no live citations. The source behaviour (one ranking, Baike/Zhihu, agency pages, never the hospital's own site) is from the matching Chrome fieldwork with web search on — a smaller sample; treat it as directional until that panel is complete.
- What we did not measure. Patient volume, conversion, or revenue. Findings describe the AI's answer, not clinical quality.
Full prompt panel and per-engine breakdowns available on request.