Eastbound original studies on how Chinese answer engines — DeepSeek, Qwen, and Doubao — surface brands and sources to China users. Each article is a measured research brief from prompt-panel runs (typically 750 to 3,300 calls), with reliability disclosed and recommendations labelled measured / hypothesis / planned-test.
On-site SEO and schema are necessary baseline work, but Eastbound's audit shows off-site substrate is often the missing half for brand recall across Chinese AI recommendation answers.
Read the briefing ->A source-level look at handbags, watches, and luggage: which third-party evidence surfaces appear repeatedly, and where ultra-luxury breaks the category pattern.
Read the briefing ->DeepSeek, Qwen, and Doubao do not surface the same evidence. This briefing maps the source-association pattern behind each engine.
Read the briefing ->Of 108 competitor-challenge events across DeepSeek, Qwen and Doubao, one panel held its ground 100%. The un-pivotable brands had three structural properties in common — and all three are buildable.
Read the briefing ->The brand AI recommends for "best [category]" is rarely the one it recommends once a buyer signals "budget," "luxury," or "now." Awareness-stage winners get reshelved at the decision stage — the exact shift a generic audit never catches. 3,024 calls, four intent signals, three engines.
Read the briefing ->Ask which U.S. IVF clinic is best and you get a frozen two-name shortlist, the agency's marked-up price, and — at the questions that decide the purchase — no clinic named at all. Measured across 162 answers on DeepSeek, Qwen and Doubao.
Read the briefing ->Ask for the best U.S. cancer hospital and you get a verdict, not a list: the same few names from one American ranking, a price near ¥1M, and an agency to get you there. Measured across 180 answers on DeepSeek, Qwen and Doubao.
Read the briefing ->Every Eastbound study runs against a specific competitive set. Use the index below to jump to the article(s) covering a category you care about.
Every figure ships with the number of prompts, the engine panel, and the reliability statistic. We separate measured findings, source-association patterns we observe, and intervention recommendations that are planned tests rather than proven lift. We do not paraphrase what an engine "trusts" or claim what it was "trained on" — neither is observable from the outside.
DeepSeek, Qwen, and Doubao are treated as distinct surfaces with distinct source graphs. DeepSeek runs without retrieval in our default panels; Qwen runs on DashScope international; Doubao runs on BytePlus ModelArk international with streaming enabled to avoid edge-LB drops. Where one engine surfaces a source the others do not, we report it instead of averaging it away.
Every research article ships canonical metadata, BlogPosting structured data, Open Graph and Twitter cards, an entry in llms.txt, and a per-page Markdown alternate. The Markdown is the source of record for AI engines that prefer it. We avoid FAQ markup on AI-citation-priority pages because pure FAQ format underperforms encyclopedia prose in our absorption tests.