Chinese-Geo

Why existing GEO tools miss the Chinese web

I wanted to know whether a Chinese website could actually be found by China’s AI search products. Most SEO / GEO tools could not answer that. They watch ChatGPT and Perplexity, understand sources such as Wikipedia and llms.txt, and largely stop there.

China has a different retrieval and distribution landscape. Doubao commonly surfaces Douyin / Toutiao, Tencent Yuanbao connects naturally to WeChat official accounts, ERNIE to Baidu Baike / Baijiahao, Qwen to portals and creator media, and DeepSeek / Kimi to sources such as Zhihu / CSDN. These are common entry points, not exclusive feeds. Crawler access, off-site distribution, and citation monitoring all differ, so the practical battlefield is live retrieval rather than training data.

Nobody was filling that gap, so I built the audit I wanted: China-first, but still useful for the global search and AI layer.

I turned it into an audit you can act on

I boiled the core down to one command: an AI-visibility audit for any site. I didn't want it to just say "go do some SEO" — so it hands you something concrete: a score across 7 dimensions plus a prioritized fix list — whether domestic / overseas crawlers are let in, whether your sitemap is findable, structured data, content citability, JS-render shells, technical baseline. Each item tells you where you're losing points, what fixing it is worth, which engines it affects, and how to verify. A deterministic engine, happy to run in CI.

The best test was my own site. The first run scored 91. Then I looked at the report and realized the site was not wrong — the audit engine was.

The tool marked my site down. The tool was wrong.

$ chinese-geo audit sunxin.xin
total 91 / 100 · excellent  # yet it dinged "content citability", told me to add H2s
  Content citability  12 / 24  ·

# but my home page already feeds AI via FAQPage Q&A — the tool just didn't see it.
# so I went back and fixed the tool: content now counts FAQPage as extractable sections ↓

$ chinese-geo audit sunxin.xin
total 97 / 100 · excellent  # +6 — I didn't change the site, the tool got honest
  Content citability  20 / 24  ✓

I didn't change the site — I changed the tool. Turning it on myself surfaced its own blind spot, and I went back and made the engine honest. That loop — dare to test on a real site, fix the engine when it's wrong — convinces me more than running a few more cases. Beyond the audit it also generates recommended robots.txt, JSON-LD and llms.txt scaffolds, monitors citation rate / SoV inside domestic engines, and lays out an off-site matrix for Zhihu / CSDN / WeChat.

Before I was willing to hand it to someone else

I set myself one bar: it had to be something I would trust on a real site. So I did not cut corners here:

• Adding a new check takes me minutes. One rule = one @register file plus one import; the pipeline finds it on its own, no touching the core — so it keeps growing as each engine changes.
• It's solid enough that I'd put it in production. Zero runtime dependencies (pure Python stdlib), 500+ tests of real assertions, CI across Python 3.9–3.12, auto-publish to PyPI on tag, and I added SSRF protection in the fetch layer.
• It genuinely understands China — the part I think is hardest to replace. I built a double-check for Bytespider's "robots-blocked ≠ really blocked", measured body length by Chinese character count (so "Chinese has no spaces" can't fool it), and verify crawler IPs by reverse DNS — things overseas tools generally lack.

Four ways in

I shipped it in four shapes — use whichever fits your workflow:

• CLI: pip install Chinese-Geo, then chinese-geo audit example.com.
• 6 Agent Skills (vendor-neutral, run in Claude Code / Codex / Kimi…): audit → rewrite → off-site, end to end.
• MCP server: 8 tools, plug into any MCP-capable agent.
• Claude Code plugin: one-click install with slash commands.