A 401-star Chinese GEO tool: how many of those stars are real

GEO Doctor (Fangx-AI/geo-doctor) is an open-source AI assistant skill that finds out why Doubao and DeepSeek don't recommend your product. Built 23 days ago, it has 401 stars on GitHub. But on Oct 23, its star count dropped from 156 back to 22. Going through its fork list shows that most of the curve was built from fake accounts, with only a handful of real users. Below: what it is, where the stars came from, whether this space still has room, and what to learn from it and avoid if you want to build a similar open-source project.

A simple "please give me a Star"

2026 Oct 15: Fang Xin (online handle 「方鑫三个金」) wrote two things in the 191nd entry on his personal site. One: Xiaohongshu cover images are finally no longer a headache. Two: he open-sourced Geo-doctor. The entry says: "I'm frantically learning GEO. I've already landed a few clients, but since I'm not confident in my results yet, I haven't taken any jobs." The last line: "If everyone could give me a Star, that'd be even better. I'd be very happy and deeply grateful."

Fang Xin makes content on WeChat Official Accounts, Douyin, and Bilibili, and also keeps an open-source repo documenting his one-person company. Its stars have climbed slowly day by day and have passed 700. GEO Doctor is a repo he created in Oct 14 and released as v20 in Oct 0.9.0.

What it is

A skill you install into AI assistants like Claude Code, Codex, Kimi Code, and WorkBuddy, plus an npx command-line tool and a local web dashboard. The README's front page says: "Make Doubao, DeepSeek, ChatGPT, and Claude willing to recommend your product." Installation is also simple: copy one sentence and send it to your AI:

Help me install this skill: github.com/Fangx-AI/geo-doctor/tree/master/skills/geo-doctor

Once installed, say "help me do GEO," and it runs eight steps: diagnosis → audit your website (can AI crawlers get in, is the body text in the raw HTML) → organize fact cards → generate a set of sample questions → ask Doubao, Yuanbao, Qwen, DeepSeek, and ChatGPT → compute a "competitor recommendation ranking" and your recommendation rate → write content in batches to fill the gaps → publish (your website and dev.to are auto-posted; for Zhihu and Juejin it fills in the article and you click publish yourself) → recheck two weeks later. If the repo is on GitHub, you can also turn on "autopilot": every Monday it samples and writes automatically and opens a PR for you to review.

It does one thing well: it insists you ask inside the apps yourself. The README explains: "The answers you see in the App are what your users see; API answers don't log in, mostly don't browse the web, and can differ from the App." Each question is asked 2 times, and it tracks the rate of appearance, not a single ranking.

That curve

From the GitHub numbers captured each day:

(The 11-06 row was looked up live on GitHub that day.)

A normal open-source project usually has only a fraction of its stars as forks. Fang Xin's one-person company repo works this way: it gains 11 to 26 stars a day, with forks at about 7% of stars. GEO Doctor had equal stars and forks on day one, both at 28, and then on Oct 23 both nearly dropped to zero together. Of Fang Xin's repos, this shape appears only in this one.

Going through the fork list (GitHub API, 11-06):

• The list has 268 forks, of which 164 belong to accounts that no longer load, either banned or deleted. 268 minus 164 gives the 104 shown on the page.
• Of the remaining 104 accounts, 97 were registered in October of 2026 years ago, and only 4 are older accounts.
• They arrived in batches: 10-17 brought 28 within one hour, 10-18 brought 50 within one minute, 10-21 brought 50 within four minutes, and 10-29 brought 100 within two minutes. The first three batches total 128 accounts, and none of them remain.
• Two of the surviving accounts I opened were registered in late October, have 0 followers, own repos named dapper-moss-lite and fuzzy-anchor-mini, and fork a few popular projects as cover. Bot accounts usually look like this.

The list of stargazers isn't visible to outsiders. But stars and forks rose and were purged together, so the stars most likely came from the same batch of accounts. This is an inference. There's no evidence of who paid for it. People who inflate counts sometimes also star unrelated projects as cover, but they usually pick popular ones and rarely keep returning to a repo that's three days old and barely known.

The same goes for the 51 PRs in the repo: 50 were opened and merged by the author himself, most are branches opened by an AI agent, and outside contributors submitted only 1 English translations, none of which have been merged yet.

Where the real users came from

Setting aside the inflated part, the visible buzz from outside is very small:

• An aggregator account on X, Geek Lite (@QingQ77), posted an intro 10-22 with 1,833 views. Search X for "geo-doctor" or "GEO 医生" and you'll find nothing else.
• A few skill navigation sites and "awesome" lists have listed it, including agentskillshub, awesome-seo-geo-skills, and Trendshift.
• There's a real user's article on Juejin, posted by 11-05, titled "Shiji Evidence Desk GEO Test: After 100 Points on the Technical Audit, Doubao Still Doesn't Find It Independently." The author used GEO Doctor on his own small tool as a fact-check site. The audit scored 100 points with all 14 checks passing, but when he asked Doubao by tool name alone, it found the tool 7 times out of 0; when he described the need, it recommended it 10 times out of 0. His own conclusion: "These checks support the judgment that 'the current request can retrieve the body text and the technical configuration is basically in place,'" but that doesn't mean AI will come looking for you.

This article is probably the most valuable public feedback on the project: someone really uses it, and uses it seriously. It also shows that between "passing the audit" and "being recommended by AI" there is still a long way to go. HN, Reddit, and the media have no coverage of it.

Why the author built it

The log is blunt about it. The 10-23 post says: "For the vast majority of small and medium businesses, they simply can't be found on the internet," and "comparing with today's mini-program promotion, this charges 1000 yuan a month. Before, Doubao couldn't find his mini-program at all; now it can." The next day he started preparing a 200k-word GEO white paper and a 24-episode course.

So this repo is the storefront for his GEO service. If you're thinking about open source for stars, be clear on this: the star count here can't be used to prove that Chinese GEO tools have real demand.

Is this opportunity crowded?

Very crowded, and the Chinese and English sides are far apart:

(Star counts are GitHub numbers between late October and 11-06.)

The last line is the most telling. Another project also pitched as "Doubao, Qwen, DeepSeek, Yuanbao first" only got 87 stars. In other words, "Chinese AI first" alone doesn't earn stars. Among Chinese GEO skills, a few have hundreds of real stars and only niubigeo reaches thousands. On the English side, several top projects have long passed 10,000 stars.

If it were you

C You can learn from it, but copying a "Chinese GEO skill" to chase stars isn't worth it: most of its stars look bought, comparable projects have only tens to hundreds of real stars, and the English side already has several top projects above 10,000.

Who it suits: people already doing GEO, SEO, or content ops who have real client cases and can show data like "before the change Doubao recommended us 0 times, after it N times"; or people who already have a Chinese developer following and can promote the project without relying on GitHub organic traffic.

Who it doesn't suit: people who can only code and hope a polished README and trending will bring stars. There are already a dozen projects like that, all with nice READMEs. People who want to jump straight into the English GEO space, where 27k-star and 12k-star projects are already in the way. People who want to buy stars to build momentum: in October 23 GitHub wiped 134 stars in a single day, and the traces are visible to anyone in the fork list.

How to do it: don't build another eight-step all-in-one skill. Go after the segment the others do badly. The hands-on test on Juejin already names the gap: a diagnostic score of 100 and Doubao still can't find the brand. What's missing is a small tool that automatically and repeatably samples Doubao, Yuanbao, and Qwen in their apps, archives the answers, and compares before and after (right now people do this by hand), or a public dataset of "what we changed and how much the recommendation rate moved." Other GEO projects would be happy to cite either one.

How to get users: write hands-on test articles in Chinese for Juejin, Zhihu, and WeChat Official Accounts, with numbers in the headline, in the style of "After a technical check scored 100, Doubao still didn't find us on its own." Then submit to lists like awesome-seo-geo-skills and skill directories. For the English version, post on X and dev.to and ride the discussion around big projects like claude-seo.

Window: before the end of 2026. If Doubao or Yuanbao open a backend for "brand visibility in AI", or a big project like claude-seo adds Chinese AI sampling, this gap closes.

What would mean I'm wrong: ① GEO doctor keeps gaining dozens of stars a day on its own, even without the batches of new accounts showing up in the fork list; ② a Chinese peer like HeiGe-GEO-SEO or recomby-geo organically passes 2,000 stars within a month or two, which would mean Chinese developer demand is bigger than it looks; ③ Doubao or Yuanbao ships an app-side interface that can be called in bulk and matches app answers, at which point "app sampling" is no longer a gap.