NiubiGEO is a self-hosted checker for how AI models describe your product, plus a ¥99 API report and a separate human test of the real chat sites.
Three different things under one name
The about page says the work started as the ProGEO open-source project, and that the question has stayed the same: when someone looks for a product like yours, can they find it and understand it. The public repository now is Albert-Weasker/niubigeo. The README lists NiubiStar as the sponsor, and the site sells diagnosis, human tests, publishing, and quoted campaigns through that network. You can run the checker yourself, or pay the team to do a slice of the work.
- The checker is Apache-2.0, tagged v0.2.0. It wants Node.js 22 or newer and your own OpenRouter key. Clone that tag, run npm ci, put the key in .env, then npm run server and open http://localhost:8787. The source is free. Model, search, and hosting bills are yours.
- The limitations note shipped with that tag says the self-hosted app has no login, access control, or TLS, and that parts of the interface are hardcoded in Simplified Chinese. It is a tool for a machine you trust.
- On the website, confirming the profile card is free. The homepage says that card only confirms product facts. It is not the AI diagnosis. The ¥99 test starts after you confirm the card.
- When this page was written, the GitHub repository listed 4,855 stars and 0 forks.
What the ¥99 report actually contains
The diagnosis page prices one report at ¥99. That fee covers API testing and the write-up. The test is at most 3 questions times 3 models, once per combination. You get the original model answers, and the links those answers cited. If an answer has no source, the report says so. Human tests on a website or an app are a separate purchase.
- Did an AI mention the product?
- Did it describe the product wrong?
- Which other products showed up?
- Which sources did the answer cite?
- Which product facts are still unclear?
- What should be fixed first?
The pricing FAQ says you do not top up a balance before a report or a human test. Sign in, fill in the requirements, and pay with Alipay or NewPay. A custom campaign needs a confirmed quote before work starts. Terms version 2026-09-18.1 say there are no automatic charges. They also say a mention applies only to that question, that time, and that platform, and that NiubiGEO does not promise a fixed ranking, a sales figure, or a citation that stays.
The checker keeps the answer, not a score
The open-source page says the tool watches provider API responses, which can differ from what a person sees in a consumer chat website or app. The README says a few minutes of repeated tests do not establish long-term growth, and that publishing an article does not guarantee an AI recommendation. One answer is not a permanent ranking. Failed runs stay on record.
Twenty public cases can be read without installing anything. Each domain has at least one analyzable answer, and 11 of the cases also ran a keyword test. In a Prototyping test that did not name Figma, two offline answers mostly explained the concept. The answer that requested web search named Figma. The README quotes GPT-4.1 mini: "Figma's prototyping tools make it easy to build and share high-fidelity, no-code, interactive prototypes." It treats a brand mention, a positive description, and an explicit recommendation as different events.
In the notion.so domain test, models described the product as notes, a workspace, or collaboration, and they named different competing products. The README says recognizing a domain after being asked about it is not the same as recommending it when nobody typed the name.
The screenshot on this page is their archived PostHog domain-recognition run from September 8, 2026. The limitations note says that study used offline openai/gpt-4o-mini, offline google/gemini-2.5-flash-lite, and web-enabled openai/gpt-4.1-mini, with a budget cap of 2 US dollars for preflight, failures, retries, and search. In the Feature Flags run, one cited source was a Splunk blog post. Provider citations are stored apart from ordinary links in the answer text. Three measurements were taken close together, one of them on a schedule. The README says those records show repeated testing, not long-term growth.
People clicking the real chat app are a different order
Human AI testing sends a person to the website or the app. You choose the platforms, the locations, and the questions. The delivery is the answer, a screenshot, the sources, and the conditions: platform, place, language, and whether it was the site or the app. The price is the number of questions times platforms times regions times repetitions. Checkout shows the current unit price, how many tests you will receive, and the total, before you pay. The pricing page does not print that unit price.
A TechMoon article last updated on 2026-09-10 described a different menu: ¥5 for one human AI test, ¥350 for one written and published article, ¥8,888 for 30 articles, and ¥28,888 for 90 articles plus 5 videos of real people. It also described a credit top-up, minimum 50, with no online payment and orders by email. The pricing page checked for this listing does not print those amounts. It lists the ¥99 API report, a human-test total calculated at checkout, Alipay or NewPay, and no top-up for those two orders.
Publishing is priced per site. Writing and revisions are included, or you can supply a draft. The article goes out after you approve it. A campaign that mixes trials, communities, creators, and content is quoted for that plan. The terms say the service must not be used to impersonate an independent customer, invent trial feedback, or submit false facts.
Two headcounts, both theirs
The homepage prints 150K+ human participants, 62K+ customers, 5.1M+ task deliveries, and 5,029 publishing sites to choose from. The network page says NiubiStar's global network is 100,000 people. Those figures are theirs. The homepage also shows an anonymized participation record and says not to read that example as sales growth, or as growth in AI recommendations. The about page names no founder. Account and order questions go to the project dashboard, or to [email protected].