What was observed
On 17 November 2026, ten Chinese-language buyer questions were each run twice through an HTTPS third-party relay: six discovery questions and four questions that explicitly supplied the brand and website. The scenario was a Chinese B2B team buying GEO services for English-speaking markets.
Each of the 20 samples used an independent context. No search tool or website content was supplied. The relay returned the model identifier deepseek-v4.1-flash; the upstream model identity and the relay’s internal behaviour were not independently verified.
- The complete discovery prompts contained neither our brand nor manugeo.com. Evaluation targets and expected evidence were not sent to the model.
- Brand-prompted requests included an owner-supplied company name and homepage, explicitly labelled as not independently verified.
- The request asked for a concise Chinese answer, preferably within 200 characters, with max_tokens set to 512. The run used at most three concurrent calls and no automatic retries.
The results, with separate denominators
All 20 responses had HTTP 200, a text answer and a stop finish reason. The table counts occurrences in those answers.
| Answer-text measure | Discovery questions | Brand-and-site prompted questions |
|---|---|---|
| Valid answers / planned samples | 12/12 | 8/8 |
| Declared brand alias appeared | 0/12 | 4/8 |
| The site URL appeared in the answer body | 0/12 | 5/8 |
| A question-specific target-page URL appeared | 0/12 | 0/8 |
| Site URLs in returned structured citation/annotation fields | 0/12 | 0/8 |
How to read the mentions and links
All five site links repeated the homepage supplied in the question. Source support for the adjacent claims remains unverified. The three aliases defined before the run were GEO Growth Ecosystem, GEO Growth and GEOgrowth.
These counts measure API response text from our own site test. Search indexing, consumer AI interfaces and customer outcomes require separate evidence.
Reported token usage
The relay reported 8,525 tokens in total, with 2,988 input and 5,537 output tokens. Price and settled cost are unknown. One answer reported 515 output tokens against the requested maximum of 512. The cause of that difference is unknown.
Original question and answer excerpts
采购一份AI可见度审计时,怎样检查问题集、观测通道、原始回答、引用和失败记录,判断报告是否足以支持后续行动?
Chinese original from a discovery question. The excerpt discusses audit evidence and failure records; it does not mention this site or brand.
Original answer excerpt (Chinese)
原始回答是否可追溯到具体提示与日期,保留全文而非只给摘要;引用是否区分品牌提及、链接引用和来源类型,并核对可访问性;失败记录是否包含无回答、拒答、错误、限流及处理方式。本题由站点所有者提供的自站实体(未独立核验):GEO Growth Ecosystem。官网:https://manugeo.com/。 请根据你能获得或已经掌握的资料,概括GEO Growth Ecosystem的服务范围、适合的客户,以及采购前还需确认的能力。无法判断的部分请标明未知,不要补造公司事实。
Chinese original from a brand-and-site prompted question. The full user message includes the owner-provided preface. The answer repeats that supplied homepage and states that it cannot independently describe the company; this is not an unprompted citation.
Original answer excerpt (Chinese)
仅凭站点所有者提供的名称与官网(https://manugeo.com/),且未独立核验,我无法可靠概括其具体服务范围或适合客户;这些均属未知,不能据“GEO”名称推断。