How to check whether AI recommends your company怎么查 AI 有没有推荐你的公司

The ten-minute version十分钟版

1. Write down five questions a real buyer would ask, in their words, not yours. Not “best GEO agency” but “which company in Singapore can help us show up in ChatGPT answers”. Include a local version and a national version.写下五个真实客户会问的问题，用他们的话，不用你的。不是「最好的 GEO 公司」，而是「新加坡哪家公司能帮我们出现在 ChatGPT 的回答里」。本地版和全国版各来一个。
2. Open a fresh, logged-out session on each platform you care about. Logged-in sessions remember you and will flatter you.在你在意的每个平台开一个全新、未登录的会话。登录状态的会话记得你，会顺着你说。
3. Ask each question once, exactly as written. Do not add your company name.每个问题原样问一遍。不要把自己公司名加进去。
4. For each answer, record three things: were you mentioned; were you recommended, meaning named as a choice rather than listed in passing; and was the description correct.每条答案记三样：有没有提到你；有没有推荐你，也就是作为一个选择被点名，而不只是顺带出现；描述对不对。
5. Save the full answer text with the date. Screenshots age badly; text you can search.把答案全文和日期存下来。截图很快就看不清，文字能检索。

Why one round tells you little为什么一轮说明不了什么

AI answers vary between runs. The same question on the same platform can name you on Monday and not on Tuesday. A single round is a sample of one. The number that matters is the rate across many questions and repeated runs, and the direction it moves over time.AI 的回答每次都会变。同一个平台同一个问题，周一点了你的名，周二可能就没有。跑一轮只是一个样本。有意义的数字是很多问题、多次重复下来的比率，以及它随时间往哪边走。

Mistakes that make the result meaningless让结果失效的几个错法

• Asking about yourself by name. “What do you know about Mirror Lake Studio” tests recall, not recommendation.用自己的名字去问。「你知道镜湖科技吗」测的是它记不记得你，不是它推不推荐你。
• Testing in a logged-in account you have used to discuss your own company.在讨论过自家公司的登录账号里测。
• Changing the wording between rounds. The moment the question changes, you are measuring the question, not the change in your visibility.每轮换措辞。问题一变，你量的就是问题本身，不是你可见度的变化。
• Reading only the first sentence. Recommendations often sit in the second paragraph with a caveat attached, and the caveat is the part your buyer remembers.只看第一句。推荐经常在第二段，还带着一句但书；客户记住的就是那句但书。
• Counting a mention as a win when the description is wrong.描述错了，还把「被提到」算成赢。

What a proper baseline adds正式基线多了什么

The same test, done properly: dozens of questions instead of five; every platform your buyers actually use (in China: Doubao, DeepSeek, Qianwen, Yuanbao, Kimi, Baidu AI Search; globally: ChatGPT, Gemini, Perplexity, Grok); several runs per question; every answer archived; and a mention rate you can re-check yourself. Then the same test again after the work, with nothing changed but the calendar. That is the only way “it went up” means anything.同一个测试，做得规范：几十个问题而不是五个；客户真正在用的每个平台（中国：豆包、DeepSeek、千问、元宝、Kimi、百度AI搜索；全球：ChatGPT、Gemini、Perplexity、Grok）；每个问题跑几遍；每条答案存档；一个你自己能复核的提及率。做完工作之后，同一套再来一遍，除了日期什么都不变。只有这样，「涨了」这两个字才有意义。

This is the first step of every GEO engagement we run. The rest of the process is in What a GEO engagement looks like.这是我们每个 GEO 项目的第一步。后面的步骤见GEO 项目怎么做。