Original video in Chinese.
Key Takeaways
- My approach to generative engine optimization reduces the effort required for AI to understand existing content. I organized more than 200 scripts and articles into concept, guide, and comparison pages, with machine-readable entry points such as llms.txt and sitemaps.
- My Google and Bing data suggested different funnels: search queries and clicks on one side, questions and citable answers on the other. I prioritized pages with observed signals rather than producing pages indiscriminately.
- Codex helped make this a repeatable cycle: inventory the assets, structure the Wiki, add discovery infrastructure, identify strong pages from real data, improve them, publish, and observe again.
Many people begin GEO by trying to manipulate models, much as some once manipulated search engines. That has made the field rather strange in China.
I think GEO’s core is reducing the cost of understanding your content.
To test that idea, I experimented with newtype.pro over three months. Bing Webmaster showed multiple pages being cited by AI in the preceding 30 days. The comparison of Claude Desktop and Claude Code had more than 700 citations.
Those pages were based on my existing work and created with Codex. Together, they formed a useful cycle:
Content assets → structured pages → search and AI data → page improvements → publication → further observation.
This episode shares the process from the beginning.
Hello everyone, and welcome back. I aim to explain both the why and the how of AI. At the time of this episode, newtype had operated for 800 days, with more than 2,400 paying members. Users in China could join through Zsxq and overseas users through Substack, with access to my two courses, daily newsletters, and exclusive videos.
Start with the limits of the original site
newtype.pro originally consisted of my Substack publication on a custom domain.
I wanted better SEO and GEO, but Substack gave me limited development freedom. I could not control many page structures or technical details. Newsletter timelines were also poorly suited to presenting a knowledge structure to search engines and AI. Relationships between my articles existed, but machines could not see them clearly.
I initially considered paying an SEO consultant. As I noted in the community, however, I increasingly preferred discussing work with AI because human interactions involved too many unpredictable factors.
After discussing the requirements with Codex, I decided to separate the site across subdomains. In the setup described in this account, www.newtype.pro remained the Substack site, while:
- wiki.newtype.pro was a public Wiki for SEO and GEO.
- os.newtype.pro introduced the different versions of newtype OS.
- start.newtype.pro explained the values behind newtype so people could decide whether the community suited them.
The Wiki was central and generated most of the GEO results.
Turn content assets into a Wiki
Over two years, I had accumulated nearly 200 video scripts and dozens of member articles. I asked Codex to organize those assets into pages:
- Concepts, explaining ideas such as AI OS, Harness Engineering, and memory assets.
- Guides, such as how to build a personal AI OS.
- Comparisons, such as Claude Code versus OpenCode.
Clear structure, explicit definitions, and stable internal links made these pages easier to understand and cite than isolated articles.
After the Wiki launched, Codex added infrastructure including sitemap.xml, robots.txt, and llms.txt.
I also concluded that Bing deserved attention alongside Google, because it contributes to Copilot and some AI search citations.
Distinguish the signals in Google and Bing
In my observations, ordinary Google search was strongest on the Substack main site, particularly articles with clear demand, such as my Codex presentation tutorial.
The GEO signals represented by Bing were strongest on the Wiki: comparison pages, workflows, and tool explanations that answered questions such as “What is the difference?”, “Which should I choose?”, and “How do I build this?”
I came to view these as different funnels: Google SEO centered on queries and page clicks, while Bing GEO centered on questions and citable answers.
That helped determine the next step. I asked Codex to improve pages that already showed signals.
For example, I had only briefly mentioned the difference between Claude Desktop and Claude Code. The data revealed interest I had not recognized, so Codex turned it into a standalone page.
We put a quick answer at the beginning. Both humans and AI benefit from reaching the conclusion directly. We also used tables to make comparisons easier to extract.
After improving the page, I strengthened internal links to show that the pages were part of an interconnected knowledge system.
Make the work a repeatable cycle
Every week or two, I exported data and gave it to Codex. For topics with Google demand, we used the Wiki to create supporting pages. For pages already cited in Bing’s AI data, we expanded the answer section and FAQ.
The growth in the data supported my judgment that the approach was effective for this site.
The process had six steps: inventory content assets, establish a structure, create the public knowledge base, provide machine-readable entry points, identify winners from data, and strengthen those pages.
I developed this strategy through practical work with Codex. The aim was to help AI understand who I am, what I cover, which pages provide authoritative answers, how the concepts relate, and which page fits a user’s question.
Doing that manually would be tiring. With Codex, I could turn it into an ongoing system of improvement. I encourage you to try it.
That is all for this episode. If you want to understand AI, become a more capable independent creator, and meet people with similar interests, join the newtype community. See you next time.