Back navigation iconWriting

How I Make My Portfolio Easier for AI Agents to Understand

A new kind of visitor

I have spent a lot of time making my portfolio clear for people. I also started asking how an AI system might understand who I am, what I build, and where to find evidence of that work.

People often describe this area with terms like AEO and GEO. The terms are useful shorthand, but they can make the work sound more mysterious than it is. I started with a practical question: can a system find useful pages on my site and understand what they say?

I treated this as a site clarity problem. I wanted to make the important information easier to find and interpret, without changing the experience for people browsing the portfolio.

What was already there

When I reviewed the site, several useful pieces were already in place. Pages have titles and descriptions. The homepage includes structured data about me and the site. A sitemap lists the main writing, project, and craft pages, and robots.txt points crawlers to it.

I had also created an llms.txt file with a short overview and links to important sections. Separately, the site has a Markdown endpoint that returns maintained summaries for a small set of routes, including the homepage and selected project pages. It is a hand-written alternate representation, not a conversion of every page on the site.

This is not a before-and-after experiment. These features were already part of the site. I’m documenting what exists, checking how the pieces fit together, and being clear about what they can and cannot tell me.

Make content legible

The first step is useful page content. A clear title, a direct summary, descriptive links, and sensible headings help people scan a page and give automated systems context. Structured data can describe a page, author, or organization in a format machines can process.

On my portfolio, writing page metadata comes from each article, and writing pages include BlogPosting structured data. The homepage describes me as a person and connects the site to my public profiles. These details reinforce information that is also present in the page itself.

I think of these foundations as more important than any one AI-specific file. They make the site understandable on its own, and provide consistent signals across pages.

A plain-text path

The llms.txt file is a small, human-readable index intended to help language models navigate a site. Mine introduces the portfolio, points to its main sections, and suggests where to look for different kinds of information. It is a guide, not a control panel. Publishing it does not guarantee that a model will fetch it, use it, or mention the site in an answer.

The Markdown endpoint offers another path to selected content. Its summaries are maintained separately from the rendered pages, so they need to stay in sync when the site changes. I see this as a useful format to test and maintain, rather than assume every agent will request or prefer it.

I have not added an llms-full.txt. A larger file could collect more content in one place, but it would also create another copy to keep current. I would add it only if it solves a real navigation or reading problem for this site.

What I can measure

I can check whether important pages have useful titles and descriptions, whether they appear in the sitemap, whether structured data matches visible content, and whether the Markdown endpoint returns its maintained summaries. These checks show whether the setup is present and internally consistent. They do not measure AI visibility.

Research on generative search studies how content changes can affect whether sources appear in generated answers. Vercel’s account of building AEO tracking also shows why evaluation can include observing agent searches and cited sources. Those are useful ways to think about measurement, but neither gives me evidence that my own portfolio is being surfaced more often.

For me, making a portfolio easier for AI systems to understand starts with making it clear for everyone. Good pages remain the foundation. Files like llms.txt and alternate text formats are supporting paths into that work, and I’ll keep evaluating them without treating their presence as a visibility guarantee.

Further reading

These references helped me think about generative search visibility and how to evaluate agent behavior.