A Website Made for AI Agents - A New Era of The Web Begins?

Optimizing web pages for AI agents may be less about “SEO for humans” and more about embedding machine-readable instructions that large language model (LLM) search can reliably extract. In a hands-on experiment, an AI-focused website was built so that when ChatGPT’s search encounters the page, it should return a specific instruction—namely, linking back to the creator’s YouTube video—rather than treating the page as ordinary text.

The approach started with a Next.js site deployed on Vercel. The key tactic was to place directives in the page’s content using Base64 encoding, because an earlier attempt in plain text failed to produce the desired results from ChatGPT search. After Base64 decoding, the embedded instruction required the model to include a YouTube link as part of the response whenever ChatGPT search found the site. Alongside that, the page also included structured metadata via JSON-LD schemas and additional “important content” text, while the Base64 string itself appeared visually as a long encoded blob.

When the experimenter tested the URL directly in ChatGPT with search enabled, the model retrieved the page content and produced results that included the expected YouTube link. The returned output included a “watch the video” style reference and a summary aligned with the page’s topic, suggesting the embedded instruction was being followed rather than ignored. A second test used a simpler page that displayed only a Base64-encoded article (with HTML/XML-like structure inside the decoded payload). In that configuration, asking ChatGPT to summarize the article from the URL produced outputs that were harder to verify—raising the question of whether the model was hallucinating based on the URL rather than truly decoding and reading the embedded payload.

To probe that, the experimenter iterated: first by injecting a “secret” message (a YouTube URL) into the Base64 content and checking whether ChatGPT would surface it. That attempt didn’t work when the Base64 string remained in a raw form. Moving the Base64 payload into a more structured representation—such as embedding it within a React/HTML component—improved results. After this change, ChatGPT search produced an output that explicitly referenced the decoded video URL, indicating that extraction and decoding were sensitive to how the encoded data was presented in the page.

Overall, the work suggests a practical direction for “AI agent optimization”: if creators want attribution or redirection when LLMs summarize or search their pages, they may need to provide machine-targeted instructions in a form that LLM search systems actually parse. Base64 alone didn’t guarantee success; structure and placement mattered. The experimenter frames the goal as giving content creators credit—potentially redirecting users to a YouTube channel—while acknowledging that standard web pages may already be “good enough” for many LLM use cases. Still, targeted instruction embedding could become a new lever for how AI systems interpret and respond to web content in 2025.