John Mueller and Martin Splitt, members of the Google Search Relations team, discussed mood coding websites on a recent episode of Unofficial research.
Both discovered that AI coding tools could quickly produce working sites. But successful SEO still required specific technical direction, the same kind you would give to a human developer.
Telling the AI to “add SEO”
Mueller compared the experience of vibe coding to working with a developer who doesn’t specialize in research.
Mueller said on the podcast:
“You can always tell the AI system, now add SEO to it. But how it works, if you go to a developer and add SEO and it’s like, what do you mean. Sprinkle in some meta tags and add some structured data.”
Vague instructions produce vague results, whether the constructor is a human or an AI. Mueller said he got better results by telling the system what it wanted up front. This included the domain name, canonical configuration, sitemap files, and a robots.txt.
It checked whether the pages used reasonable HTML and were linked correctly. He also implemented pre-publish checks to verify that URLs returned content and that JavaScript files were not blocked by robots.txt.
What they built
Mueller created test websites to see how Googlebot handles requests. He deployed them to Firebase hosting using Hugo as a static site generator, with GitHub for version control.
He recently moved from VS Code with Copilot to command line tools. He named Claude Code and Gemini CLI as what he currently uses.
Splitt tried Google AI Studio to create a client-side tool with JavaScript. He described the result as readable, looking like a standard Next.js application. But he found himself in a loop where the AI continued to use a library it didn’t want.
Split said:
“I asked him for half an hour. I tried to stop him from doing what he wanted to do and wanting to do what I wanted to do. And it was weird.”
The question of technical knowledge
Both recognized the tension in vibe coding’s promise that you don’t need to know how to code.
Mueller noted that technical understanding is helpful at every step. Knowing what type of site builder you want and how to structure pre-publish checks produced better results. Without this context, the AI will make assumptions. They can choose a static site generator, a JavaScript-heavy setup, or a full CMS with a database backend.
Mueller said:
“All of these assumptions are reasonable. If you talk to a developer, they will also make these assumptions. But if you just tell the AI system I want a website, it will choose one.”
For personal projects and low-risk static sites, the stakes are low enough to experiment. But for anything involving user data or a production department, Mueller added that you’ll want someone who understands what they’re doing.
Vibe-encoded sites and search visibility
The sites that Mueller built produced reasonable HTML that wouldn’t stand out for its flavor code.
“In practice, no one can really recognize that it’s an ambiance-coded website,” he said, adding that common ambiance coding frameworks can leave recognizable patterns.
He also highlighted a content risk. Once a site looks polished, it’s tempting to let the AI write the content as well. Mueller acknowledged that the tool can do that, but said that’s not where he sees the most value.
Splitt agreed. AI-written content raises the question of why someone would visit a site instead of talking directly to the AI.
Mueller has previously reported similar deficiencies in mood-coded sites.. He reviewed an ambient-coded Bento Grid Generator on Reddit. It identified crawling issues, outdated meta tags, and content stored in JavaScript files that search engines couldn’t access.
Looking to the future
The podcast did not include formal advice or policy positions on the mood-coded sites. Mueller and Splitt shared what they tried and what they encountered.
For people testing these tools, the message is that AI can handle parts of code generation well, especially for low-risk projects. He does not make SEO decisions himself. For this, you always need someone who knows what to ask.
Featured image: YouTube.com/GoogleSearchCentral, May 2026.
