Google just published its official guidance for optimizing for generative AI features in Search.
Most people will read it as: “keep doing what you’re doing”. That’s true, but it’s not the interesting part.
The more important signal is that Google wants AI search understood as an extension of Search, not a new discipline. AI Overviews, AI Mode, generative responses, all grounded in the same index, crawling systems, and quality signals SEO has always depended on.
These central blog posts are directional, not operational. They communicate principles, not system behavior. Criticizing this one for not revealing the algorithm misses the point, because Google never does that.
What’s worth noting is the gap that the guidance can’t close. AI search still has no equivalent to the Search Quality Rater Guidelines ecosystem. No visibility diagnostics. No transparent citation systems. No retrieval reporting. No AI-specific quality frameworks.
That’s why the industry still feels messy. The operational understanding will come from community experimentation, patent analysis, citation tracking, and eventually, I suspect, from proceedings like the DOJ trial.
The real shift is retrieval, not rankings
The old SEO question was: Can I rank for this keyword?
The emerging one is: Can my content be retrieved, understood, reused, and cited across many different query paths and interfaces?
That’s a meaningful shift. What matters now is whether a site becomes reliably associated with an entire topic space, not whether a single page holds a position for a single query variation. A site that’s difficult to crawl, weakly structured, or generically over-optimized is harder to rank. It’s also harder for AI systems to retrieve and reuse with confidence. That second part is what’s new.
The llms.txt conversation is still early
Google’s guidance is unambiguous: llms.txt is not required for Search, not a ranking factor, not needed for AI Overviews or AI Mode.
For Google-specific SEO, that’s fair.
But the same guidance opens a different door, namely agentic experiences. Google explicitly acknowledges browser agents that access websites by analyzing visual renderings, inspecting DOM structure, and interpreting the accessibility tree, and points to UCP as the direction that’s developing.
Shopify is already rolling out /llms.txt, /agents.md, and /sitemap_agentic_discovery.xml across stores. Then there was the moment Google Search Central itself briefly exposed an llms.txt file before walking it back. That captures where things stand: standards still forming, platforms experimenting in public, and even Google isn’t always sure which side of the line it’s on.
For Google Search, the file is unnecessary. For the wider agent ecosystem, the process of building one is worth exploring, not as a shortcut, but because it forces cleaner retrieval structure across important pages, category logic, content hubs, and entity relationships.
I’ll be testing this on my own site and will share results.
To make the process easier, I built llms.txt Architect, a custom GPT available in the ChatGPT directory. Feed it a sitemap, crawl export, or URL list and it selects the most important pages, removes low-value URLs, and organizes everything into a clean, structured llms.txt and llms-full.txt that helps AI systems better understand, navigate, and retrieve content from your site.
Worth doing regardless, because the process of cleaning and prioritizing your URL structure is useful on its own.
Till next time 👋
Ilias