In short: Google's core message is one idea. Its AI features run on its usual Search systems, so SEO stays the foundation. Many "GEO/AEO hacks" are pointless for Google: no need for llms.txt, no need to chunk content, no need to over-invest in structured data. And the guide names the Universal Commerce Protocol for agents.
What Google published, and why it is rare
On May 15, 2026, the Google Search Central team published an official guide: how to optimize a site for Search's generative AI features, namely AI Overviews and AI Mode. On a topic where most online advice comes from SEO vendors, a first-party stance signed by Google sets the reference. The core message fits in one line: these features are rooted in Search's ranking and quality systems, so SEO is not dead, it is the bedrock.
How Google's AI search works
The guide names two mechanisms worth understanding before optimizing anything:
- RAG (retrieval-augmented generation), or grounding: to answer, the model relies on Search's ranking systems to retrieve relevant, up-to-date pages from the index, then generates a grounded response with clickable links to those pages. In other words, being well indexed and well ranked stays the prerequisite.
- Query fan-out: the model generates several related queries in parallel to fetch additional results. A question about "how to fix a lawn full of weeds" may trigger sub-queries about herbicides, chemical-free weeding, prevention, and so on.
What Google recommends doing
The best practices are SEO best practices, reframed for AI:
- Valuable, non-commodity content. Google stresses this will matter most over the long run: a unique point of view, first-hand experience, people-first content that does not just recycle what already exists or what a model could produce on its own.
- A clear technical structure. The page must be indexable and snippet-eligible, crawlable, offer good page experience, reduce duplicate content, and follow JavaScript SEO best practices. Semantic HTML helps (notably for accessibility) without being mandatory.
- Local and ecommerce details. For products and businesses, Merchant Center (and its feeds) and Google Business Profiles help you appear in AI responses as well as other results.
The mythbusting: what is pointless to do (for Google)
This is the most interesting part, because it contradicts several common "GEO recipes." For Google's AI search:
- No need for llms.txt or "special" AI files. Google may crawl them, but gives them no special treatment. Important: llms.txt keeps its value for other LLM crawlers (ChatGPT, Perplexity, Claude); it is simply not a Google lever.
- No need to "chunk" content. Breaking content into tiny pieces for AI is not required; Google understands multiple topics on one page. There is no ideal page length.
- No need to rewrite "for AI." The systems understand synonyms and intent; no need to chase every long-tail variation.
- No inauthentic "mentions." Seeking inauthentic mentions across the web does not help; ranking favors quality and filters spam.
- No over-investing in structured data. No schema.org is required for AI search. Keep it for classic rich results, not as an AI lever.
Agents and UCP: Google opens the door
The guide devotes a section to agentic experiences. AI agents, including browser agents, access sites by analyzing the visual rendering (screenshots), the DOM structure, and the accessibility tree. Google points to "agent-friendly" website best practices and, notably, names the Universal Commerce Protocol (linking to ucp.dev) as an emerging protocol that will let Search agents do more. For a merchant, the signal is clear: agent legibility is no longer speculative, Google writes it into its own SEO documentation.
Our read, for an operator
Three concrete takeaways. First, SEO foundations beat GEO recipes: non-commodity content, clean technical structure, well-kept product data. Second, be honest about llms.txt: useful for third-party LLM crawlers, but not for Google; presenting it otherwise targets the wrong engine. Third, the agentic layer (UCP, MCP, AP2) is the real forward investment: Google now cites it explicitly, and that is where tomorrow's advantage is built.
Frequently asked questions
Does Google use llms.txt?
Not for its AI search. Google says you do not need to create files like llms.txt. They can be crawled, with no special treatment. llms.txt stays useful for non-Google LLM crawlers.
Is structured data required for AI?
No. No schema.org is required for generative AI search. Keep it for eligibility to classic rich results.
Is SEO dead?
No. Google's AI features rely on its ranking systems. SEO remains the foundation of visibility, AI included.
Sources
- Google Search Central, "A new resource for optimizing for generative AI in Google Search" (May 15, 2026)
- Google, official guide "Optimizing your website for generative AI features on Google Search"