Google's new AI optimization guide. Mostly right.

Google quietly published its AI optimization guide recently. The headline position is direct.

Optimizing for generative AI search is optimizing for the search experience, and thus still SEO.

At the headline level they are right, and I have spent most of 2025 saying the same thing. AI Overviews and AI Mode share similar ranking pipeline as classic Google search. The same fundamentals (content quality, indexing health, freshness, links, entity clarity) carry over. Most of what works in classic SEO carries directly into AI surfaces.

Where the guide reads thinner is once you remember Google is one of eight major AI organisations crawling the open web. Their advice is right for Google’s surfaces and does not always generalise.

Schema is called out in the guide as “not required for generative AI search”. Technically correct at retrieval time. But schema is still entity infrastructure feeding Google’s Knowledge Graph, which sits upstream of Google’s own AI Overviews. The full argument is in The Three Lives of Schema Markup. It is registration, not advertising. You add it for the entity foundation, not the citation boost.

Their position on llms.txt is similar. Right for Google. The actual observed use case as of mid-2026 is narrower than most agent-readiness advice implies. The file is fetched today by coding agents whose system prompts have a URL hardcoded, with Claude Code fetching platform.claude.com/llms.txt as the canonical example. Chatbot products grounding answers at query time in your llms.txt is not observably happening, regardless of what the spec or the visibility tooling claims. The fuller breakdown is in How to Make Your Website Agent-Ready under the llms.txt section.

The bit that made me actually laugh is the line that you do not need to create new machine readable files, AI text files, or Markdown. The AI optimization guide page itself is served as Markdown at developers.google.com/search/docs/fundamentals/ai-optimization-guide.md.txt. Google’s own infrastructure produces and ships a Markdown version of the same content their guide tells you not to bother with. Their public guidance and their internal infrastructure are pulling in different directions.

The last claim worth examining is that natural human writing is enough. That works as advice against keyword-stuffing for robots, but it understates what happens before any AI answer gets generated. The retrieval step that runs first is a similarity calculation. Pages that commit to one idea per section and name their subject explicitly tend to land closer to the user’s query than pages that hedge across three topics in the same paragraph. Any open embedding API will show you this on your own content. Writing well for human readers and writing in a way the retrieval step can resolve cleanly are not the same job, even if they overlap.

My Cloudflare logs show requests from 8 distinct AI crawlers in the last 7 days. OpenAI, Anthropic, Perplexity, ByteDance, Meta, Apple, Manus, plus Google. Google’s guidance may be correct for Google. The work on the other seven is what AI SEO actually means as a discipline above and beyond classic SEO.

This is the framing in our Snippet Digital play on AI SEO strategy. AI SEO is a thin layer on top of classic SEO, not a replacement. Google’s guide covers the classic SEO part well. The protocol layer (robots.txt with explicit AI bot rules, llms.txt, Markdown negotiation, MCP and A2A discovery, WebMCP, Content Signals) plus retrieval-aware writing is the additional layer.

The longer breakdown of the full protocol layer is now live in How to Make Your Website Agent-Ready.