On May 15, 2026, Google published "Optimizing your website for generative AI features on Google Search" under a brand-new "Generative AI fundamentals" section of Search Central. It is the company's first consolidated, on-record statement about what actually works for visibility inside AI Overviews and AI Mode, consolidating positions that had previously been scattered across conference talks, podcast appearances, and blog posts into one reference document. It has been updated as recently as June 29, 2026, meaning Google is actively treating it as a living document, not a one-time announcement.
Two things happened simultaneously that made this guide land harder than a typical documentation update. First, it explicitly names a set of tactics as things site owners can stop worrying about. Second, it does this while an entire industry of AEO and GEO vendors has spent two years selling exactly those tactics as differentiators. Reading the guide as "Google says AEO is fake" is the wrong takeaway, and reading it as "nothing has changed" is equally wrong. This article walks through what the guide actually says, section by section, and reconciles it honestly against the technical AEO practices covered throughout the rest of this series.
What Does Google Actually Say About AEO and GEO as Terms?
The guide addresses the terminology directly: "AEO" stands for answer engine optimization and "GEO" for generative engine optimization, both describing work focused on improving visibility in AI search experiences. Google's position is that, from Search's perspective, optimizing for generative AI search is optimizing for the search experience, and is therefore still SEO — not a separate discipline requiring a separate playbook, because Google's generative AI features are built on the same core ranking and quality systems as classic Search.
This is a real, substantive claim worth taking seriously rather than dismissing as marketing. AI Overviews and AI Mode use retrieval-augmented generation to pull content from Google's existing Search index. A page that cannot pass Google's core ranking systems — crawlability, indexability, basic quality signals — is not a meaningfully different citation candidate just because someone labels their optimization work "AEO" instead of "SEO." The GEO vs AEO vs SEO guide in this series makes a similar point from a different angle: the disciplines overlap enormously, and framing them as three separate programmes competing for budget is usually a mistake regardless of what Google says about it.
What Four Tactics Does the Guide Say to Ignore?
This is the section generating the most industry reaction, and it deserves to be quoted precisely because paraphrases have already distorted it in both directions.
llms.txt and AI-specific markup files. Google states plainly that you do not need to create machine-readable files, AI text files, AI-specific markup, or Markdown versions of pages to appear in generative AI search. Google's crawler may discover an llms.txt file, but it treats it like any other text file — no special treatment, no preferred indexing pathway. The guide adds an important caveat that most summaries drop: other crawlers may make use of such files. This is not a blanket statement that llms.txt is worthless everywhere. It is a statement that Google specifically does not privilege it. The llms.txt mechanics guide in this series already made this exact distinction before Google's guide existed: llms.txt has documented value for Bing-based retrieval (which powers a meaningful share of ChatGPT's live search) and for agentic evaluation flows, and no documented value for Google specifically. Nothing in the May 15 guide contradicts that; it confirms it from the other side.
Content chunking. There is no requirement to break content into small, AI-digestible pieces. Google's systems can understand multi-topic pages and extract the relevant passage without the author pre-fragmenting the article. Danny Sullivan made comparable remarks in January 2026, citing conversations with Google engineers who actively recommended against chunking. This is a meaningful, specific claim about Google's own extraction pipeline, and it should genuinely inform how you write for Google's AI surfaces: you do not need to slice a comprehensive guide into artificially short, isolated sections purely to make it "chunk-friendly" for Google.
AI-specific rewriting for long-tail keyword variants. AI systems understand synonyms and general meaning. You do not need to obsessively capture every long-tail phrasing a user might type. This tracks directly with how the embedding and semantic-matching mechanisms covered in the RAG pipeline article in this series actually work — semantic similarity scoring means content written naturally around a concept already matches a wide range of query phrasings without needing exhaustive keyword variant coverage.
Special schema or Markdown versions of pages. Not required for inclusion in Google's generative AI search specifically. This is the single most consequential line in the guide for anyone who has read the rest of this series, and it deserves a direct, honest response rather than a defensive one.
How Does This Square With the Schema Guidance Elsewhere in This Series?
Directly, and without hedging: Google's official position is that schema is not required for AI Overview or AI Mode inclusion specifically. The extensive body of independent research cited throughout this series — including the correlation data showing FAQPage, HowTo, and Review schema associated with higher AI citation rates across ChatGPT, Perplexity, and Google's own AI surfaces — is describing a correlation, not a confirmed causal ranking factor that Google has ever claimed to apply. Google's own guide is explicit that its systems extract meaning from unstructured page content directly, without depending on structured data to disambiguate it.
Two things can be true simultaneously. Google's own generative AI features may not require schema to identify and extract a citable passage. And schema may still measurably improve citation rates on other engines — Perplexity, ChatGPT, Claude, Gemini — whose retrieval architectures are less mature, less capable of unstructured extraction, or built on entirely separate indexes (Bing, in ChatGPT's case) with different quality signals. The correlation studies referenced throughout this series were never claiming to isolate Google's AI Overview behavior specifically; several were measuring citation behaviour aggregated across multiple engines, and several were specifically measuring Perplexity and ChatGPT. Schema work should be understood, going forward, as a bet on the non-Google half of the AI search landscape, plus a genuine trust and disambiguation signal at the entity level covered in the entity disambiguation guide, rather than as a lever pulling on Google's AI Overview ranking specifically.
What Does the Guide Say Actually Matters?
The guide's central, repeated message is that the foundational SEO work you were already supposed to be doing is what determines generative AI visibility: technical crawlability, adherence to Search Essentials and spam policies, and creating content that is genuinely helpful, reliable, and people-first. Google states this can be simplified to one test: would your visitors find this content satisfying? If the answer is yes, the guide suggests you are on the right track, since Google's systems are designed to connect people with exactly that kind of information.
The guide places specific, repeated emphasis on non-commodity content — a genuinely important thread that deserves its own dedicated treatment given how directly it validates the first-party research and original-analysis approach covered elsewhere in this series. It also confirms, in its own words, that both AI Overviews and AI Mode use a query fan-out technique, issuing multiple related searches across subtopics as they build a response, and that this fan-out process surfaces a wider and more diverse set of supporting links than a classic web search would. This is Google's own documented confirmation of a mechanism that determines exactly how a page ranking outside the top ten for a broad query can still earn a citation through a narrower fan-out sub-query — worth understanding on its own terms, covered in a dedicated companion piece in this series.
The guide also confirms that spam policies apply fully to generative AI responses, not just to classic Search results — a clarification specifically added to remove ambiguity about whether AI Overview content operates under a separate, more permissive standard. It does not.
What Does the Guide Say About AI Agents and Agentic Commerce?
A forward-looking section references agent-friendly best practices and emerging standards like the Universal Commerce Protocol and WebMCP, which let AI agents take actions on behalf of users directly from search results. Google frames this section explicitly as optional: "if this is something that's relevant to your business and you have extra time, check out the available agentic experiences." This framing — optional, forward-looking, not urgent — is a useful calibration for the agentic web checklist covered elsewhere in this series. Google is signalling the direction of travel without claiming these standards affect current-day AI Overview or AI Mode visibility.
What Should a Real AEO Programme Actually Change After Reading This Guide?
Nothing drastic, but three genuine recalibrations are warranted. First, stop treating llms.txt, chunking, and schema as if they are Google-ranking levers, and start being explicit internally about which engine each tactic is actually intended to influence — schema and llms.txt for Perplexity, Claude, and ChatGPT's Bing-based retrieval; foundational SEO and genuine content quality for Google's AI surfaces specifically. Second, treat any vendor or internal pitch that frames AEO purely as a checklist of markup and files with real scepticism; the highest-leverage work, according to both Google's guide and the independent multi-engine research cited throughout this series, is producing content with a unique point of view and first-hand evidence, not accumulating more machine-readable files. Third, keep the schema and structural work already covered in this series in place for the engines where independent data shows it correlates with citation lift — the guide only speaks to Google's own systems, and four other major engines are not addressed by it at all.
How NotionCue Helps You Separate Engine-Specific Guidance From Universal Guidance
The single hardest part of operating an AEO programme in mid-2026 is knowing which piece of advice applies to which engine. Google's May 15 guide is authoritative for Google's own AI Overviews and AI Mode. It says nothing about Perplexity, ChatGPT, Claude, or Gemini, each of which runs its own retrieval architecture with its own documented preferences.
The NotionCue Citation Tracker is built specifically to make this distinction visible in your own data rather than in the abstract. By tracking your citation rate separately across all five engines on the same weekly cadence, you can directly observe whether a structural change — adding FAQPage schema, for instance — moves the needle on Perplexity and ChatGPT while producing no measurable change on Google AI Overviews specifically. That pattern, if you see it in your own tracked data, is exactly what Google's own guide predicts, and it tells you the schema investment is working as a non-Google-specific tactic rather than a wasted effort.
The NotionCue AI Answer Gap Finder helps you act on the guide's core emphasis on non-commodity content by surfacing the specific queries where competitor content — not just any content, but content demonstrating first-hand experience or original data — is currently winning citations you are not. That gap data is a more useful guide to where your next piece of content should go than any generic markup checklist.
Start your free NotionCue trial and compare your citation rate trend across Google AI Overviews against your citation rate trend on Perplexity and ChatGPT over the same period. If your structural AEO work is producing a visible lift on the latter two and a flat line on the former, that is not a failure — it is exactly what Google's own May 2026 guidance predicts, and it tells you precisely where that investment is paying off.
Google also published a related but separate document on June 5, 2026 — guidance specifically for evaluating third-party SEO advice and tools, which explicitly names AEO and GEO as legitimate services an SEO practitioner might offer, while giving businesses a framework for auditing whether a given vendor's specific claims are grounded in Google's own documentation or in unverifiable proprietary "AI ranking factors." That document deserves its own dedicated read before hiring or renewing any AEO vendor relationship, and it is covered in full in the next article in this series.
Frequently Asked Questions About Google's May 2026 Generative AI Optimization Guide
Does this guide mean FAQPage schema is now useless?
Not useless — just not a Google-specific ranking lever for AI Overview inclusion, according to Google's own stated position. FAQPage schema retains documented correlation with citation rates on Perplexity, ChatGPT, and other engines covered throughout this series, and it retains its E-E-A-T-adjacent value as a disambiguation and content-structure signal. Google's own guide is explicit that it may crawl and index such files without penalty; it simply does not confirm they receive preferential treatment for its own AI features specifically.
Should I stop building llms.txt files given Google's guidance?
Only if your traffic and citation goals are exclusively about Google's AI surfaces. If ChatGPT, Perplexity, or agentic evaluation flows matter to your business — and the data throughout this series suggests they should for most brands — llms.txt retains documented value there. Google's guide is precise and limited in scope: it describes Google's own crawler behaviour, not the behaviour of every AI system that might visit your site.
Is this guide likely to be updated again, and how should I stay current?
Yes — it has already been revised at least once since its May 15 publication, with a "last updated" timestamp of June 29, 2026 visible on the live document. Treat it as a living reference rather than a one-time announcement, and check it directly at developers.google.com/search periodically rather than relying solely on secondhand summaries, since exact wording changes have already occurred and are likely to continue as Google's generative AI features evolve.