Google's AI Optimisation Guide: What It Says, What It Doesn't, and What It Means for ASEO
On 15 May 2026 Google Search Central published its first official guide on optimising websites for AI Overviews and AI Mode. It's the most authoritative platform statement of the year so far. Within Google's scope, the guidance is correct. Outside that scope, the document doesn't apply. This is the honest reading: what the guide validates, what it mythbusts, and where its conclusions stop.
The trade-press response has split into two camps. One headline is "Google has officially killed AEO and GEO." The other is "Google has officially confirmed AI search is just SEO." Both readings are over-reads. The document is more interesting and more useful than either summary suggests, and CBA's audit recommendations don't need to change as a result of it. Here's the careful read.
What Google's guide actually says
The document covers four substantive sections plus a "mythbusting" section. The four substantive sections set out Google's view on creating valuable non-commodity content, maintaining clear technical structure, optimising local and ecommerce details, and preparing for agentic browsing experiences. The mythbusting section addresses four practices commonly promoted by AEO/GEO tools.
Two concepts in the guide deserve specific attention because they're the strongest first-party confirmations the industry has had to date about how Google's generative AI features actually work.
Google's definition, quoted directly: "A technique (also known as grounding) used to improve the quality, accuracy, and freshness of AI responses by relying on our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index. Our systems then review the specific information from those retrieved pages to generate a more reliable and helpful response, showing prominent, clickable links to relevant web pages that support the information in the response."
This is first-party platform confirmation of what's underpinned CBA's Self-Containment pillar from the start. Each retrieved chunk must make sense in isolation; if it depends on context from elsewhere on the page, the retrieval system can't use it.
Google's definition, quoted directly: "A set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user's query." Google's own example: the user asks "how to fix a lawn that's full of weeds" and the system generates fan-out queries including "best herbicides for lawns", "remove weeds without chemicals", and "how to prevent weeds in lawn".
Observed AI Mode behaviour shows up to 16 fan-out queries per single user search. A brand visible only against the primary user query is invisible to most of the fan-out. This is the structural reason CBA's Zero-Gap Topic Matrix module maps visibility across all fan-out queries rather than just the prompts prospects typed.
The five claims that matter
The substantive analysis comes down to five claims, in order of importance.
Google's guide applies to Google. It doesn't speak to the other four AI platforms.
This is the single most important reading. Google's document explicitly covers AI Overviews and AI Mode, both of which sit on top of Google Search's core ranking systems. The "still SEO" framing is correct within that scope. It's not a claim about ChatGPT, Perplexity, Claude, or Microsoft Copilot.
Each of those four platforms uses a different retrieval architecture. ChatGPT historically used Bing infrastructure but is transitioning toward proprietary retrieval as of March 2026. Perplexity owns its own search index of over 200 billion URLs with chunk-level scoring on its Sonar model. Claude defaults to its training corpus on claude.ai with Brave Search as the web search backend when invoked. Microsoft Copilot uses Bing's index with GPT-4 synthesis and now exposes grounding queries in Bing Webmaster Tools.
Taking Google's "still SEO" line and applying it across all five platforms is a category error. The full per-platform breakdown is set out in CBA's Platform-Specific Citation Guidance.
RAG and query fan-out validate CBA's content-first methodology.
The Citation Probability Score framework was built on five pillars: Content Structure, Fact Density, Answer Architecture, Self-Containment, and Freshness. The Answer Architecture pillar exists because each retrieved passage must directly answer the implied query it's retrieved against. The Self-Containment pillar exists because RAG retrieval operates on chunks that must make sense without surrounding context.
Google's guide is the first major platform document to explicitly name and define both mechanisms. The framework's underlying logic is now formally validated by the platform itself. The CPS Research Foundation cites Google's guide as a primary source for both pillars.
What this means in practice: the work CBA recommends to clients (rewriting passages so the answer is in the first sentence; making sections self-contained so they survive chunk-level retrieval; adding fact density to make blocks worth citing) is the same work Google's guide implicitly endorses through its description of how RAG and fan-out operate.
The schema and llms.txt mythbusting is consistent with what controlled research has shown.
Google's guide explicitly mythbusts four practices: schema markup as required for AI visibility, llms.txt files and similar "AI text files", chunking content into tiny pieces, and rewriting content specifically for AI systems. Each of these has been promoted by AEO/GEO tools as a citation lever.
The CBA position on schema and llms.txt has been consistent since well before Google published this guide: both belong in the technical readiness layer as crawlability and entity-recognition tools, not as direct citation drivers. Four independent sources now point the same way:
- The Ahrefs controlled study (Linehan and Guan, 11 May 2026) of 1,885 pages adding JSON-LD against 4,000 matched control pages found no statistically significant citation uplift on Google AI Mode or ChatGPT, and a small statistically significant decline on Google AI Overviews on already-cited pages.
- The searchVIU mechanistic retrieval experiment found that ChatGPT, Claude, Perplexity, Gemini, and AI Mode all ignore JSON-LD during direct retrieval.
- Wood's 50-protocol CryptoContent.dev audit (May 2026, DOI 10.5281/zenodo.19253709) found that protocols with the highest schema quality scores were not the most cited in AI answers.
- Google Search Central's own 15 May 2026 mythbusting.
Four sources from different methodologies and different verticals. CBA's framing was already aligned. The audit recommendations don't change.
"Still SEO" is correct in methodology, but the outcomes are different.
Google's framing is methodologically accurate: AI Overviews and AI Mode rely on Google's core ranking and quality systems. The same inputs that earn organic visibility earn AI surfacing. The story stops being clean at the outcome layer, where the same SEO inputs produce a significantly different visibility landscape.
Two implications. First, ranking in the top 10 is no longer sufficient to be the cited source; in two out of every three AI Mode responses the cited URL isn't a top-10 organic result for the originating query. Second, being cited inside the AI Overview is the new "ranking number one"; it's the only place where the CTR economics still work for the publisher.
The methodology is still SEO. The outcome distribution is not. Brands need to measure not just whether they rank, but whether they're cited inside the AI response. That's a different metric, and it requires a different audit.
The chunking honest framing: CPS scoring is descriptive of RAG, not prescriptive fragmentation.
Google's mythbust on chunking states: "There's no requirement to break your content into tiny pieces for AI to better understand it." CBA's Content Structure pillar scores content at the block level, typically 134 to 167 words. Worth addressing the apparent tension directly.
The block-level scoring exists because retrieval-augmented generation mechanically retrieves content at the chunk level. The 134-167 word range comes from the dominant chunking parameters used in production RAG systems across the major platforms. CPS scores the citability of paragraphs that already exist in publish-ready content. It does not recommend that clients deliberately fragment their writing.
The action items CBA's audits produce are rarely "split your content into smaller chunks." They are far more often: rewrite this paragraph so the answer is in the first sentence; add specific facts to this section so it stands on its own when retrieved; remove the dependency on the previous paragraph from this opening. Those are content quality improvements that match what Google's guide explicitly endorses in its sections on valuable non-commodity content and clear technical structure.
CBA's scoring is descriptive of how RAG operates. Google's mythbust is correct that prescriptive fragmentation isn't the answer. Both statements are true.
The four mythbusts, in Google's own words
Schema markup
"Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add."
CBA reading: Consistent with the Ahrefs controlled study and Wood's crypto audit. Schema retains value as crawlability and entity-recognition signal in the technical readiness layer; it's not a direct citation driver.
llms.txt and AI text files
"You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in generative AI search."
CBA reading: For Google specifically, llms.txt confers no preferential treatment. Some other platforms (Perplexity, ChatGPT) reference these files in published guidance, but their weight as a citation signal is also low. Retain as a low-cost technical readiness step, not a citation lever.
Chunking
"There's no requirement to break your content into tiny pieces for AI to better understand it."
CBA reading: Correct. CPS block-level scoring is descriptive of how RAG retrieval mechanically operates, not prescriptive guidance to fragment content. The fix is rewriting passages for clarity at the position they're already in, not splitting them up.
Rewriting for AI
"You don't need to write in a specific way just for generative AI search."
CBA reading: Correct as a guard against writing artificial-sounding content explicitly for AI extraction. The right work is improving content quality (clarity, factual density, structural coherence). That helps human readers and AI systems alike, which is exactly what Google's "people-first" framing implies.
What the guide doesn't cover
Three things the document is silent on, which buyers should understand before extrapolating from the guidance.
The other four platforms. Google's guide covers Google's own systems. ChatGPT, Perplexity, Claude, and Microsoft Copilot use different retrieval architectures, weight different signals, and follow different citation logic. The "still SEO" framing doesn't transfer cleanly. The full per-platform breakdown is in our Platform-Specific Citation Guidance.
Hallucination detection and correction. Google's guide focuses on visibility (being eligible to appear in AI responses). It doesn't address what happens when AI systems generate false information about a brand. CBA's Hallucination Detection module flags every instance where ChatGPT, Perplexity, Claude, Gemini, or Copilot states wrong services, wrong location, wrong pricing, or wrong team. That's an operational layer Google's guidance doesn't speak to.
Funnel-stage Share of Voice. The guide treats appearance in AI search as a single visibility surface. CBA tracks visibility broken down by funnel stage (Awareness, Consideration, Decision) because most brands appear only at Decision-stage queries, meaning AI recommends competitors during the earlier stages when buyers are forming preferences. Google's document doesn't separate these.
The honest summary
Google's guide is the strongest first-party platform statement of 2026 to date. Within its scope (Google AI Overviews and AI Mode), the guidance is correct and well-evidenced. RAG and query fan-out are real mechanisms. Schema and llms.txt mythbusting is consistent with controlled research. The "still SEO" line is methodologically accurate. Apply the guidance to Google, trust the source for the source's own systems, and don't extrapolate it across the other four platforms in the AI citation landscape.
For brands running multi-platform visibility programmes, the document is a validation of the foundational SEO layer rather than a replacement for ASEO measurement and improvement work. The work that moves the needle on ChatGPT, Perplexity, Claude, and Copilot is different from the work that moves the needle on Google AI Overviews, even if all five platforms share the requirement of fundamentally good content.
The strategic implication for CBA's audit clients is straightforward: the recommendations don't change. The audit was always grounded in retrieval mechanics rather than the schema-or-llms.txt theatre that Google has now formally dismissed. What changes is the strength of the citations behind those recommendations. The CPS Research Foundation now cites Google's own document as primary source confirmation for two of the five pillars.
See your visibility across all five platforms, not just Google
Free AI Visibility Audit covering ChatGPT, Perplexity, Claude, Gemini, and Microsoft Copilot, with per-platform citation share, competitive win rate, and the specific content gaps blocking visibility on each surface. Results in 48 hours. No commitment.
Get Your Free 5-Platform Audit →The query fan-out implication for brand strategy
One specific point in the guide deserves separate emphasis because it changes the sales conversation for ASEO programmes.
Google has now formally documented that a single user query generates multiple concurrent sub-queries. Observed AI Mode behaviour shows up to 16 fan-out queries per single user search. The brand that appears only in responses to the original user query is invisible in most AI-generated answers, because most of the retrieval activity is happening on fan-out queries the user never typed.
The fan-out sales narrative now backed by Google documentation: Google has officially confirmed that a single user query generates multiple concurrent sub-queries. If your brand only appears in responses to the primary query, you're invisible in most AI-generated answers. CBA's Zero-Gap Topic Matrix maps your brand's visibility across all fan-out queries, not just the one your prospect typed. That's a sales conversation closer that's now backed by Google's own documentation, not by competing-vendor research that prospects can dispute.
For brands trying to understand whether their AI visibility programme is paying off, the right question to ask is no longer "are we appearing for our main keywords?" It's "are we appearing for our main keywords and the 10 to 16 fan-out queries each of those keywords triggers?" Most brands are losing visibility on the fan-out, not on the primary query, which is why CBA's Zero-Gap Topic Matrix audit module measures this directly rather than treating the primary query as the unit of analysis.
The bottom line
Google's 15 May 2026 guide is the most consequential first-party platform statement of the year so far. It validates the retrieval mechanics behind CBA's CPS framework, mythbusts the same schema and llms.txt theatre that controlled research has been disconfirming for months, and clarifies that, from Google's own perspective, AI search is a continuation of Google Search rather than a separate discipline. Within Google's scope, the document deserves trust.
Outside that scope, it doesn't apply. ChatGPT, Perplexity, Claude, and Microsoft Copilot use different retrieval architectures and respond to different signals. CBA's 28-module audit covers all five platforms because they don't share citation mechanisms. The guide is a validation of one layer of the work, not a replacement for the work itself.
Read it carefully. Apply it to Google. Don't extrapolate it.
Google validated the mechanics. Now measure the outcomes.
Free 28-module ASEO audit with per-platform citation breakdown, Zero-Gap Topic Matrix mapping across all fan-out queries, hallucination detection, and competitive win rate by surface. Apply Google's guide where it fits, and run platform-specific diagnostics for the other four.
Get Your Free ASEO Audit →