Google’s AI Optimisation Guide (May 2026): An Honest Multi-Platform Reading

Published: 16 May 2026. Responds to Google’s official guidance dated 15 May 2026.

On 15 May 2026, Google published its first official guidance on optimising for AI features in Search. The document lives at developers.google.com under the search fundamentals section, which means it has been through legal and policy review and is now the canonical position of Google Search. It will be cited as authoritative by every SEO commentator, every agency, every vendor white paper for at least the next 18 months — whether or not it answers the questions site owners actually need answered.

This guide is an honest reading of that document. Where Google is right, the guidance is repeated and reinforced. Where the document is Google-specific in ways that affect site owners whose audiences span other AI surfaces — ChatGPT, Claude, Perplexity, Microsoft Copilot, Gemini — the nuance is named. Where the framing reflects Google’s commercial interests rather than neutral technical reality, that is also named, because site owners deserve to know they are reading motivated communication, not a textbook.

Nothing here contradicts the document’s technical claims about Google Search itself. The point is that Google Search is one surface among several where buyers now make decisions, and site owners need a framework that works across all of them.

1. What Google Just Published, and Why It Matters

The document is titled Optimizing your website for generative AI features on Google Search. It explains how Google’s AI Overviews and AI Mode work, what site owners should do to appear in them, and what they explicitly do not need to do. Five sections cover the foundations: applying SEO best practices to generative AI search, building clear technical structure, optimising local and ecommerce details, a “mythbusting” section addressing common misconceptions, and a short pointer to emerging agentic experiences.

Two technical concepts are worth understanding because the document defines them clearly and they are central to how Google’s AI features actually work.

Retrieval-augmented generation (RAG), which Google also calls “grounding”, is the technique that allows AI features to produce reliable, fresh responses by retrieving relevant pages from Google’s Search index and generating an answer grounded in that content. The retrieval step uses Google’s existing ranking systems. The generation step produces a synthesised answer with clickable links back to the source pages. This is the architecture behind AI Overviews and AI Mode, and it is why the foundational SEO work that gets pages indexed and ranked remains directly relevant to AI visibility within Google.

Query fan-out is the practice of generating multiple concurrent related queries from a single user query, then retrieving content for each. Google’s own example: a user searches “how to fix a lawn that’s full of weeds” and the system generates fan-out queries like “best herbicides for lawns”, “remove weeds without chemicals”, and “how to prevent weeds in lawn”. The model then synthesises an answer from results across all of these. This is why topical depth matters — not as keyword stuffing, but as genuine coverage of the sub-topics a user might be implicitly asking about when they pose a single question.

Both of these mechanics are accurate, useful, and worth understanding. The document earns trust by explaining them clearly. The question is what the document chooses to recommend on the basis of these mechanics, and what it chooses to dismiss.

2. Where Google Is Right and Useful

Most of the document’s technical advice is sound and aligns with what experienced SEO practitioners have been saying for years. Three substantive points are worth reinforcing because Google’s validation gives them additional weight.

Non-commodity content is the long-term win

The document’s clearest illustration is its contrast between commodity content (“7 Tips for First-Time Homebuyers”) and non-commodity content (“Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line”). The first is content anyone with internet access could produce. The second is content only that specific buyer, with that specific experience, could publish. Google’s argument is that AI systems — which look across many sources — will increasingly prefer the second category because the first adds nothing new to the corpus.

This maps almost exactly onto the Footprint vs Fingerprint framework: footprint content is reproducible by competitors; fingerprint content is content only your specific entity could credibly publish. Google has now, in effect, validated the underlying principle. The sewer-line example is a good one to keep in mind. Most businesses default to footprint content because it is easier to produce and feels safer. Google’s guidance and the structural mechanics of AI retrieval both push in the opposite direction.

Technical foundations still matter, exactly as they did before

The document is unambiguous that pages must be indexed and eligible to be shown in Google Search with a snippet to appear in AI features. Crawlability, indexability, JavaScript rendering, page experience, duplicate content reduction, the technical SEO baseline — all of it remains directly relevant. Google’s AI features use the same underlying ranking systems that determine traditional results.

There is no separate technical foundation to build for AI. The work site owners already do for traditional SEO is the work that makes AI visibility possible within Google. This is genuinely useful pushback against the “AI changes everything” narrative that has dominated parts of the SEO discourse for the past 18 months.

The “SEO is still SEO” framing has substantive truth in it

Google’s point that optimising for generative AI search inside Google Search is “still SEO” is, in a technical sense, accurate for that surface. The same ranking systems retrieve the content. The same quality signals influence selection. The same indexing pipeline determines eligibility. From Google’s perspective, AEO and GEO are not separate disciplines — they are the same SEO discipline applied to a different output format.

This is technically defensible for Google Search. Whether it is the right framing for site owners thinking across the wider AI ecosystem is a separate question, addressed below.

3. Where Google’s Framing Is Google-Specific

The document’s “mythbusting” section lists four things site owners are told they do not need: llms.txt files, content chunking, AI-specific rewriting, and over-focusing on structured data. Each of these statements is true for Google Search. Each requires nuance when applied to the broader AI ecosystem.

llms.txt: not needed for Google, but Google is not the entire AI ecosystem

Google’s position is that site owners do not need llms.txt files to appear in generative AI search. The document notes that Google may discover, crawl, and index many file types including llms.txt, but does not treat them in any special way. This is accurate for Google Search specifically. Google has a sophisticated retrieval architecture and does not need an additional external signal pointing at curated content.

The standard, however, was never aimed primarily at Google. llms.txt was proposed for AI agents and LLM systems whose retrieval architecture is genuinely different from Google’s — systems that benefit from a curated, structured pointer to canonical content rather than crawling a full site from scratch. Anthropic has engaged with llms.txt positively. Perplexity uses it. Several agentic browsers and emerging AI tools have indicated support. Google saying “we don’t use this” is consistent with Google’s existing capability. It is not a statement about the wider standards landscape.

The honest framing for site owners: llms.txt is not a Google ranking signal. It is a multi-platform standard relevant to non-Google AI systems and emerging agents. Whether implementing it is worth the time depends on which AI surfaces your audience uses. For a B2B business whose buyers research extensively in ChatGPT or Perplexity, the calculation is different from a local services business whose discovery is almost entirely through Google.

Content chunking: not needed for Google, but a useful writing discipline regardless

Google’s position is that site owners do not need to break content into small chunks for AI features to understand it. Google’s systems can parse multiple topics on a page and surface the relevant section. This is accurate.

The deeper point is that “chunking” in the structural sense — writing content with clear paragraph boundaries, explicit definitions, extractable answer blocks, and topic-tight sections — helps both human readers and AI extraction. Not because Google requires it, but because it is good writing practice. The Princeton GEO-Bench study found statistics with full context improve AI citation rates by 41%; that effect is about extractability and self-containment, not about chunking as a tactic. Write paragraphs that can be quoted on their own and you optimise for both Google and every other system that might extract them.

Structured data: useful but not required, per Google

Google’s position is that structured data is not required to appear in generative AI features, although it remains useful for rich results in traditional Search. This is consistent with Google’s position on structured data generally — helpful, not magical.

For non-Google AI systems, structured data plays a slightly different role. It helps machines disambiguate entities, verify claims against entity graphs (Wikidata, Crunchbase, Knowledge Graph), and connect statements to authors and organisations. The work documented in the schema architecture for the AI era guide covers this in depth. Structured data is a corroboration tool more than a ranking signal — useful for any retrieval system that has to decide whether a claim on a page is trustworthy.

AI-specific content rewriting: correct rejection, often misapplied

Google’s position that site owners do not need to write differently for AI is correct and worth repeating. The wave of “rewrite all your content for AI” consultancy that flooded the discourse through 2025 was largely unnecessary. Good writing remains good writing.

The legitimate nuance: AI retrieval systems do reward certain structural patterns more reliably than traditional search did. Direct answers near the top of a page, explicit definitions, statistics with full context, named entities, claim attribution. Not because AI requires special writing, but because AI is better at extracting from well-structured content than humans are at scanning it. The work documented in Anatomy of an AI-Citable Page covers what changes structurally without changing the underlying writing discipline.

4. Reading the Framing: What Google’s Document Serves

This section is the one most SEO commentary on the document will avoid. It is worth including because site owners deserve to read the document as motivated communication, not as a neutral technical reference. Doing so is not an accusation of bad faith. It is recognition that every publisher has interests, including Google.

Google has a substantial commercial interest in keeping the conversation about AI optimisation inside the vocabulary of “SEO”. The reasons are straightforward.

First, Google’s documentation, developer tools, certification programmes, advertising products, and ecosystem of practitioners are all organised around “SEO” as the canonical category. A fragmented field where “AEO”, “GEO”, “LLM optimisation”, and other terms become recognised disciplines distinct from SEO is a field where Google’s organisational dominance is less complete. Asserting that these are all just SEO is consistent with Google’s incumbent position.

Second, the document positions Google Search as the natural home for AI optimisation discussion. The discussion of AI features is framed almost entirely around Google’s own products: AI Overviews and AI Mode. Other AI surfaces are not the document’s concern. Site owners reading it as “the official AI optimisation guidance” rather than “Google’s position on AI optimisation within Google Search” will, by absorption, treat Google as the centre of the AI discovery problem. This is a positioning move whether or not it is consciously intended.

Third, the specific dismissal of llms.txt, content chunking, and AI-specific rewriting weakens the commercial case for the AI-specific consultancies, tools, and methodologies that have emerged around these techniques. Some of those tools and methodologies are flawed; some are valuable. Google’s blanket dismissal does not differentiate. The framing benefits Google’s ecosystem position regardless of which side of that line any specific tool falls on.

None of this makes the document wrong on its technical merits. It does mean site owners should read it as Google’s preferred framing of the AI optimisation conversation rather than as the neutral textbook. The actual technical claims about Google Search are accurate. The framing of those claims as “the AI optimisation question, settled” is a commercial choice.

5. What Google’s Document Does Not Address

The document’s scope is explicitly “generative AI features on Google Search”. That is a legitimate scope. It is also a substantially narrower scope than “how buyers actually discover and choose vendors in 2026”. The territory the document does not cover is where most of the practical AI visibility work for B2B and considered-purchase categories now sits.

The fragmented AI discovery landscape

As of May 2026, ChatGPT processes approximately 250–500 million search-intent queries per week according to Similarweb’s 2026 AI Search report. Perplexity handles around 50 million. Microsoft Copilot handles 80–120 million per week with a heavy skew toward workplace usage. Claude has 157 million monthly visits per Similarweb’s 2026 data. Combined, the non-Google AI discovery surfaces represent a structural category that did not exist four years ago and now influences planning for every serious content programme.

Similarweb’s 2026 Market Research Panel measured how US consumers use AI tools versus search engines at each stage of the purchase journey: AI tools are used by 35% of consumers at the discovery stage versus 13.6% for search engines. AI holds a 2:1 or greater advantage at every stage from discovery through evaluation. The consumer journey, particularly for considered purchases, no longer starts in a search bar.

Google’s document does not address any of this. The omission is not dishonest — the document’s scope is explicitly Google Search — but a site owner reading only this guidance would substantially underestimate the share of buyer activity now happening outside Google’s own surfaces.

Microsoft Copilot and the enterprise Bing pipeline

Microsoft Copilot is built into Microsoft 365 apps — Word, PowerPoint, Excel, Outlook, Teams — and uses Bing as its retrieval index. For any B2B business whose buyers work in Microsoft 365 (which is most enterprise buyers in finance, law, healthcare IT, manufacturing, government, and professional services), Copilot is a work-surface, not just a chatbot. When a procurement team or general counsel asks Copilot about vendor options inside their day-to-day workflow, the retrieval layer is Bing, not Google.

Bing indexing is therefore not optional for B2B businesses targeting Microsoft 365 environments — it is the retrieval layer for Copilot-influenced discovery. This is one of the most under-discussed dynamics in the entire AI visibility conversation, and it is entirely absent from Google’s document. A site owner who follows Google’s guidance perfectly and ignores Bing has solved half of their B2B discovery problem at best.

Entity corroboration across non-Google systems

Google’s document does not engage with entity corroboration as a concept, beyond a passing reference to structured data. For Google itself, the Knowledge Graph and proprietary entity systems handle most of the corroboration work invisibly. For other AI systems, entity facts are typically triangulated across multiple public sources — Wikidata, Crunchbase, Clutch, LinkedIn, Companies House, named editorial mentions — and the consistency of these signals determines whether the system trusts what it reads on a vendor’s own pages.

The Entity Corroboration Model describes the three levels of corroboration: entity-supplied only, partially corroborated, fully corroborated. AI systems tend to retrieve content but cite only entities they can corroborate independently. For Google AI Overviews, the corroboration mostly happens within Google’s own systems. For ChatGPT, Claude, Perplexity, Copilot, the corroboration happens across the open web. Site owners who only optimise for Google’s internal entity recognition will appear in the AI Overview but vanish from ChatGPT and Perplexity citations.

Selection versus ranking

Traditional SEO is fundamentally a ranking problem: given a query, which pages rank highest. AI retrieval is increasingly a selection problem: given a query and an answer, which sources does the system choose to ground and cite. These are different problems with different mechanics.

Pages can rank highly and never be cited. Pages can rank moderately and be cited frequently. The selection criteria — extractability, attributability, entity clarity, evidence quality, structural self-containment — are partially correlated with traditional ranking signals but not identical to them. The CITATE framework exists to make selection readiness measurable independently of traditional ranking. Google’s document treats selection as a downstream consequence of good SEO, which is true within Google’s integrated system but does not describe how non-Google AI systems make their selection decisions.

Agentic readiness beyond protocols

Google’s document closes with a single paragraph on agentic experiences, pointing at the Universal Commerce Protocol and a guide to agent-friendly best practices. This is the most substantively underweighted section of the document relative to its commercial importance. Autonomous AI agents acting on behalf of buyers — researching vendors, comparing options, producing shortlists, executing purchases — are the layer where the next wave of commercial change is happening. The protocol-level work matters; what site owners need beyond protocols is structured signals that allow an agent to reason about capabilities, eligibility, decision criteria, and trust. That is a much larger surface than current Google guidance addresses.

6. Audience-Segmented Guidance: Know Where Your Buyers Are

The right balance between Google guidance and broader AI ecosystem work depends entirely on where a site’s audience makes its decisions. The blanket “optimise for AI” advice common in the SEO discourse fails because it ignores audience segmentation. The following segmentation is the honest version.

B2B SaaS, enterprise software, professional services

Buyers in this category increasingly research extensively before contacting vendors. They use ChatGPT for early-stage exploration, Perplexity for sourced research, Copilot for in-workflow questions, Claude for analytical work, and Google for verification and final checks. The buying journey may pass through three or four AI surfaces before a single page on the vendor’s site is visited.

For this audience, following Google’s guidance is necessary but solves perhaps 30–40% of the discoverability problem. The high-leverage work sits in entity corroboration, structured evidence for non-Google AI systems, Bing indexing for Copilot, and content that is structurally extractable in conversational responses. Microsoft Copilot in particular is under-served by most B2B vendors — the Bing index lag relative to Google is often substantial, and it directly suppresses visibility inside Microsoft 365 workflows.

Law firms and regulated professional services

Legal buyers research extensively before instructing. Increasingly, that research happens in ChatGPT for client-facing questions (“what does this clause mean”), in Perplexity for citation-grounded case research, and in Copilot for in-workflow document analysis. Google still handles a substantial share of immediate-need queries (“criminal solicitor London”), but the upstream research that determines which firms make a shortlist increasingly happens in conversational AI.

For regulated services where trust signals matter disproportionately, entity corroboration across Companies House, Law Society, named editorial mentions, and structured author credentials is high-leverage. Google’s guidance covers the foundation; the trust-and-corroboration layer that AI systems use to decide which firms to cite is largely absent from the document.

Local services and SMB-targeted businesses

Google still dominates local discovery, particularly through the local pack and Google Business Profile. For a plumber, dentist, or accountant serving a specific geography, Google’s guidance covers the majority of practical visibility work. ChatGPT and Perplexity handle local queries less reliably than Google. Voice assistants tied to AI systems are growing but remain secondary for local intent.

For this audience, the document’s advice can be followed closely without much amendment. Bing presence still matters for Copilot users, but the relative emphasis on non-Google AI work is smaller than for B2B categories.

Knowledge workers, technical buyers, AI-native categories

Audiences who use AI tools throughout their working day — developers, analysts, product managers, technical decision-makers — make decisions across a fragmented AI surface as a default mode. For these buyers, Google is one input among five or six. Optimising primarily for Google leaves substantial visibility on the table.

For categories targeting these audiences, the broader multi-platform work — llms.txt implementation, cross-platform entity corroboration, extractability optimisation, agent-readable service descriptions, public structured evidence — is closer to required than optional.

Consumer ecommerce and product discovery

Google Shopping, Merchant Center, AI Overviews with product listings, and Business Agent are the surfaces Google’s document focuses on for commerce. ChatGPT’s Instant Checkout, Perplexity’s shopping integrations, and emerging agent-mediated commerce are not covered. For consumer product businesses, Google’s guidance is the foundation but the agent commerce layer is where the next wave of change is happening — with Universal Commerce Protocol adoption and emerging agentic checkout flows changing how products are discovered and bought outside of search-based journeys.

7. The Practical Action Map

The action map below separates Google-aligned work (where the document’s guidance applies directly) from broader AI ecosystem work (where the document is silent or partial). Site owners should select from each section based on audience segmentation.

For Google AI features specifically — follow Google’s guidance

• Ensure pages meet the Search technical requirements and are eligible to appear with a snippet.
• Follow crawl budget best practices for large sites.
• Use semantic HTML where it serves readability, without obsessing over perfect markup.
• Apply JavaScript SEO best practices if using a JS framework.
• Maintain page experience quality across devices.
• Reduce duplicate content.
• For local businesses: maintain an accurate, complete Google Business Profile.
• For ecommerce: use Merchant Center feeds where appropriate; consider Business Agent for conversational commerce.
• Create non-commodity, first-hand, expert-led content. Avoid recycling common knowledge.
• Organise content with clear sections, headings, and a navigable structure that helps human readers.
• Support textual content with high-quality images and video where they add value.
• Verify the site in Search Console to identify and diagnose issues quickly.

For the broader AI ecosystem — where Google’s guidance stops

• Bing indexing. Verify the site in Bing Webmaster Tools. Compare Bing indexed page counts to Google: significant gaps directly reduce Copilot and ChatGPT Search visibility because both use Bing as the retrieval layer for substantial portions of their grounded responses.
• Cross-platform brand checks. Search the brand name in ChatGPT, Claude, Perplexity, and Copilot. Note where the entity is described accurately, vaguely, or incorrectly. Inconsistencies indicate corroboration gaps. The Entity Corroboration Model is the framework for prioritising fixes.
• Provider visibility checks. Search “best [your category] in [your geography or sector]” in Perplexity, ChatGPT, and Copilot. If competitors are named and the site is not, the gap is in provider-level entity signals, not in content quality. Wikidata, Clutch, Crunchbase, named editorial mentions, and authoritative directory listings are the levers.
• Structured evidence on the site. Named clients (where contractually permitted), measurable outcomes, named authors with credentials, dated case studies, transparent pricing where possible. AI systems triangulate trust across these signals; sites with none are visibility-capped regardless of content quality.
• Extractability audit. Run the CITATE criteria across the most important commercial pages. Where pages fail extraction tests — no explicit definition, statistics without source attribution, author not named, claims without evidence — these are the structural gaps that block citation regardless of ranking.
• llms.txt decision. If audience research indicates substantial ChatGPT, Claude, Perplexity, or agentic browser usage, implement llms.txt as a curated pointer to canonical content. If audience is almost entirely on Google, deprioritise. The standard is not a Google ranking signal; it is a multi-platform signal whose value depends on which platforms matter to the specific audience.
• Schema architecture as corroboration tool. Use structured data not for Google rich results but as the entity-corroboration layer that helps AI systems verify claims. Person, Organization, Article, Service, FAQPage, HowTo where applicable. The schema architecture for the AI era guide covers the full pattern.
• Agent-readiness for vendor sites. For B2B categories where buyers may delegate research to AI agents, ensure capability surfaces are explicit: what you do, what you do not do, who you serve, who you do not serve, pricing or pricing methodology, named outcomes, named integrations. Agents cannot infer; ambiguity gets converted to omission.

8. What to Do This Week, This Month, This Quarter

The volume of recommended action above is substantial. For most site owners, working through it sequentially produces better results than attempting it in parallel. The following sequencing is appropriate for a typical business of small to medium scale, working alongside ongoing content and marketing activity.

This week

• Verify the site in both Google Search Console and Bing Webmaster Tools. Compare indexed page counts. If Bing is significantly lower, this is the first remediation priority because it directly suppresses Copilot and ChatGPT visibility.
• Run brand-name searches in ChatGPT, Claude, Perplexity, and Copilot. Document where the entity is named accurately, named partially, or not named at all. These are the corroboration gaps to address over the coming quarters.
• Identify the top three commercial pages on the site. Read each as if encountering it for the first time: is the offer clear, the audience named, the evidence visible, the author attributed.

This month

• Apply the CITATE criteria to the top commercial pages and remediate the highest-impact gaps: explicit definitions, statistics with source attribution, named authors, dated claims, structured FAQs where appropriate.
• Address the most significant Bing indexing gap (typically a robots.txt issue, sitemap misconfiguration, or canonical chain problem suppressing crawl).
• Audit the Google Business Profile (for local businesses) or organisation-level entity signals (for B2B), filling gaps in named addresses, services, hours, primary category, attributes, and named contact people where appropriate.
• Identify two or three corroboration sources missing from the entity’s public footprint — Wikidata entry, Clutch profile, named press mention, directory listing — and begin the work to add them.

This quarter

• Build at least one non-commodity, first-hand, expert-led piece of content that no competitor could plausibly publish. Document a specific client outcome, a project that produced unexpected learnings, a methodology with measurable results. This is the long-term moat.
• Decide on llms.txt implementation based on audience research. If implementing, follow the standard as it currently exists and update as it evolves.
• Begin structured schema implementation across organisation, person, services, and key content. Treat schema as the entity-corroboration layer for AI systems, not as a Google ranking signal.
• For B2B categories where agent-mediated research is plausible, audit the site for agent-readability: clear capability statements, explicit pricing methodology, named outcomes, accessible contact pathways. Agents cannot infer what is ambiguous.

The Honest Summary

Google’s May 2026 AI optimisation guide is technically accurate for Google Search, useful as a baseline reference, and worth reading carefully. Its framing reflects Google’s commercial position. Its scope is narrower than “how buyers actually discover vendors in 2026”. Site owners whose audiences sit substantially inside Google’s ecosystem can largely follow the document. Site owners whose audiences span ChatGPT, Claude, Copilot, Perplexity, Gemini, and emerging agentic systems need the broader framework the document deliberately does not provide.

The right reading of the document is neither uncritical acceptance nor dismissal. Take what is useful, name what is Google-specific, recognise the framing as positioning rather than neutral truth, and build the visibility strategy your specific audience actually needs. That is the practical bridge between the SEO foundations Google rightly says still apply and the multi-platform AI visibility reality Google’s document does not engage with.

If you would like this analysis applied to your specific site, audience, and category — including the cross-platform brand-mention audit, the entity-corroboration assessment, and the CITATE-scored remediation plan — the starting point is the AI Visibility Audit.

References: Google Search Central, Optimizing your website for generative AI features on Google Search, 15 May 2026. Similarweb, 2026 Generative AI Brand Visibility Index, January 2026 data. Similarweb 2026 AI Search report. BrightEdge AI Search citation data, 2026. Princeton/Georgia Tech/IIT Delhi GEO-Bench study, 2026.