Google says: no llms.txt needed — and why that is only half the truth for an AI-ready CMS

TL;DR — the 90-second summary

What does Google say on 15 May 2026?

An official guide: “Optimizing for generative AI features on Google Search”. Core message: AEO/GEO are still SEO from Google’s perspective. No llms.txt needed, no chunking needed, Schema.org optional. Foundational SEO + non-commodity content is enough.

Where does that hold?

For Google Search including AI Overviews and AI Mode: yes. Google’s retrieval-augmented generation hits Google’s own index, which uses the standard search-ranking systems as a base. RAG runs over the classical crawl + index.

Where is it only half the truth?

Outside of Google. Claude (Anthropic), Perplexity, OpenAI crawlers, voice agents, telephony bots, browser-native MCP consumers, the upcoming Universal Commerce Protocol (UCP) — they do not ask Google. They crawl llms.txt manifests, read Schema.org JSON-LD directly in the DOM, fetch MCP tool listings. Anyone catering only to Google is invisible to those consumers.

What does that mean for our AI-ready CMS approach?

Our four layers (Structured Content, Semantic Delivery, Agent Interaction, Trust) build exactly the discovery and delivery endpoints Google does not need — but that Claude, Perplexity and a mid-market organisation’s voice agent definitely do. The Cloudflare Agent-Readiness Score of 100/100 measures this stack — not Google’s ranking.

An important point of agreement

Google explicitly says: explore the agentic experiences, browser agents are becoming relevant. The Universal Commerce Protocol is named directly. This is where the MCP tool API and agent interaction become relevant for Google-mediated use cases as well.

Recommendation for mid-market

Build both. Google basics (foundational SEO, non-commodity content, page experience) are and remain mandatory. Add the AI-ready layers for the growing non-Google world — including llms.txt, which Google rules out for itself but which becomes a standard elsewhere.

 

What Google actually says on 15 May 2026

The new guide at developers.google.com/search/docs/fundamentals/ai-optimization-guide is relatively slim — and that is part of its message. The key statements, almost verbatim:

SEO is still relevant for generative AI search

Google’s AI features in Search lean on the classical search-ranking and quality systems. Two mechanisms are named explicitly:

• Retrieval-augmented generation (RAG) — answer quality is improved by having the search-ranking systems pull relevant, fresh web pages from Google’s index and hand them to the model as context.
• Query fan-out — the model generates several parallel sub-queries (“how to fix a lawn that’s full of weeds” → “best herbicides”, “remove weeds without chemicals”, “prevent weeds in lawn”) to build more comprehensive answers.

Both mechanisms sit on top of the Google index. If you are not in the index, you do not surface in AI Overviews. If you are well-indexed, the chances are the same as in pre-AI times.

AEO and GEO are still SEO from Google’s perspective

Verbatim: “From Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.” Google’s explicit point: no separate disciplines, no separate conference tracks, no “GEO framework” alongside SEO. For Google.

The SEO basics that are unchanged

• Non-commodity content with a unique point of view, experience and depth — not recycled common knowledge.
• People-first content that answers the question satisfactorily, instead of serving every possible fan-out query individually (which Google explicitly flags as spam).
• Clean technical structure — indexable, crawl-suitable, with valid HTML (“perfect not required, but readable helps”).
• Page experience — Core Web Vitals, mobile, fast delivery.
• Images and videos of good quality with alt text and video markup.
• Local-business data via Google Business Profile / Merchant Center.

What Google explicitly marks as unnecessary

Here the guide gets unusually direct:

• llms.txt files and other “special markup”: “You don’t need to create new machine-readable files, AI text files, markup, or Markdown to appear in generative AI search.” Google processes HTML as it always has.
• Chunking: “There’s no requirement to break your content into tiny pieces for AI to better understand it.” Google reports it understands multi-topic pages.
• Special rewriting for AI: not needed.
• Inauthentic mentions: not a lever.
• Over-focusing on structured data: not required, but still useful as part of the SEO strategy (rich results).

Where Google explicitly looks “into the future”

At the end of the guide, almost in passing: agentic experiences. Browser agents (DOM inspection, accessibility tree, visual rendering) are named directly. The Universal Commerce Protocol (UCP) is linked as an emerging protocol that will let search agents do more. Google’s recommendation: “If this is something that’s relevant to your business and you have extra time, check out the available agentic experiences.”

Even Google sees: the next step is no longer just “index and answer”, it is “agent calls a tool”. That is where it gets interesting.

Where Google’s guide holds — and where it shows only part of the truth

The guide is clean, honest and internally consistent. But it has one strong precondition Google itself never spells out: it describes the AI behaviour of Google. Other consumers behave differently — structurally, not cyclically.

What holds for Google

Google’s AI Overviews and AI Mode hit the Google index. Google has a mature crawler that understands HTML semantically, chunks multi-topic pages, can evaluate Schema.org optionally and works without an llms.txt manifest. From that position the guide is correct: no special-markup obligation, no special chunking, no AI rewriting.

What does not hold for the other consumers

The AI world in 2026 is not just Google. The most important other consumers of your content:

• Anthropic Claude with web browsing and an MCP tool chain — uses llms.txt, JSON-LD, MCP discovery routes.
• Perplexity as its own AI search engine — uses Schema.org JSON-LD as a priority and has explicitly supported llms.txt since Q1 2026.
• OpenAI/ChatGPT with browsing and Operator — runs its own crawler (OAI-SearchBot), respects llms.txt directives.
• Voice agents (smart speakers, telephony bot platforms) — need speech-suitable excerpt versions that ordinary web HTML does not deliver.
• Browser-native MCP consumers — browser extensions, in-app agents that query navigator.modelContext, expect tool discovery, and want to trigger user actions.
• Internal search / RAG pipelines of your own customers, suppliers, partners — who may ingest your content into their own vector databases.

The asymmetry

Anyone who follows only Google’s guide is optimally placed in the Google universe. Anyone who also wants to be found, correctly cited and integrated into tool calls by the other consumers needs additionally the layers Google explicitly “doesn’t need”. That is not criticism of the Google guide — it is the answer to a question Google did not address.

The trend line

Google Search’s market share in DACH in 2026 still sits around 85%. But the relevant indicator is not market share, it is the distribution of answer sources: when a user asks their smart speaker by voice, when a mid-market organisation has its sales agent draft a briefing, when a customer asks a WhatsApp bot a product question — the answer does not come from Google. It comes from a direct LLM call against your content. And precisely there, whether you delivered llms.txt, MCP tools and Schema.org JSON-LD decides the outcome.

The only hard technical requirement: IPTC DigitalSourceType for AI images

Reading the Google guide carefully reveals an almost hidden passage where Google becomes unusually specific: image labelling for AI-generated content. Here the guide turns a recommendation into a technical requirement.

What Google requires

For images created with generative AI, Google expects the IPTC metadata field DigitalSourceType with the value trainedAlgorithmicMedia — embedded in the XMP sidecar data of the image file. This is not a Schema.org tag in the DOM, not an llms.txt hint, but a file-internal metadata property every image must carry on upload.

For ecommerce there is an additional rule: AI-generated product titles and descriptions must be labelled separately as AI-generated in the Merchant Center feed. For classic content sites without product data, this is not relevant.

Why this is the only hard requirement

In the entire guide this is the only point where Google speaks of a concrete technical obligation. Everything else — llms.txt, chunking, Schema.org — is explicitly marked as optional or not necessary. Image labelling is different because Google uses it as a regulatory bridge: EU AI Act Article 50 mandates AI content transparency, and Google renders the image label visibly in Search results.

What we have built for this at Moselwal

Our content-provenance package has written an IPTC-XMP sidecar with the DigitalSourceType tag on every image upload since Phase 3 (Q1 2026). When the editor flags an image as AI-generated in the backend (field tx_provenance_origin = ai_generated), the IptcXmpWriter writes the tag into the file automatically. The requirement is fulfilled without editor burden.

What Google gives us here goes beyond the obligation: the image label is also machine-readable for downstream systems — voice agents, provenance validators, reuse trackers. One and the same tag serves the Google-mandatory path and the wider trust layer (see our pillar on the AI-ready CMS, Layer 4).

What Google’s guide concretely means for our compliance stack

We checked the guide against our existing AI-ready stack. The result: no code changes needed, the only hard requirement has been fulfilled since Phase 3, everything else is a strategic hedge for the non-Google world.

What “strategic hedge” means here

We do not build llms.txt because Google needs it (it does not), but because Anthropic, OpenAI and Perplexity use it — and because the effort is near zero once the infrastructure is in place. Should Google reverse its position in 12 months (not out of the question, since the guide explicitly addresses ‘agentic experiences’ as emerging), we will already be there. Should Google stick to its position, we still serve the non-Google consumers.

Edge case: ecommerce

Anyone working with Merchant Center product data has an additional mandatory point from the Google guide: “AI-generated product data such as title and description attributes must be specified separately and labeled as AI-generated.” For Sylius or Shopware stacks this means a field extension in the product model and an adjusted feed export. For classic content sites without Merchant Center integration, not relevant.

A small polish idea from the audit pass

If an image is flagged as tx_provenance_origin = ai_generated in the backend and the alt text is empty, a BE hint in the file-edit form would make sense — AI images without alt text are doubly suboptimal: bad for accessibility and bad for Google Image Search. Polish item, not a mandatory backlog entry.

How our Cloudflare Agent-Readiness Score of 100/100 fits this context

Cloudflare launched the Agent-Readiness Score in April 2026 — a Lighthouse-style benchmark for how well a website can be used by AI agents. Our moselwal.de scored 100 out of 100 in April 2026 (see our post on the 100/100 result) — and every one of those points would have been marked “unnecessary” by Google’s new guide.

What does the score actually measure?

The Cloudflare score checks seven categories. Three of them overlap with Google’s SEO recommendations (crawlability, page performance, semantic HTML). Four of them address what Google explicitly marks as unnecessary:

• /.well-known/llms.txt with a correctly formatted manifest — a prerequisite for a 90+ rating.
• Schema.org JSON-LD in the DOM with consistent type annotation per content class — not only Article but Service, Product, FAQ, HowTo, Person, Organization, fitting the page.
• MCP tool discovery via /.well-known/mcp.json or a comparable discovery route — telling the agent which operations it can trigger.
• Stable section spine with retrieval-suitable heading hierarchies — so RAG pipelines form clean chunks instead of slicing through walls of text.

The direct tension with the Google guide

Google says: “nope, you don’t need llms.txt.” Cloudflare says: “without llms.txt you will at most score 70/100.” Who is right? Both — each for their world:

• Google measures what works for Google Search. llms.txt is not the lever there.
• Cloudflare measures what works for agents outside Google. llms.txt is a hard requirement there.

What our 100/100 means in practice

Concretely: if a user asks Anthropic Claude “what does Moselwal actually do”, Claude finds our llms.txt manifest, fetches the service detail routes, reads the Schema.org annotation, and formulates an answer with correct citations. If the same user asks the same question in Google AI Mode, the answer comes from Google’s index — same content, different path. Both work because our stack serves both.

Anyone doing only Google optimisation will see roughly 50–70/100 in Cloudflare’s scoring test. Is that enough? It depends on whether the other consumers are relevant to the business model.

What mid-market organisations should do now

The combination of Google’s guide and our stack check yields four concrete recommendations.

1. Do not neglect the SEO basics

Google’s point stands and is correct: without a clean SEO foundation, no appearance in Google AI Overviews. Non-commodity content, clear technical structure, page experience, image and video SEO remain the foundation. Anyone distracted by AI hype loses both — classic search and generative.

2. Set the IPTC tag for AI-generated images

This is the only new technical obligation from the May 2026 guide. If you run your own image pipeline tools: check that DigitalSourceType=trainedAlgorithmicMedia lands in the XMP sidecar for AI-generated assets. If you run TYPO3 with our content-provenance package: already done. If you manually upload images from Adobe Firefly, Midjourney or DALL-E: the tag is usually present in the source file but is lost on re-encoding (resize, WebP conversion) — verify.

3. For Merchant Center feeds: plan AI labels

If you push Sylius, Shopware, Magento or a custom product data pipeline to Google Merchant Center, you should plan the AI label field for title and description as a feed extension — not because it blocks hard today, but because Google has announced it will enforce here.

4. Build llms.txt + Schema.org + MCP tools anyway — for the non-Google world

This recommendation is more contested because Google says the opposite. Our position: anyone building the four layers of an AI-ready CMS is not optimising against Google but in addition for Anthropic, OpenAI, Perplexity, voice agents and upcoming MCP consumers. The effort is low once the platform is in place. The omission is risky because the non-Google world becomes more important every month.

What not to do

• No special content rewrites for AI — Google says the same thing, by the way.
• No content spam with fan-out variations for every possible user question — Google explicitly flags this as spam.
• No inauthentic-mention building — not a lever with Google or the other consumers.
• No “AEO/GEO framework” as a separate discipline alongside SEO — the worlds overlap too much.

Conclusion

Google’s guide is clean, internally consistent and says an important truth: SEO is and remains the foundation. Foundational SEO + non-commodity content + technical clarity is enough to surface in Google AI Overviews and AI Mode. The only new technical obligation is the IPTC image label for AI-generated assets — which we write by default at Moselwal since Phase 3 of our content-provenance package.

What the guide does not answer — and does not need to answer — is the AI world beyond Google. Anthropic Claude, Perplexity, OpenAI, voice agents, browser-native MCP consumers: they do not ask Google. They reach directly into discovery endpoints, JSON-LD annotations and MCP tool listings. If you want to be found there, you build llms.txt, Schema.org and MCP tools — not against Google, but in addition.

Our Cloudflare Agent-Readiness Score of 100/100 measures exactly this additional layer. In combination with Google’s foundational SEO it serves both worlds. The only sensible answer to the guide is therefore: “Everything Google says, we already do. Plus the four layers it does not mention.”