From GEO-Ready to AI-Ready: the next evolution of CMS platforms

What GEO-Ready actually meant

GEO — Generative Engine Optimization — was the label the SEO world used between 2024 and 2026 to organise its transition into the generative era. The task was no longer to get a keyword to position three on a SERP, but to be cited as a source in an AI-generated answer. From that, a checklist emerged of what a CMS had to do:

• Structured answers — every assertion in question-and-answer blocks, FAQ Schema.org markup, short definitions at the start of every section.
• Clean Schema.org — not only the mandatory WebPage and Article types, but HowTo, FAQPage, BreadcrumbList, Organization, Person and Service as declared markup so generative engines understand the relationships.
• Clear reading time and authorship — every post with an author, update date, a reading-time figure and one clear thesis in the first 200 words.
• Cite-friendly writing — in short, easily quotable sentences instead of long paragraphs; claims underpinned with source links.
• llms.txt debate — whether to make the crawler’s entry easier with an extra file. The topic has cooled down with Google’s May 2026 AI Optimization Guide: structured data and clean HTML suffice; a separate file is not required.

That stage was mandatory, and it remains so. But it rests on an assumption that has been crumbling for the past twelve months: that consumption runs through a crawler that renders HTML.

Where GEO hits its limits

Three fault lines have become visible over the past twelve months:

1. Agents want tools, not pages

A modern AI agent — Claude in Chrome, an internal Mittelstand tool, a service bot — is not a crawler. It poses a task and looks for an API it can call. A nicely structured HTML page with FAQ schema is a second-class data source for it. What it wants is a function it calls with arguments and whose output it can rely on.

2. Generative answers without provenance get rightly rejected

Until 2024 “cited by an AI answer” was a marketing success. With EU AI Act Art. 50 from August 2026 and the growing scepticism towards hallucinated sources, the question “did this really come from the original publisher, and in which version?” becomes a mandatory check. A Schema.org annotation will not do; what is needed is a cryptographic statement.

3. Content must serve multiple consumption layers at the same time

A service description should work as a marketing page in the website variant, as a concise tool answer in the AI agent variant, as a speakable block in the voice variant, as a carousel in the social variant. GEO-Ready thinks in one output (HTML for the SERP). AI-Ready must think in several outputs at once — without the editorial team maintaining four versions.

Four layers that make a CMS AI-Ready

The answer to those three fault lines is a platform build that does not replace a layer but adds four additional layers on top of the Schema.org markup that is already there.

Layer 1: structured content with AI context

Schema.org is the must. AI-Ready extends it with three fields per content block that an agent can read directly: audience (“IT lead, Mittelstand”), tone (“matter-of-fact, terse”), channels (“website, ai-agent, voice”). These fields cascade through the page tree so an agent knows whether a piece of content is meant for it without any editor intervention.

Layer 2: retrieval — vector plus MCP tool API

The actual break with GEO. Content is chunked, vectorised and indexed for semantic retrieval (RAG). In parallel the CMS exposes tools via the Model Context Protocol: search, navigation, page-content, plus domain-specific tools like product-search. The agent calls a tool, the server respects TYPO3 access rules (fe_group, hidden, starttime, endtime) and returns structured JSON.

Layer 3: provenance — who wrote, where did the image come from

Every published piece carries an Ed25519 signature, every asset an IPTC XMP tag (DigitalSourceType, AISystemUsed). AI-generated images are detected on upload from their C2PA manifest and IPTC header, marked in the TCA as “AI Provenance” and have the metadata written back to the file on save. That covers EU AI Act Art. 50 — and gives every agent a verifiable authenticity statement.

Layer 4: channel-aware output

One piece of content, several output formats. The channels fields from layer 1 decide which fields end up in which output. A detection middleware identifies the consumer from the request (user agent, Accept header, dedicated channel header) and serves exactly the representation the consumer needs — HTML with hero image for the website, JSON with a summary plus a source link for the agent, a speakable block without tables for voice, carousel blocks for social.

Migration path — turning GEO-Ready into AI-Ready

If you have finished GEO-Ready you already have the substrate for AI-Ready. The way there breaks down into four stages that are walkable in this order:

1. Add the AI context fields. Expose three fields (audience, tone, channels) on pages and content elements as a TCA extension. Editorial work starts with the 20 most-used pages and cascades through the page tree. Effort per site is small, the effect is visible immediately in Schema.org output and in controllable agent answers.
2. Stand up the retrieval layer. Wire a vector store (e.g. pgvector next to the existing database) and an MCP server into the CMS. The first three tools (search, navigation, page-content) cover 80 % of agent requests and need no domain-specific development. Important: respect TYPO3 access rules — an agent must only see what an anonymous frontend visitor would also see.
3. Introduce provenance. Roll out Ed25519 keys through a central provider (Vault, file-based with Step-CA), wire a signing middleware into the publish workflow, register an IPTC XMP listener on AfterFileAddedEvent. The hardest part is organisational — who rotates the key, where the audit log lives.
4. Wire channel-aware output. Define the four channels (website, AI agent, voice, social) per content type as Twig templates, put a detection middleware in front of routing. The temptation here is to do everything at once; the sensible move is to start with AI agent (the direct consumer pressure) and pull voice and social in afterwards.

If you are still stuck at the SEO audit or have done GEO half-heartedly: please finish that stage first. AI-Ready without structured content underneath is marketing theatre — layers 2–4 need the data base only layer 1 (GEO) provides.

What we build at Moselwal for this

We have spent the past two years building TYPO3 extensions that bring these four layers together as one platform — because we run our own platforms this way. The core building blocks are all available as Open Source under moselwal/*. The content layer (Schema.org generator stack, AI context inheritance, channel fields) lives in structured-content. The retrieval layer with MCP tools in webmcp. The provenance layer in content-provenance (Ed25519, IPTC, C2PA). The channel layer in semantic-delivery. The conversational layer (RAG pipeline, web chat) in business-agent.

The packages are independently usable — if you only need the Schema.org layer, leave the rest out. For the full AI-Ready platform we combine them as a stack; the Open Source pages document each package separately.

Conclusion

GEO-Ready has not become wrong — it is foundation. What is changing is that the consumption path no longer runs only through the crawler; increasingly it runs through the agent. Taking that seriously means building four additional layers on top of Schema.org: structured content with AI context, a retrieval layer of vectors plus MCP tools, provenance via Ed25519 and IPTC, and channel-aware output. These layers are independently usable, they work on a TYPO3 estate that has done its GEO homework, and they are the answer to agent-driven consumption that already drives more SERP-equivalent traffic than classic search.

If you are still stuck at the SEO audit: finish that stage first. AI-Ready without structured content underneath stays marketing theatre — and the agents will notice within the first two queries.