Your CMS is a GEO decision

Somebody on your team asked an AI assistant a question your company should own the answer to, and your company was not in the answer. Now there is a line item on your desk called GEO, or AEO, and everything you can find about it is written for search practitioners: prompts to test, tools to buy, dashboards to watch. None of it answers the question underneath, which is why the assistant could not use your site in the first place.

That question is an infrastructure question, and our position on it is simple: AI visibility is architecture-first and tactics-second. An assistant can only cite a source it can retrieve, parse, and attribute, and whether your site is retrievable, parseable, and attributable is decided by your content platform. A tracking tool pointed at a site an assistant cannot read will measure a problem it has no way to fix. The useful order of work is structured content, then clean delivery, then an explicit crawler policy, and then measurement. This guide is about the first three, because they are the part a tool cannot buy you.

The demand is real and the field is young. EMARKETER’s FAQ on GEO and AEO (2 April 2026) describes a category where “the terminology is unsettled, and vendor hype runs high,” and notes the concept barely existed a year earlier. At the same time, the infrastructure worry is widespread: in Contentstack’s 2026 Agentic Enterprise Report, a May 2026 survey of 621 digital leaders, 85% said they are actively evaluating whether their current CMS or content infrastructure can support agentic AI use cases. Every platform vendor has noticed, too — Sanity repositioned as “the Content Operating System for the AI era” in March 2026 — and every vendor is fighting over the narrative, while the architecture is what actually decides whether you appear in the answer.

What happens when an AI assistant answers about your category

When someone asks an assistant a question in your category, three things happen in quick succession, and your brand has to survive all three.

• First, retrieval. The assistant gathers candidate sources, from what it learned in training and, increasingly, by fetching live web pages at the moment of the question. If your page returns nothing useful to a machine that fetches it, you are out before the answer starts forming.
• Second, synthesis. The assistant reads what it retrieved and composes an answer from the parts it can actually parse. Content it cannot interpret is content it cannot use, however good it looks to a human in a browser.
• Third, attribution. Where the assistant cites or names a source, it points at something it can identify and trust. Clear, consistent, well-structured pages are easier to attribute than a tangle of unlabelled markup.

This is the whole of what the acronyms describe.

• GEO, generative engine optimisation, is the practice of getting your brand mentioned in AI-generated answers.
• AEO, answer engine optimisation, is getting your content surfaced as the direct answer.

As EMARKETER puts it, in practice the two describe the same underlying approach. Underneath the vocabulary, the job is to be retrievable, parseable, and attributable, which is a property of how your site is built, and the foundation any marketing tactic sits on.

Why machines cannot read most websites

Plenty of sites that look polished to a person are close to blank to a machine, for a few common reasons.

Many are rendered in the browser with JavaScript, so the initial HTML a fetcher receives is an empty shell, and the content only appears after code runs that the fetcher never executes. Many are built in visual page builders that flatten content into deeply nested, unlabelled containers, where a heading is just big text and a data table is a picture. And on many, the facts that matter are baked into a layout, with no structured version stored anywhere, so there is nothing to extract.

A concrete example makes the gap obvious. Picture a product specification. On a page-painting tool, that spec is often a designed image, or a stack of styled div elements with the numbers scattered through them for visual effect; a machine sees an image, or shapes, and no specification. Modelled properly, the same spec is a set of structured fields, each attribute with a name, a value, and a unit, rendered into a real HTML table with real headings. The first is invisible to synthesis. The second can be read, quoted, and attributed. The two look identical to a person, and only one can be used in an answer.

The six architecture checks

Here is the part to forward to whoever owns your platform. These are the properties that decide whether your content can be used by an AI assistant, and none of them is exotic.

1. Semantic HTML delivered server-side. The content is present in the HTML that arrives on first request, marked up with meaningful tags, headings, lists, tables, so a fetcher gets the substance without running your JavaScript.
2. Structured data on key entities. Schema.org markup (typically JSON-LD) describes your organisation, products, and articles in machine-readable terms, giving assistants explicit facts about those entities to work from.
3. Stable canonical URLs. One durable address per piece of content, declared with a canonical tag, so a citation resolves to a single place and your authority does not fragment across duplicate URLs.
4. A heading hierarchy that survives copy-paste. A real h1, h2, h3 structure that still communicates how the page is organised when the text is extracted from its styling, which is exactly what synthesis does.
5. An explicit crawler policy. A robots.txt that names the AI crawlers you allow or disallow — and the list is longer than it looks, because the major vendors each run several crawlers with different jobs and separate user-agents: OpenAI’s GPTBot, OAI-SearchBot, and ChatGPT-User, Anthropic’s ClaudeBot, Claude-User, and Claude-SearchBot, Google-Extended, PerplexityBot, and others. A policy that names only the training bots leaves the search and user-triggered fetchers ungoverned. Optionally add an llms.txt file that points assistants at your important content. llms.txt is an emerging convention, proposed in 2024, and as of August 2026 the major AI companies have not committed to reading it in production; Google’s AI optimisation guidance (10 July 2026) says outright that Search and its generative features do not use such files. It is fine to add as a low-cost extra, though it is not a ratified standard, and no strategy should depend on it.
6. Consistent entity facts across pages. The same company name, the same claims, the same numbers everywhere they appear, so an assistant synthesising across several of your pages arrives at the same coherent answer wherever it looks.

Clear all six and your site is legible to the systems now mediating discovery. Miss the first two and most of the rest cannot help you.

What this means for CMS choice

Any properly structured content platform can clear these bars. Contentful, Sanity, and Payload all store content as structured data and let you control how it is delivered, so all three can produce semantic, server-rendered, schema-marked pages. The platforms that make every check harder are the page-painting tools, the visual builders that store content as layout, because the structure a machine needs was never captured in the first place. This is why the Contentstack survey found content-readiness such a common regret: among leaders already running AI in production, 88% wished they had invested in their content and data infrastructure earlier, and the most-cited content problem was structured data that is insufficiently tagged or contextualised (32%), with a further 28% pointing to content that is inconsistently formatted for machines to interpret.

We build on Payload, and it fits this work well. Content lives in structured collections you define in code, you have full control over the markup and server-side rendering, and because it is open source and self-hosted, you own your crawler policy and your delivery outright, a maintained, secure platform with no lock-in. That is our preference and our practice. We are not claiming Payload is the only tool that can do this; a well-run Contentful or Sanity build clears the same bars.

If your current platform cannot deliver structured, server-rendered content, no tool bought on top will change that, and the right path is to move to one that can. Our guide to migrating from Contentful to Payload covers a move between two structured systems, and if you are on a visual builder, the same principles apply to a heavier lift off a page-painting tool.

The access question is separate

There is a decision this guide has taken for granted so far: whether you want AI crawlers reading your site at all. Making your content legible to assistants is a separate question from whether you want to let them in, and some organisations block AI crawlers for commercial, legal, or editorial reasons, at a real cost to visibility. The network layer is beginning to take part in that decision: from 15 September 2026, Cloudflare blocks its Training and Agent crawler categories by default on ad-displaying pages of domains newly onboarding to Cloudflare, while Search stays allowed and existing customers keep their settings — so a crawler policy increasingly lives in your CDN configuration as well as your robots.txt. That trade-off deserves its own treatment and depends on your commercial and legal position more than on your architecture.

Measure before and after

Once the architecture is right, you will want to know whether it moved anything, and you cannot manage a citation share you do not measure. This is where a measurement tool earns its place, at the end of the sequence, once the foundations are set. Prelio tracks how AI assistants represent your brand: what they say about you, how often you are cited, and how you compare with competitors across a consistent set of questions, so you can see whether an infrastructure change actually shifted the result. It is built for teams that have done the structural work and want evidence of the payoff. We make Prelio, so treat this as disclosure.

GEO is young, and nobody can promise you a ranking or a citation, ours or anyone’s. What the architecture buys is eligibility, the precondition for being used at all. The tactics on top, community presence, fresh content, third-party mentions, matter, and they work far better on a site an assistant can actually read.

How WAYF can help

WAYF builds structured, server-rendered content platforms, most often on Payload, and we audit existing sites against the six checks above so you know exactly where you stand before spending on tactics. For an AI-visibility architecture review of your site, book a call, and see our work for platforms we have built.

FAQ

What is the difference between GEO and AEO? Generative engine optimisation is about getting your brand mentioned in AI-generated answers; answer engine optimisation is about being surfaced as the direct answer. In practice they describe the same underlying approach, and the terminology across the industry is still unsettled. Both depend on your content being retrievable, parseable, and attributable.

Can a GEO tool fix my AI visibility? A measurement tool tells you how AI assistants currently represent you, which is valuable once your site is legible to them. It cannot make an unparseable site parseable. If your content is trapped in client-side rendering or visual layouts, the fix is architectural, and a dashboard will only document the problem until you address it.

Is llms.txt a standard I need to implement? It is an emerging convention, not a ratified standard, and adoption across AI platforms is uneven. It is cheap to add and reasonable to include, but it sits well below semantic HTML, structured data, and a clear crawler policy in priority. Treat it as a useful extra on top of the foundations.

Should I block AI crawlers? That is a separate decision with real trade-offs on both sides, and it depends on your commercial and legal position more than on your architecture. Blocking is a legitimate choice for some organisations; it simply comes at a cost to visibility in AI-mediated answers.

Do I need to migrate my CMS to be visible in AI answers? Only if your current platform cannot deliver structured, server-rendered content. Any properly structured CMS can meet the checks in this guide. Visual page builders that store content as layout are the ones that make this hard, and for those a migration to a structured platform is usually the right answer.

Sources

• EMARKETER, FAQ on GEO and AEO: Where AI search and SEO overlap in 2026 (2 April 2026)
• Contentstack, The 2026 Agentic Enterprise Report (survey of 621 digital leaders, May 2026)
• Sanity, Sanity Launches the Content Operating System for the AI Era (4 March 2026)
• Schema.org (structured data vocabulary)
• llms.txt (the /llms.txt proposal)
• Google Search Central, AI optimization guide (10 July 2026)
• OpenAI, Overview of OpenAI crawlers
• Anthropic, Does Anthropic crawl data from the web? (7 April 2026)
• Cloudflare, Content Independence Day: new crawler defaults (1 July 2026)