Beyond SEO and GEO: Making Websites Fit for AI Agents

For a long time, the division of labor on the web was clear: humans visit websites, search engines crawl and index them. AI systems like ChatGPT, Gemini, or Perplexity added another layer. Websites are now meant to be not just found by search engines, but also understood, evaluated, and used as a source by AI systems.

The next generation of AI agents takes this development a step further. They're meant to not just read websites, but operate them independently.

An AI agent could, for example, search for a suitable provider, compare products, fill out a form, schedule an appointment, or complete an ordering process – on behalf of a human.

This gives rise to another important discipline of website optimization alongside SEO and GEO: optimization for agentic use.

From SEO to GEO to the Agent-Ready Website

Classic search engine optimization is mainly concerned with one question:

Can a search engine find, understand, and correctly categorize our content?

With Generative Engine Optimization (GEO), or Answer Engine Optimization (AEO), another question was added:

Can an AI system understand our content and use it for its answers?

With AI agents, the decisive question instead becomes:

Can the AI operate our website too?

This changes the requirements considerably. An agent must, for example, be able to recognize:

A website can therefore look excellent and still be difficult for an AI agent to operate.

AI Agents See Websites Differently Than Humans Do

Humans orient themselves strongly by sight. We recognize a button by its styling, understand a navigation by its position, and recognize an unavailable product variant, for example, by the fact that it's shown in gray.

Machines need more explicit information.

Modern browsers provide an additional structured representation alongside the actual HTML document for this purpose: the so-called accessibility tree.

Originally, this mainly serves to make websites accessible to assistive technologies like screen readers. It doesn't primarily describe a website's appearance, but its meaning and the ways it can be operated.

A complexly designed website turns, through this, into a much more compact structure of elements such as:

This representation is especially interesting for AI agents because they can immediately recognize which elements exist and what function each one has.

Accessibility Becomes an AI Topic Too

This leads to a remarkable development:

Many measures that make a website accessible simultaneously improve its usability by AI agents. These include, in particular:

A simple example shows why this matters.

The website is understandable for humans – but ambiguous for machines.

Semantic HTML Is Becoming More Important

In recent years, websites have increasingly been built using complex JavaScript frameworks and visual components. In the process, an HTML element like a <div> is often programmed via CSS and JavaScript to behave like a button.

For a human, this works without any problem. For screen readers and AI agents, though, it can be difficult.

A real HTML button

already carries its meaning within itself. A correspondingly programmed <div>, by contrast, only gets this meaning conveyed through additional attributes.

For the agent-ready website, a surprisingly classic rule therefore applies:

HTML should once again express exactly what an element actually is.

This helps humans, search engines, and AI agents alike.

Structured Data Alone Isn't Enough

Structured data based on Schema.org remains important. It helps search engines and AI systems categorize content unambiguously.

But for an agent that actually wants to operate a website, structured data alone isn't sufficient.

The agent, for example, doesn't just need to know that a page describes a product. It also needs to be able to recognize which button adds the product to the cart.

That's why important information shouldn't exist exclusively in JSON-LD or other machine-readable supplementary formats.

Prices, availability, properties, and possible actions should also be present clearly and semantically correctly in the actual HTML.

And What About llms.txt?

In the context of AI optimization, there's currently a lot of discussion about llms.txt and Markdown content specifically provided for language models.

Such approaches can certainly make sense when an AI system wants to read and process a website's information. For an agent that's meant to actually act on a website, though, they solve a different problem.

A Markdown version can, for example, explain excellently to an AI system which services a company offers. But it can't provide the agent with a working button for submitting a contact form.

Two tasks should therefore be distinguished:

Machine-readable content helps AI systems understand a website.

Machine-operable structures help AI agents act on a website.

The two complement each other – but don't replace one another. We describe how llms.txt can be generated automatically for TYPO3 websites in our article on the llms_txt extension.

APIs and Agent Interfaces as a Second Layer

Not every AI agent has to operate a website the way a human does.

In the long run, a second access path is therefore likely to establish itself: websites make certain functions directly available through APIs or dedicated agent interfaces. An agent would then, for example, no longer need to click through a booking process, but could use a defined interface instead.

This creates two possibilities going forward:

Browser-based access: The agent operates the regular website.

Structured access: The agent communicates directly with the system via APIs or protocols specifically developed for agents.

For professional content management systems, the ability to support both paths is becoming increasingly interesting.

Why TYPO3 Offers Good Conditions for This

TYPO3 in particular offers some interesting properties for this development.

TYPO3 doesn't produce a fixed, predetermined website structure. Developers have extensive control over templates, HTML output, structured data, and interfaces. This means TYPO3 websites can be built consistently semantically.

At the same time, TYPO3 is open source. Both the CMS itself and its extensions and data structures are transparent and can be analyzed and extended as needed. On top of that comes the system's pronounced extensibility.

If standards for communication between websites and AI agents establish themselves in the future, corresponding interfaces can be implemented as TYPO3 extensions or project-specific integrations.

Staging and workflow concepts also play a role here. Agentic processes don't necessarily have to work directly on a live system. Actions can be controlled, logged, reviewed, or first carried out in workspaces or staging systems.

This makes TYPO3 particularly well-suited for scenarios in which AI doesn't just consume content, but increasingly interacts with a CMS – more on that in our article "TYPO3 & AI".

The Website Gets a New Audience

Web design and web development have so far mainly targeted two audiences: humans and search engines.

Now a third is joining them: AI agents.

These agents don't need particularly spectacular designs. They need clear structures, understandable content, and reliably operable functions.

Interestingly, this doesn't mean websites need to be built specifically "for machines." Quite the opposite.

A cleanly structured, accessible, and semantically correct website is, at the same time, a good foundation for humans, search engines, and AI agents.

From Mobile First to Machine First?

A few years ago, companies had to learn to stop developing websites exclusively for desktop PCs. "Mobile first" became a fundamental principle of modern web development.

A similar development could now be beginning with AI systems.

"Machine first" doesn't mean building websites primarily for machines in the future. It means, rather, taking into account from the start, in a website's technical architecture, that content and functionality are increasingly being consumed by machines too.

Four central requirements can be distinguished here:

SEO and GEO already cover large parts of the first three areas. AI agents now make the fourth area relevant in particular.

Conclusion

Optimizing for AI doesn't end with ChatGPT answers or good placements in AI Overviews.

With autonomous and personal AI agents, a new form of web usage is emerging. Humans will increasingly delegate tasks to AI systems rather than operating websites exclusively themselves.

For website owners, that means: a website will in the future need to not just be found and understood, but possibly also operated by machines.

Accessibility, semantic HTML, structured data, clearly defined interfaces, and a clean technical architecture are therefore growing together into a shared topic.

It's worth it for companies, then, to ask an additional question today already when planning relaunches and technical upgrades:

Is our website optimized only for humans and search engines – or can an AI agent work with it meaningfully too?