Preparing for the Agentic Web: Key Insights from R.V. Guha and Schema App

Key Takeaway: The Agents Are Coming. Is Your Business Ready?

We are entering the era of agents. These agents will research products, evaluate vendors, compare alternatives, book appointments, and complete tasks. In many cases, they’ll become the primary interface between customers and businesses.

How will an agent decide whether to recommend your business, surface your content, or take action on behalf of a customer?

That was the central theme of our webinar, Preparing for the Agentic Web: Semantic Infrastructure for the AI Era, featuring Schema App co-founders Martha and Mark van Berkel alongside Microsoft’s R.V. Guha, creator of Schema.org and RSS, and one of the pioneers behind the emerging NLWeb initiative.

Guha described this moment as the fourth major shift in how people interact with technology, following graphical user interfaces, the internet, and mobile computing. Today, we’re entering an era where people increasingly interact with computers through natural language rather than navigating websites and applications themselves.

We’re in the early parts of the fourth revolution, where rather than deal with the computer in its vocabulary, we are able to talk to it in our vocabulary, in natural language.”

When customers visit your website, you control the experience. You decide what information they see, how it’s presented, and what actions they can take. But when an AI agent becomes the intermediary, that dynamic changes. The agent interprets your business, decides what information is relevant, and determines how to respond to the user’s request, sometimes without the user visiting your site at all.

The organizations that succeed in the Agentic Web will be those that make their data easy for machines to understand, trust, and act on. Because the future isn’t just about being found by AI, but also understood by it.

Key Questions & Answers From the Webinar

The following questions were asked by attendees during the live session.

Q1: Why was Schema.org originally created?

One of the most interesting moments in the discussion was R.V. Guha’s explanation of why Schema.org was created in the first place.

As search engines evolved, it became clear that websites needed a way to provide structured information directly to machines. The challenge was that every website stores its data differently. If search engines had to interpret millions of unique database schemas, the web simply wouldn’t scale.

It became clear that the right thing to do was to make it possible for any website that wanted to, to give the search engines the actual database, not go through the web page, the HTML shenanigans and so on.”

But there needed to be a common language:

If every website gave us a database in their own schema, it would have the power of double. And so we needed Schema.org.”

That common vocabulary became Schema.org, enabling websites to communicate information in a consistent format that search engines could understand at scale.

Today, Guha noted, the same principle applies to AI systems. Structured data provides a shared framework that helps machines interpret information more accurately, making Schema.org as relevant to the AI era as it was to the search era.

Q2. What is NLWeb?

NLWeb is designed to make it easier for both people and AI agents to interact with website content using natural language. Rather than creating separate experiences for humans and machines, NLWeb provides what Guha described as “a single endpoint” that can serve both.

At its core, NLWeb helps turn a website’s underlying data into a conversational experience. Guha explained that websites already contain valuable information, but users are often limited by the navigation and interfaces a site provides. Natural language allows people and agents to ask questions directly and get answers from that data. As he put it, “Natural language is… a hell of a lot more expressive” than traditional website navigation.

NLWeb builds on structured data standards like Schema.org, which AI systems already understand well. By making website content easier for machines to interpret, organizations can prepare for a future in which AI agents don’t just read web pages but actively interact with and act on behalf of users.

Q3. What is the relationship between Schema Markup, AI bots, and agentic commerce? Do these bots even consume JSON LD, and if so, at what point?

Schema Markup provides a common language that helps AI systems understand the products, services, entities, and business information on your website. When asked why Schema.org matters in an AI-driven world, Guha explained that AI systems already understand it exceptionally well because they have been trained on data from millions of websites that use it.

As Guha shared: “These chatbots understand Schema.org incredibly well… They get less confused when they have the structured data with precise semantics.”

This becomes especially important as we move toward agentic commerce. Whether an AI agent is comparing products, answering questions, recommending vendors, or helping complete a purchase, it needs accurate, machine-readable information about your business. Structured data helps provide that clarity by reducing ambiguity and making it easier for AI systems to understand what is true about your products, services, and brand.

The better AI understands your business, the more likely it is to accurately represent, recommend, and act on your behalf in AI-driven experiences.

Q4. For organizations that already have Schema Markup on their website, what should they be thinking about/prioritizing next?

Schema Markup is an important first step, but preparing for the Agentic Web requires organizations to think beyond individual pages and focus on the data foundation that AI agents can understand, trust, and use.

Three priorities stand out to us:

1. Move from page-level markup to connected data.

Most schema markup implementations focus on describing individual pages. The next step is connecting the entities across your website, such as products, services, people, locations, and topics, into a coherent Content Knowledge Graph. This helps AI systems understand not just what’s on a page, but how information across your organization relates, effectively building context and content coherence.

2. Create a governed source of truth.

Organizations should establish a semantic data layer that serves as a trusted source of truth for both search engines and AI systems. This includes ensuring data remains accurate, synchronized with website content, and governed over time.

3. Build a modern data stack for AI consumption.

Historically, websites were designed primarily for human visitors. In the Agentic Web, organizations will need infrastructure that also makes their content and knowledge accessible to machines. This means considering how structured data, Knowledge Graphs, APIs, content systems, and emerging standards like NLWeb work together to make information discoverable and actionable for AI agents.

As AI agents become more influential in how customers discover and evaluate businesses, organizations need more than page-level markup. They need connected, trustworthy data that helps machines understand what is true about their business and how key concepts relate to one another.

Q5. Looking at Schema.org for preparation in the agentic web and using NLWeb, are there specific properties you would prioritize?

During the webinar, Mark highlighted the Action portion of the Schema.org vocabulary as an area organizations should pay particular attention to as they prepare for the Agentic Web. While traditional schema markup helps AI understand information about your business, Actions help AI understand what can be done with that information, whether that’s searching your site, registering for an event, scheduling an appointment, or completing a transaction.

As AI agents evolve from simply answering questions to helping users accomplish tasks, these action-oriented signals become increasingly important. That said, Schema.org Actions work best when they’re built on a strong semantic foundation. Organizations should continue to prioritize accurately describing and connecting their core entities, including products, services, people, locations, and contact points, through a Content Knowledge Graph.

In short, every Schema.org type and property improves machine understanding, but Action-related markup is particularly valuable for agentic experiences because it helps agents understand not just what your business is but also what users can do next. As agents become more involved in customer journeys, the ability to understand and act will become increasingly important.

Q6. Do you think every organization will eventually need an AI-facing layer of their website, similar to how they needed a mobile strategy a decade ago?

At Schema App, we refer to this AI-facing layer as a Content Knowledge Graph: a semantic data layer that gives machines the context they need to understand your business, including your products, services, expertise, and brand.

As we discussed in the webinar, we’re moving toward a future where AI agents play a larger role in how people discover information, evaluate options, and complete tasks online. Guha described NLWeb as a way to create “a single endpoint” that can serve both humans and agents, rather than building separate experiences for each.

The bigger takeaway is that organizations should focus less on building AI interfaces today and more on preparing the data behind them. The interfaces, protocols, and agent experiences will continue to evolve. Mark van Berkel captured this perfectly when he said, “Standards will come and go, but the data will stay constant.”

Q7. What evidence do we have that agents use Schema markup? I’d love to read some comprehensive data on the relationship.

There is strong evidence that AI systems understand and benefit from Schema.org markup. Guha explained that because AI systems have been trained on data from millions of websites that use Schema.org, they can interpret structured data effectively and “get less confused when they have the structured data with precise semantics.”

This isn’t just a theory. Both Google and Microsoft have publicly documented the role of structured data in helping search engines and AI systems understand content. Google continues to describe structured data as a critical mechanism for helping machines interpret web content and power modern search experiences.

Microsoft has similarly published guidance on optimizing content for AI search answers, including the use of structured data to improve machine understanding and retrieval. Given that Schema.org is already deployed across tens of millions of websites, it has become one of the most widely adopted machine-readable standards on the web.

We’re also seeing this reflected in real-world outcomes. In our own research, improving entity relationships through structured data and Knowledge Graph optimization increased AI Overview visibility by 19.72%. We’ve also worked with enterprise organizations like Wells Fargo to strengthen how AI systems understand and represent their content, using schema markup to resolve AI hallucinations.

Q8. If using NLweb on a site, is there an upside to adding this technology to an llms.txt as guidance?

This question came up during the webinar, and Mark’s answer was essentially: there’s probably some upside, but the jury is still out.

As Mark explained, llms.txt is designed to make content easier for AI systems to consume, and there’s likely value in providing information in a more machine-friendly format. At the same time, he cautioned that publishing an llms.txt file isn’t the same as knowing that AI systems are actively using it:

I would like some better indication from the big LLM providers that they are formally supporting it on the consumption side.”

What Mark was more excited about was the opportunity to make your content and capabilities directly discoverable to agents through approaches like NLWeb. He described NLWeb as a way to define agentic endpoints, expose actions, and create more efficient interactions between websites and AI systems.

It’s a way in which you can guide a more efficient use.”

If you’re experimenting with llms.txt, there’s little downside. But the bigger opportunity is investing in structured, connected, machine-readable data that can support emerging standards like NLWeb and future agent interactions.

Q9: During the webinar, a new Agentic Resource Discovery Specification was announced. What is it, and why is it important?

One of the most timely moments in the webinar came when R.V. Guha announced the newly released Agentic Resource Discovery Specification live during the session.

As Guha explained, the agent ecosystem is rapidly expanding. Organizations are creating new tools, skills, agents, and services that AI systems can use, but there is currently no standardized way for agents to discover those resources.

We’re now getting to a stage where there are so many companies, so many businesses that are creating different agentic resources, things for agents to use. Tools, skills, a whole bunch of new things.”

Today, most AI tools require users to manually find and connect resources. Guha compared this to the early days of the web, when people relied on lists of bookmarks before search engines emerged.

What we need is a search engine for agents that they can use to figure out what tools are available.”

Rather than creating a proprietary solution, Guha and collaborators across the industry chose to develop an open protocol that organizations can adopt broadly.

We decided to do this as a protocol with some reference implementations from Hugging Face, GitHub, and so on.”

GitHub announced support for agent discovery with its new Agent Finder for GitHub Copilot, demonstrating how these emerging standards are moving from concept to implementation. As AI agents become more capable and more widely adopted, discovery mechanisms like the Agentic Resource Discovery Specification could become foundational infrastructure for helping agents find the tools and resources they need to complete tasks effectively.

Learn more about the release of Agentic Resource Discovery Specification here.