Martha van Berkel opened her Tech SEO Connect presentation with a challenge: stop thinking like an SEO and start thinking like a data architect. Coming from the co-founder and CEO of Schema App—someone with 14 years at Cisco and degrees from MIT and Queen’s—that’s not just motivational talk. It’s a strategic reframe that every SEO professional needs to hear right now.
Her core thesis? Schema markup isn’t just about rich results anymore. It’s the foundation for how AI systems—and soon, autonomous agents—will understand your brand. If you’re still thinking of structured data as an SEO tactic, you’re missing the bigger picture.
From Schema Markup to Content Knowledge Graphs
Van Berkel started with a demonstration I’ve seen her do before, but it never gets old: she introduced herself using her own knowledge graph. Not a bio—a graph. She’s a rower, she attended MIT and Queen’s, she’s Canadian, she co-founded Schema App, and she once owned a 1965 Austin Healey Sprite that Kevin Bacon drove in a movie he directed.
The point? By understanding those relationships, you can now win the Six Degrees of Kevin Bacon game. That’s what knowledge graphs do—they enable inference. And that’s exactly what we’re trying to give AI systems when we implement schema markup properly.
She drew a clear distinction between basic schema markup (optimizing individual pages) and a true content knowledge graph (defining relationships between entities across your entire site). The difference matters because it’s not just about getting a rich result on one page—it’s about building a complete, connected data layer that represents your brand.
When you make that leap, she argued, you unlock more than SEO value. You get insights into topic authority. You get AI readiness. Your data becomes something that can be consumed and accessed by systems that go far beyond traditional search.
The Semantic Web in Motion
Van Berkel referenced a 2001 Scientific American article by Tim Berners-Lee about the semantic web—the original vision of machines taking action on our behalf using structured, interlinked data. Her framing: the agentic web is the semantic web in motion.
And it’s not future tense. Google is already building shopping experiences where you never leave to complete a purchase. OpenAI is doing native integrations in their chat interface. The infrastructure for agents to take action is being built right now.
This is where things got really interesting. Van Berkel introduced NLWeb, a Microsoft project announced in May that most of the room hadn’t heard of. NLWeb aims to be an open standard for the agentic web—think of it as HTTPS for agent interactions.
Here’s the kicker: NLWeb is led by Arvi Guha, the founder of schema.org. It uses structured data and RSS to inform a vector database, essentially creating a natural language interface—like an out-of-the-box chatbot—for your website. The schema markup we’re already doing becomes the data layer that powers it.
As van Berkel put it: “SEOs, the structured data that we’re doing is actually the data layer that’s going to help our organizations prepare for the agentic web.”
Building a Content Knowledge Graph: The Checklist
Van Berkel provided a tactical framework for building a content knowledge graph. This isn’t the “add some FAQ schema” approach. It’s systematic data architecture.
Be Specific About Types
Schema.org has over 840 classes. Don’t call everything a WebPage. Get specific. As Krishna Madhavan noted in his talk earlier that day, AI engines use schema as a steering mechanism to understand what a page is about. Generic typing wastes that opportunity.
Go Beyond Required Properties
Google’s documentation typically covers about 10% of the properties available in schema.org. The rest still matter for AI understanding. Be articulate about the structure and content of each page.
Go Deep
Every element on a page should be defined in relationship to the main entity. Van Berkel showed an example where the page was about one thing, and everything else on the page was defined in relationship to that one thing. Not multiple disconnected schema blocks—one connected graph.
Think Breadth
Five years ago, prioritizing schema for rich results made sense. Now the question is different: What does your brand need to be fully understood? Make sure you’re translating that completely in your data layer.
Connect Everything
This is entity linking—defining relationships between entities so AI can traverse the graph and make inferences. Don’t let your product pages, blog posts, and people exist as isolated islands. Define the relationships explicitly.
Entity Linking: The Data Behind the Results
Van Berkel shared data from two case studies on entity linking. When Schema App worked with Brightview Senior Living to implement entity linking around locations, non-branded queries for those locations increased—both clicks and impressions. Similar results with Syncrometer when disambiguating products: increased clicks and impressions on non-branded queries tied to those entities.
There are two types of entity linking: external (connecting to authoritative sources like Wikidata, Wikipedia, or Google’s Knowledge Graph) and internal (identifying the “entity home”—the one page on your site that definitively represents that entity). Think of internal entity linking as backlinking, but with context.
Van Berkel’s team has started analyzing Search Console data by entity, which lets you track how your content investments are performing at the entity level rather than just the page level. That’s a meaningful shift in how we measure topical authority.
Agent Readiness: MCP, NLWeb, and the Access Layer
The second half of van Berkel’s talk focused on what comes next: preparing for autonomous agents. Her framework has three components: your knowledge graph needs to be accessible, correct, and complete. You need AI governance. And you need to think about agentic endpoints.
She walked through two emerging standards. MCP (Model Context Protocol) has been around for about a year and is becoming the de facto standard for how agents access data. It’s essentially APIs for AI—you’re being directive about what an agent can access and how.
NLWeb builds on this. It includes an MCP server, and within a week of Google’s Atlas announcement, the NLWeb open source code included an Atlas SDK. What’s powerful about NLWeb is that your schema markup essentially trains the vector database that powers the natural language interface. You don’t have to build separate MCP infrastructure—NLWeb handles it.
Van Berkel emphasized that Microsoft is pushing hard for NLWeb to be an open standard, similar to HTTPS. Schema App is already running NLWeb on their own site as early adopters, though she noted it’s still buggy.
Knowledge Graphs Solve Hallucinations
Van Berkel made a compelling case for governance, citing research showing that knowledge graph grounding achieves 91% accuracy compared to 43% from GPT-4 alone. Other studies show 300% more accuracy in enterprise LLM responses when using graph data versus traditional relational databases.
Then she shared a case study that got the room’s attention. Wells Fargo had a location that Google’s AI Overview was incorrectly showing as “permanently closed.” They reached out to John Mueller, tried various fixes—nothing worked. The page didn’t have schema markup connected to the rest of their knowledge graph.
When Schema App added robust location schema and connected it to Wells Fargo’s broader knowledge graph, AI Overview started citing the location page instead of a 25-year-old news article—within days.
Schema markup solved the hallucination because it provided grounded, trusted data. That’s not an SEO win. That’s a brand reputation win.
Thinking About Agentic Endpoints
Van Berkel closed with something I haven’t heard many people discuss yet: agentic entry points. We need to start thinking about what a conversion looks like when it’s an agent taking action rather than a human clicking a button.
This requires collaboration beyond SEO and content teams. IT and DevOps become critical partners because you’re now defining rules and conditions for when actions can happen, which agents are authorized, and how to maintain a registry of all the actions your business wants to enable.
She mentioned booking appointments as a simple example—passing agent data to the appropriate endpoint. But the principle extends to any conversion action. If an agent is going to book, buy, or schedule on behalf of a user, your data layer and access points need to be ready.
My Takeaways
Van Berkel’s talk reinforced something I’ve been telling clients: if you ever needed a business case for comprehensive schema markup, now is the time. But she expanded my thinking on where this is headed.
Here’s what I’m taking away:
1. Schema markup is brand architecture now. It’s not a technical SEO task—it’s defining how AI systems understand your brand. Treat it with that level of strategic importance.
2. Entity linking drives measurable results. The Brightview and Syncrometer case studies show real gains in non-branded visibility. This is worth prioritizing.
3. NLWeb is worth watching closely. The founder of schema.org is leading it. Microsoft is pushing for it to become an open standard. And it uses the schema markup we’re already implementing. This could be huge.
4. Knowledge graphs solve hallucinations. The Wells Fargo case study is powerful. If your brand is being misrepresented in AI outputs, robust connected schema markup might be the fix.
5. Start thinking about agentic endpoints. What happens when agents—not humans—are completing conversions? That’s not science fiction. It’s the near-term roadmap.
Van Berkel closed by saying SEOs are going to lead the charge into the agentic web. I think she’s right. The work we’re doing on structured data today is building the data layer that agents will rely on tomorrow.
Time to start thinking like data architects.
