Microsoft Just Made NLWeb the Standard. Salespeak Customers Already Have It.

At Build 2025, Satya Nadella stood on stage and said: "Today we showed you how we are building the open agentic web."

Then Microsoft released NLWeb, an open-source protocol that turns any website into a conversational, AI-queryable MCP endpoint. The person behind it? R.V. Guha. The same mind that created RSS, RDF, and Schema.org, three protocols that shaped how the web works. NLWeb is his fourth swing at rewriting internet infrastructure.

The early adopter list tells you this isn't a research paper. Shopify, Snowflake, O'Reilly Media, Tripadvisor, Eventbrite, and Hearst are already building on it. Shopify launched "Agentic Storefronts," selling products wherever AI conversations happen, not just on shopify.com.

Microsoft compared NLWeb to HTML: "Just like HTML made it easy for anyone to create a website, NLWeb makes it easy for any web publisher to create an intelligent, natural language experience."

That's a big claim. But when the company that runs Azure, Copilot, and Bing throws its weight behind a protocol at its flagship developer conference, the market listens. Every B2B website will need to speak agent-to-agent. And if you're a Salespeak customer, yours already does.

What NLWeb actually is (and why it matters now)

Strip away the press releases and NLWeb does something specific: it turns a website into an MCP server with two endpoints.

The first is /ask, a REST endpoint that accepts natural language queries and returns structured answers. A buyer types "What's your pricing for enterprise?" and gets a Schema.org-formatted JSON response with actual data, not a link to a contact form.

The second is /mcp, an agent-to-agent endpoint built on the Model Context Protocol. This is how AI agents like Claude, ChatGPT, and Copilot communicate directly with your website. No scraping. No HTML parsing. Structured conversation between machines.

NLWeb uses Schema.org structured data as its content source. Your product information, FAQs, pricing, anything marked up with Schema.org vocabulary, becomes the knowledge base that powers the conversational endpoint. The protocol is model-agnostic, so it works with any LLM behind the scenes.

Here's what that means practically: your website stops being a static document that waits to be crawled. It becomes a participant in AI conversations. When a buyer's AI agent researches vendors in your category, your NLWeb endpoint can answer directly with verified, current, first-party information instead of whatever the LLM hallucinated from its training data.

We've been writing about this architectural shift since early 2026. What changed at Build 2025 is that Microsoft gave it an official stamp and an open-source implementation that anyone can deploy.

The agentic web just got real

The concept of an "agentic web" has been circulating for over a year. We published our own open specification for it. But there's a difference between a concept and a movement. Build 2025 was the tipping point.

Look at what's converged in the past twelve months:

• MCP adoption exploded. Over 97 million monthly SDK downloads. Anthropic donated MCP governance to the Linux Foundation in December 2025, cementing it as an industry standard rather than a single-vendor play.
• Roughly 90% of organizations are expected to use MCP by end of 2025, a number that seemed absurd eighteen months ago.
• Shopify launched Agentic Storefronts, letting merchants sell wherever AI conversations happen. Not on a website. Inside the conversation itself.
• OpenAI shipped Agent Mode in the Atlas browser. Agentic browsers like Fellou and Atlas are autonomously navigating websites, comparing products, completing transactions, without a human clicking anything.
• AI referral traffic is up 357% year-over-year as of June 2025. That growth isn't slowing down.

Each of these developments matters on its own. Together, they paint a clear picture: AI agents are becoming primary research tools for B2B buyers, and the websites that can't communicate with those agents will be invisible to the next generation of buying decisions.

This is especially acute in B2B. Enterprise buying cycles increasingly start with "ask Claude to research vendors in this space" or "have ChatGPT compare our shortlist." If your website can't answer those queries directly through an agent-readable endpoint, you're relying on whatever the LLM scraped from a blog post eighteen months ago. That's not a strategy. That's a gamble.

Why most B2B websites aren't ready

NLWeb is open-source and available on GitHub. In theory, anyone can deploy it. In practice, the gap between "available" and "implemented" is where most B2B companies will stall.

The first problem is Schema.org markup. NLWeb depends on it; structured data is the content source that powers the conversational endpoint. But most B2B SaaS websites have minimal Schema.org implementation. Maybe some basic Organization markup. Maybe a few FAQ schemas for SEO. Rarely the comprehensive product, pricing, and feature markup that NLWeb needs to generate useful answers.

The second problem is infrastructure. Standing up an MCP server requires vector database integration, LLM connectivity, and ongoing maintenance. That's not a weekend project for a marketing team. It's an engineering initiative that competes with product roadmap priorities, and in most B2B companies, it loses that competition.

The third problem is content strategy. Even with the right infrastructure, most companies haven't thought about what an AI agent querying their site actually needs to know. Agent queries aren't like Google searches. An agent asks "What's the total cost of ownership for a 200-seat deployment including implementation?" not "enterprise pricing page." Your content needs to be structured to answer the real question, not just rank for the keyword.

And there's a fourth problem that's easy to miss: most B2B companies are still optimizing exclusively for Google's PageRank. Their SEO teams haven't connected the dots to Answer Engine Optimization, let alone direct agent queryability. They're playing a game that's being disrupted while the new game has already started.

The result? Microsoft announces NLWeb, and for the average B2B company, actually having a working MCP endpoint is still a multi-month project they haven't scoped.

What Salespeak ships out of the box

Every Salespeak deployment includes an NLWeb-compatible MCP endpoint. Not as an add-on. Not as a premium tier. It's part of the core product.

Here's what that means concretely:

Your website content becomes queryable. Product information, pricing, features, FAQs, documentation. Salespeak automatically structures this content and exposes it through a conversational endpoint. You get both the /ask endpoint for direct REST queries and full MCP compatibility for agent-to-agent communication.

Schema.org structured data is generated automatically. This is the part that trips up most NLWeb implementations. Salespeak handles the markup generation from your existing content, creating the structured data layer that NLWeb requires without your team needing to become Schema.org experts.

Any AI agent can discover and query your endpoint. Claude, ChatGPT, Copilot, Perplexity, custom enterprise agents. When they research your company, they can interact with your Salespeak-powered endpoint directly. The response they get is current, verified, and authored by you. Not a hallucination.

It works alongside the human-facing AI sales agent. This is what makes the architecture clean. The same infrastructure that powers your Intelligent Front Door for human visitors also serves AI agents researching on behalf of buyers. One deployment. Two audiences. No duplicate systems.

The practical scenarios matter here:

• A buyer asks ChatGPT "tell me about [your company]'s pricing for mid-market". The AI agent queries your MCP endpoint and gets a verified, first-party answer. Not a guess from 2024 training data.
• An agentic browser like Atlas compares vendors in your category. Your endpoint provides structured, accurate information that the agent can reason about.
• A procurement team's custom AI agent evaluates your security posture. Your endpoint returns current certifications and compliance data, not a scraped paragraph from an old blog post.

You don't need to hire engineers to implement NLWeb separately. You don't need a vector database project. You don't need an MCP server rollout. It's part of your Salespeak deployment, maintained and updated as the protocol evolves.

The full architecture is documented in the Agentic Web specification, the open standard we published alongside these capabilities.

What this means for SEO and AEO

Traditional SEO optimizes for Google crawlers. AEO (Answer Engine Optimization) optimizes for AI-generated answers. NLWeb adds a third layer: direct agent queryability.

These three layers now coexist, and each serves a different function:

1. SEO: Google indexes your pages, organic search traffic. Still important. Still declining as a share of discovery.
2. AEO: LLMs cite your content, AI-generated recommendations. Growing fast. But you can't fully control accuracy; you're influencing, not dictating.
3. NLWeb/MCP: AI agents query your endpoint directly, real-time verified answers. The most control. The most accuracy. The newest layer.

Companies that only do SEO are becoming invisible to AI. Companies that add AEO show up in AI answers but can't guarantee accuracy; the LLM might still get your pricing wrong or conflate your features with a competitor's. Companies with MCP endpoints give AI agents the authoritative answer directly. No interpretation. No hallucination. First-party data.

There's an ironic twist here. Schema.org markup, previously dismissed by many B2B marketers as an SEO nice-to-have that maybe helped with rich snippets, is now the foundation for all three layers. Search Engine Land called it "your greatest SEO asset" in the agentic web. They're right. Schema.org is the shared vocabulary that connects Google crawlers, LLM training, and NLWeb endpoints.

The companies that invested in structured data early are now positioned for the agentic web almost by accident. Everyone else is starting from scratch.

Key takeaways

• NLWeb is real and backed by Microsoft. Build 2025 moved it from open-source project to industry initiative. Shopify, Snowflake, O'Reilly, Tripadvisor, Eventbrite, and Hearst are already on board.
• Every B2B website will need an MCP endpoint. AI agents are becoming primary research tools for buyers. Websites without agent-readable endpoints will be invisible to agent-mediated purchasing decisions.
• Implementation is the bottleneck. NLWeb is open-source, but deploying it requires Schema.org markup, vector databases, LLM integration, and ongoing maintenance. Most B2B companies haven't started.
• Salespeak includes it out of the box. NLWeb-compatible MCP endpoint, automatic Schema.org generation, and dual human/agent serving, all part of every deployment.
• SEO, AEO, and NLWeb are three layers of the same stack. Schema.org markup is the foundation for all of them. Invest there first.

The standard just arrived

Microsoft validated at Build 2025 what we've been building toward: the web is becoming conversational, and every website needs a machine-readable front door. The agentic web isn't a prediction anymore. It's a protocol with major enterprise adopters and 97 million monthly SDK downloads behind it.

NLWeb is still early. The protocol will evolve. Best practices haven't solidified. But the direction is locked in. AI agents will query websites directly, and websites that can't respond will lose deals they never knew existed.

The question for B2B companies isn't whether to make their website agent-ready. It's how fast they can get there, and whether they want to build the infrastructure from scratch or ship it as part of their existing stack.

If you want to see what an NLWeb-compatible MCP endpoint looks like in practice, we'd be happy to show you.