NLWeb

Your website was built for humans clicking buttons. AI agents don't click buttons. They need structured data, clear answers, and machine-readable actions. That's the gap NLWeb fills.

Here's the thing: AI agents are already driving real traffic. When someone asks ChatGPT or Perplexity "what's the best sales automation tool for mid-market SaaS," an agent goes out and queries websites on the user's behalf. If your site can't respond to that query natively, you're invisible to an entire channel that's growing 10x year over year.

NLWeb turns your website into something agents can actually have a conversation with. It's the difference between your site being a static brochure and being a live participant in AI-driven discovery. For B2B companies running inbound motions, this isn't optional anymore — it's table stakes.

NLWeb operates through a straightforward request-response cycle built on open standards:

1. Discovery. Your website publishes an NLWeb endpoint (typically at /.well-known/nlweb). AI agents find it the same way browsers find robots.txt — it's a convention, not magic.

2. Query. An agent sends a natural language question over MCP's JSON-RPC 2.0 transport. Something like: "What pricing plans do you offer for teams under 50 people?"

3. Intent mapping. NLWeb uses schema.org/Action vocabulary to map that query to a specific action type — SearchAction, AskAction, or a custom action you've defined. This isn't keyword matching. It's semantic understanding.

4. Response. Your server processes the query against your content and returns structured data — not raw HTML. The agent gets exactly what it needs to answer the user's question or complete a task.

The whole thing runs statelessly. No sessions, no cookies, no authentication headaches. Each request is self-contained.

A B2B SaaS company selling marketing automation implements NLWeb on their docs and pricing pages. A prospect asks Claude: "Compare marketing automation tools that integrate with HubSpot and cost under $500/month."

Claude's agent hits the company's NLWeb endpoint with that query. The endpoint maps it to a SearchAction, filters their pricing data, confirms the HubSpot integration, and returns a structured response with plan names, prices, and feature comparisons. Claude includes them in its answer — with a direct link to the pricing page.

No scraping. No outdated cached data. The company controlled exactly what the agent saw, and the prospect got accurate, real-time information. That's the shift NLWeb enables.

• Returning raw HTML instead of structured data. NLWeb responses should be clean JSON with schema.org types. Dumping your page's innerHTML defeats the purpose.
• Ignoring the schema.org/Action vocabulary. If you map everything to a generic SearchAction, agents can't distinguish between "tell me about" and "help me buy." Define specific action types.
• No fallback for unrecognized queries. Your endpoint should gracefully handle questions it can't answer. Return a helpful "I can help with X, Y, Z" response — not a 500 error.
• Treating NLWeb as a replacement for your website. It's a complement. Humans still visit your site. NLWeb is an additional interface layer for agents.
• Skipping rate limiting. AI agents can be aggressive. Without rate limits on your NLWeb endpoint, a single agent could hammer your server with hundreds of requests per minute.