NLWeb: Schema.org and MCP as a Conversational Site Interface

NLWeb is a protocol for giving a website a conversational interface built from structured data it already publishes. It was developed at Microsoft by RV Guha, who previously created RSS, RDF and Schema.org, and it is deliberately continuous with that lineage. The mechanism is a reuse argument. Language models already understand Schema.org vocabulary, because it is abundant in their training data. So rather than each site building bespoke natural-language processing over its own catalogue, NLWeb ingests the site's existing JSON-LD markup — plus RSS feeds or database exports — and serves queries against it. Every instance also exposes an MCP endpoint, so the same interface answers a human's typed question and an agent's programmatic call, with multi-turn context preserved across a conversation. The appeal for site owners is that it does not require re-platforming. Markup added years ago for rich results becomes the substrate for an agent interface, which is an unusually direct payoff for structured data work that many treated as pure SEO overhead. The caveats are worth stating. NLWeb is software to deploy and operate, not a file to publish — it is a running service, with the hosting and maintenance that implies. And its value is bounded by markup quality: a site with thin or inaccurate Schema.org annotations gets a thin or inaccurate conversational interface, because there is no separate source of truth. It is also the most concrete counter-argument to the usual reading of Why the Semantic Web Underperformed Expectations. The classic complaint was that nothing consumed structured markup well enough to justify producing it. Language models turned out to be that consumer — arriving roughly two decades after the markup effort began, and from a direction nobody in that effort predicted. Related: JSON-LD, Rich Result (Search), MCP Server Discovery: Two Competing .well-known Proposals.