What is NLWeb?

NLWeb is an open-source project Microsoft announced at Build in May 2025 that lets a website offer a conversational interface on top of its own content, reusing structured data the site already publishes: Schema.org, RSS, and similar semi-structured formats. Rather than requiring a rewrite of the site, NLWeb reads what already exists (product pages marked up with Schema.org, a publisher's RSS feed) and exposes it behind a layer that answers natural-language questions. Every NLWeb instance is also, natively, an MCP server: an external agent can query the site through the same protocol it uses to connect to any other tool. Conceived by RV Guha, one of the creators of RSS and of Schema.org itself, it remains an open project in progress rather than a universally adopted standard: a handful of publishers and retailers, including Shopify, Snowflake, and Tripadvisor, are experimenting with it, but adoption remains early.

See also: MCP (Model Context Protocol) · Open protocol connecting AI models and agents to external tools and data through one common standard, instead of a custom integration per source.

What it actually does and what it does not solve

The mechanism is simple to describe and harder to execute well. NLWeb builds an index from the site's semi-structured data, combines it with a language model, and returns natural-language answers with citations back to the source pages, whether the requester is a human typing a question or an AI agent querying the site on a user's behalf. Microsoft describes it as a possible HTML-equivalent for the agentic web: a common semantic layer any site can adopt without depending on a single AI vendor. What NLWeb does not do is generate structured data out of thin air: if a site does not already publish correct, up-to-date Schema.org markup, or publishes content rendered only through JavaScript with no static equivalent, NLWeb has very little to build a reliable answer from. It is distinct from agentic SEO and GEO/AEO, which cover how to become readable and citable to agents and generative engines in general: NLWeb is a specific tool that implements that readability as a queryable conversational interface on the site side. It is also distinct from AI crawler access control (who is allowed to read the site) and from agentic resource discovery (how an agent finds which endpoints a site exposes): NLWeb assumes access has already been granted and provides the query layer, not the discovery or the gate.

An enterprise example

A retailer with a catalog of tens of thousands of items, each marked up with Schema.org for price, availability, and specifications, can expose an NLWeb instance that answers "which waterproof trail shoes do you have under 120 dollars" by drawing directly on that markup, instead of forcing the shopper to filter the catalog by hand. The real work, though, is not installing NLWeb: it is verifying that the Schema.org markup is complete, correct, and synced with actual inventory, something almost no enterprise catalog guarantees today without a dedicated audit. An NLWeb instance connected to incomplete or out-of-sync data produces wrong answers with the same confidence it would produce right ones, which is worse than having no interface at all.

Why it matters for decision makers

For anyone who owns an enterprise site, the practical question is not whether to adopt the NLWeb protocol itself, which remains a young project rather than a market requirement: it is whether the structured data the site already publishes is good enough to support a natural-language query, from a human or from an agent acting on one's behalf. In most cases it is not: the markup is partial, out of sync with inventory, or missing on the pages that actually matter. That structured-data cleanup work is worth doing regardless of NLWeb, because it is the same foundation any future conversational interface will need, including ones built on other protocols entirely.