NLWeb is Microsoft's attempt at an "HTML for chat": a standard way for a website to answer natural-language questions from people and from AI agents. It was announced at Build 2025 and designed by R.V. Guha, who created RSS, RDF and Schema.org. Sixteen months on, the question for brands is whether it is worth implementing, or whether llms.txt and a good MCP server are enough.
What NLWeb Actually Is
- A protocol. The core method is ask. You send a natural-language question and get back JSON built on Schema.org types.
- A reference implementation. An MIT-licensed Python framework on GitHub that turns a site's Schema.org markup and RSS feeds into a searchable index and a conversational endpoint.
- An MCP server by default. Every NLWeb instance also serves the Model Context Protocol, so the same index answers people through /ask and agents through /mcp.
Because it builds on Schema.org and RSS, which the NLWeb project says over 100 million websites already publish, a site with good structured data already has most of what NLWeb needs.
Adoption: What Can and Cannot Be Measured
No public count of NLWeb deployments exists. Unlike llms.txt, which lives at a fixed URL and can be counted by crawling (see the state of llms.txt), NLWeb endpoints have no single standard location. Any "X thousand sites" figure you see is an estimate. What can be verified:
| Signal | Status, September 2026 |
|---|---|
| Launch partners named by Microsoft | Chicago Public Media, Allrecipes, Eventbrite, Hearst (Delish), O'Reilly Media, Tripadvisor, Shopify, Snowflake, Inception Labs |
| GitHub repository | About 6.3K stars and 700 forks, MIT licence |
| Supported vector stores | Qdrant, Snowflake, Milvus, Azure AI Search, Elasticsearch, Postgres, Cloudflare AutoRAG |
| Supported model providers | OpenAI, Anthropic, Gemini, DeepSeek, Inception, Hugging Face |
| Managed deployment | Cloudflare AutoRAG (now AI Search) offers an "NLWeb Website" quick deploy that crawls, indexes and serves a site |
The Cloudflare route matters most for adoption. It turns NLWeb from an engineering project into a dashboard setting for any site already on Cloudflare.
Security History
In 2025 a researcher disclosed a path-traversal flaw in the reference implementation that could expose files such as environment configs holding API keys. Microsoft patched it. Anyone self-hosting should run a current build and treat the endpoint like any other public API.
NLWeb, llms.txt and WebMCP Compared
| llms.txt | NLWeb | WebMCP | |
|---|---|---|---|
| What it gives agents | A curated reading list | A queryable answer endpoint plus an MCP server | Actions exposed by the page in the browser |
| Effort | One file | An index and a service, or one click on Cloudflare | Front-end code |
| Best for | Docs and content sites | Catalogs, recipes, events, listings | Transactional flows with a browser agent |
They work together rather than compete. See WebMCP and MCP ecosystem statistics.
Brand Visibility Implications
An NLWeb endpoint lets an agent ask your site directly instead of guessing from scraped HTML. The answer is built from your own structured data, so it is as accurate as your Schema.org markup. Sites whose catalog is the product, such as retailers, marketplaces, travel and recipes, gain the most. For a B2B brand with a small site, strong Schema.org markup and llms.txt capture most of the benefit today. NLWeb can wait until agent demand for it becomes measurable in your own logs.
Methodology
Figures on this page are drawn from primary announcements and documentation wherever they exist, with established trade and legal press used only to corroborate or to fill gaps a primary source leaves open. Sources: NLWeb GitHub repository, Microsoft announcement, Cloudflare blog, Microsoft NLWeb pioneer profiles. Where outlets disagree on a date or number, the page gives the range or names the disagreement rather than choosing one. Status as of September 2026; this area changes monthly, so re-check before quoting.
How Presenc AI Helps
Presenc AI logs which AI agents call your site and what they fetch, including /mcp and /ask style endpoints. You can measure whether an NLWeb deployment brings agent traffic before committing more engineering time.