Most websites still answer questions the old way: a search box that matches keywords, or a menu that hopes visitors find the right page. Meanwhile, people have grown used to asking AI assistants full questions in plain language. NLWeb is Microsoft’s open project to bring that conversational experience to individual websites, using data many sites already publish.
It is also designed for machines. Every NLWeb instance doubles as a Model Context Protocol (MCP) server, which means AI agents can query a site’s content in a structured way rather than scraping pages. That combination has made NLWeb a frequent talking point in AI search circles since its launch in May 2025.
This guide explains what NLWeb is, how it uses schema.org and RSS-style data, how it relates to MCP, and whether it is worth experimenting with in 2026.
Key Takeaways
- NLWeb (Natural Language Web) is an open-source Microsoft project, announced on 19 May 2025, for adding natural-language querying to websites.
- It was conceived by R.V. Guha, known for creating RSS, RDF and Schema.org.
- It builds on schema.org markup and semi-structured feeds such as RSS, and returns answers as schema.org JSON.
- Every NLWeb instance is also an MCP server, so AI agents can query it.
- The code is described by its maintainers as a proof of concept; treat it as an experiment, not a ranking tactic.
What NLWeb is
According to Microsoft’s launch announcement, NLWeb lets site owners add a conversational interface to their site using the language model and data of their choice, so visitors can ask questions of the site’s content the way they would ask an AI assistant. It runs on major operating systems and works with several models and vector databases.
The project’s README describes two parts: a simple protocol for asking a site questions in natural language and receiving JSON responses that use schema.org vocabulary, and a reference implementation that works with structured markup sites already have for products, recipes, attractions, reviews and similar content. The code is open source under the MIT licence.
“In short, NLWeb is to MCP/A2A what HTML is to HTTP.”
— NLWeb project, NLWeb README
Microsoft frames the ambition in similar terms, saying NLWeb could play a role in the emerging agentic web comparable to the role HTML played in the early web. That is a vision statement, not a measure of adoption, and it is worth reading that way.
Who was involved at launch
Microsoft named a small group of early publishing and ecosystem collaborators: Chicago Public Media, Common Sense Media, DDM (Allrecipes and Serious Eats), Eventbrite, Hearst (Delish), Inception Labs, Milvus, O’Reilly Media, Qdrant, Shopify, Snowflake and Tripadvisor. We found no official, dated adoption figures for 2026 on the pages we reviewed, so be wary of claims about how widely it is deployed today.
How NLWeb uses schema.org and RSS
NLWeb’s key idea is to reuse data that already exists. The README notes that schema.org and semi-structured formats like RSS are used by over 100 million websites and act as a semantic layer for the web. NLWeb loads that data into a vector database, uses a language model to interpret a question, retrieves relevant items, and returns structured answers.
| Component | Role in NLWeb | What you provide |
|---|---|---|
| Schema.org markup | Describes items (products, recipes, events, articles) in a standard vocabulary | Accurate JSON-LD on your pages or exported data |
| RSS / feeds | A semi-structured list of your content | A complete, up-to-date feed |
| Vector database | Stores embeddings for retrieval | Qdrant, Snowflake, Milvus, Azure AI Search, Elasticsearch, Postgres or Cloudflare AutoRAG (per the README) |
| Language model | Interprets questions and composes answers | OpenAI, DeepSeek, Gemini, Anthropic, Inception or Hugging Face models (per the README) |
| ask method | The core protocol call for natural-language questions | Exposed by your NLWeb instance |
| MCP server | Lets AI agents and other MCP clients query the site | Included in every NLWeb instance |
This is why NLWeb sits naturally alongside good structured data practice. If your schema.org markup is sparse or inaccurate, NLWeb has less to work with. Our schema markup guide and the new article on structured data and AI search cover how to get it right. Google’s introduction to structured data is a good primer on the vocabulary.
NLWeb and MCP
The Model Context Protocol is an open standard for connecting AI applications to external systems. NLWeb makes every site that runs it an MCP server, so an AI agent can call the site’s ask method instead of crawling pages and guessing. The README also lists A2A (agent-to-agent) support as planned.
| NLWeb | Custom MCP server | Plain website | |
|---|---|---|---|
| Main purpose | Conversational Q&A over site content | Any tools, data or actions you choose | Human browsing and crawler indexing |
| Data source | Schema.org and feeds | Your APIs and databases | HTML pages |
| Agent access | Via built-in MCP server | Via MCP | Via crawling or browser agents |
| Effort | Moderate (set-up, hosting, model costs) | Higher (design, security, maintenance) | Already exists |
| Best for | Content-rich sites: recipes, events, catalogues, publishers | Live data and transactions | Everyone |
If you want to go beyond Q&A, our new guide on whether to publish an MCP server covers custom builds, and the existing guide to WebMCP explains the browser-side approach.
Should you experiment with NLWeb?
Be clear about what NLWeb is and is not. It does not change how Google ranks you; Google says there is no special markup needed for its AI features. No major AI assistant has said it prefers NLWeb-enabled sites. The README itself describes the included code as proof-of-concept rather than a definitive implementation, and some deployment guides were still marked "coming soon" when we reviewed it.
It may still be worth a pilot if:
- You run a content-rich site (recipes, listings, events, product catalogue, documentation) where on-site search is weak.
- Your schema.org markup and feeds are already in good shape.
- You have developer time to host, monitor and pay for model and vector database usage.
- You want hands-on experience with agent-facing interfaces before competitors.
Step-by-step: a low-risk NLWeb pilot
- Audit your schema.org markup and RSS feeds for completeness and accuracy.
- Pick one section of the site (for example recipes or events) as the pilot scope.
- Set up NLWeb from the official repository in a staging environment, choosing a model and vector database you already use.
- Load the pilot data and test 50 to 100 real customer questions; record wrong or missing answers.
- Fix the underlying data rather than patching answers.
- Launch to a small share of visitors and compare engagement and conversions with your existing search.
- Decide whether to expand, keep as an internal tool, or retire.
Common mistakes
- Treating NLWeb as an SEO shortcut. It is an interface, not a ranking signal.
- Feeding it poor data. Incomplete schema.org markup produces weak answers.
- Ignoring running costs. Model calls and vector storage are ongoing expenses.
- Skipping security review. An MCP endpoint is a new public interface and should be monitored.
- Overstating adoption. Launch partners are not evidence of widespread use in 2026.
- Neglecting the website. Crawlable, fast, well-written pages remain the main route into AI answers (see can AI crawlers read your JavaScript?).
Related Guides
- Freshness Signals for AI Search: Dates, lastmod and Real Updates
- Syndicated and Duplicate Content in AI Answers: Who Gets the Citation?
- Paywalls and AI Search: Visibility Without Giving Content Away
Frequently Asked Questions
What does NLWeb stand for?
Natural Language Web. It is an open-source Microsoft project for adding natural-language interfaces to websites.
Who created NLWeb?
Microsoft says NLWeb was conceived and developed by R.V. Guha, who is known for creating RSS, RDF and Schema.org.
Is NLWeb the same as MCP?
No. MCP is a general protocol for connecting AI applications to systems. NLWeb is a project for conversational site search, and every NLWeb instance is also an MCP server.
Does NLWeb improve rankings in Google or ChatGPT?
There is no official evidence of that. Google says no special markup is needed for its AI features, and AI assistants have not announced a preference for NLWeb sites.
Do I need schema.org markup to use NLWeb?
NLWeb is built around schema.org and RSS-style data, so good markup and feeds make it far more useful.
Conclusion
NLWeb is a thoughtful, open attempt to make websites conversational for both people and AI agents, built on standards many sites already use. For most businesses in 2026 it is an experiment worth watching, and a worthwhile pilot only if your structured data is strong and you have developer capacity. Our web development team can assess whether a pilot makes sense, and our AI search audit will show whether your structured data is ready. Request a custom quote to talk it through.
