In one line: from “read” to “asked”
For AI to use your content, it used to send a crawler to grab the whole HTML page, then guess for itself which part was the price, which was the spec, and which was filler. NLWeb flips that — you actively expose your organized data as an endpoint you can query in natural language, so AI doesn’t guess, it asks.
It’s a project Microsoft open-sourced in 2025. It invents no new format. Instead it consumes your site’s existing Schema.org structured data plus RSS/feeds, wraps a retrieval and language-model layer around them, and turns your content into a conversational interface. The key word is existing: the structured data you built for SEO/GEO isn’t redone here — it’s fed in directly as fuel.
How it works: the path from your page to the answer
Break it into three stages and it’s clear:
- Ingest — NLWeb reads the Schema.org on your pages (Product, Article, FAQPage, LocalBusiness, and the like) plus RSS. The cleaner the structure and the more consistent the entities, the better the data this stage gets. The quality of this stage equals the quality of your site health — garbage in, garbage out.
- Retrieve — it builds that data into a semantically queryable index (the vector-retrieval kind). A user or agent asks a question, and the system first finds the relevant pieces inside your content.
- Answer — the language model generates an answer grounded only in the retrieved pieces of your content, not from model memory it makes up. In other words, the answer is tied to your data — exactly the “sourced and accountable” shape everyone wants.
The crucial difference: a traditional crawler reads your page as a blob of text; NLWeb lets AI query the facts you’ve already structured. For the same question — “are you open on Sundays?” — the former has to guess from a paragraph; the latter hits your openingHours field directly.
Every instance is also an MCP server — that’s the real point
NLWeb’s real weight isn’t “a chat box on your website.” It’s that every instance is also an MCP server.
MCP (Model Context Protocol) is the interface that lets AI agents call external tools and data sources. When your website is an MCP server, it means: in the coming agent ecosystem, AI doesn’t just “happen to crawl you” — it can treat you as a tool it actively calls — “go ask this shop about stock.”
That’s the signal of a shift from “web pages get browsed” to “endpoints get called.” As agents start running errands for users, the sites that can be called directly and answer with structure get wired in before the ones left as a pile of static HTML.
Why it’s the same road as GEO
GEO has only ever been one sentence: make AI treat you as a trustworthy, well-structured, easy-to-consume source.
NLWeb pushes that sentence to its limit. It doesn’t ask you to learn something new — it cashes in the work you’ve already done for GEO:
- You cleaned up your Schema.org → NLWeb ingests accurately.
- You made your entities (company name, product names, locations) consistent → retrieval doesn’t mix them up.
- Your content was already “one clear identity per page” → exposed as an endpoint, each identity answers cleanly.
The flip side: a site in poor health is hollow on NLWeb too. Patchy structured data and entities scattered everywhere only get amplified into bigger chaos once exposed as an endpoint. NLWeb isn’t a shortcut — it’s a mirror that shows how solid your foundation really is.
Who should care now, and who shouldn’t touch it yet
The honest part first: NLWeb is not “install it and mainstream AI cites you.” It’s an open-source experiment, not an industry standard, adoption is early, and mainstream AI engines have made no promise to “always prefer NLWeb endpoints.” Treat it as a silver bullet and you’ll be disappointed.
So where’s the value? In direction. It marks the road — “structured data + a conversational interface = wired in first” — very clearly. The pragmatic split:
- Worth watching now: content-heavy, many-SKU, query-dense sites (e-commerce, directories, media, knowledge bases). These have the most “being asked” scenarios, and the biggest upside as agents start querying sites directly.
- No need to self-host yet: the typical small-business site. Spending effort standing up an NLWeb server now pays off far less than getting your site health right first — because health is the precondition for all of this, and it stays useful no matter how many interface generations come and go.
Note the division of labor: standing up your own endpoint is engineering work that can wait; but getting your structured data and entities right is something to do now, and regardless. The former is optional; the latter is the foundation.
What it means for you
You don’t need to stand up NLWeb today. But it confirms the same thing again, more bluntly than before:
Clean structured data + consistent entities are the foundation. New interfaces are just outlets that cash it in.
NLWeb, MCP, and whatever interface surfaces next all stand on the same premise — how structured your content is, and how consistent your identity is. Without the foundation, any interface you bolt on is hollow; with a solid one, every new outlet automatically becomes your upside.
That’s exactly why we put the effort into site health rather than chasing every new toy. Toys get swapped out. The foundation doesn’t.