NLWeb for Local Businesses / Brenro

A client forwarded us a headline last week: Microsoft wants every website to talk to AI. The question underneath it was the honest one. "Is this something I need to buy?"

Almost certainly not, and not for a while. But the reason why is worth understanding, because it explains something more useful than NLWeb itself: how AI actually finds a local business right now, and what that means for where your money should go.

Checked against the project as it stood in August 2026. This moves quickly, and we will say so here when it changes.

What NLWeb actually is

NLWeb is an open source project from Microsoft, announced in 2025 and MIT licensed. You run it alongside your website. It reads the structured data your site already publishes, the Schema.org markup and RSS feeds, builds a searchable index out of it, and then exposes a new door: a visitor, or more to the point a visitor's AI assistant, can ask your website a question in plain language and get a structured answer back.

Every NLWeb instance also acts as an MCP server. MCP is the emerging standard for how AI assistants connect to outside tools, so in practice this means someone could point their assistant at your site and have a conversation with it, rather than reading it.

That is a genuinely interesting idea. It is not the idea most people think they are buying.

The part that gets misread

Here is the sentence that matters, and it is the one the headlines skip:

NLWeb is not how ChatGPT, Perplexity or Google's AI find you.

Those systems send crawlers. GPTBot, PerplexityBot, Google's indexing, and the rest fetch your raw HTML, read your Schema.org markup, and move on. They do not go looking for an NLWeb endpoint, and they do not ask it anything. Installing NLWeb today does not make you more likely to be named when a customer asks an assistant who fixes transmissions in Palm Desert.

What NLWeb gives you is a door for agents that someone has deliberately connected to your site. That is a real thing, and it may well matter later. It is not search visibility, and anyone selling it to you as search visibility either has not read the specification or is counting on you not having.

What it costs to run

NLWeb is not a plugin or a snippet you paste in. It is a Python service plus a vector database, which is a specialized store for the numerical representations of your content that make conversational search possible. The project supports several: Qdrant, Postgres, Elasticsearch, Azure AI Search, Cloudflare AutoRAG and others. You also connect a language model provider, which has its own running cost per question asked.

So the real shape of it is: a server you now maintain, a database you now maintain, and a per query bill, attached to a website that until now was probably just files. For a business whose site is a handful of pages, that is a lot of new machinery bolted to something that was working.

Cloudflare added a managed path in early 2026 that makes this considerably less painful if you are already hosted there. That lowers the cost. It does not change what the thing is for.

Who it is actually built for

Look at who is using it: Allrecipes, Tripadvisor, Shopify, Eventbrite, O'Reilly Media, Hearst, Chicago Public Media.

They have one thing in common, and it is not size. It is that each of them holds a large catalog where "find me the one that matches these conditions" is a question a real person genuinely needs to ask. Thousands of recipes. Millions of hotel reviews. A store with a deep inventory. An events calendar. A publisher's back catalog.

Conversational search earns its keep when your content is too large to browse. That is the test. It is not a test about how modern you are.

Three questions to tell if you are the exception

Some local businesses genuinely do have this shape. Ask yourself:

Does someone have to filter to find what they want? A restaurant with a long menu and real dietary variation, a parts supplier with thousands of SKUs, a property manager with a rotating list of units. If a customer's question is "which of your things matches my situation," that is a filtering problem, and filtering problems are what this technology is for. If your customer's question is "are you open and can you take my car Thursday," it is not.
Is your catalog already structured data? NLWeb reads what you publish. If your menu is a PDF, your inventory lives in a spreadsheet, or your listings are hand typed into a page, there is nothing for it to index. The work of getting your catalog into clean structured markup is the real project, and it is the part that pays off whether or not NLWeb ever enters the picture, because that same markup is what the AI crawlers read today.
Do you have somewhere to run it? A static site on simple hosting is a feature, not a limitation: it is fast, it is cheap, and there is very little of it to break. Adding a Python service and a database to it is a decision to start operating software. That is fine if you were going to anyway. It is a strange first step if you were not.

If you answered no to any of these, you are not the exception, and you have not missed anything.

What to do instead, right now

Everything that actually determines whether an AI assistant recommends you is available to you today, and most of it is unglamorous:

Structured data that describes your business accurately. What you do, where you serve, how to reach you, what you charge if you are willing to say. This is what the crawlers read.
Content a machine can lift a clean answer out of. Real headings, plain sentences, the answer near the question. The same thing that makes a page readable to a person.
A Google Business Profile that is complete and current. Still the single largest lever in local AI answers, and still the one most often left half filled.
An llms.txt file. A short plain text summary of your site and what each page is for. Cheap, standardized, and it is the current, working version of the idea NLWeb is reaching toward.
Crawlers that are actually allowed in. Worth checking your robots.txt, because plenty of sites block the assistants they are hoping will recommend them.

None of that is exciting. All of it works now.

Our position

We think NLWeb is pointed at something real. The web is going to need a way for agents to query a site rather than scrape it, and an open standard beats every company inventing its own. We are watching it, and we will implement it for clients whose catalogs justify it.

What we will not do is sell it to a business with six pages and a phone number as an AI upgrade. If that changes, if the assistants start consuming these endpoints and it starts affecting who gets recommended, this page will say so, with the date we checked.

Curious where you actually stand with the assistants today? Run your site through our free visibility check. It takes about a minute, it gives you a plain English report, and it costs nothing whether or not you ever call us.