Microsoft’s plan to fix the web: letting every website run AI search for cheap

“So the 10,000-foot-level view is that we’ve had three big revolutions in personal computing.” That’s how Ramanathan V. Guha, a technical fellow at Microsoft, begins his explanation of what I had thought was a relatively minor AI announcement coming at this year’s Build developer conference. But Guha continues to make his case that what he has created — a new open protocol for the web called NLWeb — is actually an important part of something truly enormous.

Oh, the three revolutions: graphical user interfaces, the internet, mobile. Guha says we’re in the middle of the fourth, but doesn’t just chalk it all up to artificial intelligence. For him, the new revolution is “being able to communicate with applications, and computers in general, with free-form language.” He loves the trend, but not the way it’s shaping up. Too much of that new communication, Guha thinks, is mediated by products like ChatGPT, Claude, and yes, even Bing. He doesn’t like the idea that the web will be utterly consumed by chatbots, which take all their knowledge and return no value. And he thinks he knows how to fix it.

Guha’s big idea is to make it easy for any website or app owner to add ChatGPT-style interaction features. With a few lines of NLWeb code, your choice of an AI model, and whatever data you supply to the model, you can have a custom chatbot up and running in just a few minutes. “It’s a protocol,” Guha says, “and the protocol is a way of asking a natural-language question, and the answer comes back in structured form.” Basically, NLWeb handles all the logistics of turning a question into an answer, and all you do is supply the data. By limiting things to your platform, and your area of expertise — plus some of the general world knowledge most models have now — you might be able to quickly make something far better than a general-purpose chatbot.

Guha shows me a few examples. He opens up a demo of Serious Eats that includes NLWeb, opens up the search box, and types “give me spicy and crunchy stuff that I can use as an appetizer.” A couple of seconds later, it pops up a few links. Then he refines the search, responding that he’s looking to make something for Diwali. He also mentions he’s a vegetarian. “This thing understands that it needs to remember that,” Guha says. From now on, the idea goes, Serious Eats will only ever recommend vegetarian dishes.

Then he goes to an outdoor retailer, and runs a similar search: he’s looking for a jacket warm enough to wear in Quebec. Not a typical clothing company search result, but the model works, and displays a bunch of product options, complete with photos and related information, that fit the bill. “It not only brings in knowledge about the weather,” Guha says, “it allows them to render it in a UI that is conducive to things like sponsored links.”

The point here is not that these searches are possible, it’s that almost any developer or website owner can deploy them. They don’t have to rely on some external AI product to answer people’s questions, and just hope and pray the bot sends the link to them. They can run the bot themselves. Microsoft is working with a number of publishers, along with companies like TripAdvisor, Eventbrite, and Shopify, to build natural interaction into the platform directly. “The idea is that rather than having these custom, one-on-one deals like Shopify now has with OpenAI, it’s an open protocol,” Guha says.

The other upside of all this? It’s cheap. Much cheaper than traditional search, Guha says. “In order to have a web search index,” he says, “you need to crawl the web and have an index that contains it.” That’s expensive for both the search engines and the websites involved. With NLWeb, Guha says, “I just take an RSS feed, put it in a vector database, and it runs off that.” It can call an inexpensive model — Guha’s demos are using GPT-4o Mini, but he says there are even cheaper options — and the whole thing is fast and easy. “It allows for the remixing, and the back and forth, at an incredibly low price.”

I should say, by the way, that all of this would sound incredibly rich coming from just about any tech executive, let alone someone at Microsoft, which is hell-bent on getting you to use Azure and Copilot for everything forever. But Guha has earned his credibility here: over the course of his career, he was instrumental in creating RSS, which helped make the web easily accessible and readable, and created the Schema.org project that turned the web into something structured and searchable. Both are now de facto standards around the web. Guha has worked at Netscape, Google, Apple, and elsewhere, and has done as much for the open web as just about anyone.

Still, Guha’s paycheck comes from somewhere. When I ask Guha what’s in it for Microsoft, he grins. The bigger picture seems to be that Microsoft wants everyone to be thinking about and implementing conversational interfaces and AI tools, on the belief that eventually most of those companies will pay for some agentic, Azure-based Microsoft tool or another. NLWeb is connected to the Model Context Protocol, an Anthropic-created open project also designed to help developers integrate AI website features, and Microsoft is heavily invested in connecting MCP to its AI offerings. It all makes a certain kind of business sense.

But if it works, NLWeb is much bigger than Microsoft. At first, most companies and developers supporting the protocol will use it to roll out better search to their websites. (Guha also created Google Custom Search, which lots of publishers used for their own site search, so he knows the challenges here.) Beyond that, well, he’s not entirely sure how it’s all going to work. If Serious Eats knows you’re a vegetarian, should it share that information with other sites? Should these models be able to do truly agentic stuff, like navigate sites and even buy stuff on your behalf? How should all the data be managed? He’s not sure. Ultimately, he’s not the one with the answers anyway; if this whole thing works, the web’s publishers and users can decide for themselves.

A project like NLWeb requires a huge amount of support, and there’s certainly no guarantee of getting it. Guha thinks companies like Meta and Google have incentives to play along, but acknowledges they might not. Lots of companies want to do their own thing, while others are happy to just take the paycheck and license to OpenAI. You can’t brute-force a protocol; all you can do is hope everyone sees a reason to get on board.

Maybe the biggest challenge facing NLWeb, though, is that the history of the web suggests that everything eventually centralizes. Over the years we’ve had aggregators like Yahoo, search engines like Google, and now chatbots like ChatGPT. There’s not actually much evidence to suggest that this kind of massive decentralization can work at all. But Guha’s argument seems to be that maybe the aggregators have just always had the better tech. Now he’s hoping to be able to give the best to everyone, and see what the web builds.