The Agentic Web Is Divided Into Two Bettings: Identity and Awareness
The agent web protocol layer is split into two bets, and many websites have set one of them unknowingly. The first bet is about who you are: a file called llms.txt that tells AI models who you are and what your content consists of. The second is about capability: a browser standard called WebMCP that tells the agent what to do with your website once it arrives. They sound like the same concept. They answer contradictory questions. And the difference determines where your effort should go in the next year, because one of these bets is automatically opened for you, while the other, which allows the agent to complete the work, takes deliberate work.
2 Files, 2 Questions
An agent who lands on your website has two things they may want to know. First is what this place is and what it includes. The second is, now that it is here, how it completes the task that its user has sent it to do. Ownership, then skill. The web agent creates a separate response for each, and the two responses come from two different files supported by two different camps.
The ID response is llms.txt, a plain text file that you publish at the root of your domain. Featured map: here’s who I am, here’s my most important pages, here’s what each one is about. The answer to the skill is WebMCP, a browser standard that allows your website to expose clickable tools to the agent, so the agent can search, sort, price, or book by calling a function instead of guessing at the interface.
Llms.txt and WebMCP are bundled as two flavors of the same thing, agent readiness. They answer two different questions. Identity is a notebook. Skilled cash register. And right now the industry is pushing every website towards the notebook, a bet with weak evidence behind it, while the cash register, a bet that takes the work forward, is left to the few teams willing to build it.
Identity Bet: LLMs.txt
Llms.txt is the markup file it resides in yourdomain.com/llms.txt. It was suggested by Jeremy Howard, founder of answer.AI, on September 3, 2024, as a way to give language models a clean, human-curated index of your content instead of having them crawl and rebuild it from HTML full of navigation, ads, and text. The tone makes sense on its face. Models work best with structure, so give them structure.
The problem is that the evidence that it does anything is slim. Responding to people on Reddit in June 2026, John Mueller of Google called llms.txt “presumptive at this point (the file has been around for years, yet no AI program uses it)” (reported by Search Engine Journal). That’s a Google Search advocate speaking the silent part out loud. The file has been there for over a year. A file that no major AI program has confirmed as being read. A file that asks you to keep a second copy of your content in sync.
Despite this small evidence, llms.txt is automatically opened by websites. AIOSEO, a plugin used on over 3 million WordPress websites, generates llms.txt automatically. A very large number of website owners are now publishing llms.txt they never decided to publish, describing their website in a file they never read.
To be clear, I am not against the file. A well-maintained llms.txt is an indicator of good content. The default problem. The identity bet wins at adoption, not because it works, but because the plugin’s checkbox was sent open. A file that you didn’t write, that no AI system is guaranteed to read, that comes out of sync with a website that says it explains the time you forgot it was there, is a rare thing to have in the middle of an agent’s strategy. But that’s where most websites live now, by default rather than choice.
Betting Skill: WebMCP
WebMCP, short for Web Model Context Protocol, is a browser standard that allows your website to register callable tools that an agent can use. navigator.modelContext API. It starts from a different question than an identity file: not who you are, but, given that the agent is already on your website, how does he complete the task he came to do. That’s close to how Mueller laid out his will in the same interview. He said he likes WebMCP’s approach to commercial integration because they have “clear goals and processes.”
Instead of an agent scanning your page and guessing where to click, your website tells you directly: here are the actions I support, here’s the input each one needs, and here’s what you get. The controller distorts. The website reveals its capabilities instead of forcing the agent to take them back.
The standard is written by the W3C Web Machine Learning Community Group by developers from Google and Microsoft, and was published as a draft on February 10, 2026. It is now in public origin testing starting from Chrome 149 to Chrome 156. The origin trial is how Chrome allows a website to open the test feature for its real visitors in a limited window, by registering it behind the developers. So you can use WebMCP tools on live traffic now, not just on your machine. The agent that uses those tools in Chrome today is Gemini.
WebMCP is a strong bet because it has a clear mission. An ID file describes you and you hope someone will read it. A power tool is invoked, returns structured data, and moves the task one step closer to execution. This brochure is waiting to be read. The cash register rings.
Why Two Bets Are Unchangeable
The cleanest way to capture the difference is by using two design constraints that I keep coming back to in Machine-First Architecture: ownership and interoperability. Identity makes your brand and content machine-readable. Cooperation allows the agent to complete an action with a predictable outcome. Llms.txt is a play on the proprietary layer. WebMCP is an interactive layer game. They are tools for two different jobs, and a website may need both, one, or the other, depending on what it wants agents to do.
This is very important now because agents are not wrong to collect on your traffic. Cloudflare CEO Matthew Prince said in June 2026 that automated traffic surpassed human traffic for the first time, at 57.3% of requests directed to web pages against 42.7% of people, a crossover he predicted at SXSW will not come until 2027. Whether the agent can do what your customer has asked for becomes the whole game. Identity without skills is a machine that reads your sign and stops at a door you can’t open.
That’s a reverse automatic detection conversion. The open file answers an important question once the agent has arrived. The level that answers the most important question is the one that almost no one has opened.
I Put Both Bets On My Website, On Purpose
I use both at No Hacks, and the difference between how much it costs to put each is a bit of an argument.
For ownership, I save llms.txt, and generate it from the website content each time I update the site, rather than leaving the plugin to write something I’ve never read. That’s the intentional part. Because I rebuild whenever the site changes, it’s always up to date with what’s actually here, and if the AI program starts reading llms.txt, mine will be fine when that day comes. I’m not betting on it. I keep it current and cheap and not disruptive as a strategy.
By default, I used WebMCP. When a WebMCP-capable browser loads a website, it registers four callable tools through the application navigator.modelContext. Two include lists of words: one enumerates each word so the agent can find the definition, and the other returns the canonical definition of the term and its source link. Two cover my active area of agent browsers and agent products: one lists all tracked products by category, and one returns full details for a single product by name. The agent doesn’t need to write my pages to answer what I mean by power betting or what agent browsers I’m tracking. Call the tool and get a clean, organized response.
The most important detail is where the tools are read. Each draws on the same data that powers the landing page. Glossary tools read the glossary. Product tools read the product list. So the agent’s response and the person’s response can never contradict each other, because one source feeds both. That’s the difference between expressing a skill and keeping a different definition of yourself, and that’s why the power bet, created this way, doesn’t have a separate copy to keep in sync.
So I grab both files. One took an afternoon and did the real work. The rest I reproduce in my own content and treat it as a fence, not a program.
What This Means for Your Website
First, find out what you’ve already published. Open it yourdomain.com/llms.txt in the browser. When something loads, your stack puts it there, which may be the default for a plugin you never set up. Read it. Ask if it really describes your website, because when an AI program starts reading llms.txt, an inaccurate one is worse than none at all. This costs five minutes, and most website owners never do it. If you want to keep it, generate it in your content so it can’t drift.
Second, decide if a power layer is worth investing in. If agents have any reason to complete work on your website, search listings, check price, start booking, start returns, then WebMCP is a paying bet, and the original trial means you can use it on real visitors now rather than waiting for it to be sent. If your website is something the agent reads instead of acting on, the power layer can wait, and your effort is pure, server-rendered content that no agent can produce.
What you should not do is assume that the file opened by your plugin is your agent strategy. It’s an identity claim no AI system is guaranteed to read, and it says nothing about what the agent can do once it arrives.
What’s Still Unsettled
Honest uncertainty: Identity betting is not the same; there is no evidence. If a giant AI system announces tomorrow that it reads llms.txt and scales the weights, the calculus changes, and the file that everyone automatically entered suddenly gains its place. I’ve looked for that signal and haven’t seen it yet.
WebMCP has its own open questions. It is a Community Group Framework, not a certified standard. Gemini on Chrome is the main agent it uses so far. And cross-browser support is less than the second coverage suggests. Microsoft has co-authored the standard, but I couldn’t find WebMCP in the official release notes for Microsoft Edge as of version 147, so I’d treat any claim that Edge ships natively as unconfirmed at this point.
Until the signal arrives, my bet is clear: by the end of 2026, energy is the key bet, and ownership will end on the fence. A file that hopes to be read loses the tool it uses, because agents are measured by whether they completed the task, not by whether they read your sign first. Place the ability to bet deliberately. Create an ID file for your content, if you have saved it. And read the one your plugin already writes before it speaks for you.
Additional resources:
This post was originally published on No Hacks.
Featured image: Master1305/Shutterstock
