Most websites now carry an llms.txt file at their domain root, and most of their owners never chose to publish it, according to a July 10 analysis by Slobodan Manic published in Search Engine Journal. The unread file is one of two agent-facing protocols websites are adopting, and only one of the two actually lets an agent complete anything once it arrives. Manic separates them into an identity bet, llms.txt, and a capability bet, a browser standard called WebMCP, and argues the industry has piled onto the one with weaker evidence.
Llms.txt is a plain-text index proposed by Answer.AI co-founder Jeremy Howard in September 2024. Manic reports that Google’s John Mueller called it “purely speculative for now” in a Reddit thread in June 2026, noting the file has existed for years without confirmed use by any major AI system. Manic reads that as a quiet admission from a Google advocate. Despite that, AIOSEO, a WordPress plugin installed on more than three million sites, generates an llms.txt by default, according to Manic. The result, per his analysis, is a large population of sites publishing a self-description nobody at the company wrote or reviewed.
WebMCP starts from a different question: not who a site is, but what an agent already on the page can do there. Per Manic, the standard is being drafted through the W3C Web Machine Learning Community Group by engineers from Google and Microsoft, published on February 10, 2026, and is now running as a public origin trial spanning Chrome 149 through 156. Sites in the trial expose callable tools through a navigator.modelContext API so an agent can search, price, or book directly instead of screen-scraping. Gemini in Chrome is, per Manic, the only agent currently consuming those tools in production.
Manic ties the stakes to a traffic claim from Cloudflare chief executive Matthew Prince. In June 2026, Prince said automated requests had passed human requests for the first time, at 57.3 percent of web traffic, a crossover he had earlier projected for 2027. Manic’s argument is that once machine traffic is the majority, whether an agent can identify a site matters less than whether it can transact on it.
Both protocols assume the agent has already found the page. Neither addresses discovery itself, which still runs on the same crawling, rendering, and internal-linking mechanics that have driven organic visibility for two decades. An unreadable or unindexed page gains nothing from either file, no matter how well it is described or how many tools it exposes. That distinction should shape budget decisions this quarter. An llms.txt costs almost nothing to audit and correct. A working WebMCP integration requires engineering time that only pays off on pages with real transactional actions to expose.
Manic’s own site reflects that asymmetry. He regenerates an llms.txt from his content on every update and calls it a hedge rather than a strategy. His WebMCP implementation exposes four callable tools, a glossary lookup and a product catalog among them, both reading from the same data that renders the human-facing pages.
He is explicit that this is his own read of an unsettled situation, not confirmed adoption data. WebMCP remains a Community Group draft, and Microsoft Edge’s support is unconfirmed as of version 147. Llms.txt could still earn its place if a major AI system confirms it reads the file, a signal Manic says he has not yet seen.
For search teams, the practical move is a five-minute audit: open the domain’s llms.txt, confirm whether a CMS or plugin published it without review, and correct or remove it. Save engineering budget for WebMCP on pages where an agent completing a real transaction, a booking, a return, or a price check, is plausible within the next year.
Based on analysis by Slobodan Manic, published by Search Engine Journal on July 10, 2026.
