The Web Is Growing A Second Layer – Almost A Third Head

What Happened

Google has quietly published two new specifications within weeks of each other, both aimed at making web content more legible to machines rather than humans. Reporting on the releases frames them as evidence that the web is developing a genuine second delivery layer — one built specifically for AI agents and automated consumers, sitting alongside (or on top of) the traditional HTML layer designed for browsers and human readers. The two specs are being compared and contrasted with two other efforts that have gained traction over the past year: the informal LLMs.txt convention and Anthropic's Model Context Protocol (MCP).

Why This Matters

For the past few years, the industry has debated how AI crawlers and agents should access site content: scrape rendered HTML, rely on ad hoc text files, or something more structured. LLMs.txt emerged as a grassroots, low-effort solution — essentially a plain-text index pointing language models toward the most useful pages. MCP took a more ambitious path, giving AI agents a standardized protocol to call tools and retrieve structured data from servers, effectively turning a website or backend into something an agent can operate rather than just read.

Google entering this space with its own specifications signals that the largest gatekeeper of web traffic doesn't see LLMs.txt or MCP alone as sufficient or authoritative. That's significant. When Google ships a spec, it tends to become a de facto standard simply by virtue of the company's reach into search, crawling infrastructure, and now AI-driven answer surfaces. If these new formats govern how content gets surfaced in AI Overviews, agent-based browsing, or future assistant products, they could reshape technical SEO in the same way structured data and schema markup did over the last decade.

Context: A Layer Built for Machines

The framing of a "second layer" or "third head" for the web captures something real: sites increasingly need to serve at least three audiences — human visitors, traditional search crawlers, and now autonomous AI agents with different parsing needs and different economic incentives (agents don't view ads, don't linger, and often summarize rather than link back).

MCP servers matter directly here because they represent the more heavyweight, interactive end of this spectrum — not just describing content but exposing callable functionality. If Google's new specs are positioned as complementary to, rather than competitive with, MCP, that suggests a layered future: lightweight discovery formats for what content exists, and protocol-level servers like MCP for what an agent can actually do with it.

What SEOs Should Watch

Practitioners should treat this as an early-stage but consequential shift: monitor adoption signals, avoid over-investing in any single convention prematurely, and start thinking about content architecture in terms of machine-actionable access, not just crawlable pages.