For twenty years, "having a website" meant one thing: pages for people. The entire stack — HTML, CSS, design, UX — assumed a human at the other end, reading and clicking. That assumption is now only half the picture. A second web is being built underneath the first, and its users aren't people. They're agents, and they don't want your page. They want your capabilities.
What's actually happening
The plumbing for an agent economy got laid astonishingly fast. The Model Context Protocol (MCP) — created by Anthropic and open-sourced in November 2024 — defines a standard client-server interface through which agents access external capabilities: API calls, database queries, file operations, web search, transactions, any tool a developer exposes. By early 2026 it had crossed 97 million downloads, and in December 2025 Anthropic donated it to the Linux Foundation's Agentic AI Foundation, with OpenAI and Block as co-founders and AWS, Google, Microsoft, Cloudflare, GitHub and Bloomberg backing it.
Read that again: a protocol for letting agents act on external systems became a neutral, cross-vendor standard in about a year. Around it, a whole commerce stack emerged — ACP (OpenAI + Stripe), UCP (Google + Shopify), AP2 and Visa TAP for agent payments, the Coinbase-led x402 protocol (which processed ~165 million agent transactions in its first months), and A2A for agent-to-agent coordination. This isn't a research demo. It's an infrastructure buildout.
The demand is already here
And the agents are arriving. Adobe Analytics measured a 4,700% year-over-year jump in generative-AI traffic to US retail sites in a single year. As we covered in the guerrilla war for agent attention, autonomous agents now browse, decide and buy on users' behalf. The question has flipped from "can a human use my site?" to also "can an agent?"
Why HTML-only loses the agent
Picture an agent tasked with "find and book the best option." To use an HTML-only site, it must scrape a human interface — parsing layouts built for eyes, fighting client-side rendering, tripping over interstitials. It's brittle, slow, and increasingly blocked. Now picture a competitor that exposes a clean endpoint: here are products, prices, availability; here's how to book. The agent completes the task there in one reliable call.
When an agent can't read your prices, it doesn't email you. It buys from the site whose machine door was open.
This is the same logic as being crawlable for citation, taken one step further: being crawlable gets you mentioned; being callable gets you transacted with. The first is AIO's table stakes; the second is the next frontier.
The layered site: one source of truth, three audiences
The practical model for 2026 is a site that serves three audiences from a single source of truth:
| Audience | What they need | What you ship |
|---|---|---|
| Humans | A usable experience | Server-rendered HTML, good UX |
| Search & AI retrieval | Readable content | Crawlable HTML, structured data |
| Acting agents | Callable capabilities | API + MCP server + structured endpoints |
Concretely, the machine layer means:
- A clean, documented API — even read-only to start (products, prices, availability, content). The minimum viable machine door.
- An MCP server exposing safe, well-scoped capabilities, advertised at
/.well-known/mcp.jsonso compatible agents discover it. - Agent-commerce rails (ACP/UCP and the payment protocols) as they mature, if you sell — so an agent can complete a purchase, not just read about one.
- One source of truth feeding all three layers, so your page, your retrieval surface and your API never drift apart.
The catch: a machine door is still a door
Opening a machine layer reintroduces every trade-off from governing crawler access. Exposing capabilities to agents means deciding which capabilities, to whom, and under what limits. Sensible defaults:
- Start read-only. Expose information (catalog, content, availability) before actions. Expand to transactional once auth and guardrails are solid.
- Authenticate and scope. Treat agent access like any API surface — keys, rate limits, least privilege. Delegated identity (an agent acting for a verified user) is the pattern to design for.
- Don't expose everything. A machine door is an attack surface and a cost center as much as an opportunity. Design it deliberately, the way you'd design a public API.
The strategic frame
The web is bifurcating. The human layer of pages isn't disappearing — but a machine layer is taking its place beside it, and the two will be governed and optimized differently. Being machine-readable (crawlable, citable) is becoming table stakes. Being machine-actionable (callable, transactable) is the edge the early movers are building now, while most sites still ship pages and nothing else.
You don't have to build all of it tomorrow. But you should stop assuming the only visitor that matters has eyes. The next decade of the web has a second audience — and it's reading your API, not your hero section.
Frequently asked questions
Does my website need an API or MCP server now?
Increasingly, yes — at least a clean, documented way for software to read your key data. Agents need to query prices, availability or content without scraping a human UI. A site exposing structured access (an API, ideally an MCP server at /.well-known/mcp.json) is usable by the agent economy; an HTML-only site forces brittle scraping or loses the agent to a competitor with a clean door.
What is MCP and why does it matter for websites?
The Model Context Protocol is an open standard, created by Anthropic and donated to the Linux Foundation's Agentic AI Foundation in December 2025, defining how agents access external tools and data through a uniform interface. It matters because it's becoming the common way agents discover and use capabilities — so an MCP server lets cooperative agents act through a defined door instead of scraping.
Is the machine layer replacing the human web?
Not replacing — bifurcating. The human web of pages remains; a parallel machine-facing layer forms beside it, where agents call APIs and MCP servers to query and transact. You now serve three audiences from one source of truth: humans (HTML), search/AI retrieval (crawlable HTML), and acting agents (APIs/MCP). Neglecting the third is the new way to be invisible.