Your Website Has a Second Audience Now: The Agent

A person visits your website and reads it. An assistant visits your website and tries to get something done. Those are two different visitors asking two different things, and most sites are only built to answer one of them.

Ask the same question two ways. A person types "how much does a custom app cost" into a search bar, skims three results, and lands on a pricing page. Someone else asks their assistant "find me a company that builds custom software and tell me what it costs," and the assistant reads a handful of sites in seconds, decides which ones actually answer the question, and reports back a short list. The person saw your page. The assistant either quoted it or skipped it, and you don't get to watch it decide.

We expect a growing share of your traffic to arrive this way, as a person's stand-in rather than the person. Three things decide how well a site does with that visitor: whether it's readable, whether it's quotable, and whether it's usable. Most sites still stop at the first one.

TL;DR

A website now has two audiences: people, and the assistants acting for them. Each reads it differently.
Readable means an assistant can parse the page at all: structured data, clean HTML, an llms.txt file.
Quotable means an assistant will actually repeat what you said, in plain, direct, checkable language.
Usable is the newest and rarest level: tools an assistant can call directly, on your terms, through a protocol called MCP.
Adoption of that third level is real and growing. One 2026 census counted more than 21,000 MCP servers in the public registry, most of them published in the last year.

Level one: readable

This is the floor, not the ceiling, and it's largely a function of how a site is built rather than what it says. An assistant reading your site needs the same things a search crawler needs, done properly: fast pages, semantic HTML instead of a wall of generic divs, structured data (schema.org markup) that states plainly what a page is and what it's about, and an llms.txt file that gives an assistant a map of the site instead of forcing it to guess from navigation. Every site we launch ships with all of this on day one. It isn't an add-on; it's what "built correctly" already includes.

Level two: quotable

Readable gets an assistant into the page. Quotable is what makes it want to repeat what it found. This is AEO, answer engine optimization, and the discipline is simpler than the acronym makes it sound: write the direct answer first, in the same words a person would ask the question, before the explanation and the caveats. A paragraph that opens with a real number or a plain yes-or-no is far more likely to get quoted than one that opens with three sentences of throat-clearing before the actual answer shows up. An assistant summarizing five sites for someone will favor the one that made its answer easy to lift.

Level three: usable

This is where most sites stop entirely, and where the next real advantage sits. Usable means an assistant doesn't just read about your business, it can act, inside limits you set.

MCP stands for Model Context Protocol, an open standard that lets AI assistants use real tools and real data on a business's behalf. Anthropic introduced it in 2024, OpenAI adopted it in 2025, and it's now governed under the Linux Foundation's Agentic AI Foundation. It isn't a plugin or one company's feature. It's shared plumbing every major assistant can use, and it's growing fast: one 2026 security census of the public MCP registry counted more than 21,000 unique servers, the large majority published within the prior year.

In plain terms, an MCP server is a list of things an assistant is allowed to do on a business's behalf, like "check availability" or "book a consultation," along with the data it's allowed to read. The business decides which tools exist, what each one can change, and which ones need a person to say yes first. We've written about what this looks like for apartment leasing, home service businesses and real estate, and that last one isn't theoretical: three real-estate sites we've built now have a live, public door where a buyer's assistant can search real listings and ask for a showing. The pattern holds across every one of them: read freely, act behind approval.

What changes for a business

Two doors, one source of truth. Your website still exists for people, and your MCP server exists for their assistants, but both should be reading and writing the same underlying data. A price, an availability window or a piece of content shouldn't be able to say two different things depending on which door someone came through.

Every agent needs an identity, not a shared key. If ten different assistants are using your booking tool, you want to know which one made which request, the same way you'd want to know which employee made a change. And every action an assistant takes, especially anything that touches money or sends a message, needs a log a person can actually read afterward.

What are the risks?

Spam and abuse at a public door. Anything reachable gets tested, and bots find a new door within days of it opening. Rate limits, identity requirements and a way to shut off a misbehaving caller aren't optional extras on a public MCP server, they're the first thing you build.

Bad data served confidently. An assistant repeating a wrong price or a stale fact does it just as fluently as it repeats a correct one. Anything a tool returns should carry a real confidence level where one exists, and nothing should be presented as certain that isn't.

Privacy. Whatever an assistant can read, it can potentially repeat back to whoever's asking. Don't put anything behind a public tool that you wouldn't want summarized to a stranger.

Where do you start?

Run the self-check first, in this order:

Is your site readable? Structured data, clean markup, an llms.txt file. If this isn't already true, nothing past it matters yet.
Is your site quotable? Pull up your own key pages and ask: does the direct answer appear in the first sentence, or is it buried after the throat-clearing?
Is anything on your site usable? Not yet, for most businesses, and that's fine. This is the level worth planning for once the first two are solid.

What we've learned building our own

Upforge's MCP server is listed on the official MCP Registry, and it lets an assistant explore our services, find open times and book a consultation. The night it launched, we tested it the way a stranger's assistant would, and one lesson stood out immediately: error messages matter more for an assistant than they do for a person. Our first version told an assistant the service was unavailable when it really meant "that specific request was wrong." A person reading that error tries something else. An assistant reading it just concludes you're down and moves on to whoever's next. We rewrote every error to say exactly what was wrong and what a correct request looks like, and that one change did more for reliability than anything else in the first week.

Try it: have your assistant book a consultation with us

If your AI assistant supports MCP connectors, you can try all three levels on us right now. Add our server at https://upforge.io/api/mcp (also listed on the official MCP Registry as io.github.upforge-dev/upforge), then ask something like:

"Find a time next week for a consultation with Upforge about making our website usable by AI assistants, and book it."

Your assistant will show you open times in your own timezone, let you pick one, check your details with you, and book it. You'll get a confirmation email with links to reschedule or cancel. Prefer to pick a time yourself? Book a consultation here.

What does it cost?

Readable and quotable are built into every site we launch at no extra line item; a website that's ready for both runs $4,000 to $15,000. Usable is its own project, scoped to what your business needs an assistant to be able to do: our AI visibility and automation work runs $3,000, $7,000 or $25,000 and up depending on scope, and ongoing AI visibility work runs as a standing engagement alongside SEO.

If you're not sure which level your site is actually at, that's a good first conversation. Start with AI visibility, or read what an MCP server actually is if you want the fuller technical picture before you commit to anything.