Six files an AI agent looks for on your website, and the three a local business should actually have.

01Why a website now has an audience of software

A business website used to have one audience, people, and one intermediary, Google’s crawler, which read the page so people could find it. In 2026 the intermediaries multiplied and they do more than index. ChatGPT, Gemini, Claude, Perplexity and Google’s AI Mode read pages to answer questions, and increasingly to act: to quote a price, to check whether you serve a ZIP code, to book a time. Each of them reads a website the way a very fast, very literal assistant would. They do not infer. They do not squint at a badge image to find your license number. They read what is there, in the form it is in, and if the thing they need is missing or ambiguous they move on to a site where it is not.

That is the reason a set of machine-facing files has grown up alongside the human-facing pages. Some are old (robots.txt dates from 1994), some are two years old, some are months old. The names are intimidating and the ideas are not. Every one of them answers one of three questions an agent asks on arrival: Am I allowed in? What are the facts? What can I do here?

Before the file-by-file tour, one grounding fact. Whitespark’s 2026 Local Search Ranking Factors report (47 experts, published November 6, 2025) found that for AI search visibility “mentions (citations) are the new link,” and its top factors were presence on curated best-of lists, a dedicated page per service and prominence on industry domains. None of the six files below replaces that. They make a business that already has substance legible to software. They do not create substance.

02File 1: robots.txt, the door

What it is. A plain text file at yoursite.com/robots.txt that tells crawlers which parts of the site they may read. Every search engine honors it, and so, by their own statements, do the major AI companies’ crawlers: GPTBot (OpenAI), ClaudeBot, Google-Extended (the control for Google’s AI training uses, separate from normal Google search) and PerplexityBot, among others.

Who reads it. Everyone, first. It is the only one of the six files you can be certain every agent checks.

Why it matters for a local business. Because it is the easiest way to disappear by accident. We regularly find Denver business sites whose hosting provider, security plugin or a well-meaning web designer added a blanket block on “AI bots” in 2024, when the conversation was about protecting content from training. For a publisher that trade-off might make sense. For a plumber whose service page exists to be quoted when someone asks “who can fix a water heater in Lakewood,” blocking the bots that answer that question is self-harm. Semrush’s April 2026 guidance on AI Mode booking put “make sure AI crawlers can access your site” at the top of the list for a reason.

What to do. Open yoursite.com/robots.txt. If you see User-agent: GPTBot or ClaudeBot or Google-Extended followed by Disallow: /, you are shut. Remove the block for the pages you want cited. Keep blocks on anything private (admin, cart, account pages). Separately, in Google Search Console, leave the AI Overviews and AI Mode toggle on; we explained the data behind that in should you opt out of Google AI Overviews.

03File 2: schema, the facts

What it is. Structured data, usually JSON-LD inside the page, using the schema.org vocabulary. It is not a separate file; it rides inside every page. It is how you say, unambiguously, “this is a Plumber, at this phone number, open these hours, serving these towns, with this price range, bookable at this link.”

Who reads it. Google has read it for a decade for rich results. The AI assistants read it because it is the most reliable statement of fact on a page; ambiguity in prose becomes a field with a value. Of the six files this is the one with the highest certainty of being read and the highest payoff per hour.

Why it matters for a local business. Because it carries exactly the inputs an agent needs to recommend and book you: identity, hours, area, price, proof, and a booking action. We set out the fifteen checks that depend on it in the AI readiness checklist. The minimum set on a service page is LocalBusiness (or the specific subtype such as Plumber, HVACBusiness, Dentist), a Service with an Offer that has a minPrice and maxPrice, openingHoursSpecification, areaServed listed as towns, sameAs links to every profile, FAQPage for the customer-language questions, and a potentialAction of type ReserveAction pointing at your booking link.

What to do. Add it to every service page, and make sure it agrees with the visible text; an agent that finds schema saying 24/7 and a footer saying 8 to 5 trusts neither. Do not mark up a Google rating on your own site; Google’s review-snippet guidelines, updated July 2026, are explicit that review markup must reflect reviews on your own pages and must not be incentivized or undisclosed. If you want the markup done without learning JSON, schema and FAQ schema are single tasks on our SEO à la carte menu.

04File 3: llms.txt, the summary

What it is. A Markdown file at yoursite.com/llms.txt, proposed in September 2024 by Jeremy Howard of Answer.AI, that gives a language model a short, plain-language description of the site and a curated list of its most important pages, so it does not have to crawl everything to understand what you are. It is the agent-era equivalent of the one-page company overview you would hand a new employee.

Who reads it. This is the honest part: nobody has confirmed that the major assistants fetch it as a matter of course. Google’s John Mueller said in 2025 that he was not aware of any AI system using it, and we have not seen a statement from OpenAI, Google or Anthropic since that changes that. What is true is that thousands of sites publish one, that some AI crawlers do request it, and that tool-building agents and smaller assistants use it when it exists. We publish one ourselves (eyetoad.com/llms.txt) and keep it current, because it costs an hour and the downside is zero.

Why it matters for a local business. Modestly. Its real value is discipline: writing llms.txt forces you to state, in one place, what you do, where, for whom, at what price, and which pages matter. That document then becomes the single source of truth you sync your listings and schema to. The file is cheap; the thinking is what pays.

What to do. Write it in plain English: name, founding year, address or service area, phone, hours, the services with price ranges, the honest positions you hold (we say outright that nothing visible happens in month one of SEO), and a list of your key pages with one line each. Keep it under a page. Update it when prices change. It is also a task on the à la carte menu.

05File 4: the agent card

What it is. An agent card is a JSON file that describes an AI agent: its name, what it can do (its “skills”), how to reach it, and how to authenticate. It comes from the Agent2Agent protocol, A2A, which Google launched in April 2025 and donated to the Linux Foundation in June 2025. Version 1.0 shipped on March 12, 2026, and its headline change was cryptographically signed agent cards, so a receiving agent can verify that a card was issued by the domain’s owner rather than forged. The Linux Foundation reported on April 9, 2026 that more than 150 organizations were involved and that v1.0 also added multi-tenancy and an AP2 extension for agent payments. The spec places the card at a well-known path on the domain (check the current spec at a2a-protocol.org for the exact filename; it has changed between versions).

Who reads it. Other agents. That is the point of A2A: it is for agent-to-agent discovery and conversation, as opposed to MCP, which is for an agent calling tools. A travel agent built by one company can find and talk to a hotel’s agent built by another. Several analysts describe the two as complementary rather than competing, and both now sit under the Linux Foundation.

Why it matters for a local business. Today, mostly as a thing to understand rather than a thing to build. An agent card describes an agent. If your business does not run an agent, there is nothing for the card to describe. The scenario where it matters, a customer’s assistant discovering your business’s assistant through a signed card and negotiating a booking, is real in enterprise travel and is beginning to appear in the products that make local businesses bookable, where the product publishes the card on the business’s behalf. If you are told you need to “create an agent card” for your dental practice by hand, ask what agent it will describe.

What to do. Nothing by hand, yet. If you adopt a booking product that runs an agent for you, make sure it publishes a signed card and that the card’s facts (name, services, area) match your schema and listings, because consistency checks now run across these files too.

06File 5: the MCP server

What it is. MCP, the Model Context Protocol, is the open standard assistants use to connect to tools and data. It was introduced in late 2024 and is now governed under the Linux Foundation’s Agentic AI Foundation. An MCP server is a small program that exposes a handful of named functions, with their inputs described, so that an assistant can call them. For a business the functions are obvious: check_service_area(zip), get_quote(service, details), get_availability(date_range), book(slot, customer, approval). The assistant reads the list of functions, decides which one answers the customer’s question, calls it, and reads the answer back in plain words.

Who reads it. Any assistant that supports MCP, which in 2026 is all of the major ones in at least some surface: desktop and browser agents, developer tools, and the app directories the assistants are building. It is the layer underneath the booking rails we described in how AI agents book Denver service businesses: where ChatGPT uses a partner integration with OpenTable or Booksy, a tool-reading assistant uses an MCP endpoint directly.

Why it matters for a local business. Because it is the only one of the six files that turns “named” into “booked” for a trade that has no partner rail. Restaurants have OpenTable. Salons have Booksy. A Denver electrician has a website, a phone, and, if they are on Jobber or Housecall Pro, an online booking page that an agent can follow but not query. An MCP endpoint lets the agent ask the three questions that matter (do you cover 80205, what does this cost, when can you come) and get answers it can read to the customer, then book with the customer’s approval. The assistant never guesses at your prices because your price book answered. It never double-books because your calendar answered.

What to do. You will not write one, any more than a restaurant writes OpenTable. You will adopt a product that runs one for you and connects it to your price book and calendar. Our sister product Doorhanger.ai does this for trades; we describe how it fits the rest of the setup on our agentic booking setup page. If you have an unusual workflow, a custom-coded agent with its own MCP tools is what we build on the AI agents for business page, and our sister site My AI Agent Army builds single-job agents to order. Whatever route you take, insist on three things: the endpoint only ever offers real open slots, the customer hears the price and time read back before anything is booked, and you can choose between auto-confirm and approving each request. Those are not technical niceties. They are the difference between a booking tool and a liability.

07File 6: WebMCP, the buttons

What it is. WebMCP is a proposed browser standard, developed by Google and Microsoft through the W3C Web Machine Learning Community Group, that lets a web page tell a browser-based agent which actions it supports and what inputs they take, through a new interface called navigator.modelContext. Instead of an agent taking screenshots of your booking page and guessing which button is “confirm,” the page declares: here is a tool called book_appointment, it takes a service, a date and a phone number. It shipped in early preview in Chrome 146 behind an experimental flag, and one report describes an origin trial opening in Chrome 149 (a single source, so treat the version as unconfirmed). Safari and Firefox were still evaluating it at the time of writing.

Who reads it. Browser agents: the assistants that drive a browser on the user’s behalf, such as the Chrome-based agents from the big assistants and tools like Perplexity’s Comet. Not search crawlers, not the chat interface itself.

Why it matters for a local business. Eventually, a lot: it is the version of an MCP server that lives inside your existing website and requires no separate endpoint. Today, not yet, because it is behind a flag in one browser. The reason to know about it now is that scheduling and website platforms will add it, and when your scheduler’s booking page registers its own tools, your site becomes agent-operable without you doing anything, provided the page was built on a platform that bothered.

What to do. Ask your website and scheduling vendors whether WebMCP is on their roadmap, and prefer vendors who say yes. If you have a developer, the early examples show existing booking functions being registered in about thirty lines of JavaScript. Do not rebuild your site for it.

How one agent uses all six in a single visit

It helps to see the files in sequence rather than as a list. A customer in Wheat Ridge tells an assistant: “Find me an electrician who can add a 240-volt outlet for an EV charger this week, and book it.” The assistant runs its searches and lands on your site. It reads robots.txt; the door is open. It fetches your EV charger installation page and reads the schema: an Electrician in Wheat Ridge, serving a named list of suburbs that includes the customer’s, open 7 to 6, an Offer on the EV charger service with a range of $600 to $1,400 depending on panel distance, a ReserveAction pointing at your booking link. It may glance at llms.txt and find the same facts in prose, which raises its confidence. It checks whether you have an agent card; you do not, and nothing is lost, because it is not looking for an agent to talk to, it is looking for a tool to use. It finds your MCP endpoint, calls check_service_area with the customer’s ZIP, gets a yes, calls get_quote with “EV charger outlet, detached garage, 40 feet from panel” and gets $1,100 to $1,300, calls get_availability and gets Thursday 9 am and Friday 1 pm. It reads all of that to the customer, the customer picks Thursday, the assistant calls book with the customer’s approval, your phone buzzes with a text, and the office confirms. If your site had WebMCP on its booking page and the customer was using a browser agent, the same sequence could have run through the page itself. Six files, one visit, one job. Every step that was missing would have ended the visit at that step.

Now run the same visit against a typical Denver electrician’s site. Robots.txt open: fine. Schema: an Organization with a name and phone, nothing about services, hours, area or price. The agent can confirm you exist and little else. There is no llms.txt, no endpoint, and the EV charger page ends in a contact form. The assistant reports a name, a phone number and a hedge, and offers to book the competitor whose site answered. The difference between the two outcomes is one afternoon of schema and a booking product, not a bigger marketing budget.

08The three a Denver business should have

Strip the list to what moves a phone or a calendar in Denver this quarter and you get three.

One: schema that carries the five inputs. Identity, hours, area, price range, proof, and a ReserveAction on every service page. This is what AI Overviews and AI Mode are reading today when they decide whether to name you for “furnace repair Arvada,” and it is what any booking agent reads first. It is the single highest-leverage afternoon on this list.

Two: an open door. A robots.txt that lets the AI crawlers reach the service pages, a Search Console toggle left on, and prices and hours in the HTML rather than in images or scripts. Zero cost, and the most common silent failure we find.

Three: a booking endpoint, if you want the job and not just the mention. For a restaurant that means being on a Reserve with Google partner; for a salon, Booksy or a similar partner; for a trade, an MCP endpoint connected to your price book and calendar, which you adopt rather than build. This is the one that pays for the other two, because it is where a recommendation becomes a ticket.

llms.txt is a cheap fourth. Agent cards and WebMCP are the ones to watch, and to ask vendors about, rather than to build. If a salesperson tells you that your HVAC company urgently needs a hand-written agent card, that is a tell.

09What not to do

Do not block the bots to “protect” a service page. The page exists to be quoted. If a page should not be quoted, it should not be public.

Do not let schema and prose disagree. The consistency checks that agents run now span the page, the markup, the listings and the machine files. A contradiction anywhere costs you trust everywhere.

Do not publish an endpoint that offers fake slots or hidden prices. The booking platforms require that their own agents identify themselves and read the price and time back to the customer before booking; your endpoint should hold itself to the same standard, or the first bad booking will become a review that every future agent reads.

Do not buy complexity. A hand-built agent card for a business with no agent, a WebMCP integration on a page nobody can reach through a flagged browser, an “AI SEO” package that is a robots.txt edit and an llms.txt file at a thousand dollars a month: these are the 2026 versions of selling a business a hundred directory submissions. The useful work is boring. One identity, one page per service with a price range, an open door, a booking endpoint, and a human who answers. Our entity signals guide and how to get cited by ChatGPT cover the substance these files point at.

If you want to see what an agent sees on your site right now, the AI assistant in the header of this page reads the same public signals and will tell you what it found and what it could not. The free audit does the full version with a person, including the schema check and the endpoint test.

Keep going