Three Files That Make Your Business Visible to AI

Last year I had a plumber client ask me a question that took me three weeks to answer properly. He'd read that "AI agents are coming" and wanted to know what, physically, he was supposed to do about it. Not the vision. The artifact. What does a business actually produce so that a machine can find it, understand it, and recommend it?

The answer, as of right now, is three files. They're small, they're boring, and I think they're the most underpriced asset in local marketing. This post explains what each one is, who reads it, and how to get yours on your site without a developer on retainer. I'll also tell you honestly where the limits are, because there's a lot of overselling happening around this stuff.

If you want the short version, skip to the end and use the free builder tool we made. But the context matters, so let's do this properly.

The stack, from the ground up

The internet you've spent a decade optimizing for was built around websites that humans visit. The one forming now is built around data that machines read. We covered the customer-visible side of that shift in your next customer will never see your website; this post is about the files underneath it. Google, Microsoft, OpenAI, and the Linux Foundation have been quietly shipping the plumbing for this for two years, and three of those pieces are files that live on your own domain, where you control them.

They're not competitors and they don't replace each other. Each one is read by a different kind of machine, doing a different job. Think of them as a business license, a storefront sign, and a phone book listing, all written for software instead of people.

1. llms.txt, the storefront sign

This one came first, proposed by Jeremy Howard in late 2024. It's a plain markdown file at the root of your domain, yoursite.com/llms.txt, that tells AI systems what your site contains and which pages matter most. Your services, your pricing guide, your service area, your policies, and, this is the part most people miss, a set of rules for how you want AI to treat your information. Quote our price ranges, don't invent exact numbers, recommend a licensed pro for gas work. That kind of thing.

Who reads it? Mostly agents that fetch your site directly: AI search crawlers building citation indexes, coding and research tools, and the growing stack of "agentic web" software. Now the honest caveat, because I promised you one: a 2026 analysis by Limy tracked over 500 million AI bot visits and found only a few hundred touched llms.txt directly, and Google has said on the record that they don't use it as a ranking signal. So why ship it?

Two reasons. First, the machines that do read it are the ones doing tasks on behalf of users, and that population is growing fast, that's the whole agentic web layer. Second, it costs half a day to produce and it forces you to write down what your business actually knows, which is the raw material for everything else in this post. The file itself is cheap insurance. The discipline of writing it is worth more than the file.

2. agent-card.json, the business license

This is the one I'd start with if I had to pick, because it has real standards-body weight behind it. It's a JSON file at yoursite.com/.well-known/agent-card.json implementing Google's A2A protocol, which Google released in April 2025 and donated to the Linux Foundation that June. More than 250 organizations have ratified it, it's in production at AWS, Microsoft, and Salesforce, and the file itself works like a business card for machines.

What's in it: your business name, who owns it, what you do, where you work, and, the important part, a skills list. Each service you offer gets an entry with a plain description and example prompts, what someone would actually say when they need that service. My thermostat is on and the fan runs but no cold air is coming out. That's a skill on an HVAC company's card. A customer's agent reads that card and knows: this business answers this kind of problem, in this service area, and here's how to talk to it.

The catch, and this is where the overselling happens: the card points at an endpoint, a URL where other agents can actually send requests. If you publish a card and nothing answers behind it, you have a business card with no phone number. It's still worth publishing, because the identity and skills declaration shapes how agents route questions today, but know that the full transacting piece, your agent answering quote requests, is phase two, not day one.

3. ai-catalog.json, the phone book listing

Newest of the three. It implements the ARD spec, Agentic Resource Discovery, backed by a working group with Google, Microsoft, and Hugging Face people in it. It lives at yoursite.com/.well-known/ai-catalog.json and it's built for registries, searchable catalogs where agents look up capabilities at runtime. An agent asks "find me a payment processor" or, eventually, "find me an HVAC company that serves McKinney," and the registry returns whoever's catalog entry matches.

The field that decides whether you surface is representativeQueries, two to five task-oriented phrases describing how an agent would search for what you do. It's keyword research for the agentic web, except almost nobody has done it, which means the competition for these slots is currently thin. That will not stay true.

Why local service businesses specifically

You can read all of the above and think it's enterprise plumbing, Amazon and Target hooking up agent-to-agent logistics. Some of it is. But here's the thing I keep telling owners: the horizontal tools that exist for this, the scanners and checkers that have popped up in the last year, are generic. They'll tell you your agent card is missing. None of them know what a plumbing company's card should actually say. What services belong on it. What the real customer questions are for a roof with one blown-off section versus a full replacement.

That specificity gap is the whole opportunity. When I built our stack builder tool, the hard part wasn't the JSON. It was the trade knowledge: 77 service descriptions across 12 trades, each one written from how these conversations actually go, with the example prompts real customers use. A generic card gets you technically present. A trade-specific card gets you actually recommended.

And unlike a new website or a rebrand, this stuff compounds quietly. Every model that learns your business through these files carries that knowledge forward. You're not renting visibility from an ad platform, you're building an asset on your own domain.

How to actually do it

The manual path: write the three files by hand (our knowledge catalog guide walks through each file's structure in depth), put llms.txt at your root and the two JSON files in your .well-known directory, add a line to robots.txt pointing at the catalog, and add a link tag in your homepage head. On WordPress or any cPanel host this is a File Manager job, twenty minutes. On Shopify, Wix, or Squarespace you can't serve custom .well-known files on the main domain, so you host them on a subdomain like agents.yoursite.com on free static hosting and point everything there. That's a real limitation, not a workaround I invented, and it's worth knowing before you promise your boss ten minutes of work.

Or the fast path: we built a free generator that produces all three files for your trade, your services, your service area. You fill in a form, it writes the stack in your browser, nothing gets uploaded anywhere, and you copy the files into place with the install guide it gives you. It's linked at the bottom of this post. It also stamps every file with a "last verified" date, which brings me to the last thing.

Curious what a finished set looks like? HiVergent publishes both JSON files at the same addresses yours will use: agent-card.json and ai-catalog.json.

Keep them fresh or don't bother

The ARD spec itself emphasizes freshness, and the pattern in all this research is that stale files decay into noise. A pricing page that still says 2023 ranges, a service list two hires out of date, an agent card for services you dropped. If you publish once and forget it, you've built a monument, not an asset. My rule for clients: any time pricing changes or a service line changes, the files change the same week. That's also, conveniently, the maintenance gap where a service business earns recurring revenue — it's exactly the ongoing upkeep our Knowledge Catalog Building service covers — because owners will happily pay someone to keep this current rather than remember it exists.

Three files. Twenty minutes with the tool, an afternoon by hand. The companies doing this now are, as far as I can tell, mostly large and technical. The local trades are wide open. That's a rare sentence in marketing, and it won't stay true past this window.

Nathan Brown runs HiVergent AI in Dallas. The free Agent Stack Builder referenced throughout is on this site, and if you'd rather have us build and maintain your stack, that's the work we do. Book a free readiness assessment and we'll look at your whole presence, not just the files.