The short version
Over the past few hours, I went through a portion of my website portfolio and rebuilt the parts that AI agents, not people, read. The pages look identical to those seen by a human visitor. But underneath, each site now hands an AI agent a clean, machine-readable map of what the site is, what it offers, and how to interact with it programmatically. The outcome: when an AI assistant is asked about one of our sites or its subject matter, it can find, trust, and use our content correctly rather than guessing.
Here’s the plain-English account of what changed and why it matters.
Why bother at all
The way people find information is shifting. More and more, the first thing that “reads” a website isn’t a person in a browser. It’s an AI agent acting on someone’s behalf. Someone asks their assistant a question, and the assistant goes out, reads a handful of sites, and synthesizes an answer.
Traditional websites are built entirely for human eyes: nice layouts, images, menus. An AI agent has to reverse-engineer meaning from that visual soup, and it often gets it wrong — missing key facts, mangling numbers, or skipping the site entirely because it couldn’t tell what the site was for.
There’s now an emerging set of open standards that let a website explicitly describe itself to machines — a kind of “back door” of structured signals that sit alongside the normal human-facing pages. Adopting them is the difference between hoping an agent understands you and telling it, in a format it’s built to consume. I wanted our sites on the right side of that line early, while most of the web still isn’t.
What I actually did to each site
Think of it as adding a machine-readable layer on top of each existing site. Nothing about the human experience changed. The additions fall into a few buckets:
1. A self-description an agent can read. Each site now publishes a plain-text and structured summary of what it is, who’s behind it, and what it covers all served in the formats agents expect (including a “give me the Markdown version” mode, so an agent can ask for clean text instead of parsing a full web page).
2. A directory of standard discovery files. There’s a growing convention for where machines look for a site’s “specs.” Those are a set of well-known addresses that host resources such as an API catalog, a description of the tools the site exposes, and metadata on how to interact with it. I created that whole directory for each site so an agent visiting for the first time finds a signposted set of entry points instead of a dead end.
3. Machine “endpoints” the agent can actually call. Beyond static description, each site now answers structured requests directly via an agent that can ask a question or request specific information and get a clean, structured answer back, rather than screen-reading a page. For our knowledge and informational sites, this means an agent can pull the real facts (what we do, our track record, our offerings) straight from the source.
4. Clear usage preferences. I added explicit signals stating how our content may be used by AI. For example, it’s fine to reference and answer questions from our material, but set our own boundaries around training. This is stated in a standard place, so any well-behaved agent honors it.
5. Response “headers” that point the way. When any page loads, it now quietly includes machine-readable pointers (“here’s the content, here’s the documentation, here’s how to talk to me”), so even an agent that lands on a random page immediately knows where the structured resources live.
The two housekeeping steps at the infrastructure level
Two changes weren’t about the site’s code at all. They were about the plumbing that sits in front of it. These are the kinds of settings that live with your domain registrar and DNS provider rather than in the website itself, and they differ from provider to provider (some make them one-click; others bury them or format them differently), which is why they require a bit of care.
Domain-level discovery + security signing. Some agents look up a site’s capabilities at the domain-name level, before they ever load a page — so I added the appropriate DNS entries advertising the site’s agent endpoints. Paired with that, I turned on the DNS security feature that cryptographically signs the domain’s records, so an agent can trust that those signals are genuine and haven’t been tampered with in transit. Provider by provider this ranged from a single toggle to a multi-step form that had to be entered exactly right.
Making both “front doors” behave the same. A site is usually reachable in two ways, with and without the “www” prefix. It’s easy for one of those to quietly point somewhere different from the other. On one site, the plain domain served the fully upgraded site, while the “www” version still pointed to a stale target, so an agent (or a checker) hitting the “www” address saw none of the improvements. The fix was a simple redirect, so both addresses resolve to the same fully upgraded site. Worth flagging as a lesson: after any upgrade, check both front doors.
How I verified it
Rather than take it on faith, I ran each site through an independent “agent-readiness” check that grades it on how well it exposes itself to AI, from “basic web presence” to the top tier, “agent-native.” The site, from Cloudflare, is a godsend. You can find it at IsItAgentReady.com
I used that as a scorecard, fixed whatever it flagged, and re-checked until each site reached the top level. I also confirmed the machine endpoints actually respond correctly to live requests, not just that the files exist.
Did I score 100 ? Nope. A couple of checks I intentionally left “unmet.” They relate to publishing a full agent login/registration system, which only makes sense for a site with private, protected data behind a sign-in. Our sites are public, so faking that machinery would be dishonest to the very agents we’re trying to serve. Leaving those as an honest “not applicable” is the correct answer, not a gap.
The outcome
Every site in the portfolio now presents itself to AI agents at the highest readiness tier. In practical terms, except for my blog, which being that it’s on WordPress, requires a bit more work, but now for the rest:
- When an assistant is asked about us or our subject areas, it can find us, read us accurately, and cite us instead of skipping us or garbling the details.
- Agents can pull real, structured facts from the source, which means fewer errors and stale numbers floating around about what we do.
- The signals are signed and trustworthy, so a careful agent knows the information is genuinely ours.
- We stated our own terms for how AI may use the content in a place where machines actually check.
None of this changes what a human sees. It’s an investment in being legible to the next generation of “readers.” The automated ones that increasingly sit between our content and the people looking for it.
As that shift accelerates, the sites that told machines who they are will be the ones that get found, understood, and represented correctly. Ours now do.