Every agency now tells clients to get ready for AI search. Almost none will show you their own homework. This note is the homework: how agencyechelon.com is actually built, and why each decision was made with two readers in mind, the human with a scroll wheel and the machine assembling an answer.
Start with what is not here. There is no CMS, no framework, no template farm, no third-party font call, no tag manager firing before you have said a word. The site is hand-built static HTML, and that is not nostalgia. Static pages load fast enough that speed stops being a conversation, they contain nothing that can break at 2 a.m., and every byte that ships was put there on purpose. When your business is telling clients what belongs in their stack, the discipline has to start at home.
Then the machine editions. There is a feed for the readers who still bless RSS, and an /llms.txt file, a plain-text edition of the entire library, summary first, all posts listed with one-line descriptions, built for the crawlers that feed ChatGPT, Gemini, Claude, and Perplexity. If generative engine optimization is the discipline of being quotable by machines, the first move is handing the machine a clean copy.
Structured data follows one law here: every block must parse, and every block must describe something a human can see. Each note carries article schema with a real author entity, breadcrumbs, and, only where a visible Quick Answers section exists, question-and-answer markup to match. Schema that describes invisible content is spam with a schema.org address, and the engines are learning to hold it against you. The whole library also ships as a single machine-readable index object, regenerated from source on every build rather than edited by hand, because hand-edited structured data is how sites end up serving broken JSON to the one reader that never complains and never comes back.
Measurement waits its turn. Visitors in Europe and California see a small card and exactly zero pixels until they answer it; thirteen fire after a yes, none after a no, and everyone else gets the standard set. HIPAA took the pixels away from healthcare advertisers and I called it a favor; the same logic applies here. If a measurement setup cannot be explained out loud to the person being measured, it does not belong on the site.
The library itself is a mesh, not a pile. One hundred and five notes, every editorial link pointing to a piece published before the one citing it, every post closing with three related notes chosen by shared topic. Images carry their post's exact slug as a filename, and the alt text describes the photograph first and the concept second, in that order, because accessibility text written for rankings insults both audiences.
The part that surprises developers: a validation script runs eighteen checks before any version of this site ships. It confirms the sitemap, the feed, the AI edition, the index, and the disk agree on exactly which posts exist. It parses every block of structured data on every page. It hunts nested links, broken references, orphaned images, and duplicate titles. It even scans my prose for the vocabulary I have banned from it, which means a machine now enforces the house style against its own author. There is a humans.txt too, for the people who read view-source the way I read other people's media plans.
Why go this far for a consultancy site? Because the pitch and the practice have to match. I tell clients that the AI answer layer rewards clear claims, clean structure, and facts a machine can lift with confidence, and that the right model is the cheapest one that clears the bar. A site is the one place where all of that is verifiable by anyone with a right click. The best case study for the work is the thing you are reading.