SEO/GEO and static representation
Every page is static HTML built from the same file the assistant reads, with no JavaScript on content pages beyond a 2 KB script that animates the layout, one structured-data graph per page, and the whole corpus published for any machine that wants it.
A site with an agent on it still has to work as a site. Search engines, AI crawlers, people with JavaScript off and people who never touch the assistant all read the same pages, and those pages have to be complete, fast, stable and true without any help from a model. So the static site is the foundation and the assistant is added on top, never the other way round.
Static first
Every route is rendered to HTML at build time, rebuilt when content changes and once a day, and published all at once under one build id (ADR-0012). The shortest review window on any fact is 30 days, so per-request rendering would buy nothing, and static output is the only kind that is still correct when every backend is down. A page always prints the date a volatile fact was confirmed. The daily build switches it to "not reconfirmed since" within a day of its review window passing.
The site is built with Astro from kb.json as plain build-time data (ADR-0013). Section Markdown goes through the same parser and plugins the compiler uses, so a heading or a table cannot mean one thing on the page and another to the assistant. Section ids on the page are the anchors from the artefact, never derived again from heading text.
In the first phase, content pages carry one first-party script of at most 2 KB compressed, which only animates islands moving when the visitor opens or filters something; every page reads and works the same without it, and shipping more fails the build (ADR-0019). When the assistant later moves onto content pages, its script is capped at 60 KB compressed and loads after the page is usable. Until then the only page that loads the assistant is /ask, and its script is the site's own (ADR-0020). The host is held to the same rule. The Cloudflare features that inject scripts into pages are switched off in configuration, HTML is served with a header that forbids rewriting it, and a check after every deploy asserts that production serves no injected script (ADR-0085).
The other budgets are numbers that fail the build too: Largest Contentful Paint under 2.5 s, Interaction to Next Paint at most 200 ms, layout shift under 0.1, CSS at most 50 KB compressed and fonts at most 120 KB in total. Accessibility is held to WCAG 2.2 AA.
One Cloudflare Worker on one domain serves the pages, the assistant and the contact form, and every change ships as a single upload (ADR-0069). Pages are served as static files without running the Worker. Because the site and the agent ship together, an answer can never cite a section the live pages do not have. If the assistant is off or unreachable, every page is exactly as it would be without it.
Routes and anchors that stay put
There is one route per page-bearing entity, and a page shows the sections of at most one entity, so every section id is unique on the page that serves it (ADR-0011). Every id resolves to a URL. A renamed page answers with a permanent redirect. A retired one redirects to its successor, or answers 410 with a short page saying when and why it was retired. Canonical URLs have no trailing slash. Because section anchors are locked (see Knowledge layer), a link to a decision keeps landing on that decision after the prose around it is rewritten.
Readable by machines
Each page carries one JSON-LD graph, and every node's @id is derived from the canonical URL of the entity it describes, so the same project referenced from two pages is the same node (ADR-0015). The profile is a ProfilePage over a Person. A case study is a TechArticle with one part per design decision, so the reasoning is addressable by a machine as well as by the assistant. Indexes are CollectionPages. There is no FAQPage: its rich results stopped appearing in May 2026, and the value here is a consistent graph, not a result type. In that graph, worksFor names the company that employed me, never the client the work was delivered for.
Every documented AI crawler is allowed, training included, and each is named in robots.txt so the policy is on record rather than implied (ADR-0016). Nothing here is a secret, and a portfolio a model has read is one it can recommend. There is no llms.txt. In 2026 one study found about one domain in ten publishing one, another found 97% of the files it examined were never requested, and Google says such files are not needed. The day a crawler operator documents reading it, that changes.
The one machine-readable export is the artefact itself. kb.json and its JSON Schema are published at stable URLs and linked from How it works, and the sitemap is generated from the route table. Nothing on the site reads the published copy: pages are built from the artefact directly.
The CV is a view of the site. /cv is rendered from the same artefact, current roles first with overlapping roles shown as overlapping, and the PDF is printed from that page at build time (ADR-0017). A CV and a site cannot disagree when one is generated from the other.
How far this has been proved
- Tested: in the first phase, content pages carry at most the 2 KB motion script and nothing injected. The byte budget fails the build, and the post-deploy check asserts it on production.
- Designed: the performance budgets, the accessibility floor, the structured data and the crawler policy. They have not yet been measured on real traffic.