Web discoverability

A public page is not done until its discoverability surface is complete. Search engines and generative engines both work from the same raw material: the initial HTML response, the head metadata, and sentences that survive being quoted out of context. Ship all three with the page, not as a later "SEO pass".

Procedure

1. Complete the per-page head. Every indexable page ships with:
   • <title> — unique per page, ≤ 60 characters, page-specific words first, site name last (Pricing — Acme, not Acme | Home of the best...).
   • Meta description — ≤ 155 characters, stating the answer or value, not a teaser. This is the excerpt search results and AI engines show; write it as the sentence you want quoted.
   • Canonical link — exactly one per page. Required wherever the same content is reachable under multiple URLs (trailing slash, query params, www).
   • Open Graph + Twitter card — og:title, og:description, og:image (1200×630), og:type; twitter:card = summary_large_image. Every shared link renders through these. The og:image file must actually exist and be a real 1200×630 PNG or JPG — link-preview crawlers render SVG inconsistently, and a dangling image reference silently kills every share card.
   • Robots meta only to exclude — noindex on thin, duplicate, or gated pages. Absence means index; never add an empty or permissive robots meta out of habit.
   • No <meta keywords> — dead surface; adding it only signals template cargo-culting.
2. Cover the site-level surfaces:
   • sitemap.xml — canonical URLs only, referenced from robots.txt.
   • robots.txt — allow by default; block only genuinely private paths. Blocking AI crawlers (GPTBot, ClaudeBot, PerplexityBot) is a product decision, not a default — flag it to the user rather than deciding either way silently.
   • llms.txt at the site root — a curated markdown reading list for AI agents, shipped alongside the sitemap and kept current as public pages change. ceh-seo:write-llms-txt owns its format and curation.
   • 404s return HTTP 404. An SPA fallback serving every unknown path as 200 poisons the index with phantom pages.
   • Moved pages get a 301 at the server/adapter level — no redirect chains, no client-side window.location redirects. Most AI crawlers run no JavaScript, and each hop in a chain loses signal.
3. Add structured data:
   • JSON-LD in a <script type="application/ld+json"> block — never microdata attributes. JSON-LD sits apart from the visible markup, so it is generated and validated without touching templates.
   • Match the type to the page: WebSite + Organization on the home page, Article/BlogPosting on posts, SoftwareApplication or Product on product pages, FAQPage on FAQ content.
   • Only mark up content actually visible on the page — invisible structured data is a penalty surface, not a bonus.
4. Check rendering:
   • Content must be in the initial HTML response. Most AI crawlers do not execute JavaScript, and search rendering is deferred and unreliable. A client-only SPA page is invisible to the engines this skill exists for.
   • SvelteKit: export const prerender = true on static public routes; SSR for dynamic ones.
   • React + Vite: public pages need a prerender/SSG pass — a bare client-side SPA is not a discoverable page, whatever its meta tags say.
5. Write the page text for citation by generative engines (GEO):
   • Answer-first sections. Open each section with the claim; elaborate after. Engines quote sentences, not pages — a section whose first sentence is throat-clearing gets skipped.
   • Extractable facts. State facts in standalone sentences that survive quoting alone ("X supports Y since v2.1"), with concrete numbers over adjectives.
   • Question-shaped headings for question-shaped content — the heading is the query the engine matches against.

Rules

• No keyword stuffing, hidden text, or doorway pages — engines derank all three.
• No SEO library or framework plugin for what static head tags do — a <svelte:head> or React helmet-equivalent block is a few lines.
• Never trade accessibility for SEO. Semantic headings, landmarks, and alt text serve both; when they appear to conflict, accessibility wins.