bcs-website/docs/seo-aeo-optimization.md

SEO, AEO & AI-Crawler Optimization Plan

Requirements and rules for optimizing the BlueCap Strategies rebuild for classic search (SEO), answer engines / AI citation (AEO/GEO), and AI-crawler acceptance. Written 2026-07-15 against mid-2026 best practices (see Sources). Where a rule can be machine-enforced, it is defined as a Zod schema constraint or a build-time check so the build fails when copy violates it — this is the TypeScript/Astro equivalent of Pydantic validation (do not add Python/Pydantic to a JS build; Astro Content Collections already run on Zod).

0. Guiding principles (2026 reality check)

1. HTML is the source of truth, not llms.txt. AI search crawlers (GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, Google-Extended) overwhelmingly skip /llms.txt and crawl rendered HTML directly; Google confirmed it does not support llms.txt. Keep our llms.txt (low cost, ~10% of sites have one) but invest first in semantic HTML, structured data, and crawlable copy. Our Astro output is fully static — 100% of meaningful copy is in initial HTML, which is the single biggest AEO advantage over the old JS-heavy WordPress/Elementor site.
2. Structured data drives citation. Schema lets answer engines verify who we are, what a page covers, and which facts are reliable to cite. FAQ schema has among the highest AI citation rates (question/answer mirrors how engines present answers).
3. Answer-ready structure wins. Clear headings, direct definitional first sentences, and self-contained paragraphs get extracted; buried or paraphrased marketing prose does not.
4. E-E-A-T via the principal. Boutique-advisory credibility is entity-driven — a real Person entity (Brad Rodgers) with knowsAbout, linked to the Organization, is a core GEO asset. (Gated on owner-approved bio — see parity plan Phase 5.)

1. Programmatic enforcement (Zod + build linter)

1a. Content-collection SEO schema (fails the build)

Extend src/content.config.ts. Every collection (pages, services, focusAreas, insights) shares an seo shape with enforced limits:

import { z } from "astro:content";

export const seo = z.object({
  // <title> renders as "`title` | BlueCap Strategies" — keep the full SERP title ≤ ~60 chars
  title: z.string().min(10).max(55),
  // meta description + OG description; 120–158 is the safe SERP/AI-snippet window
  description: z.string().min(120).max(158),
  // canonical is derived from slug, but allow override for cross-posted insights
  canonicalUrl: z.string().url().optional(),
  // every page must ship a social share image (absolute path under /assets)
  ogImage: z.string().startsWith("/assets/").default("/assets/og/default.jpg"),
  // primary entity/topic — powers internal linking + schema keywords
  primaryTopic: z.string(),
  noindex: z.boolean().default(false),
});

Rules encoded here (build fails on violation): title length, description length window, OG image present, canonical is a valid URL. This is the "rules are being followed" guarantee you asked about — enforced on every npm run build in CI before ArgoCD ever sees the image.

1b. Build-time SEO linter (what Zod can't see)

Zod validates frontmatter, not rendered HTML. Add a small check (a Node script run in CI, and/or Playwright assertions in tests/) that fails the build if any built page in dist/ violates:

• Exactly one <h1> per page.
• Heading hierarchy has no skipped levels (no <h4> before an <h2>).
• <title>, <meta name="description">, <link rel="canonical"> all present and non-empty.
• Every <img> has non-empty alt (parity plan already requires this).
• At least one JSON-LD block present, and every JSON-LD block JSON.parses.
• Internal links resolve (no 404s to our own routes).
• Canonical host matches the chosen production host (apex vs www — see §5).

Recommended: a scripts/seo-lint.mjs that walks dist/**/*.html with a lightweight parser, plus keep the existing Playwright suite for link/CTA coverage.

2. Per-page technical SEO requirements

Already implemented in src/layouts/BaseLayout.astro (keep + enforce):

• <title>, <meta name="description">, <link rel="canonical">
• Open Graph (og:type/title/description/url/image) + Twitter summary_large_image
• JSON-LD injection slot (Organization + WebSite globally, page schema per route)
• lang="en", viewport, skip-link, RSS <link rel="alternate">

To add/verify:

Tracked in launch-checklist.md §4. The list that used to live here went stale (breadcrumbs are done). This section keeps the reference detail below; the live task state is in the launch checklist only.

• Per-page OG images — every page currently falls back to one shared image (BaseLayout.astro emits a single og:image). Generate topic-specific 1200×630 images for services, focus areas, and insights.
• robots meta wired to a noindex frontmatter flag (for /thank-you/, staging). Not wired today — /thank-you/ is sitemap-excluded but emits no noindex.
• article:published_time / modified_time OG tags on insights posts.
• Breadcrumbs — DONE. BreadcrumbList in src/utils/schema.ts, applied on detail pages.
• Sitemap — via @astrojs/sitemap, excludes /thank-you/. Still to add: lastmod from content dates for insights.
• Trailing-slash consistency — config uses trailingSlash: "always"; ensure canonical + internal links all match to avoid duplicate-URL dilution.

3. Structured data (JSON-LD) matrix

Page / type	Required schema	Notes
Global	Organization, WebSite	Already in src/utils/schema.ts. Add sameAs (real socials), logo, contactPoint (phone/email), areaServed.
Homepage	Organization + WebSite (+ optional SearchAction)	—
/services/ + each service	Service (with provider → Organization, serviceType, areaServed)	One per detail page.
/focus-areas/ + each focus	Service or DefinedTerm/Article	Focus areas double as topic-authority hubs.
About / Brad Rodgers	Person + AboutPage; Person.worksFor → Organization; knowsAbout[]	Owner-approved facts only. Core GEO/E-E-A-T asset.
Insights posts	Article/BlogPosting (author=Person, publisher=Organization, dates, headline, image)	Powers AI citation of thought leadership.
Insights index	Blog / CollectionPage	—
Contact	ContactPage + Organization.contactPoint	Use real values, never the WP placeholders.
FAQ sections	FAQPage — only where the same Q&A is visibly rendered	Highest AI-citation rate. Never add invisible FAQ schema.
Breadcrumbs	BreadcrumbList	Detail pages.

Rule: structured data must reflect content that is visibly present on the page. No hidden-text schema, no unsupported claims, no fabricated review/rating markup.

4. AEO / GEO content rules

Applied during Phase 2 copy restoration and Phase 5+ enhancements (must not violate the no-paraphrasing parity rule on restored copy):

• Answer-first: each section's opening sentence should stand alone as a direct answer / definition (e.g., "Marine conservation finance is …"). Extractable by answer engines.
• Semantic headings phrased as the questions/topics people ask ("What makes PFAS remediation a capital formation problem?").
• FAQ modules on high-intent authority pages (PFAS economics, marine conservation finance) — visible Q&A + FAQPage schema. Owner-approved answers only.
• Entity consistency: use canonical terminology everywhere — catalytic capital, aquatic resources, marine & freshwater ecosystems, PFAS remediation, marine conservation finance, environmental markets, coastal resilience. Feeds primaryTopic + internal linking.
• Internal linking graph: services ↔ focus areas ↔ insights ↔ principal bio. Dense, relevant internal links are a top GEO signal.
• Freshness: insights carry publishDate/modifiedDate; surface "Updated" dates.
• Statistics & specifics: the homepage metrics (27 years, $10B+, $3.5B+, 275+) are exactly the kind of concrete, citable facts answer engines favor — keep them prominent and, where true, attributable.

5. AI-crawler policy (robots.txt) — owner decision required

Current public/robots.txt = Allow: / for everyone. 2026 best practice is explicit per-agent rules, splitting search/answer bots (that can cite and send traffic) from training bots (that ingest content into model weights). Decision matrix:

Bot	Job	Recommendation
OAI-SearchBot, ChatGPT-User	ChatGPT search index + live citation fetch	Allow
Claude-SearchBot	Claude web-search citations	Allow
PerplexityBot	Perplexity answer citations	Allow
Google-Extended	Gemini + Google AI Overviews	Allow (maximizes AI visibility)
Googlebot, Bingbot	Classic search	Allow
GPTBot	OpenAI model training	Owner choice — allow = max reach, disallow = opt out of training
ClaudeBot	Anthropic model training	Owner choice (same trade-off)
Bytespider, stealth scrapers	Known to ignore robots.txt	Consider edge/WAF block (Cloudflare) — robots.txt won't stop them

Recommended default for an authority-seeking advisory firm: allow all search/answer + classic bots (maximize citation and discoverability); training bots are a values call. Because BlueCap wants to be cited as an authority, allowing everything is the simplest high-visibility posture. Non-compliant scrapers are handled at the Cloudflare edge (see gitops doc §4.2), not robots.txt.

Keep llms.txt (already present, low cost) but treat it as supplemental, not load-bearing.

6. Measurement & Monitoring — recommendation, needs sign-off

Three distinct layers; today none is decided.

6a. Uptime / infrastructure — reuse the existing monylog stack (do NOT deploy Uptime Kuma)

The homelab already runs a full self-hosted LGTM + Alertmanager stack on the monylog box (10.66.15.21, Ansible role pglta): Prometheus (15d retention, remote-write receiver), Grafana, Loki, Tempo, Alertmanager → Pushover (creds from OpenBao) with ntfy fallback → phone. It runs outside k8s on purpose so it still alerts when the cluster is down. Grafana Alloy runs as a k3s DaemonSet and remote-writes cluster metrics + ships all pod logs to monylog.

Already flowing for BlueCap (zero work):

• Pod/deployment health — Alloy scrapes kube-state-metrics + annotated pods → Prometheus. kube_deployment_status_replicas_unavailable, restarts, OOMKills are queryable now.
• nginx logs — Alloy ships all pod logs to Loki. 5xx rate / errors available via LogQL now.

To add (small):

1. blackbox_exporter — the one gap (not in the monylog compose). Add as a compose service + a Prometheus scrape job. Probe the in-cluster Service (app health) and the public URL (full edge-path uptime + TLS expiry). monylog is off-cluster → probing the public URL is a genuine external vantage. Metrics: probe_success, probe_http_status_code, probe_ssl_earliest_cert_expiry, latency.
2. web-bluecap.yml alert rules (mirror the existing host-infra rules pattern): SiteDown (probe_success==0 for 2m), CertExpiringSoon (<14d), High5xx (Loki/blackbox), DeploymentDegraded, PodRestarting.
3. Alertmanager already routes to Pushover — just add a service: bluecap label; no new channel. Grafana already present → build a BlueCap uptime/latency/5xx/pod-health dashboard.

• ArgoCD health/sync status as an additional deploy-level signal.

6b. Visitor & behavior analytics — PostHog + Umami side-by-side trial (ADR-0005)

Decided: run PostHog and Umami side-by-side initially and keep whichever fits best. This is product/visitor analytics, distinct from the monylog ops telemetry in §6a. Trial setup:

• PostHog — Cloud free tier (1M events/mo) for the trial, not self-host (self-hosting needs ClickHouse/Kafka/etc.; PostHog's docs steer small users to Cloud). Reverse-proxy ingest through our domain to reduce ad-blocker loss. Gives autocapture, funnels, session replay.
• Umami — self-hosted in k3s: cookieless, ~2 KB, no consent banner, always-on baseline (pageviews, sources, devices, custom events). Aligns with the Digital-ESG positioning.
• Both wired through one config-driven Analytics.astro in BaseLayout, gated by env vars, so either can be toggled and the winner kept without code churn. Same event set in both: consultation-CTA clicks, per-service/focus CTA source (ties to the segmented-CTA strategy), insights reads, contact submits.
• Consent/privacy must cover the superset — PostHog (esp. session replay) likely triggers cookie-consent + privacy-policy obligations; add these before enabling PostHog in production. Umami alone needs no banner.
• Avoid GA4 unless Google Ads attribution is later required.
• Evaluate on insight value vs. weight/consent/ESG fit; record the winner in a follow-up ADR.

6c. Search & AI-citation visibility

• Google Search Console + Bing Webmaster Tools (free, essential) — submit sitemap, monitor impressions/queries/coverage, Core Web Vitals.
• AI-citation tracking (emerging, optional): periodically test brand/topic prompts across ChatGPT/Perplexity/Google AI Overviews and log whether BlueCap is cited. Lightweight manual tracker to start; dedicated GEO tools are immature — don't over-invest yet.
• Performance budget: Lighthouse in CI; target 95+ SEO/Perf/Best-Practices/Accessibility. Static Astro should hit this easily; guard against regressions from images (use Astro Image).

7. Pre-launch SEO checklist

Moved. This list is no longer maintained here — it duplicated, and disagreed with, reality (it listed the Zod schema, seo-lint, and breadcrumbs as open when all three are done, and called for Uptime Kuma, which ADR-0004 replaced with monylog).

The maintained checklist is launch-checklist.md. SEO items are in §3 (owner decisions: canonical host, robots policy) and §4 (post-cutover: OG images, noindex meta, article times, lastmod, Lighthouse, trailing slashes). What is already complete is recorded in §5.

Sources (verified 2026-07-15)

• The State of llms.txt in 2026 — aeo.press
• llms.txt Explained (May 2026) — Codersera
• Structured Data for AEO and GEO: Schema Markup Guide 2026 — Kurieta
• Answer Engine Optimization: Complete AEO Guide [2026] — Frase
• Are FAQ Schemas Important for AI Search, GEO & AEO? — Frase
• The AI User-Agent Landscape in 2026: A Complete Reference — No Hacks
• AI Crawlers Explained: GPTBot, ClaudeBot, PerplexityBot (2026) — Anagram
• Robots.txt for AI Crawlers in 2026 — Cubitrek