The close
You don't have forty problems. You have eight fixes — and the first one moves six planes.
Across seven dashboards the audit surfaced dozens of findings. They collapse into 8 deployable fixes. They're ranked here not by where they came from but by cross-plane leverage — and the top fix, a single Organization schema block, lifts SEO, AEO, GEO, AI-Visibility, Citations and Technical at once. Ship the queue and the modeled readiness moves 53 → ~72. Then we re-run the audit and prove the lift — the part no tracker does.
1 moves 6 planes
What this plane is
trust-spectrum remediation · the moatEvery other tool ends at the diagnosis. The Action Center is the deliverable: fixes ordered by trust-spectrum (proven levers before directional ones) and cross-plane leverage, each carrying the exact code to deploy and the Oracle signals it closes — then a causal loop that measures whether it worked.
Cross-plane leverage — one fix, many planes
Which planes each fix moves
leverage = how many light up| Fix | SEO | AEO | GEO | AI-Vis | Cite | Tech | Planes |
|---|---|---|---|---|---|---|---|
| Organization schema (@id + sameAs + foundingDate) | 6 | ||||||
| BLUF direct-answer rewrite | · | · | · | 3 | |||
| Outbound authority citations | · | · | 4 | ||||
| Expert quotes + statistics | · | · | · | · | 2 | ||
| Visible dates / freshness | · | · | · | · | 2 | ||
| Question headings / fan-out | · | · | · | · | · | 1 | |
| TTFB / Core Web Vitals | · | · | · | · | 2 | ||
| llms.txt + robots AI rules + headers | · | · | · | · | · | 1 |
This is the case for sequencing by leverage, not by plane. Fixing "AEO" and "GEO" as separate projects double-counts the work — the schema block, the outbound citations, and the BLUF rewrite each satisfy multiple planes' top factors in one deploy. Do the wide fixes first.
The queue — ranked, with the code
Ship the Organization schema — @id + sameAs + foundingDate
effort: 1 block→ 6 planesThe site emits no @id and no sameAs — it leaves its own entity graph for Wikidata to assert, ships schema on only 29% of pages, and one engine already has the founding year wrong. One homepage block fixes all of it: claims the entity, links all 11 hubs, and pins foundingDate.
<script type="application/ld+json">{ "@context":"https://schema.org","@type":"Organization",
"@id":"https://www.anthropic.com/#organization","name":"Anthropic","legalName":"Anthropic PBC",
"url":"https://www.anthropic.com/","foundingDate":"2021",
"founder":[{"@type":"Person","name":"Dario Amodei"},{"@type":"Person","name":"Daniela Amodei"}],
"sameAs":["https://www.wikidata.org/wiki/Q116758847","https://en.wikipedia.org/wiki/Anthropic",
"https://github.com/anthropics","https://x.com/AnthropicAI","https://www.linkedin.com/company/anthropicresearch/"] }</script>
BLUF rewrite — a 40–60 word declarative answer under every H2
effort: per-page→ AEO 17→~70Direct Answer Blocks score 17 — the worst AEO category. Every page's LLM verdict was "no direct extractable answer." Open each H2 with a zero-preamble declarative answer (the GEO-SFE pattern: +17.3% citation rate, 6-engine validated). This is what converts a mention into a citation.
Outbound authority citations — turn the empty network on
effort: in-copy→ GEO 15→~45The citation-network factor sits at 15/100 — the site links out to ~zero external authorities. Pages that cite authority become authority. Add inline citations to primary sources (arXiv, standards, named research) in research & product copy.
Expert quotes + statistical anchoring
effort: per-page→ +30–40% visibilityQuotation (33) and Statistical (31) are both weak. Quoted pages average 4.1 citations vs 2.4; pages with 19+ statistics average 5.4 vs 2.8. Add named expert quotes and brand-attributed stats with sources to the priority pages.
Add visible dates — freshness is at zero
effort: template→ Freshness 0→~60Freshness scores 0 — there are no visible dates anywhere on the site. The single highest-opportunity signal in the bible (opp 87). A template-level published/updated date is a one-change, site-wide win.
Question-shaped headings + sub-intent fan-out
effort: copy→ Conversational 14→~55Conversational alignment scores 14 — headings are labels, not questions. Reframe key H2s as the 6+ word questions users actually ask, with discrete sub-intent sections, to match AI-Mode fan-out.
Fix the slow rooms — TTFB & LCP
effort: infra→ Tech +, SEO +18 of 31 pages answer slower than 2s (peaking at 6.3s) and the homepage LCP is 4.9s. Slow TTFB hurts AI crawlability and ranking. Cache the heavy routes at the edge; the two clean pages prove the stack can do it.
AI discovery files — llms.txt, robots AI rules, security headers
effort: 3 files→ Tech +llms.txt returns 404, robots.txt has zero crawler-specific rules, and X-Frame-Options / X-Content-Type-Options / Referrer-Policy are missing. Completeness + future-proofing (Claude reads llms.txt) — honestly framed: this is hygiene, not a citation lever.
Ranked by the opportunity formula from the Oracle — SIS × site-gap × grade — then weighted up for cross-plane reach. Foundational fixes (1–2) unlock the most factors; quick wins (5–6) are template-level and ship today. Each "closes" line ties the fix back to the exact graded signals it satisfies in the bible.
Sequencing — quick wins vs big bets
Ship this week
high impact · low effortDates (template change → Freshness 0→60, the #1 signal) · Organization schema (one block → 6 planes) · Question headings (copy edit → Conversational 14→55). Three changes, days of work, and the two biggest factor gaps close.
Plan the quarter
high impact · real effortBLUF rewrite & quotes/stats (per-page content work across the 12 priority pages) · outbound citations (editorial pattern) · TTFB/LCP (infra). The depth work that turns mentions into citations — sequenced after the quick wins bank the easy lift.
Deployables — code, not advice
llms.txt
drop at /llms.txt# Anthropic > AI safety and research company. Maker of Claude. ## Core - [Claude](https://www.anthropic.com/claude): model family - [Research](https://www.anthropic.com/research): safety & interpretability - [API](https://docs.anthropic.com): developer platform ## Policies - [Responsible Scaling](https://www.anthropic.com/responsible-scaling-policy)
robots.txt — welcome the AI crawlers
replace the 71-byte stubUser-agent: GPTBot Allow: / User-agent: ClaudeBot Allow: / User-agent: PerplexityBot Allow: / User-agent: Google-Extended Allow: / User-agent: * Allow: / Sitemap: https://www.anthropic.com/sitemap.xml
BLUF answer block — the pattern for every H2
turns a heading into a citable chunk<h2>What is Constitutional AI?</h2> <p>Constitutional AI is a training method that aligns a model to an explicit set of written principles — a "constitution" — instead of relying only on human feedback. Anthropic introduced it in 2022 to make Claude's behaviour more transparent and steerable.</p> <!-- 40–60 words · declarative first sentence · entity + date + source in the next two -->
The causal re-verify loop — the moat
Deploy → re-audit → measure → prove
nobody else closes this loopBecause the score is deterministic and structural, the same audit re-run after a deploy isolates the fix's effect. To prove a citation-probability lift from 0.55 → 0.70 at p<0.05 takes 163 replays per arm — a number the engine computes (a power analysis), so "it worked" is a measured claim, not a vibe. Trackers run your prompt once and report binary; they can't tell you whether yesterday's change moved anything. This loop is the plat-02 trust-spectrum remediation engine — a BUILD-ONLY moat no data vendor sells.
Projected lifts on this page are directional (modeled from the factor formulas + signal effect sizes). The loop is what converts them into verified — re-audit after deploy and the projection is replaced by a measurement.
Projected outcome
Where the queue takes each plane
modeled · then proven by re-auditComposite readiness 53 → ~72. The largest single move is AEO (38 → 72), because the BLUF rewrite, dates, quotes and question-headings all land on the same plane — and because that's where anthropic.com is built best yet seen worst. These are modeled projections; the re-verify loop replaces each arrow with a measured delta.
Method & honesty
The queue is assembled from the real findings on the SEO, AEO, GEO, AI-Visibility, Citations, Technical and Oracle dashboards — deduped so a fix that serves several planes appears once, ranked by the Oracle opportunity score (SIS × site-gap × grade) weighted for cross-plane reach. Every fix carries the exact code or content pattern to deploy and the graded signals it closes. Projected lifts are directional — modeled from the factor formulas and published effect sizes, not promised. The causal re-verify loop (re-audit after deploy, Wilson-bounded deltas, the 163-replay power target) is what turns a projection into a proof — and it's the plat-02 BUILD-ONLY moat: the close that no diagnostic-only tracker ships.