The instrument
Every Bliss score is the output of an audited engine — not a vibe.
The Oracle is the scoring spine: 258 graded signals — the full 2026 Google-organic SEO canon now at parity with the AEO/GEO frontier — mined from a 45-source research corpus, wired into 39 weighted factors across 4 planes, frontier-ranked by a Signal Integration Score and locked by 123 passing tests. Every number in this entire report traces back to a row in here — with its evidence grade and source on the label. This is the page that proves the rest is real.
The four-plane factor spine
39 weighted factors — every one resolves to formula + signals + a fix
all BUILT| Plane | The question it answers | Factors | Signal density · by grade | anthropic.com |
|---|---|---|---|---|
| SEO | Built to rank on Google? | 12 | 129 · 14A·63B·40C·12D | 56.5 |
| AEO | Built to be cited by AI? | 13 | 60 · 28A·12B·19C·1D | 38 |
| GEO | Trusted entity LLMs confidently cite? | 8 | 58 · 3A·30B·25C·0D | 64 |
| AI Visibility (BSS) | Actually cited across engines? | 6 | 11 · measured outcome | 37% |
Each factor carries a weight, a confidence tier (HIGH / DIR / CTSD / UNV), a formula, the signals behind it, a recommendation, and a contract field — owned once in factor-registry.ts, imported by both the engine and the contract so a weight can never be declared twice. Signal density, merged in: the 258-signal bible = 129 SEO + 129 AEO/GEO (of which 11 score on both the AEO & GEO planes). SEO now holds 14 confirmed Grade-A drivers — title, headings, content depth, internal links, backlinks/PageRank, E-E-A-T, the indexability gate — at full parity with the AEO/GEO frontier. AI Visibility is the measured-outcome layer (11 BSS pillars), not an on-page evidence grade.
The 258-signal bible — graded evidence
Grade distribution
49 of 258 are Grade-A · grade = importance × confirmed-signalSEO reached parity: its 129 graded signals — crawl/render, Core Web Vitals, the 2024 leak, E-E-A-T, off-page — match the 129 AEO/GEO signals, and now hold 14 confirmed Grade-A drivers. Per-plane signal density is folded into the factor-spine table above. The grade letter encodes importance × how-confirmed a signal is — not citation type — so the confirmed canon (title, links, content, backlinks) reads A, and debunked folklore reads D.
Every signal, scored against anthropic.com
By grade
of the 227 scored (98 AEO/GEO + 129 SEO)Site verdict
how anthropic.com scores on eachThe bible — 227 signals, ranked by opportunity
opportunity = importance (grade) × site-gap · filter by plane| Opp | Signal | Plane | Grade | SIS | Status | Evidence |
|---|
Each signal's status is inherited from the measured parent-factor band on anthropic.com (PASS ≥70 · PARTIAL 45–69 · GAP <45); opportunity = a signal's importance (its grade) × anthropic.com's gap on it — the order to fix in. The frontier SIS rank is deliberately NOT a factor, so a confirmed foundation (backlinks/PageRank, title, internal links) is never buried beneath novel tactics; a foundation you already pass shows low opportunity, not low importance — the Grade column carries that. REFERENCE (32) marks signals with no per-site pass/fail by nature — the named algorithm systems (PageRank, RankBrain, BERT, core updates), Google-internal NavBoost click telemetry, and the leak's site-level features — bible knowledge we explain + proxy, never a fabricated band; UNKNOWN (37) is an honest measurement gap (e.g. Product/Review schema types that don't apply to anthropic.com). This is the unified 4-plane signal universe (227 scored: 98 AEO/GEO + 129 SEO; 11 score on both AEO+GEO planes); grade & SIS come from the master database, the site status from this audit. SIS itself = 100·Gate·B·(0.26E+0.18M+0.22I+0.14O+0.20A)/(1+R).
The debunked ledger — what the engine refuses to claim
The engine grades its own evidence — and downgrades the hype
the honesty moatMost of the industry repeats vendor numbers uncritically. The Oracle is built to distrust them. These are claims the engine deliberately cut, capped, or flagged — each one a place a competitor's tool would mislead you:
llms.txt drives citationsCUT to 0
SE Ranking's 300K-domain study found zero citation correlation. Re-scoped to crawlability/future-proofing only; AEO weight cut 4→0.
FAQ schema = free citationsPENALTY
Q&A format without evidence scores −5.74% (SIS #2, Grade A). FAQ is only a win when the answer carries data.
SpeakableSpecification boosts AIOCONTESTED
An 1,885-page study found a 4.6% citation decline vs control. Never recommended as a booster.
128-token chunk = pipeline specGRADE D
Single-source reverse-engineering (TryRankly). Kept as a design discipline, never quoted as confirmed OpenAI architecture.
3.2× ChatGPT freshness multiplierSINGLE-SOURCE
Traces to one Foglift.io blog with no disclosed methodology. Recency matters — the multiplier is not data.
"+115.1%" Princeton GEO liftMISUSED
A position-5-specific composite, not a blanket single-factor number. We quote the real per-intervention figures (+41% quotation, +30% stats).
Wikipedia = 27% of citationsCORRECTED
Real figure is 6–8%. The inflated number is everywhere; the engine carries the corrected one.
Reddit dominates everywhereENGINE-SPECIFIC
Reddit+Quora = 52.5% of citations — but Google AIO prefers brand domains 59.8%. Community dominance is per-engine, not universal.
This ledger is the wedge no tracker has: every rival sells you their best-case numbers; Bliss ships the corrected ones with the contradicting study cited. Confidence labels are a feature, not a disclaimer — you can see exactly what's proven vs directional before you spend a dollar acting on it.
The engine, applied to anthropic.com
Real receipts — the signals that actually fired on this audit
on-screen, not assertedEach plane's score in this report is backed by its own signal-depth pull — the actual graded signals checked against anthropic.com, with the evidence receipt visible on every plane dashboard. The engine didn't tell you GEO scored 64; it measured 64 graded signals (Wikidata QID, the 4-engine fact probe, hub presence) and showed its work. That's the difference between a number and a verdict.
The machinery — 10 agents, 12 formulas
The agents
who does the workThe MAX Orchestrator runs the deterministic pipeline: crawl → replay → competitor → citation → oracle. Each agent owns one stage; the Oracle Synthesizer joins the signal evidence to the factor scores.
The formulas
no hidden mathCombined = SEO·0.55 + AEO·0.45 BSS = 100·√(Σ pillarComposite·R^WSI·W·I·H / Σ W·I·H /100)·gate R^WSI = (center − margin)/(1 + z²/N) Wilson 95% lower bound ED = clip((0.62N+0.57D+0.55C+0.41H−0.06Q+0.20T)/2.29) citation_mass ∝ (BSS/100)² 23× AI conversion (DIR)
12 named formulas — combined, bss_master, wilson, sis, evidence_density, ordinal_lift, technical_gate, opportunity, authority_decay, revenue, prominence_blend, pillar_weights. All shown in the deliverable; category breakdowns always sum to the total.
The moat — why this beats a prompt tracker
Runs your prompt N=1, sees if your brand appears, reports binary. Re-run it tomorrow and the answer flips — LLMs are stochastic. You get a number you can't reproduce and can't act on.
Scores the structural factors that cause citation — graded, weighted, deterministic. Same input, same score, every time. Repeatable, auditable, and tied to a deployable fix — not a coin-flip you watch.
33 competitor tools, distilled into one engine. The 127-tool catalog was mapped from the entire AI-search tooling field; Bliss is the only one that couples all four planes, grades its own evidence, and closes the loop from score → cause → deployable code. The trackers tell you that you're invisible. The Oracle tells you why, with the receipt, and hands you the fix.
Method & provenance
The spec is generated from the codebase, never hand-edited — so the census, factor weights, and signal grades on this page can't drift from what actually runs. Every seed in the contract must carry a provenance {sourceUrl, fetchedAt, method} trail or CI fails it (the "no Sarah Chen" rule). The engine is reviewed, not trusted: weights and formulas are labeled HIGH / DIR / CTSD / UNV, and the debunked ledger is the proof that "directional" means directional.