Ai Citation Auditor

AI Citation Auditor: Will AI Cite This Page?

A pre-publish and live-page gate that predicts whether AI answer engines will cite a page, and tells you exactly what to fix if they will not. It runs entirely on tools any Claude has out of the box (web fetch, web search, or content you paste in). No paid SEO stack, no crawler, no connectors.

Ranking in the ten blue links and getting cited in an AI answer are now two different games. This skill scores the second one.

---

The Mental Model: Citation Is a Vote

An AI answer engine does not "rank" sources the way classic Search does. For every answer it runs a silent vote between competing sources of truth, and your page is trying to win that vote:

1. Latent - what the model already learned in training. Wins by default when nothing contradicts it.
2. Retrieval - what the model just fetched for this specific query. Overrides latent for anything current.
3. Source-quality filter - a hidden re-ranker that discounts aggregators, thin listicles, forums, and pages that read like manipulative SEO, before the vote even happens.

There is also a hard prerequisite underneath all three: the model has to be able to retrieve you at all. Retrieval rank is still the single strongest predictor of AI citation. If a page cannot rank in roughly the top ten for its target query, the layers below are close to moot. Fix retrievability first, then compete on the layers.

This skill diagnoses which vote you are losing and returns the highest-leverage fix first.

---

When to Use This Skill

Trigger on any of:

• "Will AI cite this page?" / "Will this get pulled into an AI Overview?"
• "Why isn't [URL] showing up in AI Overviews, ChatGPT, or Perplexity?"
• "Run a citation audit on this URL"
• "Is this draft AI-citable before I publish?"
• "Is this page chunk-extractable?"
• "Does this read like SEO content to a model?"
• "AEO check" / "GEO audit" / "AI visibility audit"
• "Make this section citation-ready" (use REWRITE mode)

---

Modes

AUDIT (default)

Input: a live URL, or pasted content (Markdown or HTML), plus the target query the page should win.

Steps:

1. If a URL is given, fetch it with web fetch. If content is pasted, use it directly. If a query is not given, infer the most likely target query from the H1 and opening, and state the assumption.
2. Run a quick retrieval-prerequisite check: web search the target query and note whether the page (or its domain) surfaces in the top results. This is a directional signal, not a rank tracker.
3. Score all 6 layers using the rubric below.
4. Where a layer needs external truth (consensus alignment, corroboration), run one or two web searches to check the page's claims against what is broadly established.
5. Emit the scorecard, then the top 3 fixes, then the optional JSON block.

Output: a 0 to 100 score, per-layer breakdown, verdict, and a ranked fix list.

REWRITE

Input: one section or paragraph, plus the query it should answer.

Steps:

1. Identify the single question that section should answer.
2. Rewrite it answer-first: the first sentence resolves the question completely, in plain language, so it can be lifted out of context and still be correct.
3. Make it self-contained (no "as mentioned above"), attach at least one specific, checkable fact or figure, and align it with established consensus.
4. Return the rewrite, plus a one-line note on which layers improved.

Output: a citation-ready version of the passage.

---

Scoring Rubric (6 Layers)

| # | Layer | Weight | Critical? | Evidence tier | |---|---|---|---|---| | 1 | Answer-first extractability | 22% | Yes | Confirmed + Model-observed | | 2 | Evidence and corroboration | 20% | Yes | Confirmed | | 3 | Consensus alignment | 16% | No | Model-observed | | 4 | Source-quality fingerprint | 18% | Yes | Leak-inferred + Confirmed | | 5 | SEO-slop detection | 12% | No | Leak-inferred | | 6 | Machine structure | 12% | No | Confirmed |

Verdict bands: PASS at 80 or above. MARGINAL 60 to 79. FAIL under 60. Any critical layer scoring under 40 forces FAIL regardless of total. If the retrieval prerequisite fails, cap the headline verdict and label it "Not retrievable: fix ranking first."

Evidence tiers (so you can defend every score):

• Confirmed - stated in Google's own published Search and AI guidance, or standard, uncontested practice.
• Leak-inferred - drawn from the 2024 Google Content Warehouse API documentation leak. Directional, not confirmed by Google.
• Model-observed - drawn from how answer engines behave in the field and from published model guidance. Directional and subject to silent change.

Layer 1: Answer-first extractability (critical)

Does the opening sentence answer the target query completely, before any preamble? Can each section be lifted out and still make sense on its own? For a named entity, does the first line read as a clean fact triple, for example "[Brand] is a [type] that [does X], founded [year]"? Answer engines pull individual chunks, so buried answers and context-dependent sections lose. Pass looks like: question resolved in sentence one, every H2 self-contained, no orphan pronouns depending on earlier text.

Layer 2: Evidence and corroboration (critical)

Are the page's claims backed by named, checkable sources, specific figures, and dates? Vague assertions ("studies show", "many experts agree") do not survive a model that cross-checks. Original data, cited primary sources, and concrete numbers win. Pass looks like: key claims carry a source or a specific figure, dates are present, nothing central is unverifiable.

Layer 3: Consensus alignment

Do the facts match what is broadly established across the web? A page can be well written and still lose the vote if it contradicts consensus without strong evidence, because the model's latent knowledge overrides it. Run a search to confirm the page is not stating something the rest of the web disputes. Pass looks like: claims align with consensus, or any contrarian claim is backed by unusually strong, cited evidence.

Layer 4: Source-quality fingerprint (critical)

Would the hidden re-ranker treat this as a trustworthy source or discount it? Signals: a real named author with visible expertise, consistent brand facts across the site, entity clarity, and structured data. Contradicting facts across your own pages (different founding year, different counts) is a penalty here. Pass looks like: identifiable expert author, consistent brand facts everywhere, entity and organization signals present.

Layer 5: SEO-slop detection

Does the page read like content built for a model rather than a reader? Flags: a long throat-clearing intro before the answer, repeated keyword phrasing, filler sub-sections that add nothing, and template sameness across a site. High effort and originality score well, low-effort scaled content scores badly. Pass looks like: gets to the point fast, no filler, reads as written by someone with first-hand knowledge.

Layer 6: Machine structure

Can a parser cleanly read the page? Signals: valid structured data, a clean heading hierarchy, question-and-answer formatting, a short summary near the top, and lists or tables for comparable data. This is a citation-efficiency layer for multi-engine visibility, not a Google indexing requirement. Pass looks like: relevant schema present, logical H-structure, scannable formatting that maps to how the content will be quoted.

---

2026 Field Findings (folded into scoring)

Directional patterns that hold across engines right now:

• Rank is the entry ticket. Across engines, the strongest predictor of being cited is already ranking well in classic search for that query. Layers 1 to 6 decide who wins among pages that can be retrieved at all.
• Consensus facts beat FAQ markup. FAQ rich results are fading as a signal. What moves an answer engine is fact consistency: every page, profile, and listing stating identical numbers (founding year, counts, refund window). One contradicting fact is a source-quality penalty.
• The About page wins branded queries. For "who is [brand]" style questions, the most-cited page is often the About page, not the homepage. Lead it with a clean fact triple.
• Current-year dating helps commodity and comparison queries. For "best X" style queries, models lean toward content dated to the current year. A visible, accurate date refresh can restore freshness for AI even when Google would want a deeper update. Do not fake dates; keep them truthful.
• Every section must stand alone. Answer engines quote fragments. A section that only makes sense after reading the one above it will not be pulled.

---

Output Format

Emit the human-readable scorecard first (this is the part worth screenshotting), then an optional JSON block for anyone wiring it into a pipeline.

AI CITATION AUDIT
Page:          
Target query:  ""
Mode:          AUDIT

VERDICT:  FAIL (54 / 100)      bands: PASS >=80 | MARGINAL 60-79 | FAIL   fix ranking first

LAYER                          SCORE  WT   TIER              DIAGNOSIS
1 Answer-first extractability   45   22%  Confirmed/Model   Answer buried under a 90-word intro
2 Evidence and corroboration    38   20%  Confirmed         Claims have no named sources or figures  [CRITICAL <40]
3 Consensus alignment           72   16%  Model-observed    Aligns with consensus, one unsourced stat
4 Source-quality fingerprint    60   18%  Leak-inf/Conf     No named author; brand facts consistent
5 SEO-slop detection            55   12%  Leak-inferred     Long filler intro, repeated key phrase
6 Machine structure             68   12%  Confirmed         Clean H-structure, no Article schema

TOP 3 FIXES (highest leverage first)
1  [Layer 2, critical]  Add named sources and one concrete figure to the three core claims.
2  [Layer 1]            Move the direct answer to sentence one; cut the intro.
3  [Layer 4]            Add a real author with a short expertise line and consistent byline.

Optional machine-readable block:

{
  "page": "...",
  "target_query": "...",
  "mode": "AUDIT",
  "verdict": "FAIL",
  "total_score": 54,
  "retrieval_prerequisite": { "surfaces_top_results": false },
  "layers": {
    "answer_first_extractability": { "score": 45, "weight": 22, "critical": true },
    "evidence_corroboration":      { "score": 38, "weight": 20, "critical": true },
    "consensus_alignment":         { "score": 72, "weight": 16, "critical": false },
    "source_quality_fingerprint":  { "score": 60, "weight": 18, "critical": true },
    "seo_slop_detection":          { "score": 55, "weight": 12, "critical": false },
    "machine_structure":           { "score": 68, "weight": 12, "critical": false }
  },
  "fix_list": [
    { "priority": 1, "layer": "evidence_corroboration", "fix": "..." },
    { "priority": 2, "layer": "answer_first_extractability", "fix": "..." },
    { "priority": 3, "layer": "source_quality_fingerprint", "fix": "..." }
  ]
}

---

What This Skill Does NOT Do

• It does not write full articles. It audits and rewrites single sections.
• It does not build links, mentions, or off-site authority.
• It does not track rankings or replace a rank tracker. The prerequisite check is a directional read from a live search.
• It does not guarantee a citation. Answer-engine logic is undocumented by vendors and shifts silently. Treat scores as decision support.

---

Honest Caveat

This is directional, not gospel. No AI vendor publishes its citation logic, and it changes without notice. The scores tell you where a page is weak against how these systems are observed to behave, not a promised outcome. The durable bets are the truthful ones: primary sources, consistent facts, real authorship, and a clear answer up front. Those survive the fact-checking that is already tightening across models.

---

About This Skill

This is a free, single-file, dependency-free version of a larger retrieval-layer auditing system used internally at GODRANK. It is intentionally self-contained so anyone can drop it into a Claude Project or skills folder and run it with no setup. Use it, fork it, improve it.

Author: Niro, Founder of GODRANK (Madrank Digital Ltd). X: @PGR_mx

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: madrank8
• Source: madrank8/ai-citation-auditor
• License: MIT
• Homepage: https://x.com/PGR_mx

Install and usage instructions live in the source repository linked above.