Geo Citability

AI Citability Scoring Skill

Core Insight

AI language models cite passages that meet specific structural criteria. Research on Generative Engine Optimization (Aggarwal et al., KDD 2024) found GEO tactics can lift visibility in AI answers by roughly 30–115% depending on the tactic and query set (see docs/SOURCES.md for provenance on this and the length figure). The pattern: AI systems preferentially extract passages that are ~134–167 words long, self-contained (understandable without surrounding context), fact-rich (specific statistics, dates, named entities), and directly answer a question in the first 1-2 sentences.

This score is a proxy, not a measurement. It scores how extractable your passages are against these research-derived criteria — it does not ask an AI engine whether it actually quoted you. Any "expected lift" is a research-informed estimate, never a promise. For the live check of whether the engines name you, run the free scan at areyoufoundbyai.com (two buyer questions on ChatGPT and Gemini; the trial and paid tiers cover all seven engines).

This is fundamentally different from traditional SEO copywriting, which optimizes for keyword density and user engagement metrics. GEO citability optimizes for extractability -- the ease with which an AI system can pull a passage from your content and present it as a direct answer.

Citability Scoring Rubric (0-100)

Category 1: Answer Block Quality (30% of total score)

This measures whether content contains clear, quotable answer passages that AI systems can extract verbatim.

Scoring Criteria:

What to look for:

• Definition patterns: "X is [definition]." / "X refers to [explanation]." / "X means [meaning]."
• Answer-first structure: The answer appears in the first sentence, followed by supporting detail.
• Quantified answers: "The average cost of X is $Y" rather than "Many factors affect the cost of X."
• Comparison answers: "X differs from Y in three ways: [list]" rather than "X and Y are often confused."

High-citability example:

Content delivery networks (CDNs) are distributed server systems that cache and serve
web content from locations geographically close to end users. A CDN reduces latency
by 50-70% on average by serving assets from edge servers rather than a single origin
server. The three largest CDN providers as of 2025 are Cloudflare (serving approximately
20% of all websites), Amazon CloudFront, and Akamai Technologies.

Word count: 58. Self-contained: Yes. Facts: 3 specific data points. Definition pattern: Yes.

Low-citability example:

If you've ever wondered why some websites load faster than others, the answer might
surprise you. There's this amazing technology that has been around for a while now.
It's changed the way we think about web performance. Let me explain how it works and
why you should care about it for your business.

Word count: 52. Self-contained: No (no topic identified). Facts: 0. Definition pattern: No.

Category 2: Passage Self-Containment (25% of total score)

This measures whether individual passages can be extracted and understood without needing the surrounding content.

Scoring Criteria:

Self-containment checklist for each passage:

1. Does the passage explicitly name the subject (not "it," "this," "they")?
2. Can someone understand the main point reading ONLY this passage?
3. Does the passage contain at least one specific fact, statistic, or named entity?
4. Is the passage between 50-200 words (the optimal extraction length)?
5. Does the passage avoid starting with conjunctions ("But," "However," "And") that imply prior context?

Category 3: Structural Readability (20% of total score)

This measures the structural formatting that helps AI systems parse and segment content.

Scoring Criteria:

Structural best practices for AI citability:

• Heading hierarchy: H1 (page title) > H2 (major sections) > H3 (subsections). Never skip levels.
• Question-based headings: "What is [topic]?" and "How does [topic] work?" are directly matchable to AI queries.
• Paragraph length: 2-4 sentences per paragraph. AI systems parse short paragraphs more reliably.
• Tables: Use for any comparison of 3+ items. AI systems extract table data with high accuracy.
• Lists: Use ordered lists for sequential processes, unordered lists for non-sequential items.
• Bold key terms: Bold the first use of important terms. This aids AI entity recognition.

Category 4: Statistical Density (15% of total score)

This measures the presence of specific, verifiable data points that AI systems prioritize when selecting citation sources.

Scoring Criteria:

What counts as a statistic:

• Specific percentages: "73% of marketers report..."
• Dollar amounts: "The average cost is $4,500 per month"
• Timeframes: "Implementation takes 6-8 weeks on average"
• Named studies: "According to the 2025 HubSpot State of Marketing Report..."
• Specific counts: "The platform integrates with 340+ tools"
• Comparison data: "40% faster than the industry average"

What does NOT count:

• "Many companies use..." (vague)
• "A significant percentage..." (vague)
• "Studies show that..." (no named source)
• "Experts agree..." (no named experts)

Category 5: Uniqueness & Original Data (10% of total score)

This measures whether the content provides information that AI systems cannot find elsewhere, making it a necessary citation source.

Scoring Criteria:

Signals of unique content:

• "Our analysis of [X] data found..."
• "We surveyed [N] [professionals] and found..."
• "Based on our experience with [N] clients..."
• Custom charts, graphs, or data visualizations
• Case studies with specific named outcomes
• Original frameworks, methodologies, or taxonomies

Analysis Procedure

Step 1: Fetch and Parse Page Content

1. Use WebFetch to retrieve the target URL.
2. Extract the main content area (exclude navigation, footer, sidebar, ads).
3. Preserve heading structure (H1-H6 tags).
4. Preserve paragraph boundaries, lists, and tables.
5. Calculate total word count of main content.

Step 2: Segment Content into Blocks

1. Split content at each heading (H2 or H3) to create content blocks.
2. For each block, record:
   • The heading text
   • The full text content under that heading
   • Word count of the block
   • Number of paragraphs
   • Number of lists and tables
   • Number of statistics/data points
   • Whether the block contains a definition pattern
   • Whether the first 60 words form a standalone answer

Step 3: Score Each Block

For each content block, calculate:

• Answer Block Quality sub-score (0-100)
• Self-Containment sub-score (0-100)
• Structural Readability sub-score (0-100)
• Statistical Density sub-score (0-100)
• Uniqueness sub-score (0-100)

Block Citability Score = (Answer * 0.30) + (SelfContain * 0.25) + (Structure * 0.20) + (Stats * 0.15) + (Unique * 0.10)

Step 4: Calculate Page-Level Score

1. Calculate the average of all block scores for the page-level citability score.
2. Identify the top 3 highest-scoring blocks (highlight as strengths).
3. Identify the bottom 3 lowest-scoring blocks (flag for rewriting).
4. Calculate the percentage of blocks scoring above 70 (the "citability coverage" metric).

Step 5: Generate Rewrite Suggestions

For each block scoring below 60, generate a specific rewrite suggestion:

1. Identify the primary weakness (buried answer, lack of facts, poor structure, etc.).
2. Propose a rewritten opening sentence using a definition or answer-first pattern.
3. Suggest specific statistics or facts that could be added.
4. Recommend structural improvements (add list, add table, split paragraph).

Output Format

Provenance (THL): tag the score [scan] (data fetched this run), [partial-scan], [heuristic] (judgement, no data), or [unmeasured] — and emit — instead of a number when [unmeasured] or pure [heuristic]. A number with weak provenance still reads as hard data. See the GEO Method.

Generate a file called GEO-CITABILITY-SCORE.md:

# AI Citability Analysis: [Page Title]

**URL:** [URL]
**Analysis Date:** [Date]
**Overall Citability Score: [X]/100**
**Citability Coverage:** [X]% of content blocks score above 70

---

## Score Summary

| Category | Score | Weight | Weighted |
|---|---|---|---|
| Answer Block Quality | [X]/100 | 30% | [X] |
| Passage Self-Containment | [X]/100 | 25% | [X] |
| Structural Readability | [X]/100 | 20% | [X] |
| Statistical Density | [X]/100 | 15% | [X] |
| Uniqueness & Original Data | [X]/100 | 10% | [X] |
| **Overall** | | | **[X]/100** |

---

## Strongest Content Blocks

### 1. "[Heading]" -- Score: [X]/100
> [First 2 sentences of the block]

**Why it works:** [Explanation]

### 2. "[Heading]" -- Score: [X]/100
> [First 2 sentences of the block]

**Why it works:** [Explanation]

---

## Weakest Content Blocks (Rewrite Priority)

### 1. "[Heading]" -- Score: [X]/100

**Current opening:**
> [First 2 sentences as they exist]

**Problem:** [Specific issue -- buried answer, no facts, etc.]

**Suggested rewrite:**
> [Rewritten opening 2-3 sentences with answer-first pattern and facts]

**Additional improvements:**
- [Add table comparing X, Y, Z]
- [Include statistic about ...]
- [Split long paragraph into 2-3 shorter ones]

---

## Quick Win Reformatting Recommendations

1. **[Specific recommendation]** -- Expected citability lift: +[X] points
2. **[Specific recommendation]** -- Expected citability lift: +[X] points
3. **[Specific recommendation]** -- Expected citability lift: +[X] points
4. **[Specific recommendation]** -- Expected citability lift: +[X] points
5. **[Specific recommendation]** -- Expected citability lift: +[X] points

---

## Per-Section Scores

| Section Heading | Words | Answer Quality | Self-Contained | Structure | Stats | Unique | Overall |
|---|---|---|---|---|---|---|---|
| [H2 heading] | [N] | [X] | [X] | [X] | [X] | [X] | [X] |

Reference Data

Optimal Passage Characteristics (from GEO Research)

• Optimal length for AI citation: 134-167 words (Bortolato 2025 analysis of AI Overview passages)
• Adding statistics, authoritative quotations, source citations, or fluency optimization: the GEO study (Aggarwal et al., KDD 2024 — the joint Princeton / Georgia Tech / Allen AI / IIT Delhi paper) reports 30-115% visibility improvements across these methods, varying by query category. It is one study, not three; cite it as one.

AI System Citation Preferences