Grounding Citation Analysis

Reverse engineering Google AI Mode's sentence-level citation behaviour using #:~:text= URL fragments.

---

The Core Insight

Every Google AI Mode and Gemini citation URL contains a hidden Web Text Fragment anchor:

https://example.com/page#:~:text=Exact%20sentence%20Google%20cited%20here

Decode it and you know exactly which sentence Google extracted from the source page — no guesswork. This is the first reproducible study to exploit this at scale.

---

What This Repo Does

flowchart LR
    A[queries.csv 100 queries] --> B[Bright Data AI Mode Scraper]
    B --> C[Raw citation JSON with #:~:text= URLs]
    C --> D[Text fragment parser]
    D --> E[Cited sentence per citation row]
    E --> F[Source page scraper]
    F --> G[Positional analysis]
    G --> H[Statistical analysis + Charts]

Loading

1. Collect — Use Bright Data's Google AI Mode Scraper (real SERP, not API) to gather citation URLs for 100 queries
2. Parse — Decode #:~:text= fragments to extract the exact cited sentence from every citation URL
3. Scrape — Fetch source pages and locate each cited sentence within the document (positional analysis)
4. Analyse — Statistical tests for positional bias, sentence length preferences, structured content advantage, platform divergence
5. Visualise — Publication-quality charts for the accompanying article

---

Quick Start

Prerequisites

• Python 3.11+
• A Bright Data account with API key
• ~$6–25 budget for data collection (1,000 queries + source pages ≈ $25 max)

Setup

git clone https://github.com/yourusername/grounding-citation-analysis
cd grounding-citation-analysis

# Install dependencies
pip install -r requirements.txt

# Configure credentials
cp .env.example .env
# Edit .env and set: BRIGHTDATA_API_KEY=your_key_here

Run the Full Pipeline

# Step 1: Collect AI Mode citations
python scripts/01_collect_ai_mode.py --limit 20  # start small to verify

# Step 2: (Optional) Collect Gemini citations
# Set BRIGHTDATA_GEMINI_DATASET_ID in .env first
python scripts/02_collect_gemini.py --limit 20

# Step 3: Parse #:~:text= fragments → citations.csv
python scripts/03_parse_text_fragments.py

# Step 4: Scrape source pages (positional analysis)
python scripts/04_scrape_source_pages.py --limit 100  # start small

# Step 5: Statistical analysis
python scripts/05_analyze_patterns.py

# Step 6: Generate charts
python scripts/06_generate_charts.py

---

Project Structure

grounding-citation-analysis/
├── .env                          # Your Bright Data API key (not committed)
├── .gitignore
├── requirements.txt
├── README.md
│
├── queries/
│   └── queries.csv               # 100 queries across 12 categories
│
├── scripts/
│   ├── 01_collect_ai_mode.py     # Bright Data AI Mode Scraper trigger + poll
│   ├── 02_collect_gemini.py      # Bright Data Gemini Scraper trigger + poll
│   ├── 03_parse_text_fragments.py # #:~:text= URL decoder → citations.csv
│   ├── 04_scrape_source_pages.py  # Source page fetcher + positional analysis
│   ├── 05_analyze_patterns.py    # Statistical analysis (scipy)
│   └── 06_generate_charts.py     # Matplotlib/seaborn visualisations
│
├── data/
│   ├── raw/                      # Raw Bright Data JSON snapshots
│   ├── parsed/
│   │   ├── citations.csv         # One row per citation, includes cited_sentence
│   │   ├── answers.csv           # One row per query/answer
│   │   └── source_pages.csv      # Source page positional data
│   └── analysis/
│       ├── summary_stats.json
│       ├── positional_distribution.csv
│       ├── domain_frequency.csv
│       ├── platform_overlap.csv
│       └── category_breakdown.csv
│
├── notebooks/
│   ├── 01_methodology.ipynb
│   └── 02_findings.ipynb
│
├── reports/                      # Generated charts (PNG)
│
└── article/
    ├── article.md                # Full LinkedIn article
    └── diagrams/                 # Mermaid source files

---

Key Data Schema

citations.csv (primary output)

Column	Description
platform	ai_mode or gemini
query	The search query
citation_url_raw	Full URL including #:~:text= fragment
citation_url_clean	URL without fragment
domain	Cited domain
has_text_fragment	Boolean — does URL contain #:~:text=?
cited_sentence	Decoded sentence Google cited
cited_sentence_word_count	Word count of cited sentence
fragment_raw	Raw (URL-encoded) fragment
cited_flag	Boolean from Bright Data — marked as citation

source_pages.csv (after step 04)

Column	Description
found	Boolean — cited sentence located in page
block_index	Element index where sentence was found
block_total	Total elements in page
relative_position	block_index / block_total (0 = top, 1 = bottom)
page_word_count	Total word count of source page
has_structured_content	Page contains <ul>, <ol>, or <table>

---

Research Questions & Hypotheses

#	Hypothesis	Test
H1	Cited sentences cluster in the top 30% of documents	One-sample t-test, mean position < 0.5
H2	Cited sentences are shorter than average page text	Descriptive statistics, histogram
H3	Structured pages (lists/tables) are cited more	Chi-square test
H4	AI Mode and Gemini cite overlapping but distinct URLs	Jaccard similarity
H5	Sentence length varies by query category	One-way ANOVA

---

Comparison to Prior Research

Study	Sample	Granularity	Method	This Study's Advance
Ahrefs (2024)	1.9M citations	Page level	Custom crawler	Sentence-level decoding
Surfer SEO (2024)	46M citations	Domain level	API collection	Real SERP; positional data
Seer Interactive (2024)	Variable	Query level	Gemini API	Within-page position
DEJAN AI (2024)	Conceptual	Theoretical	Manual + LLM	Empirical validation
This study	100+ queries	Sentence level	Bright Data + #:~:text=	—

---

Bright Data API Reference

AI Mode Scraper

• Dataset ID: gd_mcswdt6z2elth3zqr2
• Endpoint: POST https://api.brightdata.com/datasets/v3/trigger?dataset_id=gd_mcswdt6z2elth3zqr2
• Input: [{"url": "https://www.google.com/search?udm=50", "prompt": "...", "country": "US"}]
• Output includes citations[].url with #:~:text= fragments

Why not the Gemini API? Surfer SEO's research found that the Gemini developer API returns different answers from the actual Google Search AI Mode SERP. Bright Data's scraper hits the real google.com/search?udm=50 endpoint via residential proxies.

---

Licence

MIT — see LICENSE. Cite this repo if you use the methodology.

---

Author

[Your Name] | [Your LinkedIn] | [Your Website]

Accompanying article: How Google Actually Chooses Which Sentences to Cite in AI Mode