🧠 LLM Optimization Analyzer for Screaming Frog (with Gemini API)

A custom JavaScript-based analyzer for Screaming Frog that evaluates webpages for Large Language Model Optimization (LLMO) using the Gemini 1.5 Flash API. The script is using a Batched Pairwise approach.

Inspired by recent research:

• Batched Self-Consistency Improves LLM Relevance Assessment and Ranking
• C‑SEO Bench: Does Conversational SEO Work?

---

🚀 What It Does

This script performs a passage-level, structured audit of your pages using LLM analysis principles and Screaming Frog’s Custom JavaScript Engine.

It extracts content like:

• <title>, meta descriptions
• H1–H2 headers and surrounding text
• Primary paragraphs and list items
• JSON-LD structured data (schema.org)
• FAQ patterns and semantic blocks

Then, it generates a comprehensive prompt for Gemini to:

1. Identify target queries
2. Score the content per query (0–5)
3. Highlight strongest passages
4. Reveal content gaps
5. Suggest LLM optimization actions

---

🔧 How to Use in Screaming Frog

✅ JS Rendering must be enabled in Screaming Frog!

Step-by-step:

1. Open Screaming Frog
2. Go to Configuration → Custom → Custom JavaScript
3. Paste the full llmo-analyzer.js script
4. Enable JavaScript Rendering under Configuration → Spider → Rendering
5. Run your crawl
6. View Custom Javascript output tab for Gemini-powered LLMO insights

---

📊 Output Example

OVERALL LLMO SCORE: 4.2/5 TOP 3 POTENTIAL: yes RANKING POTENTIAL: high

TARGET QUERIES & SCORES:

1. "wireless gaming keyboard" - Score: 4.5/5 (strong)
2. "gaming keyboard mouse combo" - Score: 4.2/5 (strong)

CONTENT GAPS:

• No pricing details
• Missing comparison to competitors

RECOMMENDATIONS:

• Add structured FAQ section
• Include product comparison table

---

📐 How It Works (Brief)

• Simulates Batched Pointwise (PW) evaluation from research — passing up to 100 key passages in one structured Gemini prompt.
• Weighted scoring reflects traditional SEO prioritization (e.g., <title> > <h1> > <p>).
• Results reflect Gemini’s understanding of how LLMs rank pages semantically.

---

⚠️ Limitations

• Gemini API does not support native multi-document scoring (simulated via prompt structure)
• No true self-consistency (one-shot only)
• Limited to 4096 tokens per page
• Domain-specific variance (best for structured pages like ecommerce/docs)
• Model-specific scores (Gemini 1.5 Flash)

---

📘 Research Backing

This script is modeled after "mentioned" benchmarks:

• +7.5% NDCG@10 gain from batched vs. pointwise LLM scoring (GPT-4o) // Legal batch
• 2.77x boost in visibility when placed 1st in LLM context window (C‑SEO Bench) // Retail domains!

---

📄 License

MIT — use freely with attribution if publishing derivative work.

---

🙋‍♀️ Need Help?

Open an issue or contact me to implement this across large sites or integrate it with live LLM APIs.

---