🔬 What happens when you analyze leaked search engine source code and build an SEO tool from it?

You discover the entire industry has been optimizing wrong for years.

Ryan Jones (Head of SEO at Razorfish) joined me on Marketing Speak to share what he learned from the Yandex and Google leaks. The truth is uncomfortable: while SEOs count keywords and stack H2 tags, search engines are plotting documents in vector space using machine learning.

Here's what actually matters in 2025:

📊 Complete topic relevance (not single keyword optimization)

🎯 Semantic similarity scores using cosine distance

🤖 BERT-extracted keywords from current top-rankers

📱 Intent matching using ML models (not keyword pattern matching)

🔗 Third-party consensus for knowledge panels

📄 Research papers that still accurately describe how AI search works

Ryan built SERPrecon to measure content the way search engines actually measure it. The tool uses Yandex's exact-title-tag scoring logic. It applies Passage BERT to predict AI Overview citations. It extracts entities and keywords using the same open-source algorithms Google uses.

Ryan reverse-engineers which phrases AI Overviews will cite based on Google's patent, then writes content incorporating those exact phrases. Result? He consistently "steals" citations from competitors.

This isn't theoretical. This is production-level SEO based on actual source code.

The show notes, including the transcript and checklist to this episode, are at marketingspeak.com/534.