AI, Neural, and Voice Search Patents Reference
120+ patents covering neural ranking, BERT/Transformers, deep learning IR, vector embeddings, LLMs, and voice/conversational search.
Neural Ranking (13 patents)
| Patent | Description |
|---|---|
| US7840569B2 | Neural network document ranking |
| US8117209B1 | Feature-based ranking (Reasonable Surfer) |
| US10229166B1 | NavBoost — implicit user behavior ranking |
| US9558233B1 | Selection quality scoring (dwell time) |
| US9031929B1 | Site quality score |
| US8661029B1 | User feedback-based ranking |
| US10296642B1 | Engagement scoring for ranking |
| US8442984B1 | Website quality signal generation |
| US9684697B1 | Click data quality signals |
| US9092510B1 | Post-click behavior (pogo-sticking) |
| US8838587B1 | Session dwell duration |
| US11132700B1 | A/B testing signals for ranking |
| US8255413B2 | Dwell time analysis |
BERT / Transformers (18 patents)
| Patent | Description |
|---|---|
| US10452978B2 | Attention mechanism — Transformers foundation |
| US20230334045A1 | BERT query integration |
| US12111859B2 | Cross-attention encoder for ranking |
| US11782998B2 | Dense retrieval with transformers |
| US20190114362A1 | Entity-based embeddings |
| US10810270B2 | Search based on emotional state + browsing history |
| US20180046834A1 | Query-entity matching |
Deep Learning IR (40+ patents)
Key patents cover: learned ranking functions, neural feature extraction, deep semantic matching, representation learning for search.
Core concepts:
- Learned ranking functions replace hand-crafted feature weights
- Neural feature extraction from raw text without manual engineering
- Deep semantic matching — meaning over keyword overlap
- Representation learning — documents and queries mapped to same vector space
Embeddings / Vector Search (41 patents)
| Patent | Description |
|---|---|
| US12099533B2 | Dense document representations |
| US11782998B2 | Vector similarity search |
| US20190114362A1 | Entity embeddings |
| US12210825B2 | Image captioning embeddings |
| US12051205B1 | Multimodal embeddings (text + image + video) |
How vector search works in Google:
- Documents and queries embedded into the same high-dimensional space
- Semantic similarity calculated via cosine distance
- Dense retrieval surfaces documents that match meaning, not just keywords
- Used in AI Overviews, People Also Ask, featured snippet selection
LLMs for Search (29 patents)
| Patent | Description |
|---|---|
| US11769017B1 | PaLM/LaMDA for generative search answers |
| US11003865B1 | RAG (Retrieval-Augmented Generation) |
| US12353469B1 | LLM output verification / source checking |
| 2024 hypergraph patent | Multi-relationship entity search |
| 2025 pairwise ranking patent | ML-based document pair scoring |
RAG pipeline (US11003865B1):
Query → Retrieval (vector + keyword) → Top-k documents → LLM synthesis → AI OverviewPages that appear in AI Overviews are cited sources in the RAG retrieval phase.
LLM Verification (US12353469B1): LLM outputs are cross-checked against source documents. Citations with verifiable, accurate claims score higher for inclusion in AI Overviews.
Voice / Conversational Search (213+ patents)
Across 9 categories:
| Category | Count | Focus |
|---|---|---|
| Voice search | 13 | Speech-to-text query processing |
| Speech recognition | 32 | Acoustic models, language models |
| Conversational AI | 24 | Dialog management, context tracking |
| Voice assistants | 26 | Command interpretation, action execution |
| NLU (Natural Language Understanding) | 31 | Intent extraction, slot filling, semantic parsing |
| Voice query context | 31 | Session context, disambiguation |
| Multimodal | 25 | Combined voice + visual + text search |
| Audio content | 9 | Audio indexing, podcast search |
| Voice navigation | 9 | Turn-by-turn, local voice queries |
SEO Implications
- Write for semantic meaning, not keywords — BERT processes full context, not isolated terms
- Clear direct answers in the first paragraph — RAG extracts these for AI Overviews
- Build entity co-occurrence — entity embeddings drive vector retrieval
- Cite verifiable sources — LLM verification patent (US12353469B1) checks citations
- Optimize for voice queries — natural language, question format, conversational tone
- Include multimodal content — images with descriptive alt-text, video with transcripts
Related Learning Modules
- Module 11: Neural AI Search — Deep neural patent coverage
- Module 12: Voice Search & Conversational SEO — All 213+ voice patents
- Module 13: User Behavior Signals — NavBoost and CTR