The Bottom Line
- Google is obsessed with efficiency. Their patents show 200x computational savings. They're optimizing for billions of queries at pennies each.
- OpenAI is building agent armies. Not single AI assistants—teams of agents that negotiate with each other on your behalf.
- Your website might become an API. OpenAI's patents describe AI agents that can call your website directly. No human clicking required.
- Multi-day AI tasks are coming. Patents describe systems where you submit a task and check back in a week.
Why We Read the Patents
Everyone reads the blog posts. Everyone watches the keynotes. But blog posts are marketing. Keynotes are theater. If you want to know what a company is actually building—not what they want you to believe—you read the patents.
Patents are legally binding. They're technical. They're boring. Which is exactly why they're useful: companies can't bullshit in a patent filing.
We pulled 13 recent patent applications from Google and OpenAI. Not the entire USPTO database—just the ones filed in the last year that relate to AI search, agents, and retrieval. What we found is... not what the conference circuit is talking about.
The Numbers
First, a reality check on scale. Google has filed 46,384 patents with the USPTO. OpenAI has filed 71. Google is a patent machine; OpenAI is a startup. But that makes OpenAI's patents more interesting—every filing is a strategic choice, not bureaucratic routine.
| Metric | Value |
|---|---|
| Patents we analyzed | 13 |
| Google patents (from 46,384 total) | 6 |
| OpenAI patents (from 71 total) | 7 |
What Google Is Building
Google's patents reveal a company obsessed with one thing: doing more with less. Every filing is about efficiency—serving billions of queries without going bankrupt on compute.
Matryoshka Embeddings
US-20250384346-A1
This one's clever. Named after Russian nesting dolls, these are embeddings that contain smaller embeddings inside them. A 1024-dimensional embedding can be "sliced" down to 512, 256, or 128 dimensions—each one still usable, just less precise.
"Enabling efficient retrieval at various computational budget levels."
Why this matters: Google can now dynamically choose how much compute to spend per query. Simple query? Use the cheap 128-dimensional version. Complex query? Spring for the full 1024. Multiply that by billions of queries and you're talking real money.
200x Cheaper Object Recognition
US-20250384650-A1
Google's patent for "open-vocabulary object detection" buries the lede in the technical details:
"Approximately 200x computational savings versus competing methods."
Two hundred times cheaper. That's not incremental improvement—that's a different game. It means multimodal search (images, video, diagrams) goes from "expensive luxury" to "standard feature." If you've been ignoring visual content because "text is what matters for SEO," this patent suggests you're about to be wrong.
Query Routing
US-20250370993-A1
Here's something the SEO community doesn't talk about enough: there is no single "Google search." There are many systems, and a router decides which one handles your query.
"The system uses a trained model to assess sampling ratios and information loss to determine which source will produce superior accuracy within acceptable latency constraints."
Translation: Google has multiple search systems optimized for different things, and an AI picks the best one per query. Your content might rank differently depending on which system handles it. This explains some of the "ranking volatility" that SEOs obsess over—it's not instability, it's routing.
Privacy-Preserving Embeddings
US-20250384079-A1
This one's about sharing search data without exposing it.
"Embeddings can be shared with others via digital keys, allowing limited access to search for that person in media collections without exposing the actual embedding data."
Think about what this enables: personalized search without Google seeing your personal data. You hold a key that unlocks your preferences, and you can revoke it anytime. It's Google's answer to "personalization vs. privacy"—have both.
AI That Knows When to Shut Up
US-20250384215-A1
This patent describes AI for group conversations:
"A two-stage ML model approach where a first model evaluates whether user input targets another participant or the assistant, then a second model generates contextual responses only when appropriate."
Current AI assistants respond to everything. This patent describes AI that listens to group chats and only speaks when addressed. It knows when to shut up. If Google Workspace integrates this, expect AI assistants that participate naturally in team conversations instead of hijacking them.
Strategic Example Repetition
US-20250384666-A1
"Strategically repeats difficult demonstration examples within the context input. Simply adding more demonstration examples doesn't guarantee improved performance."
This is a prompt engineering patent. The insight: showing AI more examples doesn't always help. What helps is strategically repeating the hard examples. If you're building AI applications, this is useful. If you're creating content, the implication is that AI systems are getting smarter about learning from examples—your best content matters more than your quantity of content.
What OpenAI Is Building
If Google's patents are about efficiency, OpenAI's are about autonomy. They're building AI that doesn't just answer questions—it runs errands. For days. With friends.
Multi-Agent Workspaces
US-20250377934-A1
This is the weird one. OpenAI is patenting systems where multiple AI agents work together in a shared workspace:
"Agents have access to a common workspace and can view the state of a workspace. Agents autonomously decide whether to 'yield' (defer to others) or 'act' (respond) to workspace updates."
Think about what this means. Not one AI assistant. A team of specialized agents—one for research, one for writing, one for fact-checking—working together and deciding amongst themselves who should handle what. It's AI project management.
"Have Your People Call My People"
US-20250373574-A1
This patent literally uses that phrase. It describes AI agents representing different users and negotiating with each other:
"AI agent instances interacting with each other on behalf of different user accounts. Recipient users can block unwanted agent communications through blocklists."
Your AI schedules a meeting with my AI. Your AI negotiates a price with my AI. Your AI blocks spam from other AIs. We're entering a world where AIs transact on our behalf, and OpenAI is building the consent framework for it.
Multi-Day Tasks
US-20250378278-A1
Current AI is synchronous: you ask, it answers, done. This patent describes asynchronous AI:
"Users can submit follow-up prompts to modify, clarify, or add to the original request without waiting for completion. Particularly useful for multi-hour or multi-day computational tasks."
You submit a task. Check back tomorrow. Add a clarification. Check back in a week. The AI works on your project the way a contractor would—asynchronously, with check-ins. This is fundamentally different from chat.
Your Website as an API
US-20250110811-A1
This one matters for anyone with a website:
"Publishers can host a manifest file and specification document at their own domains. The system automatically discovers and integrates these APIs, enabling it to intelligently call external endpoints without requiring retraining."
Translation: you publish a structured description of what your website can do (a manifest file), and AI agents can call your site directly. Not "visit your website and click around"—call it programmatically. If you run an e-commerce site, AI could add items to a cart. If you run a booking service, AI could make reservations. No human clicking required.
The Personalization Notepad
US-20250200361-A1
OpenAI is patenting AI that learns about you continuously:
"Learn such facts, preferences, or contexts from conversational prompts." Stored in a "personalization notepad" for "generating tailored responses across sessions." "Users can request deletion of learned information."
Every conversation teaches the AI more about you. It remembers that you prefer concise answers. That you're vegan. That you hate sports metaphors. But you can delete anything it learned. The consent model is baked in.
Context-Aware Preferences
US-20250384222-A1
"Structured storage of user preferences in session-specific caches. Dynamic integration of instructions into system prompts based on trigger events."
This is subtler. AI that knows when to apply your preferences. You want professional tone at work, casual tone with friends. You want verbose explanations when learning, terse answers when in a hurry. The AI reads the context and adapts.
Learning from Relationships
US-20240370779-A1
"Leverages naturally-occurring pairs from unlabeled data (adjacent text, docstrings paired with code implementations)."
This is a training patent—how to make AI understand content better. The insight: AI learns from relationships in your content. Code next to its documentation. Headings next to their content. Your site structure teaches AI what things mean.
Two Very Different Visions
What's striking about these patents is how different Google and OpenAI's approaches are. They're not building the same thing.
Google is building a search engine that's 200x cheaper to run. Efficiency patents. Compression patents. Routing patents. They're optimizing for billions of queries at pennies each. They're playing defense—making sure AI search doesn't bankrupt them.
OpenAI is building AI employees. Agent patents. Task patents. Negotiation patents. They're not optimizing search—they're replacing it. Why search for a restaurant when your AI agent can negotiate a reservation with the restaurant's AI agent?
These are different bets. Google is betting AI is a feature of search. OpenAI is betting AI replaces search.
What This Means for You
If you create content or run a website, here's what these patents suggest:
Multimodal content matters now. Google's 200x efficiency gains in image recognition mean visual content is about to be as searchable as text. Stop treating images as decoration.
Consider making your site AI-callable. OpenAI's API manifest patent means websites can expose functionality to AI agents. If you sell something, an AI might want to buy it without a human ever visiting your site. Is your site ready for that?
Structure teaches meaning. Patents on contrastive learning show that AI understands content through its relationships—code next to documentation, headings next to paragraphs, questions next to answers. Well-structured content isn't just easier to read; it's easier for AI to understand.
The conversation is shifting to agents. Today, you optimize for "getting found in search." Tomorrow, you might optimize for "being useful to AI agents." Those are different problems.
Sources
All patents fetched via USPTO API, December 2025. Full patent documents available at patents.google.com.
Track Your AI Search Visibility
See where your brand appears across ChatGPT, Perplexity, and Google AI Overviews.