If the patents from the early 2000s represent the foundational blueprint of Google’s semantic search engine, then modern language models like BERT, MuM, and LaMDA are the state-of-the-art engine upgrades.
Many in the SEO community view these models as revolutionary replacements for everything that came before.
But as we conclude our deep dive into Koray Gübür’s presentation, “Semantic Search Engine & Query Parsing” we see a different truth.
These technologies aren’t a new blueprint; they are incredibly powerful components designed to execute the original blueprint with unprecedented speed and nuance.
They are the engine, not the car.
BERT: Understanding Language in both Directions
Announced in 2018, BERT (Bidirectional Encoder Representations from Transformers) was a quantum leap in natural language understanding.
Before BERT, models typically read a sentence from left to right or right to left. BERT reads the entire sentence at once, from both directions.
BERT is designed to “pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers.”
How it connects to the blueprint:
Remember the “Multi-Stage Query Processing” patent from 2004 that tried to understand words by looking at synonyms and co-occurrence?
BERT does the same thing, but on a massive scale.
It understands that the meaning of the word “bank” in “I sat on the river bank” is defined by the words “river” and “sat” that surround it.
It masters context.
This is a supercharged version of the context-building principles we saw in earlier patents.
MuM: The Multitasker and Context Consolidator
Announced in 2021, MuM (Multitask Unified Model) is Google’s answer to complex, multi-step queries.
The classic example Google gives is: “I’ve hiked Mt. Adams and now want to hike Mt. Fuji next fall, what should I do differently to prepare?”
A simple search engine sees keywords. MuM sees concepts, entities, and tasks.
It understands three different contexts at play:
-
Trekking (the activity)
-
Mountain (the general entity type)
-
Specific Mountain Trekking (comparing the attributes of two specific entities: Mt. Adams and Mt. Fuji)
How it connects to the blueprint:
MuM is the ultimate application of Context Vectors and Entity Reconciliation.
It can access and synthesize information from multiple contextual domains (geography, climate, fitness, gear) and reconcile the attributes of two distinct entities to provide a comparative answer.
It’s also multi-modal, meaning it can understand text, images, and video to build a complete picture.
This is the “structured database” envisioned by Sergey Brin’s 1999 patent, now being queried in real-time by an incredibly powerful AI.
LaMDA: The Conversational specialist
LaMDA (Language Model for Dialogue Applications) is designed to make search more conversational and intuitive.
It’s about connecting one question to another in a “Human Sensible Way” (Slide 77).
LaMDA is judged on its ability to be:
-
Sensible: Does the response make sense in context?
-
Specific: Is it detailed and not overly general?
-
Interesting: Is it insightful or unexpected?
-
Factual: Is it grounded in truth?
How it connects to the blueprint:
LaMDA is the evolution of “Midpage Query Refinements” and “Query Expansion.”
The 2003 patent aimed to turn a single search into a multi-step refinement process.
LaMDA takes this to its logical conclusion, enabling a fluid, back-and-forth dialogue where each query builds upon the last.
It uses the semantic clusters and context from the previous turn in the conversation to inform the next, creating a truly interactive search journey.
What does this mean for you
These powerful new models don’t invalidate the old principles; they amplify them.
-
Context is More Important Than Ever: BERT’s bidirectional nature means it can spot shallow, out-of-context keyword stuffing from a mile away. Your content must be semantically rich and logically structured.
-
Answer Complex, Multi-Part Questions: MuM is designed for users who are planning, comparing, and synthesizing. Create content that goes beyond simple definitions and provides a comprehensive, comparative analysis. Think like a user planning a complex task and answer all the sequential questions they might have.
-
Structure Content for Conversation: LaMDA signals a future where search is a dialogue. Organize your content with clear, logical flows. Use headings, lists, and FAQs to break down complex topics into digestible pieces that can be served up conversationally. A single H2 could answer one turn in a conversation.
The takeaway is clear: Google is not throwing out its old playbook.
It’s equipping it with an AI-powered engine that can execute its long-standing mission: to organize the world’s information and make it universally accessible and useful, at a level of depth and understanding that was once pure science fiction.
In my final article on this Déjà Vu SEO series, I will tie everything together, summarizing the key lessons from this journey through Google’s foundational patents and outlining what a truly resilient, future-proof SEO strategy looks like in practice.