An Idiots proof of the new knowledge graph · AEO

An Idiots proof of the new knowledge graph

I compared, a google brand search of a person, in Google, Google AI Mode, and then Gemni... I was trying to prove to a "potential client" that I wasn't full of crap and Google was rebuilding its knowledge graph for the new AI world... and he wasn't in it! LOL. Heres a few thoughts Gemni gave me (FYI: this will be a lot to read ) [ prompt for Gemni: I did a search for elizabeth lyons , google does not have a knowledge panel, google shows multiple people, google ai mode shows others and you gemni shows less - is this a relevant observatoin of google redoing its knowledge graph, and moving beyond what was indexed and used for the LLM in the past. ] ///. first reply [System LayerObservation in PracticeExplanation of Phenomenon1. Classic Search Index"Shows multiple people," "no knowledge panel."This is the result of an Entity Disambiguation challenge. When a common name has multiple distinct, globally relevant people (e.g., an author, an academic, a politician), the classic search algorithm recognizes that multiple entities exist but cannot confidently resolve the query to a single, canonical one. Without that high-confidence resolution, the system often defaults to showing the "messy" organic results, and an authoritative, single Knowledge Panel is suppressed.2. Google AI Mode (AI Overviews)"Shows others."AI Mode (powered by customized Gemini models) uses Retrieval-Augmented Generation (RAG). It performs a deeper, semantic search (a "query fan-out") across subtopics and data sources, including the web and the Knowledge Graph. It then synthesizes an answer. Since it is designed for exploratory and complex questions, it might surface a different "Elizabeth Lyons" that the AI model deems most relevant to its synthesized overview, potentially drawing from high-authority sources that the classic panel failed to consolidate.3. Gemini (The LLM)"Shows less."As a large language model, Gemini's responses are often driven by a highly selective and curated knowledge base to ensure factual precision. For entities with common names, the model tends to gravitate toward the one or two most authoritative or well-referenced entities it has high confidence in. By focusing on selectivity over breadth, it appears to "show less" than the raw, noisy index of the classic search results.] [ next reply That is an extremely relevant observation and it beautifully captures the current state of flux in Google's search technology, confirming your hypothesis about the evolution of its knowledge systems.

What you observed with the varying results for a common name like "Elizabeth Lyons" is a direct, real-world manifestation of the three different layers of Google's knowledge delivery system working (and sometimes competing) simultaneously:

System LayerObservation in PracticeExplanation of Phenomenon 1. Classic Search Index"Shows multiple people," "no knowledge panel."This is the result of an Entity Disambiguation challenge. When a common name has multiple distinct, globally relevant people (e.g., an author, an academic, a politician), the classic search algorithm recognizes that multiple entities exist but cannot confidently resolve the query to a single, canonical one. Without that high-confidence resolution, the system often defaults to showing the "messy" organic results, and an authoritative, single Knowledge Panel is suppressed. 2. Google AI Mode (AI Overviews)"Shows others."AI Mode (powered by customized Gemini models) uses Retrieval-Augmented Generation (RAG). It performs a deeper, semantic search (a "query fan-out") across subtopics and data sources, including the web and the Knowledge Graph. It then synthesizes an answer. Since it is designed for exploratory and complex questions, it might surface a different "Elizabeth Lyons" that the AI model deems most relevant to its synthesized overview, potentially drawing from high-authority sources that the classic panel failed to consolidate. 3. Gemini (The LLM)"Shows less."As a large language model, Gemini's responses are often driven by a highly selective and curated knowledge base to ensure factual precision. For entities with common names, the model tends to gravitate toward the one or two most authoritative or well-referenced entities it has high confidence in. By focusing on selectivity over breadth, it appears to "show less" than the raw, noisy index of the classic search results.

Conclusion: Your Observation is a Key Insight into the Knowledge Graph's Evolution

You are correct. This observation is a tangible sign of Google actively redoing (or evolving) its Knowledge Graph and moving beyond the index of the past.

The shift you described is the core concept behind Answer Engine Optimization (AEO)—the move from:

"10 Blue Links" (Classic Search, showing you a list of where to look)
To "One Direct Answer" (AI Overviews/Gemini, synthesizing the answer for you)

When you search for "Elizabeth Lyons," the system is grappling with the fundamental problem of identity resolution, and the different interfaces reveal how each component is attempting to solve it:

The Knowledge Graph is the Foundation: The KG is the database of facts and entities. For "Elizabeth Lyons," the KG contains multiple nodes (multiple people).
LLMs are the Reasoning Layer: The AI systems (AI Mode, Gemini) are now acting as a layer on top of the KG and the search index. They apply semantic reasoning and RAG to try and choose the most contextually relevant entity, which is why they sometimes diverge from the classic results.

The fact that the results differ dramatically between the three interfaces illustrates that the transition is still happening, and the AI components are not always perfectly aligned with the classic search results, which is a common finding in recent studies about LLM-SERP alignment.

3 comments