Gathering by Semantic Proximity, Not Textual Match

 

The decomposed questions are matched not only by word match but also by semantic proximity. Vectors are used here. Through the process of converting sentences into sequences of numbers (embedding), the proximity between the numbers on the question side and the page side is measured using cosine similarity, etc.

In simple terms, while keyword search looks at "whether the same words are written," vector search looks at "whether the same thing is being said." Therefore, even pages that do not write the words "AI search mechanism" verbatim may be candidates if they explain the process flow. Conversely, pages that just list keywords without substance do not reduce the semantic distance and are excluded from candidates.

In practice, a hybrid configuration combining keyword match methods (like BM25) and vector search is common. It is worth noting that decisions are not made based on just one of these methods.

Fragments, Not Pages, Are Extracted

The gathered information is handled not at the page level but at the level of smaller fragments. This division is called chunking. Long articles are cut into multiple chunks, and only chunks close to the question remain, from which rankings are reassigned (re-ranking).

This is a key point of understanding. In traditional search, the conclusion was "which page to show," but in AI search, the conclusion is "which fragment to use as material." Even if an article is excellent as a whole, if a single extracted paragraph does not make sense on its own, it is less likely to be chosen as material.

Answer Generation Layer: From Gathered Fragments to Constructed Answers

The narrowed-down fragments are passed to the generative AI along with the question sentence. This is where the answer generation layer begins.

Creating Sentences Based on Provided Materials

The generative AI reads the provided fragments as context and constructs an answer sentence to the question. This "searching for external information before answering" structure is known as the RAG (Retrieval-Augmented Generation) framework proposed by Lewis et al. in 2020. The generation itself is handled by large language models (LLMs) based on the Transformer by Vaswani et al. (2017).

It is important to note that this layer is not a place to introduce new facts. What the generative AI does is integrate and rephrase the provided materials. Content not included in the materials is filled in by the knowledge obtained during the model's training or by inference. The weight that influences the quality of the answer leans more towards the materials than the skill of the generation.

Linking Answers to Information Sources

The process of linking answers to real information sources is called grounding. It is a mechanism that associates each part of the answer sentence with the fragment it originates from and presents it as a citation. This association is why AI search answers are accompanied by reference links.

However, the linkage to information sources is not one-to-one. Since multiple information sources are integrated during the summarization process, the presented citation is not necessarily the sole origin of that sentence. Even if your company's site is displayed as a citation, it cannot be determined from the outside which part of the answer sentence originates from your company's description.

The Cause of Discrepancies Often Lies in the Retrieval Layer

The phenomenon of generating content that differs from the facts is called hallucination. While often discussed as a performance issue of generative AI, following the process flow reveals that the cause often lies in the retrieval layer. Relevant information may not be indexed, only outdated descriptions may be picked up, or fragments may be cut at incomplete positions, missing conditional parts. In such states, no matter how carefully the generation layer constructs the answer, it will collapse, leading to the perception of lying.

If you have confirmed that your company's information is incorrectly described on AI search, refer to the related article 'Risk Management and Practical Avoidance Techniques to Prevent Misinformation with AI Measures' for a breakdown of the causes and countermeasures.

Parts Overlapping with Traditional Search Engines and Parts That Have Changed

AI search has not replaced traditional search engines but has layered on top of them. The foundation of crawling and indexing is shared, and pages not indexed are not handled by AI search either.

What has changed is the evaluation unit and the user's endpoint. Previously, the endpoint was high ranking, but in AI search, the endpoint is having your company's description included in the answer sentence. These two do not correlate. Pages that rank high but are not cited and pages that are repeatedly cited despite low rankings both exist. The divergence between high ranking and citation is a phenomenon unique to AI search.

Different Depths of Processing in AI Overviews and AI Mode

In the context of Google search, AI Overviews (formerly SGE) and AI Mode are often confused. The former is a summary displayed at the top of regular search results, while the latter is a search mode based on dialogue. Both use Gemini, but the depth of question handling differs.

AI Overviews, being positioned as a summary of search results, strongly ties with the regular search index. AI Mode, through the aforementioned QFO (Query Fan-Out), expands more sub-questions and responds to multi-turn additional questions based on conversation history. Even with the same term "AI search," the former is closer to an extension of existing SERP, while the latter leans towards conversational search.

It is also worth recognizing that the internal processing of each service is not fully disclosed. What can be confirmed from public information is up to the framework, and the detailed behavior is supplemented by observation within the company.

Sources Are Narrowed Down Through Four Gateways

 

There is not just one "condition for being cited." Following the process flow, gateways to pass through are placed in stages, and the symptoms that appear differ depending on where you fall.