This explains why ranking highly for a single head term no longer guarantees visibility. A page that ranks first for “sustainable packaging solutions” might not appear for any of the synthetic queries the system actually uses. Meanwhile, a page ranking lower for the main term but performing well across multiple sub-queries has many more opportunities to be selected for the final response.

In effect, query fan-out builds on the process of latent intent projection, mapping a query into related meanings and expanding it into neighboring concepts. The retrieved passages from those expansions form a temporary custom corpus, and because every selection is probabilistic, the retrieval paths remain non-deterministic.

As a result, the competition is no longer just for “the ranking,” but for being part of the constellation of content the system may draw from when it breaks a user’s question in many possible ways

1. Query Reception and Context Retrieval The process begins with receiving a user’s query, which can be typed, spoken, image-based, or multimodal. The input is processed, based on type, including ML models applied to convert non-text input (e.g. images) to machine-readable formats (e.g. for images – captioning, object detection, or semantically rich embeddings)
2. User State Retrieval The system immediately retrieves and aggregates contextual information about the user and their device, forming a “user state”. This includes prior queries, data from previous search result pages (SRPs) and documents (SRDs), contextual user signals (including synced schedules, activity, location, and active applications), as well as stored user attributes and preferences (e.g. dietary restrictions, media preferences). This user state is continuously updated and can be stored as an aggregate embedding.
3. Semantic Fingerprinting (User Embeddings): This contextual information is converted into semantically-rich embeddings that represent the user’s “semantic fingerprint”. This allows for modular personalization, meaning two users asking the same query may receive different answers based on their individual profile alignment and semantic relevance
4. Synthetic Query Generation (Query Fan-out) Leveraging Large Language Models (LLMs), the system expands the initial query into a multitude of synthetic queries. This query fan-out mechanism allows the search engine to research deeper into content beyond the literal terms of the original query. Some of these might be: 
   • Alternative formulations: Synthetic queries like follow-up questions, rewritten versions, and “drill-down” queries, created in real-time based on the original query and contextual information.
   • Entity-based Reformulations: LLMs crosswalk entity references to broader or narrower equivalents using Knowledge Graph anchors. For example, “SUV” could be expanded to specific models like “Model Y” or “Volkswagen ID.4”. This directly incorporates the role of entities and knowledge graphs in enriching query understanding.
   • Intent Diversity and Lexical Variation: The prompt-based query generation emphasizes intent diversity (e.g., comparative, exploratory), lexical variation (synonyms, paraphrasing), and entity-based reformulations.
   • Deep Search: Google’s “Deep Search” capability can issue hundreds of these synthetic queries and reason across disparate sources to generate expert-level summaries.
5. Document Selection and Custom Corpus Creation The generated synthetic queries are then used by the search system to retrieve relevant documents. The selection of these documents forms a custom corpus, which is responsive to both the original query and the expanded synthetic queries. Ranking for inclusion in generative answers increasingly depends on language model reasoning, rather than solely on static scoring functions like TF-IDF or BM25. Dual encoder models may be used for efficient document retrieval.
6. Query Classification and Downstream LLM Selection The system processes the combined data (query, context, synthetic queries, selected documents) to classify the query into specific categories. Examples of these categories include: “needs creative text generation,” “needs creative media generation,” “can benefit from ambient generative summarization,” “can benefit from SRP summarization,” “would benefit from suggested next step query,” “needs clarification,” or “do not interfere”. This entity detection or classification helps stabilize the meaning of ambiguous terms, for example, distinguishing “Jordan sneakers” from “travel Jordan” by recognizing the entity type.
7. LLM Orchestration: Based on this classification, specialized “downstream LLMs” are orchestrated by the system for processing, each trained for a particular response type (e.g., a creative text LLM, an ambient generative summarization LLM, a clarification LLM). 
8. Multi-Stage LLM Processing and Synthesis (Reasoning) Once the custom corpus is assembled, the selected downstream LLMs process the data and generate the final natural language (NL) response
   • Reasoning Chains: AI Mode leverages “reasoning chains,” which are structured sequences of intermediate inferences connecting user queries to responses logically. Content needs to be granularly useful and align with each logical inference to be selected for these reasoning steps.
   • Grounded Generation: The generation process involves extracting chunks from relevant documents, building structured representations, and synthesizing a coherent answer62. This process includes grounding, recitation, and attribute checking from the source documents themselves to improve factuality and keep names, specs, and relationships straight.
   • Multimodal Output: Responses can be multimodal, drawing from text, video, audio, imagery, and dynamic visualizations. The system can transcribe videos, extract claims from podcasts, interpret diagrams, and remix them into new outputs like lists or visual presentations.
   • Personalised Summarisation: The NL-based summary is more likely to resonate with the user and omit content they are already familiar with, based on their user state.
9. Source Citation and Linkification To ensure accuracy and transparency, relevant portions of the AI-generated natural language summaries are linkified to their source documents. The process of linkification involves comparing the semantic embeddings of the AI-generated text with those of potential source documents to verify verifiability and closeness of content, where sources are benchmarked and excluded from citing if not sufficiently close. Links can be made to sections (passages or sentences) or to entire documents. 
10. Personalized and Multimodal Output The final output, delivered at the client device, is highly personalized due to the continuous updating of the user state. Responses can be multimodal, including text, images, 3D models, animations, and audio. The system can even omit content the user is already familiar with to make the response more efficient.