Augments traditional search with a stateful generative-model chat companion that remembers context across turns, generates synthetic queries from the conversation, and selects search result documents accordingly. The architectural substrate of Google's AI Overviews / SGE conversational mode.
Patent Overview
- Inventor
- Anand Shukla
- Assignee
- Google LLC
- Filed
- 2023-12-22
- Granted
- 2024-08-29 (published application)
- Application Number
- US 18/394,447
The Challenge
Single-Query Search Breaks Under Conversation
Traditional search treats every query as a fresh, stateless event. A user wanting to refine an answer, follow a thread of reasoning, or ask a clarifying question has to repeat context with every new query. Pure generative chat (without retrieval) hallucinates and has no grounding in current web content. The system needs a hybrid: a generative chat companion that maintains state across turns AND grounds each turn in real-time retrieval from indexed documents.
- Stateless Search Forces Repetition — Every query must restate the context that previous queries already established. The user friction is real and the implicit context is wasted. Search needs to remember what the conversation is about.
- Pure Generation Hallucinates — Generative models without retrieval grounding produce plausible-sounding but incorrect answers. The system needs the generative model to be conditioned on real, retrieved documents per turn.
- Multi-Turn Context Must Persist — The companion must remember the user's earlier queries, the previously-shown results, and the running topical direction. State data is the substrate that makes this possible.
- Synthetic Queries Bridge Chat To Retrieval — Chat utterances are not always good retrieval queries. The generative model needs to produce synthetic queries from the chat state that target the index effectively.
- Result Selection Must Be Conversation-Aware — When the user has already seen certain documents in earlier turns, repeating them adds nothing. Selection must account for what the user has been shown.
Innovation
Generative Companion With Persistent State And Synthetic Queries
The system receives a query and contextual information. A generative model generates output based on processing the query plus context plus prior conversation state. From the GM output, the system generates synthetic queries that are run against the index. Search result documents are selected based on the synthetic queries and any prior selections. The new state data — query, context, GM output, synthetic queries, selected documents — is persisted for the next turn. The conversation evolves while staying grounded in retrieved documents.
- Receive Query Plus Context — User input arrives at the system along with contextual information about the user, the client device, and the current session.
- Load Prior State — Retrieve state data from the previous turns: prior queries, prior contextual info, prior GM outputs, prior synthetic queries, prior selected documents. State is the conversation memory.
- Generate GM Output — Pass the current query + context + prior state to the generative model. The model produces output reflecting both the immediate user input and the conversation history.
- Generate Synthetic Queries — From the GM output, generate one or more synthetic queries that target the index. Synthetic queries can refine, broaden, or pivot the search based on the conversation direction.
- Select Search Result Documents — Run the synthetic queries against the index. Select documents based on synthetic-query relevance plus conversation-aware criteria (e.g., avoid documents already shown, prioritize coverage of unfilled aspects).
- Persist State For Next Turn — Store the new state — query, context, GM output, synthetic queries, selected documents — alongside prior state. The state grows turn by turn.
- Render Response To User — Combine selected documents with GM output into the user-facing response. The user sees an answer that's both generatively synthesized and grounded in retrieved sources.
Stateful Retrieval-Grounded Chat
The patent's central contribution is the persistent state that ties multi-turn chat to per-turn retrieval. Each turn updates the state; each new turn reads the state and the live index together. The generative companion is neither a stateless search nor a hallucinating chatbot — it's a grounded conversation.
State + Generation + Retrieval = Conversational Search
Three components together produce the hybrid: state for conversation memory, generation for synthesizing the response and synthetic queries, retrieval for grounding in real documents.
- Persistent State Data — Query, context, GM output, synthetic queries, selected documents — all persist across turns. The conversation memory.
- Synthetic Query Generation — The GM produces queries to run against the index. Chat utterances become retrieval targets through generation.
- Conversation-Aware Selection — Document selection considers what was shown in prior turns. Avoid repetition, prioritize coverage of unfilled aspects.
Every conversational-search surface on the modern web sits on this pattern.
<\/section>Technical Foundation
State Components And Generation Flow
The framework defines what the state contains and how generation plus retrieval plus selection compose into a turn.
- Query Input — The current user utterance. Can be a question, instruction, refinement, or off-topic pivot. Treated as input to GM, not directly as a retrieval query.
- Contextual Information — User-specific or device-specific context: profile, location, prior session info, ongoing topical interest. Conditions the GM output.
- Generative Model Output — The GM's response to the query + context + state. Includes draft response text, intent classification, and synthetic queries.
- Synthetic Queries — Queries derived from GM output, run against the index. The bridge between chat and retrieval.
- Selected Search Result Documents — Documents picked from synthetic-query retrieval, filtered for conversation-aware relevance (not previously shown, fills coverage gaps).
- State Data — The persistent record of all of the above, accumulating turn by turn.
Key Insight: Conversational search needs three things working together. Without state, every turn is a fresh search and the chat falls apart. Without generation, the user has to write retrieval-ready queries themselves. Without retrieval, the answer hallucinates. The patent's framing is that all three must coexist in a tight loop per turn, with the state being the persistence layer that makes multi-turn conversation coherent.
<\/section>The Process
End-To-End Turn Flow
Each turn runs the full loop from input to persisted state.
- User Input — User types or speaks. Input plus current contextual info enters the system.
- State Recall — Retrieve prior turns' state from the conversation store.
- GM Pass — Generative model processes query + context + state. Produces draft response and synthetic queries.
- Synthetic Query Execution — Run synthetic queries against the index. Retrieve candidate documents per query.
- Conversation-Aware Selection — Filter candidates: drop already-shown documents, prioritize coverage of aspects the conversation has not addressed, balance diversity with relevance.
- Response Synthesis — Combine selected documents with GM draft. Render with citations to source documents.
- State Persistence — Append this turn's components to the conversation state. Ready for the next turn.
Quality Control
Quality Control
Guards Against Drift And Hallucination
Multi-turn generative search can drift off-topic or amplify earlier hallucinations. Controls keep the conversation tethered.
- Retrieval Grounding Per Turn — Every turn must include at least one synthetic query and retrieve documents from the live index. The GM cannot produce a response from state alone; new retrieval grounds each turn.
- Citation Required — GM-synthesized content cites the documents that grounded it. Citation discipline prevents hallucination from compounding across turns.
- Conversation-Drift Detection — When successive turns drift far from earlier topical anchors without explicit user signal, the system flags drift. May reset state or surface a confirmation to the user.
- Already-Shown Avoidance — Documents shown in earlier turns are suppressed by default in subsequent turns. The user gets new sources rather than repeated ones.
What This Means for SEO
What This Means for SEO
This is the most consequential patent in the entire collection for the AI Overviews / SGE era. Knowing the mechanics of stateful chat changes how to think about being cited in multi-turn conversational responses.
- Being Cited Is The New Ranking — Conversational search cites source documents inline. The user reads the answer and sees attribution. Optimization is for being one of the cited sources, not for being clicked from a SERP.
- Synthetic Queries Are What Reach Your Content — Your content has to match the synthetic queries the GM generates from the conversation. These queries differ from raw user input because they're tuned for retrieval. Anticipating likely synthetic-query patterns (entity + attribute, entity + comparison, entity + how-to) is the new keyword research.
- Conversation-Aware Selection Favors Diverse Sources — Documents already shown in prior turns are suppressed. Being on the diversity-gain frontier (offering a perspective or sub-topic that other top results don't cover) is the new long-tail.
- State Persistence Means Multi-Turn Visibility — If your content gets cited in turn 1, you may be cited again in subsequent turns as the conversation deepens. Strong topical depth earns multi-turn visibility, not just first-turn.
- Coverage Of Unfilled Aspects Wins — The selection actively wants to cover aspects the conversation has not addressed. Content that fills under-covered niches of a topic gets pulled in to round out the answer.
- Citation-Friendly Content Format — Generative responses cite passages or paragraphs that are clearly bounded and factually crisp. Pages with structured, citable claim units (definition blocks, fact tables, structured paragraphs) are easier to surface as citations.
- Contextual Information Personalizes Conversations — User-specific and device-specific context conditions the GM. The same conversation produces different responses for different users. Content optimized for user-defined audiences (rather than generic search intent) compounds with this personalization.
- AI Overviews And SGE Are Built On This Pattern — Every modern conversational-search surface (Google AI Overviews, Microsoft Copilot, ChatGPT Search, Perplexity) follows the same state + generation + retrieval architecture. Understanding the pattern is the foundation for understanding the entire AI-search era.