Composes a synthesized natural-language answer for a query by retrieving grounded passages from authoritative sources and feeding them to a large language model that produces a summary with citations, the technical substrate behind AI Overviews and Search Generative Experience.
Patent Overview
- Inventor
- Srinivasan Venkatachary
- Assignee
- Google LLC
- Filed
- 2022-09-15
- Granted
- 2023-09-26
- Application Number
- US 17/946,229
The Challenge
The Challenge
Document retrieval returns ranked links; users still have to read and synthesize. For many information needs, a directly-composed answer with citations is far more useful than a list of pages. The system needs to compose such answers reliably, with grounded sources, at search-engine scale.
- Ranked Links Force The User To Synthesize — When the answer requires combining facts from several sources, ten blue links push the work onto the user. The system can do the synthesis itself if it can ground in retrieved sources.
- Pure LLM Generation Hallucinates — A language model generating answers without retrieval produces plausible-sounding but unverifiable claims. Grounding in retrieved passages is the antidote to hallucination.
- Sources Must Be Attributable — Synthesized answers without source citations fail trust and verifiability tests. Each claim should trace to a specific source so users can verify and explore further.
- Latency Budget Constrains The Pipeline — Search users expect sub-second response. Retrieval plus LLM generation plus citation alignment all must fit in the latency budget at billions-of-queries scale.
- Quality Threshold For Display — A wrong generative answer costs more user trust than a missing one. The system must confidently suppress generation when grounding is weak or contested.
Innovation
How The System Works
The patent retrieves a small set of authoritative passages from the index, formats them as grounded context for a large language model, generates a synthesized answer constrained to cite the source passages, aligns each claim with its source, and renders the answer with inline citations above traditional search results.
- Retrieve Grounded Passages — Standard plus dense retrieval surfaces top candidate passages. Authority filters prefer high-quality sources. The passage set is the grounding context for generation.
- Format Context Window — Retrieved passages plus the user query plus prompting instructions form the LLM input. The format constrains the model to ground answers in the provided passages rather than internal parametric knowledge.
- Generate Synthesized Answer — The LLM produces an answer composed from the provided passages. Generation is conditioned on the grounding constraint so the model stays close to source content.
- Align Claims To Sources — Each sentence or claim in the generated answer is aligned to its source passage(s). The alignment supports inline citations and downstream verification.
- Score Grounding Quality — The alignment quality is scored. Strong alignment across all claims produces high confidence; weak or contested alignment produces low confidence.
- Decide Whether To Display — Above-threshold confidence triggers display of the generative answer above traditional results. Below-threshold cases fall back to standard SERP without a generative answer.
- Render With Citations — The displayed answer includes inline citation markers linking to source passages. Users can verify each claim and explore the underlying documents.
Grounded Generation With Citations
The patent's load-bearing combination is retrieval-augmented generation plus claim-to-source alignment plus confidence-gated display. Each layer prevents a specific failure mode of pure-LLM generation while preserving the synthesis benefit.
Retrieval Constrains Generation
An LLM generating freely will hallucinate. An LLM generating from retrieved grounded passages stays close to the source material. The retrieval step is the manipulation-resistance layer that makes generative search safe to ship.
- Grounded Context — Retrieval supplies the passages the LLM must use. Generation is constrained to ground in provided sources rather than relying on parametric memory.
- Citation Alignment — Each generated claim aligns to its source passage. Citations are not bolted on; they emerge from the generation pipeline itself.
- Confidence-Gated Display — Low-confidence answers suppress generation entirely. Better to show standard results than to show a wrong synthesized answer.
Technical Foundation
Technical Foundation
The patent specifies the retrieval-augmented-generation pipeline, the prompting structure, the alignment model, the confidence scorer, and the display integration.
- Retrieval Layer — Hybrid dense plus sparse retrieval surfaces top candidate passages. Authority and freshness filters refine the set before generation.
- LLM Inference Backend — Large language model (PaLM, Gemini, etc.) generates the synthesized answer from the grounded context. Inference is optimized for latency at production scale.
- Prompt Template — Structured prompt formats the query, retrieved passages, and grounding instructions. The template constrains the LLM to source-aligned output.
- Claim-Source Aligner — Per generated sentence, the aligner identifies which source passage(s) supports it. Multiple alignment strategies (entailment models, citation prediction, attention-based) combine.
- Confidence Scorer — Combines per-claim alignment strength, source authority, and cross-source agreement into a single confidence score. Calibrated against historical accuracy.
- Display Integration — Above-threshold answers render above standard results with inline citation chips. Below-threshold cases produce no generative answer; users see standard SERP.
The Process
The Process
The pipeline runs in the query path with caching and parallelism to stay within the latency budget. Each user-facing response is the result of retrieval, generation, alignment, scoring, and rendering executed in sequence with optimizations.
- Receive Query — Query arrives. The dispatcher determines whether the query is a candidate for generative answering (informational, factoid, exploratory).
- Retrieve Passages — Hybrid retrieval surfaces top candidates. Filters keep the set high-authority and topically focused.
- Prompt And Generate — The prompt template formats query plus passages; the LLM generates the synthesized answer constrained to ground in the supplied context.
- Align Claims — Per claim, the aligner identifies source passages supporting it. Output is the per-claim provenance map.
- Score Confidence — Combine alignment strengths, source authority, and cross-source agreement into a final confidence value.
- Gate On Threshold — Above-threshold answers proceed to display. Below-threshold answers are suppressed; the SERP shows standard results only.
- Render With Citations — Approved answers render above standard results with inline citation chips. Users can hover or click citations to view sources.
Quality Control
Quality Control
Generative search risks hallucination, source-mismatch citations, and wrong answers at scale. The patent specifies multi-layer safeguards.
- Grounding Constraint Enforcement — The LLM is constrained to ground in retrieved passages. Outputs that drift from sources are caught by the aligner and trigger confidence reduction or suppression.
- Source Authority Filter — Low-authority sources are filtered before retrieval. The LLM only sees authoritative context, reducing the chance of low-quality content propagating into the answer.
- Cross-Source Consensus — When multiple sources agree, confidence rises. Single-source claims face higher confidence bars; contested claims fail confidence.
- Sensitive Query Handling — Health, finance, and other sensitive queries get stricter thresholds. Some categories suppress generative answers entirely in favor of curated content.
- Continuous Evaluation — Generated answers are continuously evaluated for accuracy, citation quality, and user trust. Regressions trigger rollback or threshold tightening.
Real-World Application
This patent family is the technical substrate behind AI Overviews (rebranded from Search Generative Experience in 2024) and the conversational answer surfaces across Google's search products. It marks the shift from ranked-link retrieval to grounded generative answers.
- Grounded Generation Constraint — Generation grounds in retrieved authoritative passages rather than parametric LLM memory. Manipulation resistance comes from the retrieval layer.
- Cited Citation Format — Each claim aligns to source passages and renders with inline citations. Users can verify and explore.
- Confidence-gated Display Trigger — Below-threshold confidence suppresses the generative answer. Better to show standard results than wrong answers.
Why Citation Slots Are The New SERP Real Estate
AI Overviews surface synthesized answers with citations to source pages. Earning citation slots is becoming as important as ranking in the blue-link results. Content that grounds well (clear facts, authoritative sourcing, structured for extraction) wins citation visibility.
Why Authoritative Sources Compound Disproportionately
The retrieval step filters for authority before generation. Sites established as authoritative on a topic are repeatedly chosen as grounding sources, accumulating citation traffic that less-authoritative competitors cannot break into.
<\/section>What This Means for SEO
What This Means for SEO
The patent is retrieval-augmented generation for search: retrieve authoritative passages, generate a synthesized answer constrained to cite them, align each claim to its source, and gate display on confidence. SEO implication: earning citation slots in AI Overviews is becoming as important as ranking, and authoritative, extractable content wins them.
- Citation Slots Are New SERP Real Estate — AI Overviews surface synthesized answers with citations to source pages. Earning citation slots is becoming as important as blue-link ranking. Structure content to ground well (clear facts, authoritative sourcing, extractable structure) to win citation visibility.
- Authoritative Sources Compound Disproportionately — The retrieval step filters for authority before generation. Sites established as authoritative on a topic are repeatedly chosen as grounding sources, accumulating citation traffic that less-authoritative competitors struggle to break into.
- Retrieval Constrains The Generation — The model generates only from retrieved grounded passages to avoid hallucination. To be cited, you must first be retrieved as a grounding passage. Strong retrieval fundamentals and extractable passages are the precondition for citation.
- Claim-To-Source Alignment Rewards Clear Facts — Each generated claim is aligned to its source passage. Content stating clear, attributable facts in self-contained spans aligns cleanly to claims and gets cited, while vague or hedged prose is harder to attribute.
- Confidence Gating Favors Reliable Content — Display is gated on confidence. Well-sourced, consistent, corroborated content raises the confidence that lets an answer surface, while contradictory or thin content suppresses it. Reliability is a display gate, not just a quality nicety.
- Write For Extraction — Grounding draws from extractable passages. Self-contained paragraphs with resolved references, complete claims, and clear topic sentences are pulled as grounding far more often than context-dependent prose. Extractability is a citation lever.
- Synthesis Reduces Clicks, Raises Citation Value — Users often get the answer from the summary without clicking, but cited sources gain brand exposure and selective clicks. As clicks shift, being the cited authority behind the answer becomes the strategic position to hold.