Returns complete natural-language answer sentences for intent-bearing queries, extracted from authoritative sources and rendered above traditional links, so users get the answer at a glance rather than scanning blue-link results.
Patent Overview
- Filed
- 2013-06-04
- Granted
- 2016-04-06 (published application)
- Application Number
- EP 14171008.1
The Challenge
The Challenge
Many queries express a clear intent for which a direct sentence-level answer exists in some authoritative source. Returning a list of pages forces the user to scan, when the system could surface the answer directly.
- Blue Links Are Slow For Factoid Intent — When the user asks 'how tall is Mount Everest', the answer is one sentence. Returning ten pages forces unnecessary scanning. Direct answer is faster.
- Authoritative Sources Already Have The Answer — Wikipedia, government sites, encyclopedia entries assert authoritative facts. The system can extract and surface them rather than asking users to find them.
- Intent Detection Must Be Reliable — Not every query has a clean direct answer. The system must detect intent reliably to know when to surface a direct answer versus when to default to standard results.
- Extraction Must Preserve Authority — The extracted sentence must be attributable to its source so users can verify and explore further. Authority and provenance must accompany the answer.
- Wrong Answers Cost Trust — A direct answer that is wrong is worse than no direct answer. Confidence threshold must be conservative; wrong answers erode user trust in the entire feature.
Innovation
How The System Works
The system classifies the query for intent, identifies authoritative source sentences that answer it, extracts the best-matching sentence with full context, scores its confidence, and renders a direct answer block with source attribution above the traditional result list.
- Classify Query Intent — An intent classifier determines whether the query has a clean direct-answer intent (factoid, definition, how-much, when-was). Non-intent queries skip the direct-answer path.
- Retrieve Authoritative Candidate Sources — From the indexed corpus, retrieve authoritative pages likely to contain the answer. Authority comes from PageRank, source-reputation signals, and knowledge graph linkage.
- Extract Candidate Sentences — From candidate pages, extract sentences that linguistically match the query intent. Pattern-based extractors and span-prediction models work in concert.
- Score Extraction Confidence — Each candidate sentence is scored on semantic match to the query, source authority, and freshness. Confidence is calibrated against historical accuracy.
- Pick Best Answer Or Suppress — Above-threshold candidates surface as direct answers. Below-threshold queries fall back to standard result lists. Conservatism is the rule.
- Render With Attribution — The answer renders prominently in the SERP with the source attributed. Source linking lets users verify and explore.
- Capture Feedback — User reaction to the direct answer (click through, continue scrolling, ask follow-up) feeds back into confidence calibration and source weighting.
Sentence-Level Answer Surface
The patent's load-bearing idea is to elevate retrieval from page-level to sentence-level for intent-bearing queries. The user gets the answer, not the document containing the answer.
Answer The Question, Surface The Source
Direct answer plus visible source attribution is the format. The answer satisfies the user; the source preserves verifiability and accountability.
- Intent Classification — Not every query has a direct answer. Classification gates the feature so it only fires when warranted.
- Authority-Weighted Extraction — Extraction prefers authoritative sources. The same fact stated by a low-authority source is weighted less.
- Conservative Confidence — Threshold is high. The system surfaces direct answers only when confidence justifies it. Wrong answers cost more trust than missed-opportunity suppressions cost coverage.
Technical Foundation
Technical Foundation
The patent specifies the intent classifier, the extraction models, the confidence scoring, the source authority framework, and the rendering layer.
- Intent Classifier — Neural model classifies queries for direct-answer-eligibility. Trained on labeled query-intent pairs plus log-derived weak supervision.
- Candidate Source Retriever — Retrieves authoritative pages likely to contain the answer. Uses standard text retrieval plus authority weighting.
- Sentence Extractor — Span-prediction or pattern-based extractor identifies candidate answer sentences within retrieved pages. Outputs spans plus confidence.
- Confidence Scorer — Combines semantic match, source authority, freshness, and cross-source consensus into a calibrated confidence score.
- Source Authority Index — Per-source authority scores are maintained. Authority informs both retrieval ranking and extraction confidence.
- Direct-Answer Renderer — Renders the answer block prominently with attribution. Layout adapts to device form factor and answer length.
The Process
The Process
The pipeline runs in parallel with standard retrieval. The user-perceived latency is the slower of the two paths, kept tight by caching and parallelism.
- Receive Query — Query enters the dispatcher. Both standard retrieval and direct-answer paths fire in parallel.
- Classify Intent — Intent classifier outputs direct-answer-eligibility plus confidence. Below-threshold intents skip direct-answer extraction.
- Retrieve Authority Candidates — Eligible queries trigger authority-weighted retrieval of candidate sources.
- Extract Candidate Sentences — Sentence extractor produces candidates with spans and confidences.
- Score And Pick Winner — Scoring combines all signals. Above-threshold winners surface as direct answers; below-threshold cases suppress the answer.
- Render With Source Attribution — Winning answer renders above standard results with the source linked.
- Feed Learning Loop — User feedback (clicks, scrolls, follow-ups) refines confidence calibration and source authority weighting.
Quality Control
Quality Control
Wrong direct answers erode user trust quickly. The patent specifies safeguards across the pipeline.
- High Confidence Threshold — The threshold for surfacing direct answers is conservative. Many extractable answers are suppressed because confidence is not high enough, preferring no-answer to wrong-answer.
- Cross-Source Consensus — When multiple authoritative sources agree, confidence rises. Single-source claims face higher confidence bars.
- Source Authority Auditing — Source authority scores are reviewed periodically. Sources that produce wrong facts repeatedly have their authority reduced.
- Time-Sensitive Fact Handling — Time-sensitive facts (current prices, current schedules) are validated for freshness. Stale facts are suppressed.
- User Feedback Channel — Users can flag wrong direct answers. Flags feed back into source quality and confidence calibration.
Real-World Application
Sentence-level direct answers are the conceptual ancestor of Featured Snippets, Knowledge Panel facts, and the answer layer in Search Generative Experience and AI Overviews.
- Above-fold Rendering Position — Direct answers render prominently above standard results. The user sees the answer before any blue link.
- Authority-weighted Source Selection — Extraction prefers authoritative sources. Authority is one of the strongest signals in the scorer.
- Conservative Trigger Threshold — Direct answers surface only when confidence is high. Suppression is preferred over wrong-answer risk.
Why Question-Form Content Earns Visibility
Content structured as questions plus direct-sentence answers gives the extractor exactly what it needs. H2-as-question with first-sentence answer is a pattern this patent's primitives reward consistently.
Why Becoming The Authoritative Source Compounds
Once your site is recognized as authoritative for a topic, the extractor preferentially picks your sentences. The compound effect of authority on direct-answer surfacing is one of the most powerful structural advantages in modern SEO.
<\/section>What This Means for SEO
What This Means for SEO
When intent queries are handled natively, ranking depends on whether the page satisfies the intent, not whether it contains the literal query string.
- Intent Categories Group Queries — Many distinct phrasings map to the same intent. A page that satisfies an intent ranks for the whole phrasing family. Map your content by intent, not by keyword.
- Question-Form Content Captures Intent Queries — Heading-as-question with answer-as-first-sentence-below is the pattern the model is trained to extract. Build content sections in this shape, not as flowing essays.
- Intent Drift Is A Real Risk — The intent behind a query can shift over months. Pages that fail to update lose position not because the page got worse but because intent moved. Re-check intent for your top queries quarterly.