Augments incoming queries with additional terms or constraints derived from query context, user profile, and historical reformulation patterns, increasing recall on under-specified queries while preserving precision.
Patent Overview
- Inventor
- Krishna Bharat
- Assignee
- Google LLC
- Filed
- 2012-06-15
- Granted
- 2015-09-08
- Application Number
- US 13/525,065
The Challenge
The Challenge
Users often issue queries that are too short or under-specified to retrieve their intended results. The system can augment such queries by adding inferred terms, but the augmentation must be careful, adding too much produces drift, and adding the wrong terms produces irrelevance.
- Short Queries Are Often Under-Specified — A two-word query may have meant three or four words to the user. Document retrieval against the literal query misses what the user actually wanted to find.
- Augmentation Risks Drifting — Adding terms aggressively changes the query meaning. The user wanted X, the system retrieved Y. Augmentation must preserve user intent.
- Context Predicts What To Add — Recent queries, location, time, user profile all hint at what is missing. Mining these signals provides candidate augmentation terms.
- Historical Reformulations Reveal Patterns — When users reformulate queries, they show what they meant. Mining reformulation logs reveals canonical augmentations users themselves apply.
- Augmentation Must Be Discoverable Or Reversible — Surprising augmentations confuse users. Either the augmentation should be obvious (visible chips), or the user should be able to revert easily.
Innovation
How The System Works
The system identifies under-specified queries, extracts context signals and historical reformulation patterns, generates augmentation candidates, scores them on intent preservation and retrieval likelihood, applies the best candidates, and surfaces the augmentation visibly so the user can override.
- Detect Under-Specification — Short queries, queries with no clear entity match, queries with broad ambiguity are flagged as candidates for augmentation. Clearly-specified queries skip the augmentation path.
- Gather Context Signals — Session history, location, time, user profile, and recent activity feed the augmentation candidate generation.
- Mine Reformulation Patterns — For similar past queries, retrieve the typical reformulations users applied. These are gold-standard augmentation candidates.
- Generate Candidate Augmentations — Combine context signals and reformulation patterns to produce candidate added terms or constraints. Each candidate is a specific augmentation.
- Score Candidates — Each candidate is scored on intent preservation (does it change the meaning?), retrieval likelihood (does it produce useful results?), and historical click-through (have past users found these results useful?).
- Apply Top Candidate — The highest-scoring candidate becomes the augmentation. The augmented query goes to retrieval.
- Surface Augmentation Visibly — The augmented query appears in the SERP with a visible indicator (chips, query summary). The user can edit or remove augmentations they did not intend.
Context-Driven Augmentation
The patent's load-bearing idea is to augment queries using multiple context signal sources, score the candidates carefully, and surface the augmentation transparently so users stay in control even when the system has added inferred terms.
Add What The User Meant
Naive augmentation adds what the system happens to think; smart augmentation adds what the user likely meant. Context signals plus reformulation history make the inference reliable.
- Context Aggregation — Multiple signals (session, location, profile, time) combine into a context model that informs augmentation choices.
- Reformulation Mining — Past user reformulations tell the system what augmentations have worked. Mining produces strong candidate sets grounded in real behavior.
- Visible Augmentation — The user sees what was added. Edit and override are first-class. The system stays transparent rather than silently modifying queries.
Technical Foundation
Technical Foundation
The patent specifies the under-specification detector, the context aggregator, the reformulation miner, the augmentation generator, the scorer, and the SERP integration.
- Under-Specification Detector — Heuristic plus learned model flags queries likely to benefit from augmentation. Short queries, ambiguous entity matches, and broad topics trigger augmentation.
- Context Aggregator — Combines session signals, location, time, user profile into a structured context vector for downstream augmentation generation.
- Reformulation Miner — From query logs, extracts patterns of reformulation per query type. The miner produces candidate-augmentation pools indexed by query similarity.
- Augmentation Generator — Combines context and reformulation patterns to generate candidate augmentations. Templates plus generative composition produce candidate added terms.
- Scorer — Per candidate, scores on intent preservation, retrieval likelihood, and historical click-through. Outputs ranked candidate list.
- SERP Integration — Augmented queries render in the SERP with visible augmentation indicators. The user can edit or remove augmentations.
The Process
The Process
Augmentation runs in the query path before retrieval. Added latency is small because most of the work is fast lookups against pre-computed indexes.
- Receive Query — Query arrives at dispatcher. Under-specification detector runs first.
- Detect Augmentation Candidacy — If query is sufficiently specified, skip augmentation. Otherwise enter the augmentation path.
- Aggregate Context — Available context signals are gathered into the context vector.
- Mine Reformulations — Reformulation patterns for similar queries are retrieved from the miner index.
- Generate And Score Candidates — Candidate augmentations are generated and scored. Top candidate is selected.
- Apply And Retrieve — Augmented query goes to retrieval. Results are produced as for any standard query.
- Render With Indicator — SERP shows augmented query with visible indicator. User can edit or override.
Quality Control
Quality Control
Wrong augmentation degrades user experience. The patent specifies safeguards.
- Confidence Threshold — Augmentation only applies when candidate confidence is high. Low-confidence cases retrieve against the original query.
- Intent Preservation Verification — Each candidate is checked for intent preservation. Candidates that change the query meaning are excluded.
- User Override Surfaced — Augmentations are visible. Users can remove or edit them in one click.
- Historical Click-Through Validation — Candidates are validated against historical click-through patterns. Patterns that historically produced poor outcomes are excluded.
- Per-User Calibration — User-specific patterns shape augmentation. Users who consistently override augmentations get less aggressive augmentation over time.
Real-World Application
Query augmentation appears in Google's auto-completion, related-query chips, and the implicit query expansion that powers many SERP behaviors. The patent's primitives shape how the engine handles short and under-specified queries.
- Context-driven Augmentation Source — Augmentations come from session context, profile, location, and historical patterns. Multiple signal sources produce reliable candidates.
- Confidence-gated Trigger Logic — Augmentation only applies when confidence is high. Low-confidence queries retrieve literally.
- User-visible Transparency — Users see what was added and can override. The system stays transparent rather than silently modifying queries.
Why Long-Tail Variations Capture Real Traffic
Sites that cover the augmented variants of common queries (location-specific, time-specific, modifier-rich versions) catch traffic when augmentation triggers. The variants are the real queries users issue once augmentation is applied.
Why Schema Markup Helps Augmentation Targeting
Pages with explicit context markup (location, time, audience, entity) are easier for the system to validate as good augmentation targets. The markup informs whether the augmented retrieval should rank the page.
<\/section>What This Means for SEO
What This Means for SEO
The patent augments under-specified queries with inferred terms from context, user profile, and reformulation history, scored for intent preservation and surfaced transparently. SEO implication: covering the augmented variants of common queries captures the real expanded queries users issue once augmentation triggers.
- Long-Tail Variants Capture Real Traffic — Sites covering augmented variants of common queries (location-specific, time-specific, modifier-rich versions) catch traffic when augmentation triggers, because the variants are the real expanded queries the system retrieves against.
- Schema Markup Helps Augmentation Targeting — Pages with explicit context markup (location, time, audience, entity) are easier to validate as good augmentation targets. The markup informs whether augmented retrieval should rank the page, so structured context cues aid eligibility.
- Add What The User Meant — Augmentation adds inferred intent from context and reformulation history. Content that anticipates the likely-intended specifics (the city, the recent timeframe, the qualifier) aligns with augmented queries better than content covering only the bare term.
- Context Signals Drive Augmentation — The system uses context and historical reformulation patterns to choose terms. Understanding how users typically refine an under-specified query in your space tells you which augmented variants to cover.
- Transparent Augmentation Keeps Users In Control — Augmentation is surfaced visibly so users can override. Because users see and can reject the added terms, content must genuinely match the augmented intent, not just the system's guess, to retain the click.
- Cover The Specific, Not Just The General — Under-specified queries get specified by augmentation. A page covering specific, qualified versions of a topic catches augmented retrievals that a single broad page misses. Build out the specific variants.
- Intent Preservation Is Scored — Augmentation candidates are scored on intent preservation and retrieval likelihood. Content that cleanly satisfies the preserved intent of the augmented query, rather than drifting, is what the augmented retrieval favors.