Suggests query refinements derived from result clustering and historical reformulation patterns, helping users narrow broad queries through one-click drill-downs that match the sub-topics actually present in the result set.
Patent Overview
- Filed
- 2010-02-05
- Granted
- 2015-04-28
- Application Number
- US 12/700,756
The Challenge
The Challenge
Broad queries return diverse results. Users need to narrow without retyping. The system needs to suggest refinements that are both relevant to the current result set and aligned with how users typically reformulate similar queries.
- Broad Queries Produce Mixed SERPs — A query like 'cameras' returns DSLRs, mirrorless, point-and-shoot, action cameras, mixed across results. Users want to drill into one sub-topic.
- Refinements Should Match Result Set — Suggesting 'wedding cameras' when no wedding-camera content ranks is useless. The refinements must reflect the topics actually present in results.
- Historical Reformulations Reveal Real Drilldowns — Query logs show how users actually narrow queries. Mining reformulation patterns tells the system which refinements are valuable for which queries.
- Refinements Need Topical Coherence — A good refinement adds one clean dimension (price, brand, category). Multi-axis refinements confuse rather than clarify.
- Latency Budget Limits Computation — Refinement suggestions render in the SERP. They cannot take long to compute or they delay the whole page.
Innovation
How The System Works
The system clusters top results by topic, mines historical reformulation patterns for similar queries, picks refinements that match both the result clusters and the historical patterns, and renders them as one-click suggestions in the SERP.
- Retrieve Result Set — Standard retrieval produces the top results for the user's query. These results are the corpus for refinement generation.
- Cluster Results By Sub-Topic — Top results are clustered into topical groups. Each cluster represents one sub-topic the user could drill into.
- Mine Historical Reformulations — For similar past queries, retrieve the reformulation patterns users typically applied. These patterns inform refinement candidate generation.
- Generate Candidate Refinements — Combine cluster labels and reformulation patterns to produce candidate refinements. Each candidate is one specific narrowing of the original query.
- Score And Rank Candidates — Candidates are scored on result-set coverage, historical click-through, and topical coherence. Top candidates become displayed refinements.
- Render In SERP — Selected refinements appear as one-click chips in the SERP. Clicking a refinement issues the narrowed query.
- Learn From Clicks — Which refinements users click feeds back into the candidate-scoring model. The system continuously improves refinement quality.
Refinements From Results Plus History
The patent's load-bearing combination is current results (for what is in the SERP now) plus historical reformulations (for how users actually narrow similar queries). Either alone is weaker; together they yield refinements that are both relevant and natural.
Live Data Plus Past Behavior
Result-set clustering captures the current diversity of the SERP. Historical reformulations capture user behavior over time. Combining both makes refinements feel right because they match both what is here and what users usually want.
- Result-Cluster Refinements — Each result-set cluster becomes a candidate refinement. The user can drill into the cluster that matches their interest.
- Historical Reformulation Mining — Past queries' reformulation patterns inform which refinements users typically want. Aligns with empirical behavior.
- Scored And Ranked Display — Top candidates render as one-click chips. Scoring balances result coverage and historical preference.
Technical Foundation
Technical Foundation
The patent specifies the result clusterer, the reformulation miner, the candidate generator, the scoring model, and the SERP renderer.
- Result Clustering — Top results are clustered using topic models or graph-based methods. Cluster labels become candidate refinement keywords.
- Reformulation Miner — Query log analysis identifies reformulation patterns: ways users typically narrow similar queries. The miner produces a candidate pool per query class.
- Candidate Generator — Combines cluster labels and reformulation patterns to produce refinement candidates. Each candidate is a specific narrowed query.
- Scoring Model — A learned model scores candidates on result-set coverage, historical click-through, topical coherence, and diversity. The model output is a single score per candidate.
- Diversity Filter — Selected refinements must be diverse (different drill-down dimensions). The filter prevents redundant chips that all narrow on the same axis.
- SERP Renderer Integration — Refinement chips render in the SERP. Click handling routes to the narrowed query through the standard query pipeline.
The Process
The Process
The refinement pipeline runs in the SERP composition path. Total added latency is small because clustering and lookups are cached for common queries.
- Receive Query And Results — Standard retrieval completes. The refinement pipeline takes the top results as input.
- Cluster Results — Top results are clustered into sub-topics. The clustering uses cached models for speed.
- Lookup Historical Reformulations — Reformulation pool for similar queries is retrieved from precomputed indexes.
- Generate Candidates — Cluster labels and reformulation patterns combine into candidate refinements. Duplicates are removed.
- Score And Filter — Candidates are scored. Diversity filter ensures the displayed set covers different drill-down axes.
- Render Chips — Selected refinements render as clickable chips in the SERP. Layout fits the device form factor.
- Log Clicks For Learning — Click and dwell patterns on refinements feed back into the scoring model. Continuous improvement loop.
Quality Control
Quality Control
Bad refinements waste the user's attention and pollute the SERP. The patent specifies safeguards.
- Minimum Result Coverage — A refinement must lead to a substantial set of relevant results, not a near-empty SERP. Coverage-below-threshold candidates are dropped.
- Diversity Constraint — Selected refinements must drill on different dimensions. Multi-chip groups that all narrow the same axis are filtered out.
- Historical Performance Check — If a refinement was historically a poor performer (low click, fast pogo-stick), it is downweighted or excluded.
- Topical Coherence — Refinements must combine cleanly with the original query syntactically. Awkward concatenations are filtered.
- Sensitive Query Handling — Medical, legal, and other sensitive queries get curated refinements rather than algorithmically generated ones, to prevent harmful drill-downs.
Real-World Application
Query refinement appears in Google's SERP as 'Related searches', 'People also ask', and inline refinement chips. The same primitives underlie related-search modules in other Google products.
- One-click Refinement Format — Refinements render as clickable chips. The user can drill in without retyping.
- Result-aware Selection Method — Refinements reflect both the current result set's clusters and historical reformulation patterns.
- Learning loop Continuous Improvement — Click and dwell signals feed back into refinement scoring. The system continuously improves.
Why Refinement Phrases Are Content Gaps
Every refinement is a query users actually want. Mining the refinement chips and 'People also ask' boxes for content gaps is a direct content-strategy lever. Filling those gaps captures the refinement-driven traffic.
Why Topical Coherence Matters Per Page
A page that fits cleanly into one refinement cluster outranks a page that spreads across many. The patent's clusterer rewards clear topical focus when assembling refinement candidates.
<\/section>What This Means for SEO
What This Means for SEO
Query refinement systems suggest reformulations users adopt, so capturing the refined query family is a content-mapping exercise.
- Refinements Are Pre-Indexed Intent — The refinements the system suggests are the queries it has high confidence about. Map your content to refinement labels in PAA and related-search modules.
- Negative Refinements Reveal Pain Points — Refinements like "not", "without", "alternative to" reveal user frustrations with the current SERP. Pages that answer those negative refinements often have low competition.
- Refinement Suggestions Drift — Today's refinements are not yesterday's. Monthly checks on which refinements appear catch shifts before competitors do.