Generates query suggestions seeded by which results the user clicked, building a click-driven query-rewriting layer. The signal flows backwards: clicked results inform what queries users meant.
Patent Overview
- Inventor
- Paul Haahr, others
- Assignee
- Google LLC
- Filed
- 2009
- Granted
- 2019-10-29
The Challenge
The Challenge
Forward query suggestion uses query similarity to suggest. Backward suggestion uses result clicks: which results users clicked from a given query reveal what they actually meant. That signal can power suggestions to other users issuing similar queries.
- Forward Suggestion Misses Click Signal — Suggesting based on query similarity alone misses the strongest intent signal — what users actually clicked.
- Clicked Results Reveal True Intent — When users click result R from query Q, that's evidence Q meant R-shaped content. Aggregate this across users and the signal is strong.
- Result-Backed Suggestions Generalize — From clicked results, derive query suggestions that other users issuing similar queries can use. The signal generalizes across user pool.
- Suggestion Coverage Must Scale — Per query, suggestion generation must fit latency budget. Pre-computed mapping from queries to result-derived suggestions required.
- Manipulation Resistance — Result-click signal is exploitable via click manipulation. Detection and filtering required.
Innovation
How The System Works
The system tracks per-query result clicks aggregated across users, derives suggestion candidates from clicked-result content, scores candidates by alignment and click strength, and surfaces them as query suggestions at query time.
- Aggregate Per-Query Clicks — Per query, aggregate which results users clicked across user pool. Output is per-query click distribution over results.
- Derive Suggestions From Clicked Content — Per clicked result, derive candidate suggestion queries based on result content (title, anchors, topical model).
- Score Candidates — Per candidate, score by click strength and topical alignment.
- Pre-Compute Suggestion Map — Per query, pre-compute top-N suggestion candidates. Cached for query-time consumption.
- Surface At Query Time — Per query, retrieve pre-computed suggestions and surface as SERP suggestions or autocomplete.
- Detect Manipulation — Pattern analysis flags suspicious click patterns suggesting manipulation. Filtered before suggestion derivation.
- Recompute Periodically — Per crawl or per traffic window, suggestion map recomputes against fresh click data.
Clicks Reveal What Queries Meant
The patent's load-bearing idea is that clicked results are intent signals. Reading them backwards — clicked content informs query meaning — produces suggestions that capture real user intent.
Result-To-Query Is The Signal Direction
Forward: query to results. Backward: clicked results to suggestion queries. The backward direction is what extracts intent from behavior.
- Per-Query Click Aggregation — Per query, aggregate which results users clicked. Click distribution reveals intent.
- Result-Content-Derived Suggestions — Per clicked result, derive suggestion candidates from result content. Title, anchors, topical model feed suggestion derivation.
- Pre-Computed Mapping — Per query, pre-compute top-N suggestions. Cached for query-time consumption. Latency budget respected.
Technical Foundation
Technical Foundation
The patent specifies the click aggregator, suggestion deriver, candidate scorer, suggestion-map builder, surface layer, and manipulation detector.
- Click Aggregator — Per query, aggregates which results users clicked across user pool. Output is per-query click distribution.
- Suggestion Deriver — Per clicked result, derives candidate suggestion queries from result content (title, anchors, topical model).
- Candidate Scorer — Per candidate, scores by click strength and topical alignment.
- Suggestion-Map Builder — Per query, builds top-N suggestion list. Cached.
- Surface Layer — Per query at query time, retrieves cached suggestions and surfaces as SERP chips or autocomplete entries.
- Manipulation Detector — Pattern analysis flags suspicious click patterns. Filtered before suggestion derivation.
The Process
The Process
Aggregation and pre-computation run offline; surfacing runs at query time.
- Aggregate Clicks Offline — Per query, per-result click counts aggregate across user pool.
- Derive Suggestions — Per clicked result, suggestion deriver produces candidate suggestion queries.
- Score Candidates — Candidate scorer produces per-candidate score.
- Build Suggestion Map — Per query, top-N suggestions cached.
- Receive Query — Query arrives at query time.
- Retrieve And Surface — Cached suggestions retrieved and surfaced.
- Recompute Periodically — Per traffic window, suggestion map recomputes against fresh click data.
Quality Control
Quality Control
Click-based suggestions are subject to manipulation. The patent specifies safeguards.
- Click-Pattern Manipulation Detection — Pattern analysis flags suspicious click patterns. Filtered before suggestion derivation.
- User-Pool Diversity Requirement — Aggregations require diverse user-pool support. Single-user click patterns filtered.
- Suggestion-Quality Calibration — Suggestions calibrate against held-out labeled data. Mis-calibration produces irrelevant suggestions.
- Adversarial Defense — Click manipulation (click farms, bot traffic) actively defended via pattern filtering.
- Continuous Recalibration — Scoring weights and detection thresholds recalibrate against fresh data.
Real-World Application
Result-based query suggestions underpin the related-searches chips and the 'people also search for' surface on modern SERPs. The backward-signal-flow pattern is a foundational architectural insight.
- Backward signal Direction — Result clicks inform query meaning. The reverse direction extracts intent.
- Aggregate-pooled User-Pool Requirement — Diverse user-pool aggregation required. Single-user click patterns filtered.
- Pre-computed Latency Strategy — Per-query suggestion map pre-computed. Cached for query-time retrieval.
Why Click-Worthy Results Multiply Discovery
Pages that earn clicks for a given query become candidate-content for that query's suggestions, which then surface to other users on adjacent queries. The compound effect favors pages that consistently win clicks.
Why Surface Appearance Matters
Click-driven suggestions reward results that earn clicks from SERPs. Title, snippet, and structured-data presentation directly influence whether a page becomes a suggestion-target.
<\/section>What This Means for SEO
What This Means for SEO
This patent generates query suggestions backwards from which results users clicked, so clicked content reveals what queries meant and seeds suggestions for other users. SEO implication: pages that consistently win clicks for a query become suggestion targets that surface to users on adjacent queries, compounding discovery.
- Click-Worthy Results Multiply Discovery — Pages that earn clicks for a query become candidate content for that query's suggestions, which then surface to users on adjacent queries. Consistently winning clicks compounds your exposure across related searches.
- Surface Appearance Drives Suggestion Targeting — Suggestions reward results that earn clicks from the SERP, so title, snippet, and structured-data presentation directly influence whether you become a suggestion target. A compelling SERP appearance feeds the loop.
- Clicked Content Defines Query Meaning — When users click your result for a query, your content's terms and topic inform what that query meant. Being the clicked answer effectively teaches the system to associate your content with that intent.
- The Signal Flows Backward From Behavior — Forward suggestion uses query similarity; this uses what users actually clicked, the stronger intent signal. Optimizing for genuine engagement, not just keyword similarity, taps the more powerful backward signal.
- Aggregate User Support Is Required — Suggestions require diverse user-pool support, and single-user click patterns are filtered. Broad, genuine click performance across many users is what promotes you into the suggestion map.
- Manipulated Clicks Are Filtered — Click farms and bot traffic are detected and removed before suggestion derivation. You cannot buy your way into suggestions with artificial clicks; the signal must come from real searchers.
- Win Adjacent-Query Visibility By Winning Clicks — Because suggestions surface to users on related queries, earning clicks on one query extends your reach to its neighbors. Treating click performance as a discovery multiplier guides where to invest.