A single query is almost never a complete intent. The system identifies that a query belongs to a multi-step task spanning many queries, many sessions, and many days, then ranks for the task rather than the single query.
Patent Overview
- Inventor
- Jaime Teevan, others
- Assignee
- Microsoft Corporation
- Filed
- 2010-08-23
- Granted
- December 4, 2012
The Challenge
The Challenge
A query like 'flights to Tokyo' looks atomic to the ranker, but it sits inside a trip-planning task that also includes 'Tokyo hotels', 'JR Pass cost', 'Tokyo neighborhoods to stay', 'plug adapter Japan', and 'best time to visit Tokyo'. The challenge: detect that a sequence of queries belongs to one task, then rank against the task rather than the single query.
- Single-Query Ranking Misses The Goal — Per query, the ranker optimizes for the literal query string, not for the larger goal the user is pursuing.
- Sessions End Before Tasks End — Per session, a task often spans many sessions and many days, so session-bounded analysis cannot see the full task.
- Topic Similarity Is Not Task Membership — Per query pair, two queries can share topic and not share task, or share task and not share topic.
- Task Stage Is Invisible To The Ranker — Per query, the engine cannot tell whether the user is exploring, comparing, deciding, or executing.
- Cross-User Task Patterns Are Untapped — Per user population, common task arcs repeat across millions of users and could inform future-step prediction if the system recognized them.
Innovation
How The System Works
The system observes the user's query stream across sessions, clusters queries that belong to the same task using behavioral and content signals, classifies the task type from aggregated patterns, predicts the task stage, and re-ranks results to fit the task rather than the individual query.
- Observe Cross-Session Query Stream — Per user, queries are tracked across sessions, devices, and days rather than confined to a single browsing session.
- Compute Pairwise Task Affinity — Per query pair, features including content similarity, click overlap, temporal proximity, and dwell pattern feed a pairwise classifier.
- Cluster Queries Into Tasks — Per user, pairwise affinities are clustered into task groups using graph clustering or related methods.
- Classify Task Type — Per task cluster, the task is classified into a type such as trip-planning, purchase-research, medical-investigation, or learning-arc.
- Predict Task Stage — Per task, the user's current stage is predicted from the query sequence: exploration, comparison, decision, or execution.
- Re-Rank For Task, Not Query — Per query inside a task, the ranker promotes documents that fit the active task and stage rather than only matching the query string.
- Suggest Next-Step Queries — Per task, the system proposes follow-up queries that complete the task arc, drawn from common arcs across the user population.
The Unit Of Intent Is The Task, Not The Query
The patent's load-bearing idea is that an individual query is a fragment. The user's real intent is the task, and the task spans many queries. Ranking against the task produces better results than ranking against any single query in isolation.
Task-Conditional Ranking
Per task, ranking is computed against the multi-query task context rather than the single query string. Per query, the ranker reads the task the query belongs to.
- Cross-Session Stream — Per user, queries are observed across sessions, devices, and days.
- Task Clustering — Per user, queries are grouped into tasks by behavioral and content affinity.
- Task-Aware Re-Ranking — Per query, results are re-ranked to fit the active task and stage.
Technical Foundation
Technical Foundation
The patent specifies query stream observation, pairwise task affinity scoring, task clustering, task-type classification, stage prediction, and task-aware re-ranking.
- Query Stream Observation — Per user, queries are captured across sessions with timestamps, device identifiers, click data, and dwell time.
- Pairwise Task Affinity Features — Per query pair, features include lexical similarity, click overlap on result sets, temporal proximity, and topical drift.
- Task Clustering — Per user, pairwise affinities form a graph that is partitioned into task clusters using graph-based clustering methods.
- Task Type Classification — Per task cluster, aggregated features classify the task into a known task type or label it as a novel pattern.
- Stage Prediction — Per task, the position in the canonical arc for that task type is predicted from the query sequence.
- Task-Conditional Ranker — Per query inside a task, the ranking function reads task type and stage as inputs alongside the query string.
The Process
The Process
From a stream of queries, the system clusters them into tasks, classifies each task and stage, and re-ranks query results to fit the task rather than the single query.
- Capture User Query Stream — Per user, queries across sessions and devices are accumulated.
- Score Pairwise Task Affinity — Per query pair, an affinity score is computed from behavioral and content features.
- Build The Task Graph — Per user, the affinity scores form a graph that is clustered into task groups.
- Label Each Cluster — Per task cluster, a type label is assigned based on aggregated cluster features.
- Predict Task Stage — Per task, the user's current position in the task arc is predicted.
- Re-Rank For Task And Stage — Per query, results are re-ranked to fit the active task and the predicted stage.
- Surface Task-Completing Suggestions — Per task, follow-up queries and resources are proposed to advance the user toward task completion.
Quality Control
Quality Control
Task identification can over-cluster unrelated queries, under-cluster related ones, or misclassify the stage. The patent specifies safeguards to keep task assignment honest.
- Affinity Threshold — Per query pair, affinity must exceed a confidence threshold before queries are placed in the same task cluster.
- Temporal Window Limit — Per task, the maximum temporal span limits stale queries from being attached to an active task.
- Stage Prediction Confidence — Per task, when stage prediction confidence is low, the ranker falls back toward generic task-aware ranking.
- Cross-User Validation — Per task type label, classifier outputs are validated against aggregated cross-user task patterns to catch idiosyncratic mislabeling.
- User-Driven Resegmentation — Per user signal, explicit task-switching cues trigger task boundary resets so unrelated work does not contaminate the active task.
Real-World Application
Task identification powers task-aware ranking on web search and assistant surfaces. A user planning a Tokyo trip sees neighborhood-comparison content earlier in the task arc and booking-execution content later in the arc, even though the queries are similar at both stages.
- Cross-session Observation Window — Tasks are tracked across days and devices, not bounded by browsing session.
- Pairwise affinity Clustering Signal — Behavioral and content similarity drive task cluster formation.
- Stage-aware Ranking Output — Results are matched to task type and predicted task stage.
Why Single-Query Optimization Plateaus
Per query, content optimized for the literal query string can win that query in isolation but lose the broader task. The user who lands on a query-perfect page that does not address their next task step bounces in search of better task fit.
Why Topical Clusters Beat Keyword Lists
Per task type, a content cluster that covers the full task arc captures the user across multiple queries and stages. Single pages chasing single keywords compete for fragments while task-arc clusters earn the task.
<\/section>What This Means for SEO
What This Means for SEO
Search-task identification reframes SEO from query targeting to task ownership. The engine reads the user's task across queries and sessions; the strategy implication is to map content to the task arc, not to a single query position, and to plan for stage-fit at each step.
- Map The Task Arc Before The Keyword List — Identify the multi-query task the audience is performing. Trip-planning, purchase-research, learning-arc, and medical-investigation each have known sub-queries and stages. Ship content for every stage of the arc, not just the high-volume head term.
- Stage-Fit Wins The Click — Exploration-stage users want comparison and overview content; decision-stage users want pricing, reviews, and trust signals; execution-stage users want booking, checkout, and how-to guides. The same query at different stages should land the user on different pages, and the engine increasingly knows the stage.
- Topic Clusters Reflect Task Membership — Building a cluster of pages that covers a full task arc gives the search engine permission to rank you across the task. Single keyword targeting forfeits the task to whichever site mapped the arc more completely.
- Internal Linking Should Mirror Task Progression — Link the exploration page to the comparison page, the comparison page to the decision page, and the decision page to the execution page. The internal link graph encodes the task arc the search engine is already trying to read.
- Cross-Session Recall Compounds Ranking — Users return to topics across days when a task is active. A page that gets bookmarked, returned to, or sent to other users during a task builds the cross-session engagement signal the system is observing.
- Suggest The Next Step On The Page — If the engine surfaces task-completing suggestions, your page should too. Explicit next-step navigation captures users at the moment they would otherwise leave, and keeps the task progression on your domain.
- Stop Optimizing Pages In Isolation — A page that wins one query but does not connect to the task it belongs to is a leaky entry point. The unit of SEO planning is the task, not the page or the query. Pages serve stages within tasks, and stages are what the engine is increasingly ranking against.