Classify whether the user is hunting for a specific known item or exploring an unfamiliar topic, and serve a different kind of result page for each. The mechanical foundation for matching content format to the user's actual search phase.
Patent Overview
- Inventor
- Ryen W. White, others
- Assignee
- Microsoft Corporation
- Filed
- 2009-04-30
- Granted
- October 18, 2011
The Challenge
The Challenge
One ranker serves two very different users with the same query. A user typing a topic phrase might be hunting a specific page they have seen before, or might be entering a topic they barely understand and need an overview of. The challenge: detect which phase the user is in and adjust the result set so a lookup user gets the definitive answer fast and an exploratory user gets the overview, comparison, and breadth they need.
- Phase Is Invisible To Lexical Matching — Per query, the same query string can mean known-item lookup or open exploration depending on the user's state. Lexical matching cannot distinguish the two.
- Lookup Users Need Precision — Per session, lookup users want one definitive page. A diverse result list looks like noise to them.
- Exploratory Users Need Breadth — Per session, exploratory users want multiple perspectives, comparisons, and overviews. A single definitive page leaves them under-informed.
- One Ranking Fails Half The Audience — Per query, a single ordering optimized for one phase under-serves the other half of users issuing the same string.
- User Intent Shifts Across The Session — Per user, the same person may start exploratory and shift to lookup as they learn. A static ordering misses this evolution.
Innovation
How The System Works
The system classifies each query as exploratory or lookup using query features, session features, and user history, then selects ranking strategies and result formats tuned to the detected phase, so each user receives the result presentation that matches their actual task.
- Extract Query Features — Per query, length, specificity, presence of named entities, modifiers, and rarity are extracted as classification features.
- Read Session Context — Per session, prior queries, reformulation patterns, and dwell history indicate whether the user is exploring or refining toward a target.
- Read User History Signals — Per user, longer term interest signals indicate whether this topic is familiar or new, sharpening the classification.
- Classify Query Phase — Per query, a model labels the query as exploratory, lookup, or mixed based on the combined features.
- Select Ranking Strategy — Per phase, the ranker applies a strategy tuned for that phase. Lookup rewards precision and definitive answers. Exploratory rewards diversity, overviews, and comparisons.
- Adapt Result Presentation — Per phase, the result page format shifts. Lookup leans toward direct answers, knowledge cards, and a single dominant result. Exploratory leans toward grouped categories, multiple perspectives, and topic maps.
- Update On Phase Shift — Per session, reformulations and behavior changes can trigger a phase reclassification, and the presentation adjusts on the next query.
Same Query, Different User Phase
The patent's load-bearing idea is that the search phase, not just the query string, dictates which results win. The ranker reads phase as a first-class input and rewards the format that fits the phase.
Phase-Conditional Ranking
Per phase, the criteria for a good result differ. Per query, the ranker first identifies the phase and then applies the matching criteria.
- Query Features — Per query, specificity plus entity content plus rarity.
- Session Signals — Per session, prior queries plus reformulation pattern.
- Phase-Tuned Strategy — Per phase, separate ranking and presentation logic.
Technical Foundation
Technical Foundation
The patent specifies feature extraction, classification, phase-conditional ranking strategies, format adaptation, and reclassification on session shift.
- Query Feature Extraction — Per query, length, specificity, named entity content, presence of comparison terms, and rarity are computed as classifier inputs.
- Session Feature Extraction — Per session, prior query history, reformulation patterns, and dwell patterns provide context for classification.
- User Feature Extraction — Per user, longer term topic familiarity is summarized to refine the per-query classification.
- Phase Classifier — Per query, a trained classifier outputs an exploratory or lookup or mixed label with a confidence score.
- Phase-Conditional Ranking — Per phase, the ranker selects different scoring weights and different result diversity targets.
- Format Adaptation — Per phase, the result page format shifts between definitive-answer, grouped-categories, and comparison-matrix layouts.
The Process
The Process
From the arriving query and session context, the system classifies the phase, selects the matching ranking strategy and format, and returns a result page tuned for the user's actual search task.
- Receive Query With Context — Per query, query string plus prior session activity plus user signals arrive at the classifier.
- Compute Features — Per query, query plus session plus user features are computed in one pass.
- Classify Phase — Per query, the classifier labels the phase and emits a confidence score.
- Select Ranking Strategy — Per phase, the ranker pulls the matching scoring weights and diversity targets.
- Score Candidates — Per (query, document) pair, candidates are scored under the selected strategy.
- Select Result Format — Per phase, the presentation layer picks the matching layout.
- Reclassify On Session Shift — Per session, behavior signals trigger reclassification so the next query gets a fresh phase label.
Quality Control
Quality Control
Phase classification can mislabel queries and cause user frustration. The patent specifies safeguards to keep the experience honest when the classifier is uncertain.
- Confidence Threshold — Per query, when the classifier's confidence falls below a threshold, the ranker falls back to a balanced default rather than committing to a phase.
- Mixed Phase Handling — Per query, queries labelled mixed receive a hybrid layout with a definitive answer block above a broader exploration list.
- Reformulation Watch — Per session, if a user reformulates immediately after a phase-tuned page, the system increases the probability that the phase was wrong and adjusts.
- User Override Signal — Per user, explicit user actions to expand or collapse exploratory views are read as labels that update future classifications.
- Content Relevance Floor — Per ranker, phase logic cannot override the requirement that returned documents meet a minimum content relevance bar.
Real-World Application
Phase-aware search is the mechanical foundation behind why some queries trigger direct-answer cards, others trigger grouped result categories, and still others surface comparison and overview content. The ranker is not deciding what is true. It is deciding what kind of help the user needs at this moment.
- Per-query Phase Detection — Every query receives a phase label and confidence.
- Phase-tuned Ranking Strategy — Lookup and exploratory use separate scoring profiles.
- Format-adaptive Result Presentation — Layout shifts between definitive, grouped, and comparison.
Why The Same Query Surfaces Different Result Types
Per query, the phase label varies across users and sessions, so the same query string can return a definitive answer for one user and a topic overview for another. The ranking layer is doing more than ordering documents. It is choosing the genre of answer.
Why Format Fit Compounds Engagement
Per phase, content presented in the matching format is more likely to satisfy the task, which increases dwell, decreases reformulation, and reinforces the page's standing in destination and engagement signals downstream.
<\/section>What This Means for SEO
What This Means for SEO
Phase-aware ranking means the format of the page matters as much as the keywords on it. The ranker is choosing between definitive answers, overview content, and comparison content based on the user's phase, and only one format wins for each phase.
- Match Content Format To Query Phase — Identify whether the target query is exploratory or lookup, and choose the page format accordingly. Lookup queries want a direct answer block at the top. Exploratory queries want categorized overviews and comparisons.
- Lookup Pages Reward Precision And Speed — For lookup queries, the user is hunting one answer. Open with the definitive answer, then context. Do not bury the resolution under introductory paragraphs.
- Exploratory Pages Reward Breadth And Structure — For exploratory queries, the user is mapping the territory. Provide categories, comparisons, definitions, and onward links. Depth without structure looks like a wall of text.
- Hybrid Pages Capture Mixed Phases — When the same query string carries both phases, layer a definitive answer above an exploration scaffold. Direct lookup users land at the top. Exploratory users scroll into structure.
- Reformulation Is The Failure Signal To Avoid — If users land and immediately reformulate, the ranker reads format mismatch. Audit pages where reformulation rate is high and shift the format to match the dominant phase for the query.
- Comparison And Overview Content Owns Exploratory Queries — Pages that explicitly compare options, summarize categories, and map the landscape win for exploratory queries that single-product pages cannot satisfy.
- Phase Shifts Inside The Funnel — A user often begins exploratory and becomes lookup as they decide. Build content for both ends of the funnel so the same site can serve the user across the phase transition rather than losing them to a competitor at the decision point.