Aggregate the sequence of queries and clicks across many users to find which destinations actually ended successful searches. The mechanical foundation for being the page at the end of the journey, not the page that catches the first click.
Patent Overview
- Inventor
- Ryen W. White, others
- Assignee
- Microsoft Corporation
- Filed
- 2007-06-27
- Granted
- March 8, 2011
The Challenge
The Challenge
Per-query ranking treats every search as an isolated event. A user typing a broad query usually does not stop there. They issue follow-up queries, click results, refine, and eventually land on the page that actually solved the task. Conventional ranking sees none of this trajectory. The challenge: capture the trail of queries and clicks that successful past users walked, and use that trail to rank for the queries earlier in the journey.
- Isolated Queries Miss The Journey — Per query, ranking treats each search as standalone, ignoring that real users issue multiple queries before they reach the page that satisfies them.
- First Click Is Not Success — Per session, the first clicked result is often a stepping stone rather than the destination. Counting first clicks as success rewards stepping stones over destinations.
- Reformulation Carries Hidden Signal — Per user, the way a query is reformulated after a click tells the system whether the click satisfied the user, but per-query ranking cannot read that signal.
- Destination Pages Get Under-Credited — Per document, a page that ends successful trails accumulates strong evidence of satisfaction even when it does not match the initial query closely.
- Broad Queries Lack Direct Answers — Per query, broad or exploratory queries do not have a single correct result. The right answer is the page where trails end, not the page that lexically matches the query.
Innovation
How The System Works
The system records query trails from many users, identifies trails that ended in apparent success based on dwell time and trail-end signals, aggregates the destination pages of successful trails, and uses those destinations to rank results for queries that appear earlier in the same trail patterns.
- Record Per-User Query Trails — Per user, the sequence of queries, result clicks, page dwell times, and reformulations is logged as an ordered trail.
- Identify Trail Boundaries — Per session, the system detects where one task ends and another begins so trails reflect intent units rather than arbitrary time windows.
- Score Trail Success — Per trail, end-state signals such as long dwell, no further reformulation, and absence of return-to-results indicate the trail ended successfully.
- Aggregate Across Users — Per query pattern, trails from many users are aggregated to identify common successful destinations that emerge above noise.
- Build Query-To-Destination Map — Per query, the map records which destination pages typically end successful trails that began with or passed through that query.
- Rank Using Destination Evidence — Per query, candidate documents are scored partly on how often they appear as successful destinations for that query or for trails passing through it.
- Surface Trail Suggestions — Per query, the system can also suggest reformulations or onward queries by reading the trails of users who succeeded from this query.
Destinations Beat First Clicks
The patent's load-bearing idea is that the value of a page is established by being the place where searches end, not by being the place where searches start. The ranker reads the whole journey, not the click that opened it.
End-Of-Trail Authority
Per trail, the page where the trail ended in success carries stronger evidence of satisfying the user's underlying intent than any page clicked along the way.
- Trail Recording — Per user, sequence of queries plus clicks plus dwell.
- Success Scoring — Per trail, end-state dwell and absence of reformulation.
- Destination Aggregation — Per query pattern, common successful destinations.
Technical Foundation
Technical Foundation
The patent specifies trail capture, boundary detection, success scoring, cross-user aggregation, destination mapping, and integration into the ranker.
- Trail Capture — Per user, query strings, clicked result URLs, dwell times, scroll, and reformulation events are logged in order.
- Boundary Detection — Per session, topic shifts, long idle gaps, and explicit new-task signals partition raw activity into bounded trails.
- Success Scoring — Per trail, the system uses dwell on the final page, absence of subsequent reformulation, and exit signals to score whether the trail ended successfully.
- Cross-User Aggregation — Per query and per query-pattern, destinations are aggregated across many users to surface stable successful endpoints.
- Destination Mapping — Per query, the system stores a ranked list of pages that frequently appeared as the end of successful trails passing through that query.
- Ranker Integration — Per ranker, destination evidence combines with content relevance, link signals, and population signals to produce the final ordering.
The Process
The Process
From raw user activity, the system extracts trails, scores them, aggregates destinations, and feeds destination evidence into the ranker so queries earlier in the journey surface the pages that ended successful trails.
- Log Raw Activity — Per user, queries, clicks, dwell, and reformulations are streamed into the trail log.
- Detect Trail Boundaries — Per session, the system partitions raw activity into trails representing single tasks.
- Score Trail Success — Per trail, success is inferred from end-state dwell and absence of further reformulation.
- Aggregate Successful Destinations — Per query and per query pattern, destinations of successful trails are accumulated.
- Build Destination Index — Per query, the destination index records which pages most often ended successful trails.
- Score Candidates Against Destinations — Per (query, document) pair, candidates are credited for being known successful destinations.
- Blend Into Final Ranking — Per ranker, destination scores combine with content and link signals to surface the trail-validated pages.
Quality Control
Quality Control
Trail-based ranking can be distorted by spam trails, popular but unsatisfying destinations, or thin data. The patent specifies safeguards to keep destination evidence honest.
- Minimum Trail Volume — Per query, destination evidence is only used after enough independent trails have been observed to be statistically meaningful.
- Dwell Floor — Per trail, success requires meaningful dwell on the final page so that quick exits cannot inflate destination scores.
- Reformulation Penalty — Per trail, if the user reformulates and continues searching after the candidate destination, that trail is not counted as success.
- Spam Trail Filtering — Per user, behavioral patterns inconsistent with genuine task completion are filtered out before aggregation.
- Content Relevance Floor — Per ranker, destination evidence cannot promote a document that fails minimum content relevance for the query.
Real-World Application
Trail-based ranking is the mechanical foundation for surfacing the page that ends the journey rather than the page that catches the first click. For broad or exploratory queries where no single page lexically matches, trail evidence is what lets the ranker point users to the destination that actually satisfies the underlying intent.
- Whole-trail Unit Of Analysis — Queries, clicks, and dwell across the entire task arc.
- Cross-user Aggregation — Destinations stable across many independent users.
- Destination-weighted Ranking Signal — End-of-trail evidence augments content relevance.
Why Destination Pages Earn Compound Prominence
Per query pattern, a page that ends successful trails for an upstream query gathers strong destination evidence that lifts it for the upstream query even when its content does not lexically match. The reward compounds because more visibility yields more trails, and more trails reinforce the destination signal.
Why Broad Queries Reward The End Of The Journey
Per query, broad or exploratory queries lack a single correct answer at the lexical level. The ranker resolves them by reading where users actually ended up, which converts a vague initial query into a specific successful destination.
<\/section>What This Means for SEO
What This Means for SEO
Trail-based ranking means the page that ends the journey wins, even if it is not the page that lexically matches the first query. SEO strategy must aim at being the destination of the task, not the catch-point of the first click.
- Aim For The End Of The Trail — Identify the resolution page that users actually want when they begin a multi-query task. Build that page so completely that no further reformulation is needed. The system rewards endpoints, not stepping stones.
- Catching The First Click Is Not Winning — A page that ranks high for a broad initial query but sends users back to refine is a stepping stone. The ranker learns to demote stepping stones over time and promote the destination.
- Long Dwell Without Reformulation Is The Success Signal — Design pages so the user can complete the task on the page without returning to search. Long dwell paired with no further reformulation is the explicit positive signal the system reads.
- Map The Trail That Leads To Your Page — Audit which upstream queries and intermediate pages typically precede arrivals at your page. Build internal links, suggested next steps, and topical scaffolding that mirror that natural trail.
- Reformulation Carries Negative Evidence — When users land on your page and immediately reformulate, the system reads that as failure. Surface the answer above the fold, match the query language, and resolve the task before the user gives up.
- Broad Queries Are Resolved By Destinations — For broad queries with no lexical winner, the ranker reads where users finally land. Position your page as the natural endpoint for several upstream queries rather than the lexical match for one.
- Aggregated Trails Beat Single-Session Behavior — The system reads patterns across many users, not single sessions. Sustained satisfaction across a broad audience accumulates into a destination signal that is hard to dislodge with short-term tactics.