The system models the user's current context, attention state, and predicted information need, then surfaces relevant information before a query is typed. The mechanical foundation for Discover feeds, Now cards, proactive AI assistants, and zero-query discovery surfaces.
Patent Overview
- Inventor
- Eric J. Horvitz
- Assignee
- Microsoft Corporation
- Filed
- 2001-08-15
- Granted
- August 23, 2005
The Challenge
The Challenge
Search assumes the user knows what to ask. Many of the highest-value information needs occur in moments when the user has not yet formed the query: walking into a meeting, opening a document at a deadline, arriving in a new city, starting a new task. The challenge: deliver the right information at the right moment without requiring the user to type anything, and without interrupting attention that is committed elsewhere.
- Query-First Search Misses The Moment — Per session, the user must have already noticed the need and formed a query. Many needs never reach that point and stay unmet.
- Interruption Has A Cost — Per delivery, surfacing information when attention is committed elsewhere costs more than it gives. The cost must be modeled.
- Context Is The Query — Per moment, time, location, calendar, recent activity, and active document carry enough signal to predict the information need before the user types.
- Attention Is Finite — Per surface, only a small number of proactive items can be shown without saturating the user. Selection must be ruthless.
- Wrong Proactivity Damages Trust — Per failure, a proactive surface that shows irrelevant items teaches the user to ignore the channel entirely.
Innovation
How The System Works
The system continually models the user's context and attention state, predicts the most likely information needs in the current moment, computes the expected utility of surfacing each candidate against the cost of interrupting, and delivers only the candidates whose net value exceeds the threshold.
- Observe Context Signals — Per moment, the system observes time, location, calendar, recent activity, active application, and ambient signals.
- Model Attention State — Per user, current focus, busyness, and interruptibility are estimated from the same observable signals.
- Predict Information Needs — Per moment, the system predicts the most likely information needs given the observed context.
- Retrieve Candidate Information — Per predicted need, candidate items are retrieved across documents, messages, news, search results, and structured data.
- Compute Net Value Per Candidate — Per candidate, expected user value is computed against the expected cost of interrupting the current attention state.
- Deliver Above-Threshold Items — Per surface, only candidates whose net value exceeds the surface's threshold are delivered.
- Update From Engagement — Per cycle, observed engagement and dismissal feed back into context modeling, need prediction, and threshold calibration.
Discovery Without A Query Is A Bet On Context
The patent's load-bearing idea is that the user's context already contains the query, and a well-modeled system can surface the answer before the user notices the need. The cost of interruption sets the bar that proactive items must clear.
Context-Inferred Information Need
Per moment, the most useful information is inferred from context rather than typed. Per surface, only candidates whose value exceeds the attention cost are delivered.
- Context Modeling — Per moment, time and location and activity infer the need.
- Attention Cost Modeling — Per delivery, interruption is priced against current focus.
- Net-Value Threshold — Per surface, only above-threshold items are surfaced.
Technical Foundation
Technical Foundation
The patent specifies context observation, attention modeling, need prediction, candidate retrieval, value-cost computation, surface selection, and engagement-driven update.
- Context Observation — Per moment, observable signals across time, location, calendar, active application, and recent activity are collected.
- Attention Modeling — Per user, a probabilistic model maps observed signals to a distribution over attention and interruptibility states.
- Need Prediction — Per moment, predicted information needs are scored against context and historical engagement patterns.
- Candidate Retrieval — Per predicted need, candidate items are retrieved across multiple information sources.
- Value-Cost Computation — Per candidate, expected user value is computed against expected interruption cost given the modeled attention state.
- Surface Selection — Per candidate, the delivery surface, ranging from silent log to ambient card to interruptive prompt, is selected from the value-cost profile.
- Engagement Update — Per cycle, engagement and dismissal feedback updates context modeling, need prediction, and threshold parameters.
The Process
The Process
From an observed context window, the system predicts needs, retrieves candidates, prices each against attention cost, and surfaces only the candidates that clear the threshold for the chosen surface.
- Observe Current Context — Per moment, time, location, calendar, and activity signals are collected.
- Estimate Attention State — Per user, the current attention and interruptibility distribution is computed.
- Predict Likely Needs — Per moment, candidate information needs are ranked by predicted relevance to the current context.
- Retrieve Items Per Need — Per predicted need, candidate items are drawn from multiple sources.
- Compute Net Value — Per candidate, value minus interruption cost is computed against the chosen surface.
- Deliver Above-Threshold Items — Per surface, only candidates whose net value exceeds the threshold are surfaced.
- Update From Feedback — Per cycle, engagement and dismissal recalibrate the system.
Quality Control
Quality Control
Proactive surfaces are fragile. The patent specifies safeguards to keep the channel trustworthy.
- Interruption Budget — Per session, a cap on interruptive deliveries prevents the surface from saturating attention.
- Dismissal Discount — Per item type, repeated dismissal lowers the predicted value for similar candidates.
- Quiet Mode Threshold — Per user, focus or do-not-disturb states raise the net-value threshold so only exceptional items deliver.
- Privacy Boundary — Per signal, context observation is constrained by user-controlled privacy boundaries that the system must respect.
- Surface Calibration — Per surface, ambient cards, badge indicators, and interruptive prompts use separately calibrated thresholds matched to their attention cost.
Real-World Application
Attention-aware proactive delivery is the mechanical foundation for Discover feeds, Now-style assistant cards, lock-screen previews, browser new-tab feeds, calendar-driven briefings, location-driven prompts, and proactive AI assistant suggestions. The user does not type a query. The system reads context and decides what is worth surfacing.
- Zero-query Discovery Mode — Information is surfaced without an explicit search.
- Context-driven Need Inference — Time, location, calendar, and activity infer the need.
- Attention-priced Delivery Decision — Each candidate pays an interruption cost before delivery.
Why Discoverability Extends Past The Search Box
Per surface, the search box is one of many entry points. Proactive surfaces deliver information that would never have produced a typed query. Content that earns proactive distribution reaches audiences who were not actively searching.
Why Context Signals Matter More Than Keywords
Per moment, the context that predicts a need is broader than the keywords a user might type. Entity coverage, structured timing, geographic relevance, and topical authority position content to be selected by context inference rather than query match.
<\/section>What This Means for SEO
What This Means for SEO
Attention-aware proactive delivery means discoverability is no longer bounded by the search box. SEO must produce content that proactive surfaces can confidently deliver in the right moment, without requiring the user to ask first.
- Discoverability Includes Zero-Query Surfaces — Discover feeds, Now cards, lock-screen previews, assistant suggestions, and browser new-tab feeds distribute content to users who never typed a query. SEO that optimizes only for typed queries leaves these surfaces uncaptured. Plan for both typed and proactive distribution.
- Context Match Beats Keyword Match — Proactive surfaces match content to context, not to query text. Time-of-day relevance, location relevance, calendar-event relevance, recent-activity relevance, and entity timing matter more than keyword frequency. Build content the system can place in a moment, not just retrieve from a query.
- Entity And Topic Coverage Earn Proactive Slots — Content that the system can confidently associate with entities, events, places, and topics in the user's context is selectable by inference. Generic content with weak entity grounding is not. Explicit entity markup and topical authority compound into proactive distribution.
- Timing Is A Distribution Lever — Publishing aligned with predictable context windows, scheduled events, recurring rituals, or location-driven moments raises the probability of proactive selection. The same content delivered out of context never clears the threshold.
- Trust Is The Surface Currency — A page selected for proactive distribution and then disappointing the user reduces future selection for similar items. Accuracy, freshness, and clarity protect future distribution. Misleading or stale content burns its proactive slot for the source.
- Format Affects Surface Eligibility — Proactive surfaces favor formats they can render compactly: headline plus deck, lead image plus paragraph, table plus answer, structured event card. Long-form pages without extractable structure forfeit the surface even when the content underlying is excellent.
- Zero-Query Discovery Is The Modern Distribution Layer — Discover, Now cards, assistant suggestions, and AI agent recommendations distribute content alongside the traditional SERP. The same probabilistic framework that ranks search results decides what to push proactively. SEO that ignores the proactive layer optimizes half the channel. Build for both.