The first personalized search algorithm shipped at Bing. Per-user relevance is computed by matching the query and candidate documents against a profile assembled from the user's history, behavior, and content interests, then re-ranking the global result set against that profile.
Patent Overview
- Inventor
- Jaime Teevan, Susan T. Dumais, Eric J. Horvitz
- Assignee
- Microsoft Corporation
- Filed
- 2005-11-15
- Granted
- April 6, 2010
The Challenge
The Challenge
A global ranking treats every user as the average user. The same query from a software engineer in Seattle and a high-school student in Manila returns the same ten links. The challenge: personalize the ranking so each user sees results matched to their interests, history, and context, without rebuilding the index per user and without breaking the global relevance signals the engine depends on.
- Average-User Ranking Wastes Specificity — Per query, a single global score buries the result that is perfect for this user underneath results that are average for everyone.
- Ambiguous Queries Stay Ambiguous — Per query string, short queries carry many intents, and a global ranker has no way to know which intent this user actually holds.
- User History Sits Unused — Per user, prior queries, prior clicks, and prior content interactions reveal interests that the ranker ignores when it scores the next query.
- Per-User Indexing Does Not Scale — Per session, rebuilding the global index for each user is computationally impossible, so personalization must happen at the ranking layer.
- Privacy And Latency Both Matter — Per request, the profile must be applied without exposing user history server-side as a leak risk and without slowing the SERP perceptibly.
Innovation
How The System Works
The system builds a user profile from documents the user has read, queries the user has issued, and content the user has authored or saved. At query time the engine retrieves a global candidate set and re-ranks those candidates by similarity between each candidate and the user profile, producing a personalized ordering on top of the global ranking.
- Assemble The User Profile — Per user, profile data is collected from query history, click history, viewed documents, authored content, and stored files.
- Represent The Profile As Term Vectors — Per user, the profile is converted to a weighted term vector with terms drawn from the user's content corpus.
- Retrieve The Global Candidate Set — Per query, the engine retrieves candidate documents using the standard global ranker.
- Score Candidate-Profile Similarity — Per candidate, similarity between the candidate's term vector and the user profile vector is computed.
- Combine With Global Relevance — Per (query, candidate) pair, personalized similarity is blended with the global relevance score to produce a personalized score.
- Apply User Control — Per user, the personalization weight is adjustable, including the option to disable personalization for ambiguous queries or by explicit setting.
- Return Personalized Ranking — Per query, the SERP reflects both the global signal and the user-specific match, with personalization applied on top of a known-good baseline.
Profile-Conditional Relevance Is The Default Layer
The patent's load-bearing idea is that personalization belongs at the ranking layer, not the indexing layer. A single global index can serve every user when each user's profile re-scores the candidate set against their own term vector.
Re-Rank, Do Not Re-Index
Per user, personalization happens by re-scoring the global candidate set against a user profile vector. Per engine, the index is shared across all users.
- User Profile Vector — Per user, term weights are derived from history, clicks, and content interactions.
- Candidate Similarity Scoring — Per candidate, similarity to the user vector is computed against a content vector.
- Blended Personalized Score — Per query, global relevance is combined with personal similarity into a final score.
Technical Foundation
Technical Foundation
The patent specifies profile construction, profile representation, candidate retrieval, similarity scoring, score blending, and user-control mechanics.
- Profile Source Aggregation — Per user, signals are pulled from query history, click history, document views, authored content, and saved or bookmarked items.
- Term Vector Construction — Per user, content terms are weighted by frequency, recency, and source, yielding a vector representation of the user's interests.
- Candidate Document Vectors — Per candidate, a comparable term vector is computed for similarity scoring against the user profile.
- Similarity Function — Per (user, candidate) pair, cosine similarity or an equivalent measure produces the personalization score.
- Score Blending Function — Per ranker, the global relevance score and the personalization score are combined with tunable weights.
- Client-Side Profile Option — Per privacy mode, the profile and similarity computation can run on the client to avoid sending user history to the server.
The Process
The Process
From a query, the system retrieves global candidates, looks up the user's profile vector, scores candidate-profile similarity, blends with the global relevance score, and returns a personalized ranking on the SERP.
- Receive Query And User Identity — Per query, the query string arrives with a session identifier or signed-in user identifier.
- Retrieve The User Profile — Per user, the stored profile vector is loaded from the profile store or constructed on the fly from session history.
- Run The Global Ranker — Per query, candidate documents are pulled using the standard global relevance pipeline.
- Score Profile Similarity Per Candidate — Per candidate, similarity to the user profile vector is computed.
- Blend Global And Personal Scores — Per (query, candidate) pair, the global and personal scores are combined with tunable weights.
- Apply Personalization Confidence — Per query, when the profile coverage is thin or the query is ambiguous in a way the profile does not resolve, the personalization weight drops.
- Return Personalized SERP — Per query, the user sees a ranking that reflects both global relevance and their own interest profile.
Quality Control
Quality Control
Personalized ranking can collapse into a filter bubble, surface stale interests, or break for new users with no history. The patent specifies safeguards to keep results honest.
- Profile Coverage Threshold — Per query, personalization is applied only when the user profile has enough signal to make a confident similarity judgment.
- Global Relevance Floor — Per ranker, no document can be promoted by personalization above a candidate that has substantially higher global relevance and acceptable personal similarity.
- Recency Weighting In The Profile — Per user, recent interactions are weighted higher than long-ago interactions so stale interests do not distort current rankings.
- Ambiguity Detection — Per query, when the query is highly ambiguous and the user profile pulls strongly toward one interpretation, personalization is applied with confidence; when the profile is silent, global ranking dominates.
- User Override — Per user, personalization weight is user-controllable, including the ability to disable it for a session or for specific query types.
Real-World Application
Personalized search shipped at Bing as the production deployment of this work. The same query string produces materially different rankings for two users when their profiles diverge on the topics in the candidate set. The same ten links that the global ranker returns are reordered against the user's term vector before the SERP renders.
- Per-user Profile Granularity — Every signed-in user carries their own profile vector.
- Re-rank only Architectural Choice — The global index is shared; personalization is a re-ranking layer.
- Blended scoring Final Ranking — Global relevance and personal similarity combine into the displayed order.
Why The Search Engine Sees Each User As A Vector
Per user, the profile is not a category label like demographic or persona. It is a high-dimensional term vector built from what the user actually reads, asks, and writes. The ranker compares documents against that vector directly, which means matching a real user beats matching a hypothetical persona.
Why Personalization Is The Modern Baseline
Per engine, every major search system since 2006 ships some flavor of profile-aware ranking. The single global SERP that everyone saw a decade earlier is a historical artifact. Two users typing the same query will see different rankings, and SEO has to account for both.
<\/section>What This Means for SEO
What This Means for SEO
Personalized re-ranking means the SERP that a tracker reports is not the SERP the user sees. Strategy has to be built around the user's interest profile, not just the query string. Two users typing the same query get different rankings, and the engine reaches that result by matching the user's history against your content.
- Same Query, Different SERPs — Position tracking that pulls a single SERP per query understates ranking variability. The user who has read three articles on your topic sees your page ranked higher than the user who has never engaged with the topic. Plan campaigns for both audiences and accept that one number cannot describe both rankings.
- Topical Consistency Builds Profile Match — If a user repeatedly engages with content in your niche, their profile vector pulls toward your terminology. The next time they search a related query, your page scores higher on personal similarity even without ranking better on global relevance. Returning visitors are not just retention; they are a ranking signal.
- Brand And Author Recurrence Compound — Profile vectors absorb terms the user has read repeatedly, including brand names and author names. A user who has read your author across three articles carries those terms in their vector. The next article ranks higher for them even when content overlap is modest.
- Match The Vocabulary Your Audience Already Uses — The profile is built from terms the user actually reads. If your page uses different vocabulary than the rest of the user's reading history, the similarity score is lower. Mirror the terminology the target audience already encounters, not the terminology you wish they used.
- Cold-Start Users See The Global Ranking — When the profile is thin, the engine falls back toward global relevance. New users, incognito users, and first-time visitors land on the unpersonalized ranking. Global signals like authority, freshness, and content depth are still the table stakes, especially for top-of-funnel acquisition.
- Audience Modeling Belongs Upstream Of Keyword Research — Keyword research selects queries. Audience modeling selects the profile vectors those queries will be scored against. Choose the audience first, then choose the queries that audience will issue, then build content that scores high on both global relevance and similarity to that audience's reading history.
- Personalization Is Permanent, Not Optional — Every signed-in or logged-state user gets profile-aware ranking. The strategy implication is to stop expecting one SERP and start planning for a distribution of SERPs across the audience segments that matter, with content depth and topical consistency as the levers that compound across the segments.