Scores ads by combining per-request signals (query, context) with per-user signals (profile, history) and weights items based on how previously-served ads with similar signals performed, producing relevance-aware ad selection.
Patent Overview
- Inventor
- Amit Singhal
- Assignee
- Google LLC
- Filed
- 2008-04-21
- Granted
- 2013-01-08
- Application Number
- US 12/107,124
The Challenge
Ads Need Both User And Request Context
Per-query ad targeting alone produces shallow relevance: every user sees the same ads for a given query. Per-user targeting alone produces blunt relevance: every query gets the same ads for a given user. Real relevance comes from combining both. The system needs an ad-scoring framework that reads request-side signals (current query, recent context) and user-side signals (profile, history) together, and that learns from the performance of previously-served ads to keep the combination calibrated.
- Per-Query Targeting Is Shallow — Treating every user as identical given a query misses the personalization that distinguishes ad relevance. A user about to buy versus a user just browsing should see different ads even on the same query.
- Per-User Targeting Is Blunt — Treating every query from a user as identical misses the immediate intent the current query expresses. A user’s long-term profile says little about which ad to show for this query right now.
- Combining Signals Is Non-Trivial — Just averaging or concatenating user and request signals does not produce good ad scoring. The combination has to be learned, weighted, and continuously calibrated against ad performance.
- Performance Data Must Feed Back — Without a feedback loop, the system cannot adjust which signals matter for which ad types. Past ad performance is the ground truth that calibrates current scoring.
- Decisions Are Multi-Stage — The score drives several decisions: whether to serve an ad at all, how to serve it (slot, format), which ads to order, which ads to filter. Each decision uses the score differently.
Innovation
Score Ads With User Plus Request Plus Performance Feedback
Each ad is scored using a combination of user information, request-associated information, and the ad’s own attributes. The information items are weighted based on how they performed in previous ad serving: signal types that historically correlated with successful ads weigh more in current scoring. The resulting score drives the multi-stage decisions of whether, how, and where to serve each candidate ad.
- Receive Request With Context — An ad-serving request arrives with the current query or document context and the user’s identifier. The system can fetch user information and request information from these inputs.
- Gather User Information — Pull profile attributes, interaction history, demographic signals, and inferred interests for the user. Each item is a candidate signal.
- Gather Request Information — Extract the current query terms, recent session context, geographic context, and any explicit intent signals. Each item is a candidate signal.
- Score Each Candidate Ad — For each candidate ad, compute a relevance score that combines user information, request information, and the ad’s attributes. Each signal item carries a weight.
- Weight Signals By Past Performance — Use historical data on how previously-served ads with similar signal combinations performed (click-through, conversion). Signal weights are tuned based on past correlation with ad success.
- Apply The Score To Decisions — The combined score determines whether to serve an ad at all (some queries get no ad), which ads to serve, ordering, slot assignment, and filtering of weak candidates.
- Log Performance For Feedback — Track click-through and conversion for served ads. The performance data feeds back into the signal weighting to keep the combination calibrated as user behavior and ad inventory change.
Three-Way Combination With Learned Weights
Per-user and per-request signals each cover part of the relevance picture. Combined with ad attributes and weighted by past performance, they produce ad scoring that adapts to both the immediate context and the longer-term behavioral patterns of the user.
Performance Calibrates The Combination
The weights applied to user information, request information, and ad attributes are not fixed. They are learned from how previously-served ads with similar signal profiles performed.
- User Information — Profile, interaction history, inferred interests, demographic signals. The long-term context against which immediate request signals are interpreted.
- Request Information — Current query, session context, geography, intent classification. The short-term context that anchors the ad to what the user is doing right now.
- Performance-Weighted Combination — Past ad-serving outcomes shape how much each signal contributes. The combination self-adjusts as the system observes which signals correlate with ad success.
Ad scoring is a feedback-driven combination, not a static formula.
<\/section>Technical Foundation
Signal Inventory And Weighting
The framework treats every signal as a candidate input whose contribution is learned from past ad performance.
- User Information Items — Discrete signals about the user: demographic attributes, interaction history, inferred interests, account-level preferences. Each item is a candidate input to ad scoring.
- Request Information Items — Discrete signals about the current request: query terms, session context, geography, intent class, time of day. Each is a candidate input.
- Ad Information Items — Discrete signals about each candidate ad: its targeted terms, its category, its historical performance characteristics. Each is a candidate input.
- Weight Function — How the items combine into a single ad score. Weights are learned from past ad performance: items that historically correlated with click-through and conversion get higher weights.
Quality Metrics
- Ad Score — Each f_i is a feature function over ad, user, and request information. Each w_i is the learned weight for that feature. Higher scores indicate higher predicted ad value.
score(A, U, R) = sum( w_i * f_i(A, U, R) ) - Feedback Signal — Click-through times conversion. Drives the weight updates that keep the combination calibrated.
perf(A) = ctr(A) * conv_rate(A)
Key Insight: The patent treats ad scoring as a learning-to-rank problem with explicit signal inventory and performance feedback. The contribution beyond simple combination is the loop: signals that historically produced good ads weigh more in current scoring. The system improves automatically as it observes its own outcomes.
<\/section>The Process
End-To-End Ad Selection
Per-request the runtime gathers signals, scores candidate ads, makes the multi-stage decisions, and logs performance for the feedback loop.
- Request Arrival — Ad-serving request arrives with query, user identifier, and any explicit context.
- Signal Gathering — Fetch user information and request information. Each item is keyed for later weighting.
- Candidate Ad Set — Identify candidate ads from the ad inventory based on coarse targeting (keywords, audience, geo).
- Per-Ad Scoring — Apply the weighted combination function to each candidate ad. Produce per-ad scores.
- Multi-Stage Decisions — Use the scores to decide: serve an ad or not, which ads to serve, in what order, in which slots, with what format. Filter out weak candidates.
- Render And Log — Render the selected ads to the user. Log impressions, clicks, and conversions per ad-signal combination.
- Feedback Loop — Periodically retune signal weights based on the logged performance data. The combination evolves with user behavior and inventory changes.
What This Means for SEO
What This Means for SEO
Ad scoring is adjacent to organic ranking but shares mechanics. Understanding how ads are scored reveals how the engine reads commercial intent, and gives SEO teams a window into what signals matter for paid-and-organic alignment.
- Commercial-Intent Queries Get The Most Ad Pressure — Queries the engine classifies as commercial trigger heavier ad serving. Organic results are compressed on these queries. Targeting commercial queries with organic content competes against an ad-rich SERP.
- User Signals Drive Per-Person Variability — The same query can produce different SERPs for different users because per-user signals weight into ad scoring (and influence organic ranking through related personalization). SERP analysis should account for cross-user variance.
- Request Context Includes Recent Searches — Recent session searches contribute to request information. A user’s recent activity primes which ads they see next, and by extension which kinds of content the system surfaces around them.
- Performance Feedback Favors Established Ads — Ads with strong historical performance get higher weights for related signal combinations. New advertisers face a cold-start problem. Established organic results may have an analogous advantage on signals that correlate with past success.
- Intent Classification Is The Upstream Lever — Whether a query is treated as commercial determines the ad pressure on the SERP. The commercial-query classifier (US 8046350) is the upstream decision; ad scoring (this patent) is the next step. Both shape what users actually see.
- Brand-Plus-Modifier Queries Are Ad-Rich — Queries that combine a brand with a commercial modifier ("brand discount", "brand reviews") produce high ad pressure. Organic content for these queries needs to differentiate strongly or be compressed below the fold.
- Personalization Compounds Across Surfaces — The same user signals that influence ad scoring also influence organic personalization. Investing in audience-defined content that maps to clear interest signals can pay off in both surfaces.