Sends a query to multiple search engines in parallel, scores each engine's result list, then merges the lists into a unified ranking by their representative scores, producing meta-search results from heterogeneous sources.
Patent Overview
- Inventor
- Prabhakar Raghavan
- Assignee
- Verity, Inc.
- Filed
- 2001-08-27
- Granted
- 2004-04-27
- Application Number
- US 09/940,977
The Challenge
Multiple Search Engines, One User Need
Different search engines specialize in different content (general web, news, scholarly, image, vertical). A user query may be best served by combining results from multiple engines, but their score scales are not comparable and their result lists overlap unevenly. The system needs a way to merge multiple ranked lists into a single ordered list that respects each engine's relative confidence while presenting a coherent unified ranking.
- Engines Have Different Score Scales — Each search engine produces relevance scores on its own scale. Engine A's top score may be larger or smaller than engine B's top score with no comparable meaning. Naive concatenation produces nonsense.
- Overlap Across Engines — The same document may appear in multiple engines' result lists. The merge has to handle overlap: combine ranks, prefer one source, or deduplicate.
- Need A Common Scoring Scale — To merge, results have to be evaluated on a common scale. The patent's mechanism is to assign each entry a scoring value via a configurable scoring function, then aggregate per result list.
- Per-List Representativeness Matters — Not every result in a list is equally important. The merge weights each list by a representative value computed from its entries, capturing the list's overall confidence in its results.
- Robust To Engine Variability — Engines can return different numbers of results, with different score distributions and noise patterns. The merge has to be robust across these variations without manual per-engine tuning.
Innovation
Score Per Entry, Per List, Then Merge
The system transmits the query to a set of search engines and collects each engine's returned result list. It selects a subset of entries from each list, assigns a scoring value to each entry using a configured scoring function, and computes a representative value per list from its entries' scores. The merged result list is built by ordering entries according to their per-list representative values and per-entry scores.
- Transmit Query To Engines — Send the user query in parallel to a set of search engines. Each engine processes the query independently and returns its own ranked result list.
- Receive Result Lists — Collect the returned result lists. Each carries an internal ranking and possibly per-entry scores in the engine's native scale.
- Select Subset Per List — From each list, select a subset of entries (typically top-k or a quality-thresholded subset). The selection focuses the merge on the highest-confidence portion of each engine's output.
- Score Each Subset Entry — Apply a scoring function to each selected entry. The function produces a value on a common scale that can be compared across engines and across entries.
- Compute Representative Value Per List — From the per-entry scores, compute a representative value for the list. The representative captures the list's overall confidence (mean, top-k mean, weighted aggregate).
- Build Merged List — Order entries across all source lists by combining their per-entry scores with the representative values of their originating lists. Lists with higher representative values contribute their entries earlier in the merged order.
Meta-Search Via Two-Level Scoring
The patent introduces two-level scoring: per-entry within each list, and per-list across the set of engines. The two-level structure lets the merge respect both individual entry quality and overall list confidence.
Two Levels, One Order
Per-entry scores tell you how good a specific result is. Per-list representative values tell you how good that engine's response is overall. Both signals shape the merge order.
- Per-Entry Scoring — A configurable function that maps each entry to a value on a common scale. The function can use position, native score, content cues, or learned features.
- Per-List Representative — An aggregate over the entry scores in a list. Captures the list's overall confidence and acts as a weight in the merge.
Meta-search is robust when you score both the items and the lists they come from.
<\/section>Technical Foundation
Inputs To The Merge
The merge takes a set of ranked lists and produces a single ranked list. The scoring functions and representative-value formulas are the tunable pieces.
- Source Result Lists — Each engine's ranked output. Internal ordering and per-entry native scores are inputs to the per-entry scoring function.
- Scoring Function — Maps each entry to a comparable score. Common choices: position-based decay, native-score normalization, content-quality model output.
- Representative-Value Formula — Aggregates per-entry scores into a per-list score. Common choices: top-k mean, max-score, weighted sum.
- Merge Ordering Rule — How per-entry and per-list scores combine to produce the final order. Can be a simple product, a learned combination, or a tunable formula.
Key Insight: Meta-search merging has been studied since the 1990s. The patent's contribution is making the merge two-level: not just scoring entries, but also scoring lists. The per-list dimension allows the merge to favor engines that produced confident, coherent result lists over engines that returned scattered or low-confidence results, even when individual entries from the weaker engine score well.
<\/section>The Process
Meta-Search Pipeline
End to end, the meta-search runs in parallel queries, scoring, and merging.
- Issue Parallel Queries — Send the user query to each configured engine simultaneously.
- Wait For Result Lists — Collect responses. Apply timeout policies for slow engines so the merge does not stall on outliers.
- Per-Entry Scoring — Apply the scoring function to entries in each list. Produce comparable values across engines.
- Per-List Representative Value — Aggregate per-entry scores into a representative value per list.
- Combine And Order — Build the merged list by combining per-entry scores with per-list representatives. Deduplicate documents that appear in multiple source lists, preferring the highest-scoring instance.
- Return Merged Results — Deliver the unified ranked list to the user. Source attribution can be preserved per entry if useful.
What This Means for SEO
What This Means for SEO
Multi-engine merging is foundational to universal search results (web plus news plus images plus video in one SERP). Knowing the mechanism shapes how content should be structured across formats.
- Universal Search Composes Multiple Engines — When a SERP shows web, news, image, and video results together, multiple back-end engines have run in parallel and their lists have been merged. Optimizing for one format alone leaves merge slots on the table.
- Strong Per-List Performance Helps Your Format — Engines that consistently return high-confidence results get higher representative-value scores and contribute earlier in the merge. Content that helps an engine produce confident result lists indirectly benefits from the per-list weighting.
- Cross-Format Coverage Captures Merge Slots — If you own a topic with content in multiple formats (web article, video, image, news), you participate in multiple source engines' result lists. Each appearance is a candidate for the merge.
- Duplicate Documents Are Deduped — When the same document appears in multiple engine results, the merge keeps the best instance. Hosting duplicate versions of the same content across formats does not multiply your slots; investment in one canonical version per format is sufficient.