United States v. Google LLC is an antitrust case whose public record documents Google ranking factors: how Google Search retrieves, scores, and ranks web documents. Court filings, trial exhibits, expert testimony, and internal engineering discussions describe a ranking architecture with four parts: traditional information-retrieval methods, historical user-interaction data, document-quality assessments, and machine-learning models.
The record contradicts four common assumptions about Google’s ranking algorithms:
- Google doesn’t rank webpages with a single quality metric. Its search infrastructure processes numerous raw signals, combines them into higher-level signals, and calculates scores that influence result order.
- User-interaction data contributes directly to ranking. Navboost records query-document interactions. Glue extends interaction analysis to other search-result features.
- Machine learning doesn’t eliminate traditional ranking signals. RankBrain, DeepRank, and RankEmbed complement counting-based models, document-relevance signals, and established retrieval methods.
- Search quality depends on more than the relevance of individual webpages. Query interpretation, result freshness, device context, geographic location, search-result features, and user behavior all contribute.
These points come from six primary sources published by the U.S. Department of Justice (DOJ). The sources include the plaintiffs’ proposed findings from 2024 and 2025, the January 2025 Pandu Nayak interview notes, and Professor Douglas Oard’s November 2023 rebuttal presentation.
Read the record in four separate categories: evidence presented by litigants, witness testimony, expert interpretation, and judicial findings. Plaintiffs’ proposed findings aren’t judicial determinations. Their weight comes from the testimony and exhibits they cite, particularly where independent evidence supports the same technical point.
1. Legal background
In October 2020, the DOJ and participating states brought an antitrust action against Google. The complaint alleged that Google unlawfully maintained monopolies in general search and related search-advertising markets.
The litigation focused on Google’s distribution agreements, search-engine defaults, competitive advantages, and the relationship between search volume and product quality. One economic question ran through all of it: did Google’s access to large quantities of queries and user-interaction data reinforce its competitive position?
A search engine needs extensive infrastructure to crawl and index documents, interpret queries, retrieve information, rank results, and evaluate user satisfaction. Engineers use query logs and interaction records to improve each of those processes. Google’s distribution agreements gave it substantial search traffic. The plaintiffs argued that this scale sustained Google’s quality advantages and restricted competitors’ ability to develop comparable systems.
In August 2024, the court found that Google had unlawfully maintained monopoly power through exclusionary conduct. The finding covered two markets: general search services and general search text advertising.
The remedies proceedings then addressed how to restore competition without unnecessarily impairing search quality, privacy, or innovation. The DOJ case repository records a final judgment dated December 5, 2025. Appellate proceedings and compliance activity continued in 2026.
Primary evidence: DOJ case docket.
2. Google’s search-ranking architecture
2.1 Ranking uses multiple interacting systems
Google’s ranking architecture combines retrieval algorithms, relevance signals, quality assessments, learned representations, and post-processing systems.
Exhibit PXR0357 contains notes from a January 31, 2025 interview with Google engineer Pandu Nayak. The notes describe more than 100 raw signals and identify the following higher-level components:
| Ranking component | Function documented in the evidence |
|---|---|
| Q* | Assesses document quality |
| Navboost | Measures historical interactions between queries and documents |
| RankEmbed | Represents queries and documents using learned semantic embeddings |
| PageRank | Contributes link-based information to document-quality assessments |
| RankBrain | Supplies machine-learned ranking signals |
| DeepRank | Uses deep-learning methods to evaluate relevance |
| Twiddlers | Re-rank previously selected search results |
| Glue | Measures interactions across search-result features |
The interview also describes an internal debugging interface. The interface displays top-level signal scores and a combined ranking score identified as Final IR.
A ranking signal, then, isn’t necessarily an independent ranking factor with a fixed universal weight. Some signals are aggregates. Others are outputs from trained models. A signal’s contribution depends on the ranking stage and on the system that evaluates it.
Source: PXR0357, pp. 1-2.
2.2 Retrieval and ranking are distinct operations
The plaintiffs’ 2024 proposed findings describe a multistage process. Google’s traditional systems narrow a large set of candidate documents, and later ranking models refine the results.
A webpage must first become eligible for consideration. If Google doesn’t retrieve a document for a query, a strong score on another quality measure can’t give that document a high final ranking.
In practice, the sequence runs from query interpretation to candidate retrieval, document scoring, ranking refinement, and presentation of the results page. The disclosed evidence doesn’t establish one fixed sequence for every Google Search feature or query category.
Source: Document 839, paragraphs 181-189.
3. Navboost: direct evidence of user-interaction signals
3.1 Navboost records query-document interactions
The January 2025 Nayak interview notes describe Navboost as a traditional signal. The signal measures how frequently users click a particular document for a particular query, segmented by location and device type. According to the notes, Navboost uses the most recent 13 months of data.
The plaintiffs’ 2024 proposed findings cite testimony that Navboost records clicks on search results and retains query-document interaction information.
Historical user behavior therefore contributes to Google’s ranking infrastructure. The evidence doesn’t show that a raw click automatically raises a webpage’s ranking, or that every click carries equal weight.
3.2 Navboost depends on query context
Navboost records interactions between a specific query and a specific document. It doesn’t assign a universal popularity score to every webpage.
That structure suggests historical behavior helps Google distinguish among documents that satisfy different search intents. A webpage can perform well for one query and poorly for another, because relevance depends on the information requested.
The record supports this query-specific reading. It doesn’t disclose a complete, current production formula that converts every interaction into a final ranking contribution.
3.3 Click counts require bias correction
The Nayak interview notes describe two weaknesses of click prediction: it’s vulnerable to manipulation, and it can misrepresent user satisfaction.
A result in the first position gets more visibility than a result in a lower position. A naive click-based model can mistake that visibility advantage for superior relevance. The interview discusses Google’s efforts to address this position bias in interaction-derived signals.
Google uses click-derived information. Its ranking system is not a sort of webpages by raw click-through rate (CTR).
Primary evidence: PXR0357, pp. 1, 2, and 4; Document 839, paragraphs 182-184.
4. Glue: user behavior across the entire results page
Glue is a user-interaction system that covers more than conventional organic web links.
The 2025 remedies proposed findings separate Navboost’s query-document click records from Glue’s measurements. Glue measures interactions with other elements of the results page, including rich results, information panels, and images.
Professor Douglas Oard’s UPXD105 presentation cites testimony that Glue aggregates multiple interaction types: clicks, pointer holds, scrolling, and swipes. These signals can help determine whether a search feature appears and where Google positions it.
Why Glue matters
A Google results page isn’t an ordered list of ten webpages. It can contain maps, images, videos, direct answers, knowledge panels, and conventional organic listings. Search quality therefore depends on both document ranking and the arrangement of search features.
An evaluation that covers only traditional link positions can miss significant portions of the search experience. Oard’s presentation identifies this limitation in Google’s search-quality experiments.
The evidence shows that user-interaction data influences whole-page presentation. It doesn’t prove that Glue directly controls every feature or every result position.
Sources: Document 1370, paragraphs 622-632; UPXD105, slides 9-14.
5. RankBrain, DeepRank, and RankEmbed: machine learning in search
5.1 Machine learning complements traditional ranking
The plaintiffs’ 2024 proposed findings identify RankBrain and DeepRank as important deep-learning components. The findings cite testimony that describes these models as complementary to traditional systems.
The Nayak interview notes describe a progression from conventional relevance models, including methods associated with BM25, toward learned ranking signals.
Neural ranking models didn’t replace Navboost, PageRank, or other established methods. Google combined the approaches because each captures a different kind of information.
5.2 RankEmbed measures semantic relationships
The January 2025 interview notes describe RankEmbed as a dual-encoder model. A dual encoder represents a query and a document in a learned vector space, then evaluates their relationship with a similarity calculation. This architecture helps a search engine identify relevant documents when a query and a webpage use different wording.
The notes characterize RankEmbed as fast and effective on common queries, with potential limitations on less frequent queries.
The notes also report that Google trained RankEmbed on a sample from one month of search data. That detail describes the training configuration discussed in the interview. Don’t generalize it to every version of RankEmbed or to every Google ranking model.
5.3 Neural relevance doesn’t eliminate exact terminology
Semantic retrieval changes how search engines identify relevant documents. It doesn’t make precise terminology irrelevant, because Google combines learned representations with conventional signals and other retrieval methods.
If you publish content, describe entities, topics, attributes, and relationships clearly. The evidence gives you no reason to trade accurate titles, descriptive headings, technical vocabulary, or direct answers for vague semantic associations.
Sources: PXR0357, pp. 2-3; Document 839, paragraphs 185-189; Document 1370, paragraphs 633-635.
6. PageRank and document quality
According to the January 2025 Nayak interview notes, PageRank still contributes to quality signals for webpages. That contradicts the claim that Google has abandoned link-based analysis.
The notes leave three questions open:
- PageRank’s current numerical weight
- Whether PageRank applies equally across query types
- The complete set of link-related signals used in production
The notes also identify Q* as a document-quality measure.
Google distinguishes among several forms of relevance and quality evidence. A document can be semantically relevant to a query without being the most authoritative or useful result. A well-established document can carry strong quality signals and still fail to answer a particular query. Link-based quality information and query-dependent relevance contribute to different parts of Google’s ranking assessment.
Source: PXR0357, p. 1.
7. Search data, competitive scale, and ranking quality
7.1 User-side data contributes to search improvement
The litigation examined whether larger volumes of user-interaction data help a search engine deliver better results.
Professor Douglas Oard’s November 2023 rebuttal presentation challenged conclusions from Google’s expert, Professor Edward Fox, about the incremental benefit of user-side data. Oard argued that the experiments considered too narrow a subset of Google’s search infrastructure.
His presentation makes two specific criticisms:
- Scope. Google’s experiment retrained six components. It didn’t comprehensively measure effects on indexing, spelling correction, search features, advertising, or whole-page ranking.
- Methodology. The evaluation didn’t account for differences between mobile and desktop search presentations, among other limitations.
A limited experiment can’t establish the total value of interaction data across an entire search engine.
7.2 Search-quality metrics aren’t identical to user satisfaction
The Oard presentation discusses IS4@5, a search-quality measure that covers results in the first five positions. The presentation argues that this kind of measure can miss freshness, specialized queries, and interactions with rich search features.
A measured gain in one metric doesn’t represent an equal gain in every dimension of user satisfaction. Evaluating search quality requires relevance, freshness, coverage, presentation, and real-world behavior.
7.3 Diminishing returns don’t imply zero returns
The evidence separates diminishing marginal benefit from no benefit.
A search engine can gain substantially from its initial collection of user data and gain less from each additional observation. The smaller gains can still matter for uncommon queries, geographic variations, recent events, and specialized search intents.
The Nayak testimony cited in UPXD105 describes a tradeoff between the value of additional data and the cost of processing it. Oard’s presentation argues that limited experimental results don’t make additional interaction data competitively irrelevant.
These are expert positions presented during litigation. They aren’t a universal quantitative law for all search engines.
Source: UPXD105, slides 7-9, 17-24, and 44-49.
8. The 2025 remedies evidence: ranking systems and competitive access
Document 1370 is the plaintiffs’ remedies proposed findings, filed May 29, 2025. The filing connects search-ranking technology to the economic questions behind the antitrust remedy. It addresses search-data access, competitive scale, syndication, and the potential use of ranking information by rival search providers.
Its technical discussion identifies two broad categories of ranking systems:
- Counting-based systems. Navboost and Glue depend on historical query and interaction records.
- Machine-learning systems. These systems learn relationships between queries and documents.
The distinction has competitive consequences. Counting-based systems can depend heavily on access to large quantities of real user activity. Neural models can generalize from training examples, but their quality also depends on training data, model architecture, and evaluation methods.
The filing discusses how data-sharing arrangements might help competitors improve their search services. It also raises three concerns: confidential ranking information, intellectual property, and the ability to reproduce search behavior through machine-learning techniques.
These competing concerns explain why the remedies proceedings examined data access and search syndication in addition to distribution agreements.
To determine which proposed remedies became enforceable obligations, consult the December 2025 final judgment separately.
Sources: Document 1370, particularly paragraphs 622-639; DOJ final judgment, December 5, 2025.
9. What the evidence means for SEO
The litigation provides technical evidence for search engine optimization (SEO). It doesn’t provide a complete operational manual for ranking webpages. The strongest conclusions concern system architecture and the types of information Google processes.
In the following table, SERP means search engine results page.
| Evidence from litigation | Supported SEO interpretation | What isn’t established |
|---|---|---|
| Navboost records query-document interactions | Search intent and user response matter to ranking systems | A fixed CTR target or guaranteed ranking gain |
| RankEmbed uses semantic representations | Clear topic coverage and semantic relevance matter | That keyword matching is obsolete |
| PageRank remains a quality input | Links remain relevant to document assessment | A universal backlink quantity or weight |
| Glue analyzes search-feature interactions | Search-result presentation affects how users interact with results | That Glue controls every SERP feature |
| Query interpretation uses multiple systems | Precise entities and unambiguous language support retrieval | That adding entity names guarantees rankings |
| Ranking combines multiple signals | SEO requires attention to several distinct quality and relevance dimensions | A single universal ranking-factor formula |
9.1 Build documents around specific search intents
Answer the question that your target query represents.
Distinguish informational, navigational, commercial, and transactional intents. Two queries that contain the same keyword can ask for different things.
To define the subject and purpose of a document, use clear page titles, relevant headings, specific terminology, and complete answers.
9.2 Establish factual depth and entity clarity
Identify the central entities in your content, describe their relevant attributes, and explain their relationships accurately. For example, a technical article about Navboost must distinguish Navboost from Glue, RankBrain, and RankEmbed. Those distinctions make the document more precise and easier to interpret.
The court evidence supports the importance of query interpretation and semantic relevance. It doesn’t establish that a particular entity-writing formula changes ranking scores.
9.3 Measure search performance without mistaking correlation for causation
You can monitor impressions, clicks, query coverage, organic landing-page performance, and later user engagement. These measurements help you find pages that fail to satisfy search demand.
They don’t expose Google’s internal Navboost values. They also don’t prove that a click signal caused a particular ranking change. Treat Search Console performance data and website analytics as observational evidence, not as a direct view into Google’s ranking algorithms.
9.4 Maintain content accuracy and freshness
The evidence identifies freshness as a dimension of search quality. It also describes interaction systems that help Google respond to changing user interests.
When facts, regulations, prices, product specifications, or other material details change, update the content. Nothing in the evidence supports changing a publication date without substantive revisions.
9.5 Avoid artificial interaction manipulation
The Nayak interview notes identify manipulation as a weakness of click-based prediction.
Don’t use artificial clicks, automated search-result interactions, or manufactured engagement. The litigation provides no reliable basis for claiming that synthetic clicks can consistently manipulate Google’s current production rankings.
10. Evidence quality and interpretive limitations
The six sources have different evidentiary roles.
| Source | Evidentiary role | Principal limitation |
|---|---|---|
| DOJ case page | Official litigation chronology and document repository | Doesn’t independently prove every claim in linked filings |
| DOJ trial exhibits index | Identifies publicly released trial materials | An index isn’t substantive testimony |
| Document 839 (2024) | Plaintiffs’ proposed factual findings supported by cited evidence | Advocacy filing, not the court’s adopted findings |
| PXR0357 (2025) | Contemporaneous notes of a discussion involving Pandu Nayak | Not a verbatim certified transcript, and the public version is redacted |
| Document 1370 (2025) | Plaintiffs’ remedies-stage factual proposals | Includes disputed assertions and proposed interpretations |
| UPXD105 (2023) | Expert rebuttal presentation by Douglas Oard | Expert advocacy and analysis, not a neutral technical specification |
The records also reflect different periods in Google’s engineering history. Google may have modified, retrained, or replaced a system after the 2023 or 2025 document that describes it. A historical description of an algorithm can’t establish its exact configuration in November 2026.
Public redactions further limit what anyone can reconstruct about Google’s proprietary ranking infrastructure.
To test a technical claim, corroborate it across witness testimony, contemporaneous exhibits, and judicial findings. Keep those three categories separate when you report the result.
11. Conclusion
The disclosed record supports five conclusions:
- Google’s ranking architecture combines traditional retrieval methods, user-interaction signals, document-quality assessments, and machine-learning models.
- Navboost records query-document interaction information. Glue extends interaction analysis to other search-result features.
- PageRank contributes to document-quality signals. Neural models complement traditional ranking systems and don’t universally replace them.
- User-side data contributes to search quality, evaluation, and engineering decisions. Its marginal value varies across systems and query categories.
- Search competition depends partly on access to queries, user interactions, distribution channels, and the infrastructure required to build and assess search services.
The record also limits what outside observers can claim. The public filings don’t disclose Google’s complete current ranking formula. They don’t validate a universal list of ranking weights. They don’t prove that clicks, links, semantic similarity, or any other single signal independently determines search position.
If you research SEO, use the litigation to replace unsupported ranking theories with documented descriptions of Google’s information-retrieval architecture. If you research antitrust, the same evidence explains why search data, distribution, and technical infrastructure became central issues in the case.
Primary legal sources and exhibits
1. U.S. Department of Justice: United States and Plaintiff States v. Google LLC (2020). Official case repository. It contains litigation documents, trial materials, remedies proceedings, the December 2025 final judgment, and later appellate filings.
https://www.justice.gov/atr/case/us-and-plaintiff-states-v-google-llc
2. U.S. Department of Justice: trial exhibits index. Official index of publicly released exhibits, including technical presentations, internal Google documents, and demonstrative evidence.
https://www.justice.gov/atr/us-and-plaintiff-states-v-google-llc-2020-trial-exhibits
3. Plaintiffs’ Remedies Proposed Findings of Fact, Document 1370, May 29, 2025. A 261-page remedies-stage filing. It addresses search technology, user-side data, competitive scale, generative AI, and proposed remedies.
https://www.justice.gov/atr/media/1402141/dl?LanguageId=1
4. PXR0357: January 31, 2025 call with Google engineer Pandu Nayak. A four-page public exhibit. It describes ranking signals, Navboost, RankEmbed, PageRank, query interpretation, and internal debugging methods.
https://www.justice.gov/atr/media/1398866/dl?inline=
5. Plaintiffs’ Proposed Findings of Fact, Document 839, February 2024. A 509-page filing. It addresses market definition, search competition, distribution agreements, user-interaction data, and Google’s ranking systems.
https://www.justice.gov/usdoj-media/atr/media/1340251/dl?inline=
6. UPXD105: rebuttal testimony of Professor Douglas W. Oard, November 15, 2023. A 50-slide presentation. It examines the value of user-side data, limitations of search-quality experiments, Navboost, Glue, and ranking-model evaluation.
https://www.justice.gov/d9/2023-11/418021.pdf
Additional judicial records
Final Judgment, December 5, 2025.
https://www.justice.gov/atr/media/1421546/dl?inline=
Remedies Memorandum Opinion, December 5, 2025.
https://www.justice.gov/atr/media/1421681/dl?inline=
Research note: This draft synthesizes the six supplied sources and identifies relevant sections, testimony, and exhibits. It is a source-grounded analytical draft. It is not a paragraph-by-paragraph audit of the approximately 800 pages of underlying PDFs, or of every document linked from the DOJ exhibit index.