Google Search API leak: architecture and ranking systems

The May 2024 leak of the Google Search API leak warehouse documentation exposed 14,014 attributes across about 2,500 pages. It did not expose the ranking algorithm, the source code for scoring functions, or the weight of any signal.

The documentation describes data fields, internal modules, and the relationships between components. Those fields cover documents, websites, links, entities, user interactions, and the presentation of search results. They support a multi-stage model of ranking: separate systems retrieve candidates, interpret queries, adjust the order, and assemble the results page.

This draft draws on eight supplied sources. Two have limits: the HexDocs index exposes little readable content, and the Olaf Kopp article is partly paywalled. Claims from those two sources are qualified where they appear.

1. Origins, authenticity, and scope of the 2024 API leak

1.1 How the documentation became public

Internal Content Warehouse API material appeared on March 27, 2024, in a public repository tied to Google’s API client documentation. The exposure reportedly lasted until May 7, 2024.

Rand Fishkin published the original SparkToro investigation on May 27, 2024. His report covers his contact with the source, his verification steps, and an independent technical review by Mike King of iPullRank. The source, Erfan Azimi, identified himself publicly after publication.

Fishkin counted about 2,500 pages of documentation containing 14,014 attributes. Former Google employees reviewed parts of the material. They judged its structure and naming conventions consistent with Google’s internal engineering documentation.

1.2 What the documentation contains

The Content Warehouse archive describes structured data models and API interfaces. Its fields cover document properties, quality metadata, link information, and search-related behavioral data.

The archived HexDocs v0.4.0 API reference preserves the original technical names and field descriptions. Use it to check any claim that a secondary source makes about a field.

The HexDocs documentation index identifies version 0.6.1. The index exposes too little readable material to support a version-by-version comparison.

A field in the API proves that a schema defines that field. It doesn’t prove that Google populates the field, uses it in production ranking, gives it a nonzero weight, or keeps it active in 2026.

1.3 Evidence classification

The sources supply four categories of evidence, and each category establishes something different.

Seirdy’s independent critique points out that legacy attributes, experimental fields, and evaluation-only data sit beside active production components. The documentation gives no reliable way to tell them apart, so it cannot support the claim that every attribute is a ranking factor.

2. Google Search architecture: retrieval, ranking, and reranking

Google Search processes a query through several technical functions. The evidence supports the following conceptual sequence. It is an illustrative synthesis, not a verified internal Google pipeline, and several stages can overlap or run in parallel.

1. Query understanding. Interpret entities, language, and intent.
2. Candidate retrieval. Look up the index and match documents lexically and semantically.
3. Relevance and quality scoring. Apply document signals, link signals, and quality classifiers.
4. Behavioral signals and reranking. Apply NavBoost, contextual adjustments, and Twiddlers.
5. Results assembly. Combine web results, images, videos, local features, and AI answers.

Google has not disclosed the exact production order or the boundaries between these systems.

2.1 Candidate retrieval and relevance scoring

Candidate retrieval identifies documents that might satisfy a query. Relevance systems then evaluate those candidates with signals from document content, query interpretation, and other indexed properties.

The documentation references components tied to document scoring and retrieval infrastructure. Mike King’s analysis connects those references to known Google Search systems, including Ascorer. The references disclose neither a complete scoring equation nor the relative contribution of each system.

2.2 Quality evaluation

Relevance and quality are separate concepts. A page can contain the exact words of a query and still present unreliable information, duplicated content, or a poor answer.

The leak contains quality-associated attributes at both the site level and the document level. They show that Google evaluates quality in a structured way. They don’t reveal a universal quality formula.

2.3 Reranking

Reranking adjusts an existing candidate order after earlier retrieval or scoring stages. Google’s Twiddler framework works this way: a Twiddler modifies results that a preceding process already scored, and it doesn’t replace the retrieval engine.

A page’s final position can therefore reflect several adjustments, not one relevance score. Section 8 covers the Twiddler evidence and its limits.

3. NavBoost: query-document interaction signals

NavBoost uses historical searcher interactions with results for particular queries. It adds that record to the evidence from a document’s content and link profile.

Two sources support it. The 2024 documentation describes fields for clicks, impressions, and interaction classifications. Google Search executive Pandu Nayak testified publicly that NavBoost contributes to ranking.

3.1 Click classifications

The leaked documentation includes these attribute names:

The labels don’t reveal the classification rules. A long click, for example, isn’t proof of satisfaction: a searcher might stay on a page because the information is hard to understand.

3.2 Geographic and device segmentation

NavBoost-related documentation references geographic segmentation and separate device contexts. Interaction histories can therefore differ by location and by device.

A restaurant query in California might produce a different interaction history from the same query in London. Mobile searchers might prefer directions or click-to-call actions, while desktop searchers might select menus or detail pages.

For SEO, no single universal click-through rate determines rankings. The query context and the population that generates the interaction data both matter.

3.3 Why click manipulation is unreliable

Click-based ranking signals don’t establish that artificial clicks produce durable ranking gains. The documentation describes mechanisms that filter, normalize, and interpret interaction data. Google has not disclosed the exact protections.

Synthetic engagement demonstrates neither relevance nor satisfaction. Concentrate instead on three things: attract relevant searches, present accurate snippets, and satisfy the underlying information need.

4. Glue: interaction signals across the entire results page

Glue extends behavioral analysis beyond conventional organic links. According to Nayak’s testimony, NavBoost applies to web results, and Glue covers the other features of the results page. Those features include image results, videos, local listings, and direct answers.

4.1 The 2026 Resoneo analysis

Resoneo’s October 2026 investigation reports access to 137 internal schemas containing 789 fields associated with NavBoost and Glue.

The report groups the behavioral information into six families: clicks, visibility, navigation, satisfaction, subsequent search journeys, and query refinement. It also discusses fields for result exposure, scrolling, AI Overview interactions, and position-adjusted engagement.

These are Resoneo’s findings. This review has not independently authenticated the underlying schema collection.

4.2 Why raw CTR is insufficient

Placement and presentation both affect click-through rate (CTR). A result at the top of a search engine results page (SERP) gets more exposure than a result placed after advertisements, an AI Overview, and an image carousel.

Resoneo describes fields such as position_unbiased_query_goodness, view_time_on_per_pixel, and num_ads_above_top. The names suggest that the underlying systems can represent this exposure context.

Consider two hypothetical results:

Result A receives more clicks. Result B wins a larger share of the searchers who see it. Neither metric alone shows which result better satisfies the query.

Apply the same caution when you read Search Console CTR, compare SERP features, or investigate a ranking change.

4.3 Zero-click searches and AI-generated answers

A search session can end successfully without a website visit. A direct answer can supply a weather forecast, a business opening time, or a definition, so a zero-click outcome doesn’t indicate dissatisfaction.

Resoneo reports Glue-related fields for answer consumption, abandonment, subsequent query reformulation, and AI Overview interactions. The fields suggest that Google’s interaction analysis extends to answers consumed directly on the SERP.

How much these fields contribute to live ranking and AI presentation decisions remains unverified.

5. Link architecture, PageRank, and link quality

Links remain a documented part of Google’s retrieval infrastructure. The 2024 disclosure references PageRank-related attributes, anchor text, link classifications, and several forms of link metadata.

The fields show that the architecture represents link information in several ways. They don’t show the weight of any PageRank variant.

5.1 Link evaluation is not backlink counting

A backlink count measures the number of observed inbound links. It doesn’t measure their contribution to ranking. Link systems can consider source quality, context, anchor text, and spam-related information.

Fishkin’s investigation proposed that interaction data helps Google separate higher-value link sources from lower-value ones. That interpretation came partly from the leak source. Treat it as a hypothesis, not a verified description of link scoring.

5.2 Legacy attributes require caution

Seirdy’s analysis identifies older PageRank-related fields and warns against assuming that every retained attribute takes part in modern scoring. Software systems routinely keep compatibility fields, deprecated values, and experimental data structures.

The toolbarPagerank field is one example. Its presence doesn’t show that Google’s discontinued public Toolbar PageRank is a live ranking mechanism.

Link analysis is part of the documented architecture. Its weighting in 2026 is undisclosed.

6. Document quality, site-level signals, and author information

The Content Warehouse material identifies attributes for document classification, website characteristics, and author-related information. Several drew attention because they appear to describe concepts that Google representatives had discussed publicly.

6.1 Small personal websites and commercial classification

Seirdy identifies a smallPersonalSite attribute and a commercialScore field. The first classifies small personal websites. The second records whether a document has commercial characteristics.

Neither field shows that Google penalizes personal sites or rewards commercial publishers. A classifier’s output can support analysis, evaluation, or conditional processing without affecting final rankings.

6.2 Host age and registration information

The documentation references host age and domain registration metadata, so Google-related systems can represent temporal information about websites and domains.

It doesn’t prove that an earlier registration date produces a ranking advantage. Age correlates with accumulated links, historical content, brand recognition, and user demand. Separate those correlations from a causal benefit of the registration date itself.

6.3 Authors and entity consistency

The documentation supports investigating author metadata and the consistency between representations of the same document information. An author can appear in a visible byline, structured data, publication records, and other metadata.

Inconsistent information can complicate entity resolution. The leak doesn’t establish a universal author-authority score that applies to every search result.

In practice, attribute content accurately, keep identity information consistent, and state who holds editorial responsibility.

7. Human quality raters and evaluation data

Human quality evaluation is an established part of Google’s search quality processes. The leak references EWOK, an internal quality-rating environment, and fields that hold human judgments.

Fishkin identified documentation for per-document relevance ratings and human evaluation data. Some field descriptions say that the ratings are typically populated in evaluation pipelines.

That wording limits what the fields prove. Human judgments can train models, evaluate search experiments, and calibrate quality measurements without becoming per-page ranking inputs. A rating field doesn’t show that an individual rater can change a page’s live organic position.

Human assessments contribute to Google’s quality evaluation infrastructure. Their role in individual production ranking decisions is not fully disclosed.

8. Twiddlers, Superroot, and final ranking adjustments

Twiddlers adjust search results after earlier ranking processes. Olaf Kopp’s summary of the Twiddler Quick Start Guide places them within Superroot and separates them from the earlier Ascorer process.

The summary describes two types, predoc and lazy. They operate at different processing stages and carry different computational costs.

A result can therefore be relevant at first and still move because of additional rules, contextual conditions, or other ranking considerations.

This evidence has two limits. Kopp attributes the concepts to an older leaked engineering guide, reportedly from 2018, and the supplied source doesn’t independently establish that guide. The guide’s full text is also unavailable through the public page. The summary alone cannot establish execution order, implementation details, or production use in 2026.

9. What the leak changes for SEO strategy

Use the Content Warehouse documentation to sharpen technical hypotheses, not to build a checklist of 14,014 supposed ranking factors.

Mike King’s follow-up analysis in Search Engine Land makes the same argument. He reads the leak in the context of information retrieval, technical SEO, and established ranking research. He doesn’t treat isolated attributes as instructions.

The evidence supports five operational priorities.

9.1 Improve document accessibility and interpretation

Search systems must discover, render, classify, and retrieve your content before any scoring applies. Maintain crawlable navigation, accurate canonicalization, consistent internal links, meaningful document structure, and accessible primary content.

These controls address the retrieval and interpretation stages. They don’t depend on speculative scoring weights.

9.2 Match content to specific search intents

A document can match a query lexically without completing the searcher’s task. When you plan content, separate informational, navigational, commercial, transactional, and local intent.

Match the response format to the query type:

• Comparison query. Provide structured alternatives and evidence.
• Procedural query. Provide ordered instructions.
• Local query. Provide accurate location and service information.

9.3 Establish recognizable entities and demand

The leak’s references to entities and interaction data strengthen the case for examining brand recognition and navigational search demand.

Fishkin argues that recognizable brands benefit from user demand and established search behavior. His broader conclusion, that brand is the dominant SEO factor, is an interpretation and not a universal ranking law.

You can measure branded search impressions, direct visits, returning visitors, and relevant mentions. These measurements help you assess audience recognition. None of them independently proves a ranking effect.

9.4 Evaluate search satisfaction, not just traffic

A click is not the end of a successful search. Examine three things:

• Whether each page answers its intended question.
• Whether visitors complete the relevant task.
• Whether the content meets the expectation that its search snippet creates.

Google Analytics engagement and conversion metrics can help you diagnose the website experience. Don’t present them as a measurement of Google’s private NavBoost classifications.

9.5 Treat every leak-derived hypothesis as testable

A field name can motivate an experiment. It cannot substitute for one.

To validate a hypothesis, follow these steps:

1. State one specific hypothesis.
2. Record baseline performance.
3. Control for competing explanations.
4. Measure outcomes over a period long enough to cover normal ranking variation.

For example, test whether clearer title tags and more accurate page introductions improve organic CTR and task completion for a defined query group.

If performance improves, the experiment still doesn’t identify which internal Google system caused the change.

10. Confirmed evidence, plausible interpretations, and unsupported claims

The following table separates documented infrastructure from inferred ranking behavior.

The 2024 disclosure, the archived documentation, and the 2026 Resoneo investigation complement each other. They don’t share the same provenance or verification status, so keep them separate in any technical model of Google’s ranking infrastructure.

11. A research framework for future Google ranking investigations

Start each investigation with a documented component. Identify its likely function, then determine which independent observations could confirm or contradict the proposed behavior.

Three questions remain open:

1. Which data fields do production systems populate and consume? The API archive provides names and descriptions. It provides little direct evidence of execution paths in 2026.
2. How do behavioral systems interact with relevance and quality scoring? NavBoost and Glue have stronger operational evidence than most isolated Content Warehouse attributes. Their influence varies by context and remains undisclosed.
3. How have AI-generated answers changed the relationship between rankings, exposure, and satisfaction? Resoneo’s 2026 reporting suggests that the infrastructure represents interactions with AI answers alongside conventional result engagement. Establishing the causal relationship requires independent authentication and further operational evidence.

A universal ranking formula cannot be reverse-engineered from field names alone. Each of these questions can be tested.

Sources

The following table accounts for all eight sources. Accessibility describes the material available during this review.