In May 2024, an accidental GitHub commit exposed thousands of internal Google API attributes, giving SEO professionals the clearest view yet of how Google's ranking systems actually work, and confirming several long-debated theories about clicks, domain authority, and topical focus.
In a significant revelation, leaked Google search algorithm documentation and internal API references provided SEO professionals with an unprecedented look at the mechanics behind search rankings. The leak originated from an accidental commit by yoshi-code-bot on March 13, 2024. The bot posted a link to a GitHub repository, itself a fork of the official googleapis Elixir API project.
The documents were dissected by Andrew Ansley, Director of Helium SEO, who previously analysed the Yandex leak. His findings reveal critical components and methodologies used by Google, including systems that directly contradict the company’s public statements. For SEO professionals, the implications are substantial.
u003cpu003eBoth the original and forked repositories on GitHub are public. It is highly unlikely that a Google engineer made this mistake inadvertently. The leak most probably originated from a separate internal source and was posted as an apparent original commit. The alternative, that the entire dataset is fabricated, is implausible given that it contains u003cstrongu003eover 14,000 variables and thousands of archived recordsu003c/strongu003e. A hoax of that scale borders on the impossible.u003c/pu003e
Why the leaked algorithm is a challenge for SEOs
The sheer volume of variables documented in the leak makes comprehensive tracking virtually impossible without automation and machine learning. Even accounting for only 10 to 20 percent of the disclosed signals would require significant tooling investment. That said, several high-signal components are actionable right now, and understanding them is the first step toward better rankings.
Google’s 2024 algorithm: key discoveries
Neural Search and PageRank_NS
Google has modified its classic PageRank algorithm with a neural search variant called PageRank_NS. This version enhances document understanding by incorporating neural network models, including deep learning techniques that allow Google to analyse content at a far deeper level than traditional link-graph methods. The leak documents seven distinct types of PageRank in total, including the well-known ToolBarPageRank. PageRank_NS is specifically associated with improving document comprehension and ranking accuracy.
This evolution underscores Google’s shift toward AI-driven relevance scoring. Rather than relying purely on link signals, the system now interprets the semantic meaning of content before assigning ranking weight.
Business model identification
Google’s algorithm includes specific methods to identify business models and apply tailored ranking criteria to each. The models explicitly referenced in the documentation include News Websites, YMYL (Your Money or Your Life) sites, Personal Blogs, E-Commerce stores, and Video Sites. The rationale for filtering Personal Blogs as a distinct category is not fully explained, but the mechanism confirms that Google evaluates different site types using different metrics.
| Business Model | Likely primary ranking focus | Status |
|---|---|---|
| News Websites | Content freshness, relevance, and source authority | Confirmed |
| YMYL Sites | E-E-A-T signals, authoritativeness, trustworthiness | Confirmed |
| Personal Blogs | Filtered separately; rationale undisclosed | Partial |
| E-Commerce | Transaction-related metrics, product page quality | Confirmed |
| Video Sites | Engagement, watch signals, visual quality | Confirmed |
Core ranking components: NavBoost, NSR, and ChardScores
The three most critical algorithmic components revealed in the leak are NavBoost, NSR (Neural Search Rank), and ChardScores. Each directly contradicts positions Google has taken publicly.
- NavBoost measures user engagement and rewards documents that generate better clicks. It compares performance against expected baseline values and uses internal link counts, salient terms, and click data to adjust rankings dynamically.
- NSR (Neural Search Rank) uses neural search models to rank documents based on semantic relevance and content quality.
- ChardScores predict site and page quality based on a range of signals including user behaviour and content characteristics.
Google also employs a sitewide authority metric, with traffic data from Chrome browsers serving as one of the key site-level signals. This suggests Google monitors user behaviour at a granular level to assess site credibility.
Advanced scoring: embeddings, site focus, and click data
Google’s scoring function uses page embeddings, site embeddings, site focus scores, and site radius measurements to evaluate content context and relevance. These work alongside several categories of click data to determine final rankings.
| Metric | Value | Change | What it means |
|---|---|---|---|
| Page Embeddings | Vector representation of individual page content | Used for relevance scoring | Helps Google understand semantic meaning at page level |
| Site Embeddings | Vector representation of entire domain content | Used for topical identity | Creates a topical fingerprint for the site |
| SiteFocusScore | Concentration of a site on a specific topic | Higher = better topical authority | Rewards sites with a clear topical niche |
| SiteRadius | Deviation of page embeddings from site embedding | Lower deviation = stronger focus | Penalises pages that stray from site's core topic |
| Good Clicks | Clicks followed by satisfied user behaviour | Positive ranking signal | Core NavBoost input |
| Bad Clicks | Clicks followed by immediate bounce or reformulation | Negative ranking signal | Indicates poor page-to-query match |
| Longest Clicks | Sessions where users spent significant time on-page | Strong positive signal | Proxy for content satisfaction |
Page Quality (PQ): how Google scores content effort
Page Quality (PQ) is one of the most important discoveries in the entire leak. Google uses a large language model (LLM) to estimate the effort invested in creating a page, and uses this estimate to determine whether the content can be easily replicated. High PQ scores reward pages that demonstrate genuine investment in their content.
Elements that contribute positively to PQ effort scores include proprietary tools, original images, video content, unique data, and substantive depth of information. These are also the elements proven to improve user satisfaction, which explains why the metric correlates with real-world performance.
u003cpu003eAn article that summarises findings from another source will receive a lower PQ score than the original source, even if it is longer or better written. Original research, proprietary data, and genuine editorial effort are what drive high PQ. This is a direct confirmation that thin or derivative content faces structural disadvantages in the ranking system.u003c/pu003e
Lighthouse Audit variables: 200 confirmed signals
The documentation references no fewer than 200 variables directly related to Lighthouse Audits, including LCP, INP, CLS, Accessibility, SEO, Best Practices, and AMP. This settles the long-running debate about whether Core Web Vitals and page experience signals influence rankings. They do, and the documentation confirms it explicitly.
Google has historically given contradictory public statements on this point. The leak ends that ambiguity. Page audits matter. They may not be the dominant ranking factor, but they are far from decorative. For teams looking to address these signals systematically, our Core Web Vitals service covers the full technical scope.
Topical authority: SiteFocusScore and SiteRadius confirmed
Topical authority, long theorised through reverse-engineering of Google patents, is now substantiated in the leaked documentation. Google uses four core metrics to build and measure a site’s topical identity: SiteFocusScore, SiteRadius, SiteEmbeddings, and PageEmbeddings.
The SiteFocusScore measures how concentrated a site is on a specific subject area. The SiteRadius evaluates how far individual page embeddings deviate from the site’s overall embedding. Google effectively creates a topical fingerprint for each domain, and every page is measured against it. Pages that stray significantly from the site’s core topic generate a higher SiteRadius score, which signals dilution of topical identity.
The practical implication: removing or noindexing topically irrelevant pages is not just good editorial hygiene. It is a direct lever for improving how Google perceives the site’s authority within its primary subject area.
Image quality and Host NSR
Image quality signals
Google assesses image quality using click signals that measure usefulness, presentation, visual appeal, and engagement. These metrics are part of Search CPS Personal data and contribute directly to page ranking. High-quality, relevant images that generate positive user interactions improve a page’s overall signal profile.
Host NSR
Host NSR is a site-level rank computed for host-level domain chunks. It incorporates NSR, site-level PageRank (site_pr), and new_nsr data. This chunking system allows Google to evaluate different sections of a domain independently and then aggregate them into an overall site rank. It is analogous to how Google assesses individual pages and topics, but applied at the domain architecture level.
Strategic insights for SEOs from the Mustang ranking system
Google’s new ranking system, internally named Mustang, surfaces several actionable priorities for search optimisation professionals.
Website design and NavBoost
NavBoost rewards documents that generate better user engagement through clicks. A well-structured, intuitive site architecture directly supports this by reducing friction in navigation and encouraging users to explore more pages. Investing in site design is not an aesthetic choice; it is an algorithmic one.
Content relevance and topical pruning
SEOs should audit their content libraries and remove or noindex pages that are not topically aligned with the site’s primary subject area. Each page should reinforce the site’s topical identity rather than dilute it. This applies especially to sites that have accumulated broad or loosely related content over time. Our search intent mapping service can help identify which pages belong and which should be removed.
Frequent updates and content freshness
Regularly updating content with new information, original images, and video improves both the PQ effort score and freshness signals processed by FreshnessTwiddler. Google’s indexing memory system retains the last 20 versions of a page for active consideration, meaning frequent meaningful updates can progressively improve how a page is ranked.
Indexing and memory management: Alexandria and SegIndexer
Google’s indexing system, named Alexandria, categorises documents into tiers using SegIndexer and TeraGoogle. Important, regularly updated content is stored on flash drives, while less significant content is relegated to solid-state drives or hard drives. This tiered storage structure directly affects how quickly and frequently content is re-evaluated for ranking.
Google maintains a record of every page version but primarily uses the last 20 versions for active ranking consideration. Consistent, meaningful updates can push older, lower-quality versions out of the active window, which may improve current ranking positions.
Ranking factors, demotions, and Twiddlers
Text weight and short content
Google considers the literal weight of text, meaning its visual prominence through boldness and size, as part of its scoring. Short content is not inherently penalised: it can rank well if it is substantive and not classified as thin content. Short pages operate under a distinct scoring system that rewards quality over length.
Guest posts vs. niche edits
Guest posts from newer pages receive a higher value multiplier compared to older niche edits. While niche edits remain useful, the recency of the linking page matters significantly. Fresh, relevant links from recently published pages carry more algorithmic weight than legacy insertions into older content. This has direct implications for link acquisition strategy.
Twiddlers: FreshnessTwiddler and Panda
Twiddlers are re-ranking methods applied after initial scoring to adjust results based on specific signals. The two most notable referenced in the documentation are FreshnessTwiddler, which adjusts rankings based on content recency, and Panda, which continues to function as a quality filter. These systems allow Google to dynamically refine search results without altering core ranking scores.
Three confirmed myths: domain authority, the sandbox, and Chrome data
Domain authority exists in Google’s systems
Despite repeated public denials, the leaked documentation confirms that Google calculates and uses a site authority metric. This metric is stored as part of the Compressed Quality Signals data structure. Google’s public position that it does not use domain authority as a ranking factor is directly contradicted by its own internal documentation. Building genuine entity authority therefore remains a core strategic priority.
The sandbox is real
The documentation confirms the existence of a sandbox mechanism for new websites through an attribute called hostAge, used explicitly to sandbox fresh spam. New domains are therefore subject to a probationary period during which their rankings are suppressed regardless of content quality or link acquisition.
Chrome data influences rankings
The use of Chrome browser traffic data for ranking purposes is validated in the documentation, despite previous denials by Google representatives. Chrome data contributes to sitewide authority signals and user engagement metrics used in NavBoost and related systems.
Google’s microservice architecture for search
Google’s ranking systems operate as a series of discrete microservices, each handling a specific stage of the search process. The documentation outlines components responsible for crawling, indexing, rendering, processing, ranking, and serving results. This modular architecture enables scalable handling of billions of daily queries and allows Google to update individual components without disrupting the entire system.
The leaked Content Warehouse API and the accidentally published documentation for the now-deprecated Document AI Warehouse offer insight into how Google stores and retrieves data on content, links, and user interactions at scale.
What the leak means for E-E-A-T
The leak has meaningful implications for how sites should approach Experience, Expertise, Authoritativeness, and Trustworthiness (E-E-A-T). The effects cut in both directions.
Positive implications
- Confirmed ranking factors align with E-E-A-T. Page Quality, topical authority, content freshness, and user engagement are all signals that naturally reward genuine expertise and authoritative content.
- Transparency about internal systems. Understanding how the Content Warehouse API processes content allows practitioners to structure and present information in ways that align with how Google evaluates it.
- Validation of established best practices. The leak confirms that high-quality content, consistent site architecture, and regular updates are not just advisory: they are algorithmically rewarded.
Risks to avoid
- Over-optimising for click signals. Prioritising CTR manipulation over genuine content quality can improve short-term metrics while eroding long-term trustworthiness. The leaked algorithm includes manipulation scoring.
- Misinterpreting the effort metric. The LLM-based PQ effort score is not a measure of word count. Inflating content length through padding or keyword repetition will not improve PQ scores and may actively harm user experience and E-E-A-T signals.
u003cpu003eThe impact of this leak on E-E-A-T depends entirely on how the information is applied. Used to create genuinely useful, well-structured content that satisfies user intent, it is a powerful strategic resource. Used to game individual signals in isolation, it risks triggering the manipulation scoring mechanisms that are also documented in the same leak.u003c/pu003e
How SEO professionals can act on this information
The leak provides a rare opportunity to align SEO strategy with how Google’s systems actually operate rather than how Google says they operate. The most actionable priorities are consistent across site types: invest in original content with measurable effort signals, maintain a tight topical focus, build fresh links from recently published pages, and ensure technical performance meets the 200-variable Lighthouse benchmark.
This information is most valuable for practitioners with medium-to-high technical expertise. For site owners without dedicated SEO resources, the most important takeaway is structural: a focused site with consistently high-quality content and good user experience metrics will outperform a sprawling site optimised superficially across many topics. For teams managing complex sites at scale, understanding the role of AI-driven search optimisation is increasingly essential.
For context on how these signals interact with Google’s broader quality evaluation frameworks, see our analysis of what Google’s core updates actually penalise and our guide to recovering from quality-related ranking losses.
u003cpu003eThe weight of evidence suggests it is genuine. The dataset includes over 14,000 variables and thousands of archived records, making a fabrication of this scale implausible. The most likely explanation is that real internal documentation was published accidentally via the yoshi-code-bot GitHub commit and then propagated publicly before Google could suppress it.u003c/pu003e
u003cpu003eYes, according to the leaked documentation. Despite repeated public denials, the leak confirms that Google calculates a site authority metric stored within its Compressed Quality Signals data structure. This directly contradicts official statements from Google representatives.u003c/pu003e
u003cpu003eYes. The NavBoost system explicitly uses click data, including good clicks, bad clicks, longest clicks, and site-wide impressions, to adjust rankings. This confirms what many SEOs had theorised based on patent analysis and circumstantial evidence.u003c/pu003e
u003cpu003eThe sandbox is confirmed in the documentation through an attribute called u003cstrongu003ehostAgeu003c/strongu003e, which is used to suppress rankings for new domains suspected of spam. New websites are subject to a probationary period regardless of their content quality or link profile.u003c/pu003e
u003cpu003ePageRank_NS is a neural search variant of Google’s classic PageRank algorithm. It incorporates deep learning models to improve document understanding and produce more semantically accurate ranking signals. The leak documents seven total variants of PageRank, with PageRank_NS being the most advanced in terms of content comprehension.u003c/pu003e
u003cpu003eGoogle uses SiteFocusScore, SiteRadius, SiteEmbeddings, and PageEmbeddings to construct a topical identity for each domain. Pages that deviate from the site’s core topic increase the SiteRadius score, signalling topical dilution. Sites with a tight, consistent topical focus score better on these metrics and receive ranking advantages as a result.u003c/pu003e
u003cpu003eYes. The documentation references over 200 variables directly tied to Lighthouse Audit metrics, including LCP, INP, CLS, Accessibility, SEO, Best Practices, and AMP. While these signals are not the dominant ranking factor, the leak confirms they are a meaningful and measurable part of the algorithm.u003c/pu003e
u003cpu003eAccording to the leak, guest posts from newer pages receive a higher value multiplier than older niche edits. Recency of the linking page is a meaningful variable in how link value is weighted. Niche edits retain some value, but fresh placements on recently published pages outperform legacy insertions algorithmically.u003c/pu003e