The Reality Gap: Public Guidance vs. Engineering Reality

In May 2024, internal Google API documentation, the Content API Warehouse, 2,500+ pages of Java-based protobuf specs, was published to a public GitHub repository by an automated bot called yoshi-code-bot. Rand Fishkin at SparkToro and Mike King at iPullRank published independent analyses on May 27, 2024.

Google's response was a non-denial: "We would caution against making inaccurate assumptions about Search based on out-of-context, outdated, or incomplete information." They did not call the documentation fabricated, inaccurate, or non-production. They also declined to say which elements are accurate, invalid, active today, or weighted.

Separately, the DOJ v. Google antitrust trial in October 2023 produced sworn testimony from senior engineers including Pandu Nayak (VP of Search) and Eric Lehman (Distinguished Engineer). Court documents stated that "learning from user feedback is perhaps the central way web ranking has improved for 15 years."

The table below maps the core contradictions. Each row pairs a specific public statement with the specific internal evidence that contradicts it.

Of the eight contradictions in that table, the click data row holds the widest gap. For over a decade, several Google spokespeople denied or minimized the role of click data in rankings. The leaked documentation and the sworn testimony put click data at the center of the ranking system.

NavBoost is not a machine learning model. Eric Lehman described it under oath as "essentially a large spreadsheet". The spreadsheet stores which URLs a user clicked, and how often, for each query. A long click, where the user stays, is positive. A short click, where the user returns fast, is negative. The system also tracks aging buckets, which score click performance separately by content age and show whether the engagement on a page improves or decays.

The pipeline processes click data through the CompressedQualitySignals module as crapsNewUrlSignals (URL-level), crapsNewHostSignals (host-level), and crapsNewPatternSignals (pattern-level). Google keeps both squashed (production) and unsquashed (experimentation) data. Click squashing is statistical compression. The compression stops any single dominant signal from overwhelming the rankings. The squash step gives Google its primary defense against CTR manipulation, and clicks stay a core input.

The trial also revealed Google's ranking distills to two top-level signals: Quality (Q*) and Popularity (P*). Chrome visit data and the NavBoost interaction signals power the Popularity signal. Google publicly states that it does not use Chrome browsing data for ranking.

Clicks are one half of the disclosure. Links are the other half. The public message on links moved from "one of the top three ranking factors" to "we need very few links to rank pages". The leaked documentation shows at least 11 distinct PageRank variants and a granular anchor text pipeline.

The anchor text processing is far more granular than public guidance suggests. Google maintains completely separate scoring pipelines for internal vs. external links: normalizedScoreFromOffdomain vs. normalizedScoreFromOnsite, with independent counts, volume denominators, and aggregate scores. Redirected links and fragment links each take a separate score. The system also deduplicates anchors inside a source organization, so one company cannot flood the anchor signals.

The system also analyzes full context around each anchor: fullLeftContext and fullRightContext capture all terms preceding and following the anchor text. Content position matters: inbodyTargetLink vs. outlinksTargetLink differentiates main content links from sidebar and footer links. The onsiteProminence attribute measures the importance of a page inside its own site, and Google does not take the number from a theoretical PageRank distribution. The computation propagates simulated traffic from the homepage and from the pages with high search-click volume. The simulation starts at entry points that real user behavior confirms. Google says that clicks do not matter for ranking, and clicks seed the simulation that sets internal page importance.

New pages inherit homepage authority (homepagePagerankNs) until acquiring individual PageRank. The Nearest Seeds variant seeds calculations from known-trusted pages rather than distributing uniformly. Q* measures authority as the distance from a known good source. The homePageInfo attribute then sorts a homepage into one of four trust tiers: FULLY_TRUSTED, PARTIALLY_TRUSTED, NOT_TRUSTED, and NOT_HOMEPAGE. Every page on the domain inherits from this tier, a site-wide authority gate by another name, despite Google's denials of a "domain authority" concept.

The link tiers pass one site-level judgment. The quality module passes another. Google has stated repeatedly, and through several spokespeople, that it holds no domain authority metric. The leaked documentation shows a field called siteAuthority inside CompressedQualitySignals, the pre-computed quality module that can disqualify a page before query-time ranking begins.

The contentEffort attribute carries the most weight for a practitioner. The documentation describes it as an "LLM-based effort estimation for article pages" that quantifies human labor, originality, and resources invested in creating content. The contributing factors include unique images, videos, embedded tools, in-depth content, original data, and linguistic complexity. contentEffort is the closest algorithmic proxy for the Experience dimension of E-E-A-T.

The quality system around contentEffort uses codenames that reveal a layered evaluation pipeline. chard acts as the initial content classifier, and it triggers a more rigorous E-E-A-T evaluation when it classifies a page as YMYL. rhubarb measures the quality differential between a URL and its parent site. tofu predicts site-level quality from content patterns.

Danny Sullivan stated in 2024 that Google does not have a system that says "this is a brand, let's rank it higher." The internal evidence suggests something more nuanced: Copia and Firefly monitor content velocity (the ratio of URLs generated against substantive articles produced), and directFrac measures the fraction of direct traffic to a site, a brand signal. High direct traffic may boost quality scoring. The system does not explicitly favor "brands," but the signals it measures (direct traffic, click satisfaction, authority accumulation) structurally advantage established brands.

Google publicly states they do not penalize AI-generated or automated content, only "scaled content abuse" regardless of production method. Internally, Patent US9767157B2 (N-gram Quality, granted 2017) builds phrase models from known-quality sites with 2-gram through 5-gram frequency patterns. The model measures each phrase by its relative frequency across the pages of a site. A programmatic page that produces an unnatural phrase distribution (repeated boilerplate, identical sentence structures, shallow variable substitution) takes a low quality flag. The policy ignores the production method. The detection system does not.

The Q* quality metric adds a hard limit. A site that scores below 0.4 on the 0-1 scale cannot win a rich result, whatever structured data it publishes. A Featured Snippet and a People Also Ask entry both sit behind that gate. Google publicly encourages sites to implement structured data for rich result eligibility. Internally, a hard quality gate prevents low-scoring sites from ever appearing in them.

The quality module scores the page. The E-E-A-T criteria decide what the module learns to score. Danny Sullivan and Gary Illyes have both stated that E-E-A-T is not a ranking factor or score. Sullivan: "E-E-A-T is not a direct ranking factor. It's a concept from the Quality Rater Guidelines." Illyes: Google does not have an internal E-A-T score. Mueller: "There's no single ranking factor you can point to and say it's the deciding factor."

The statement is true about direct measurement and misleading about practical impact. E-E-A-T is not a single signal. E-E-A-T is a marketing label for 80+ independent algorithmic features. Google evaluates the features at three levels: the document, the domain, and the originator entity.

The DOJ trial revealed the actual engineering pipeline: Quality Raters evaluate pages using E-E-A-T criteria from the Quality Rater Guidelines. Those evaluations become training data for Google's RankEmbed and RankEmbedBERT models. The models learn to predict rater-like quality scores at scale. E-E-A-T is not a direct ranking factor. E-E-A-T is the training objective for the quality scoring model.

Google's only white paper that confirms differential E-A-T weighting is "How Google Fights Disinformation" (2019). The paper states: "Where our algorithms detect … a YMYL topic, we will give more weight in our ranking systems to … the authoritativeness, expertise, or trustworthiness of the pages."

The API leak adds a mechanism: the internal classifier chard determines whether a page is YMYL. A YMYL classification triggers the more rigorous E-E-A-T evaluation. chard is the algorithmic gate that the white paper describes.

The YMYL gate does not stop at ranking suppression. A 16-month SE Ranking experiment tested 2,000 AI-generated articles across 20 new domains. Finance pages kept only 9 of 100 in the index at month 16. Health pages kept only 14 of 100. Broad niches (food, home, lifestyle) kept near-complete indexing over the same period.

The YMYL quality gate operates at the indexing stage, and it removes a page from the index instead of a demotion. The public framing of E-E-A-T as a ranking concept understates that operational scope.

Content effort measures the work in a page. Freshness measures when that work happened. The public guidance on freshness is short: update your content, keep it current, dates matter. The internal systems run a more careful pipeline that separates a cosmetic edit from a real content improvement.

Google keeps only the last 20 versions of a document (via urlHistory / CrawlerChangerateUrlHistory). The window of observable content evolution is finite, so a page has no permanent record of every change you make to it.

The Content Freshness Scoring patent (US8549014B2) tracks the age distribution of content inside a document: how much of it is old, and how much the author added recently. The patent also treats a sudden spike in link acquisition as a spam signal. An update that changes the age distribution of the information improves the freshness score. A date change alone does not.

Freshness measures the age of the content. A second group of attributes measures the age of the host. Google has publicly denied a "sandbox" for new sites, a deliberate trust period that holds a new domain back from its full ranking potential. The leaked documentation contradicts the denial.

The sandbox effect is real, and it works differently from the early SEO theory. No single deliberate penalty applies. Four systems disadvantage a new domain at the same time. The domain has no NavBoost history, so no click signals boost it. In the source tier system it has no accumulated backlink signals, and in CompressedQualitySignals it has no siteAuthority accumulation. The host also carries the explicit hostAge sandbox flag.

For programmatic SEO builds on new domains, the effect is direct. The first 13 months carry a structural disadvantage, whatever the quality of the content. The evidence-builder strategy answers the constraint. Win the achievable queries first, accumulate NavBoost signals, and then compete for the harder queries.

The sandbox is the last of the eight contradictions. Together they change how a practitioner allocates the budget, and they leave the public advice directionally correct: create useful content, build real authority, maintain technical health. The problem is that the advice is incomplete, and the incompleteness createsmisallocation of resources. The clearest example: "create helpful content" obscures 13 independent classifiers that each detect a specific pattern of unhelpful content.

The reality gap now extends into AI search. Google describes AI Overviews and AI Mode as complementary to organic results, but the systems cite from different source pools. Daniel Shashko's sentence-level analysis (42,971 citations, 520 queries, 6 platforms) found that AI Mode and Gemini share only 3.5% of cited domains. Kevin Indig documented a 56.6% click decline across 10 B2B SaaS sites correlated with AIO rollout. Profound's analysis of 250M+ AI responses found that traffic explains 5% of citation behavior and backlinks explain 3.8%.

The Semrush 2024 Ranking Factors study shows the same pattern on the organic side. Text relevance is the strongest ranking factor at 0.47 correlation. Domain authority follows at 0.21. The metrics that practitioners have optimized for a decade, the ones that populate every SEO dashboard, lose explanatory power across both channels.

Mike King's counter-perspective after analyzing the leak remains the most grounded summary: despite the revelations, the fundamental practice remains unchanged. "Build websites and content that people want to visit, spend time on, and link to." The leak does not change what to do. It changes why it works and how much to invest in each signal.