Topicality in Google refers to how relevant a webpage’s content and topic are to a user’s search query. In Google’s search ranking systems, topicality (often designated as T* or T-Star) is one of the two core foundational signals – paired alongside Q* (Quality) – used to decide which web pages show up in search results.
Search engine optimisation practitioners have historically treated relevance as an abstract concept, reverse-engineering an opaque algorithmic black box through empirical observation. That era of speculation ended with two watershed disclosures: the sworn testimonies and internal exhibits unsealed during the United States et al. v. Google LLC federal antitrust litigation (Case No. 1:20-cv-03010-APM) and the May 2024 leak of Google’s internal Content Warehouse API documentation.
Together, these primary sources establish what the Hobo Technical SEO framework defines as a “New Canon of Truth,” transforming search optimisation into an evidence-based systems engineering discipline.
“Topicality” was addressed with surprising specificity in the DOJ antitrust trial. Far from being an abstract concept, it was revealed to be a formal, engineered system within Google, designated internally as T* (pronounced “T-star”).
Operating alongside the Quality Score (Q*) system, the explicit function of the T* engine is to compute a document’s fundamental, query-dependent relevance. It serves as a “base score” that answers the foundational question: How relevant is this document to the specific terms used in this search query? Whilst Q* assesses query-independent trustworthiness (“Is this source trustworthy?”), the Relevance System (T*, powered internally by the Goldmine scoring engine) calculates dynamic, query-dependent relevance in real time.
Rather than operating as a single monolithic equation, Google Search functions as a multi-stage pipeline where distinct subsystems filter, score, and re-rank candidate documents. In the initial phase, after the Trawler crawling system fetches web documents and indexing tiers such as Alexandria and TeraGoogle store them, the primary scoring engine codenamed Mustang retrieves tens of thousands of candidate documents matching a query. To narrow this massive pool down to a few hundred viable results, Google computes each document’s T* base score by combining three foundational signal classes known internally as the ABC signals: Anchors (A), Body (B), and Clicks (C).
Furthermore, disclosures from the remedies phase of the DOJ antitrust trial (specifically the September 2, 2025 Remedies Opinion, Document 1436, and Trial Exhibit PXR0356) demonstrate that the components of Topicality (T*) also serve as essential building blocks for Google’s two overarching top-level pillars: Quality (Q*) and Popularity (P*). Whilst the Body (B) component supplies content-derived metrics that feed into the static Quality (Q*) score alongside PageRank (PageRank-NearestSeeds) and siteAuthority, the Anchors (A) and Clicks (C) components feed directly into the dynamic Popularity (P*) signal alongside real-world Google Chrome browsing telemetry (chromeInTotal).
How Was The Topicality (T*) System Revealed?
Whilst SEOs coined the content marketing concept of “Topical Clusters” around 2017 to describe hub-and-spoke site architecture, the actual algorithmic ranking engine inside Google that calculates topicality (T*) was built strictly by Google’s own computer scientists and software engineers.
The existence and exact mechanics of Google’s formal Core Topicality System (T*) reached the public through a three-stage chain of engineering creation, federal legal compulsion, and technical SEO translation:
1. The Engineers and Product Directors Inside Google
- Hyung-Jin “HJ” Kim (The Engineer): The key figure behind mapping and scoring algorithmic topicality is Hyung-Jin “HJ” Kim, a prominent Google Search Vice President and engineer (the creator of
Navboostand head of the Page Quality team). In internal DOJ interview notes (Trial Exhibit PXR0356), Kim explained that Google’s engineers worked out how to calculate baseline query-document relevance (T*) by combining the “ABC Signals” – Anchors (link text), Body (on-page terms), and Clicks (user dwell time/engagement) – using hand-crafted mathematical formulas. - Elizabeth Tucker & Liz Reid (The Product Executives): Whilst engineers built the underlying maths, Elizabeth Tucker, Director of Product Management at Google (and formerly of Google Maps), was the official who brought the exact phrase “Core Topicality Systems” into public SEO awareness. Speaking on Google’s Search Off the Record podcast (and echoed by VP of Search Liz Reid), Tucker openly discussed her oversight of these systems, explaining that the core topicality space is not a single isolated algorithm, but a massive constellation or “galaxy” of topic-related systems and machine learning models (like
BERT) working together to map out the meaning behind search terms and mitigate biases.
2. The Public Leak (The U.S. Department of Justice Antitrust Trial)
The public did not learn about the internal mechanics of T* because an external marketer reverse-engineered Google’s code. Instead, the U.S. Department of Justice (DOJ) forced Google to hand over internal blueprints, slide decks (such as Dr. Eric Lehman’s “Life of a Click” presentation), and engineer testimonies during the U.S. v. Google LLC antitrust trial. On May 2, 2025, the summary notes of the DOJ’s February 18, 2025 interview with HJ Kim (Trial Exhibit PXR0356) were uploaded to the public DOJ court docket.
3. The SEO Researchers
Once the U.S. government uploaded thousands of pages of complex legal exhibits and technical data to the public docket, technical SEO researchers stepped in as translators. As industry coverage documented, I originally surfaced the DOJ summary of HJ Kim’s interview (Exhibit PXR0356) from the court filings on May 2, 2025, and alongside leading algorithmic analysts, translated what Google’s engineers had built for website owners.
| Analyst & Publication | Primary Analytical Focus | Core Technical Contribution to Understanding T* |
| Shaun Anderson – me (Hobo Web) | Site Architecture, Digital Dossiers & Strategic SEO | I originally surfaced the DOJ PXR0356 HJ Kim summary. Bridges computer science and actionable SEO strategy by deconstructing the CompositeDoc and PerDocData “digital dossier” Google maintains on every URL. Frames the architecture as a strict split between Query-Independent Authority (Q*) (“Is this source trustworthy?”) and Query-Dependent Relevance (T*) (“Is this document relevant to this specific query?”), mapping leaked domain embeddings like siteRadius and siteFocusScore. |
| Natzir Turrado (Natzir.com) | Algorithmic Maths, Curves & Vector Models | Models the conceptual mathematical weighting of T* (Topicality) against Q* (Quality) and Navboost via the analytical heuristic final_score = α · topicality + β · quality + γ · navboost. Explains how Google integrates deep-learning vector models (RankEmbed) to measure query-document similarity alongside pre-computed mathematical curves (sigmoids) and thresholds to balance real-time relevance with processing speed. |
| Roger Montti (Search Engine Journal) | Conceptual Meaning & Semantic Intent Mapping | Synthesises public statements from Google’s Elizabeth Tucker and Liz Reid on “Core Topicality Systems”. Warns against “reductionist thinking” (such as mechanically stuffing keywords into headings) and explains how modern retrieval engines evaluate semantic meaning and user intent via models like BERT in a post-keyword era. |
Does Google Disclose the Topicality (T*) System on Its Public Ranking Systems Page?
A critical question for SEO practitioners is whether Google openly documents the Topicality (T*) engine on its official public documentation hub, A Guide to Google Search Ranking Systems (introduced by Google Search Central in November 2022 to catalogue its “notable” ranking systems).
An inspection of that official guide reveals a striking omission: Google does not mention T* (Topicality), Q* (Quality), P* (Popularity), Mustang, or Navboost anywhere by name on its main public disclosure page. Instead, Google’s public-facing documentation only lists specific sub-components, AI language models, and specialised Twiddler adjustments, including:
- Language and Semantic Models:
BERT(Bidirectional Encoder Representations from Transformers),MUM(Multitask Unified Model),Neural matching, andRankBrain. - Content & Domain Adjustments: The
Helpful content system(integrated into core ranking systems in March 2024),Original content systems,Passage ranking system,Product reviews system,Exact match domain system,Freshness systems(Query Deserves Freshness / QDF),Site diversity system, andDeduplication systems. - Link & Trust Systems:
Link analysis systems and PageRank,Page experience system,Reliable information systems,Local news systems,Crisis information systems,Removal-based demotion systems, andSpam detection systems(plus retired or integrated historical systems such asHummingbird,Panda, andPenguin).
As Pandu Nayak’s trial testimony confirmed, foundational systems like Navboost and Topicality (T*) are core engines you will not find named on Google’s public guide to ranking systems. However, as Roger Montti documented in Search Engine Journal’s analyses of Google’s Core Topicality Systems and Core Topicality Systems, Google’s Elizabeth Tucker and VP of Search Liz Reid publicly referenced the existence of “core topicality systems” when discussing how Google balances search biases and combines multiple systems to understand queries and topics in a post-keyword era. What Google referred to externally in passing as “core topicality systems” was proven in federal court to be the Topicality (T*) system.
The “ABC” Signals Framework: How Google Evaluates Topicality (Anchors, Body, Clicks)
Google uses topicality signals to judge how well a page’s content matches a user’s query. In a February 2025 DOJ interview (Trial Exhibit PXR0356, uploaded May 2, 2025), Google search engineer Hyung-Jin “HJ” Kim explained that Google’s “ABC signals are the key components of topicality (or a base score)”, which is Google’s determination of a document’s relevance to a query (justice.gov). Specifically, Exhibit PXR0356 records: “T* (Topicality) effectively combines (at least) these three signals in a relatively hand-crafted way. Google uses [redacted] to judge the relevance of the document based on the query terms.”
Alongside modern machine learning and semantic scoring, Google evaluates topicality across five interconnected layers:
- A – Anchors (Context from Links): Derived from hyperlinks pointing from a source page to a target document (defined by Kim as “a source page pointing to a target page (links)”), this signal evaluates what other websites say about a page, a direct legacy of Google’s original PageRank algorithm. Pandu Nayak testified that “another very important signal is the [hyper]links between pages,” known as anchor text, which provides “a very valuable clue in deciding what the target page is relevant to” (regmedia.co.uk). For example, a page heavily cited with the anchor “JavaScript tutorial” boosts its topical relevance for that query. Importantly, Google clarified that it does not mix user click data into its link analysis. When asked if click data influences the anchor signal, Dr. Kenneth “Ken” Lehman explained: “To generate the anchor signal, that’s just from links between web pages, and it doesn’t involve clicks” (regmedia.co.uk).
- B – Body (On-Page Keywords and Text): The most traditional information retrieval signal, based on the presence, frequency, and prominence of query terms within the document itself (“terms in the document”) across titles, headings, meta tags, and body text. “The most basic and in some ways the most important signal is the words on the page and where they occur,” testified Pandu Nayak, noting that the presence of query terms is “actually kind of crucial” and that term frequency and position (e.g. title vs. body) are “very important” relevance cues (regmedia.co.uk). In short, what the document “says about itself” is central to determining its topicality.
- C – Clicks (User Interaction Signals and Dwell Time): One of the most significant confirmations of the trial. A 2016 Google “Three Pillars of Ranking” slide (Trial Exhibit UPX0004 by Dr. Eric Lehman) listed User-interactions (“what users say about the document”) as the third pillar alongside Body and Anchors, encompassing clicks, attention (hover/scroll), carousel swipes, and quick returns to the SERP. In Exhibit PXR0356, HJ Kim confirmed that even historical user behaviour – specifically “how long a user stayed at a particular linked page before bouncing back to the SERP” (dwell time) – was used as a foundational topical relevance signal (justice.gov). The same user engagement data that powers the
Navboostre-ranking system provides this foundational Clicks (C) signal for T*, though Google witnesses stressed that click data is used carefully in feedback loops rather than blindly promoting clickbait. - Machine Learning and Context: Models like
BERT(Bidirectional Encoder Representations from Transformers),Neural Matching, andRankEmbedBERThelp Google go beyond exact-word matching to understand the deeper context, intent, and meaning behind a search query and a page’s text. - Topical Scores: Google uses specific algorithms and topicality scores – including
T*base scores,QBST(Query-Based Salient Terms),WebRefentity topicality annotations (topicalityE2), linktopicalityWeight, and site-level embeddings (siteFocusScoreandsiteRadius) – to measure how well a document or resource addresses a subject rather than just counting keyword repetitions.
| ABC Signal Component | Internal Conceptual Definition (Exhibit UPX0004 & PXR0356) | Primary Content Warehouse API Modules & Attributes | Downstream Top-Level Pillar Routing | Core Algorithmic Demotions & Anti-Spam Filters |
| A – Anchors | “What the Web says about the document” – Hyperlinks pointing from a source page to a target document. | RepositoryWebrefOutlinkInfo.topicalityWeight, Anchors.Anchor.fullLeftContext, fullRightContext, context2, Anchors.AnchorSource.homePageInfo, anchorPhraseCount. |
Feeds directly into the Popularity (P*) top-level signal to promote well-linked documents. | anchorMismatchDemotion, pageMismatchTaggedAnchors, phraseAnchorSpamPenalty (LINK_SPAM_PHRASE_SPIKE), penguinPenalty, BadBackLinks. |
| B – Body | “What the document says about itself” – Presence, prominence, and semantic alignment of terms within the document. | PerDocData, titlematchScore, avgTermWeight, OriginalContentScore, contentEffort, CenterPiece, QualitySalientTermsSalientTermSet, topicalityE2, siteFocusScore, siteRadius. |
Feeds content-derived metrics directly into the Quality (Q*) top-level signal and Mustang scoring. |
KeywordStuffingScore, spamtokensContentScore, GibberishScore, shingleInfo (near-duplicate detection), racterScores / AGC, pandaDemotion. |
| C – Clicks | “What users say about the document” – Historical dwell time on a linked page before bouncing back to the SERP. | NavBoostDocData, Craps click module, goodClicks, badClicks, lastLongestClicks, unsquashed vs. squashed click counts, Glue super query log, chromeInTotal. |
Feeds directly into the Popularity (P*) top-level signal and Navboost / Glue re-ranking. |
navDemotion, serpDemotion, clutterScore (intrusive ad layouts), violatesMobileInterstitialPolicy, IP priors, distinct-user filtering. |
Inside Trial Exhibit PXR0356: Why Google Hand-Crafts the T* Formulas
Google’s ranking engineers hand-crafted the formulas for these signals (rather than relying purely on ML) so they could understand and adjust how each factor contributes to relevance (justice.gov). First surfaced and analysed in the Hobo Web Strategic SEO 2025 research series, Trial Exhibit PXR0356 (USDOJ-GOOG-00195742) reveals that out of Google’s more than 100 raw signals, almost every signal – aside from LLM-based systems like RankBrain, DeepRank, and RankEmbed – is hand-crafted by engineers.
HJ Kim detailed the exact engineering process: for the majority of signals, Google takes relevant empirical data (webpage content and structure, user clicks, and label data from human quality raters logged in EWOK), performs a statistical regression, applies a sigmoid function or related pre-computed curve, and determines the operational threshold. In the extreme, “hand-crafting means that Google looks at the relevant data and picks the mid-point manually”. Kim explained the exact rationale for maintaining human mathematical control over signals like T*:
“The reason why the vast majority of signals are hand-crafted is that if anything breaks Google knows what to fix. Google wants their signals to be fully transparent so they can trouble-shoot them and improve upon them.” – Hyung-Jin “HJ” Kim, Google Search Engineer, DOJ Trial Exhibit PXR0356 (February 18, 2025)
Exhibit PXR0356 notes that Microsoft Bing uses complex machine learning techniques to optimise ranking functions, making it difficult to fix or troubleshoot when rankings break – a problem worsened by deep learning – giving Google a major operational advantage when responding to ranking challenges. This aligns directly with Pandu Nayak’s testimony explaining why Google refuses to hand over top-level ranking control to an end-to-end neural network:
“I think it’s risky for Google – or for anyone else, for that matter, to turn over everything to a system like these deep learning systems as an end-to-end top-level function. I think it makes it very hard to control.” – Pandu Nayak, VP of Search, U.S. v. Google LLC Trial Testimony
The development of this system was a major engineering undertaking, described as being in a “constant state of development” from its inception until approximately five years prior to the testimony, indicating its maturity and foundational status within the ranking stack. At its core a standard “old style” Information Retrieval (“IR”) metric, Exhibit PXR0356 notes that even today, “Ranking development (especially topicality) involves solving many complex mathematical problems,” where a dedicated team of engineers works continuously on hard mathematical problems within a given topicality project.
Topicality (T*) vs. Topical Authority: Page-Level Relevance vs. Site-Wide Expertise
A frequent point of confusion in modern search strategy is the distinction between Topicality (T*) and Topical Authority. Whilst both concepts govern whether content ranks for a subject, they operate at different architectural layers, timescales, and mathematical scopes within Google’s pipeline:
- Topicality (T*): Measures how well a single page matches a specific query in the moment. It is a fast, dynamic, query-dependent base score calculated during retrieval (
Mustang/Goldmine) using the page’s Body (B) text, inbound Anchors (A), and historical query-specific Clicks (C). - Topical Authority: Measures whether your entire website is recognised by Google as a trusted, comprehensive expert on a broader subject. Rather than evaluating a single URL in isolation, Topical Authority is governed by site-wide vector embeddings (
site2vecEmbedding), geometric focus metrics (siteFocusScoreandsiteRadius), Query-Based Salient Terms (QBST) footprint across a topic cluster, and the slow-moving domain Quality Score (Q*) (siteAuthorityandPageRank-NearestSeeds).
As Pandu Nayak confirmed in DOJ Trial Exhibit PXR0357 (January 31, 2025), Google ranks individual pages, but page-level topicality is continuously contextualised by site-wide signals:
“Site-wide signals and classifiers are also used and contribute to our understanding of pages. Having some good site-wide signals does not mean that all content from a site will always rank highly, just as having some poor site-wide signals does not mean all the content from a site will rank poorly.” – Pandu Nayak, VP of Search, DOJ Trial Exhibit PXR0357 (January 31, 2025)
| Architectural Dimension | Topicality (T*) | Topical Authority (Q* & Site Embeddings) |
| Evaluation Scope | Single Document / URL (CompositeDoc / PerDocData) matched against a specific search query in the moment. |
Entire Website / Domain / Subdomain recognised as a trusted, comprehensive expert on a broader subject. |
| Query Dependency & Speed | Query-Dependent & Fast (Real-Time): Recomputed dynamically for every query competition in milliseconds. | Query-Independent & Slow (Persistent): Pre-computed site-wide classifiers and vector embeddings updated across crawl/index cycles. |
| Primary Engine & Leaked Attributes | Goldmine & Mustang: ABC signals (Anchors, Body, Clicks), titlematchScore, avgTermWeight, topicalityE2, topicalityWeight. |
QualityNSR (New Site Rank) & TopicEmbeddingsVersionedData: siteFocusScore, siteRadius, site2vecEmbedding, siteAuthority, pagerank_ns. |
| Strategic Requirement | Comprehensive on-page coverage of salient terms (QBST), clear HTML heading hierarchy, contextual internal/external anchors, and strong dwell time. |
Deliberate pillar-and-cluster architecture, strict topical constraint (pruning off-topic drift to keep siteRadius tight), and high contentEffort. |
Deep Dive into Pillar B: The Body Signal, Semantic Embeddings, and Topical Scores
Whilst the Body (B) signal begins with classical lexical matching – verifying that the query terms appear in the <title> tag (titlematchScore), headings, and body text (avgTermWeight) as Pandu Nayak testified – Google’s Content Warehouse API leak reveals how the system computes deeper Topical Scores so that it does not rely solely on counting keyword repetitions.
1. First-Pass Scoring in Mustang, CompositeDoc, and PerDocData
Every URL in Google’s index is encapsulated inside a master data record called the CompositeDoc, which contains the PerDocData module and CompressedQualitySignals. During first-pass scoring by Mustang, Google isolates the CenterPiece of the page (the Main Content, stripping out navigation menus, footers, and ads) and evaluates its structural hierarchy. As documented in the Hobo Web On-Page SEO and Headings SEO Checklist guides, pairing a clear <h2> or <h3> heading with a focused paragraph creates a distinct semantic “passage” for Google’s Passage ranking system.
To prevent manipulation of the Body (B) signal through keyword repetition or low-effort automation, PerDocData tracks OriginalContentScore (scored 0 to 512 to flag derivative text), shingleInfo (near-duplicate fingerprinting), contentEffort (an LLM-based estimation of human labour and originality powering the Helpful Content System), KeywordStuffingScore, spamtokensContentScore, GibberishScore, and racterScores / AGC (detecting Artificially Generated Content).
2. Solving the “Cold Start” Problem: QBST, WebRef, and topicalityE2
In Trial Exhibit UPX0203, analysed in depth in Hobo Web’s “We Fake It”: Google’s Illusion of Understanding Content, Google engineer Dr. Eric Lehman made a candid admission regarding why Google historically leaned on user click reactions rather than deep text comprehension:
“A billion times a day, people ask us to find documents relevant to a query… What’s crazy is that we don’t actually understand documents. Beyond some basic stuff, we hardly look at documents. We do not understand documents. We fake it… So we watch how people react to documents and memorize their responses.” – Dr. Eric Lehman, Google Search Engineer, Trial Exhibit UPX0203
However, Dr. Lehman noted the critical flaw of relying on memorised clicks – the cold start problem:
“When our fakery fails – maybe because a document is new, recently-changed, or rarely-shown – we look stupid.” – Dr. Eric Lehman, Google Search Engineer, Trial Exhibit UPX0203
As Pandu Nayak similarly testified, “Navboost is not the only factor because there might be lots of documents that don’t have clicks”. To evaluate the Body (B) signal accurately even when a page is brand new or targets a long-tail query without click history, Google uses four semantic systems:
- QBST (Query-Based Salient Terms): Functioning as the “memorisation of relevance” alongside
QualitySalientTermsSalientTermSet,QBSTidentifies the full constellation of co-occurring words and sub-concepts that top-ranking pages share for a topic. A page lacking these salient terms is filtered out before final ranking stages. - WebRef Entity Topicality Scores (
topicalityE2): Rather than counting string repetitions, Google’sWebRefannotator extracts Knowledge Graph entities (encoded_mid) from the document and pairs them with the query’sQRefannotations. InsideRepositoryWebrefEntityAnnotations, Google assigns each entity on the page a specific topicality score stored as a two-decimal fixed-point integer intopicalityE2, converted via the explicit formula:topicality_score = topicality_e2 / 100.0. This score measures how central each entity is to the overall topic of the document, supported bytopEntity,referencePageScores,NlpSaftDocumentsyntactic parsing, andtopPetacatTaxIdtaxonomy classification. - BERT, Neural Matching, and RankEmbedBERT: As Pandu Nayak explained in “How AI powers great search results,” neural networks look at an entire query or page to understand underlying concepts – for instance, recognising that “insights how to manage a green” is about golf course maintenance. Models like
BERT,RankBrain, and LLM-trainedRankEmbedBERT(fine-tuned on human rater quality data and 13-month click logs) evaluate deep contextual alignment between query intent and page prose. - Site Embeddings (
siteFocusScoreandsiteRadius): Usingsite2vecEmbedding, Google mathematically measures how closely an individual page’s topic aligns with the website’s core theme (siteFocusScore) and whether the page stays within a tight geometric orbit (siteRadius). Furthermore, as outlined in the Hobo Web Super Topicality framework, injecting net-new Entity-Attribute-Value (EAV) triplets increases mutual informationI(E(c_j); q)between document chunks and AI query fan-outs, a process termed Creative Inference Optimisation.
Deep Dive into Pillar A: Anchor Text, topicalityWeight, and Link Context
Whilst Dr. Ken Lehman confirmed in his DOJ testimony that the raw anchor signal (A) is generated purely from links between web pages without mixing in click data, the 2024 Content Warehouse API leak revealed how Google weights, contextualises, and gates the value of those links in practice. As Google’s John Mueller noted in 2017:
“Most links do provide a bit of additional context through their anchor text. At least they should, right?” – John Mueller, Google (2017)
1. Click-Gated Link Index Tiers (Indextier / SourceType)
As detailed in the Hobo Web Definitive Guide to Link Building, although the anchor signal itself is calculated from the link graph, Google partitions its link index into Three Distinct Quality Tiers (Indextier / SourceType: low, medium, and high quality). For a link to pass full value, the source page it resides on must be validated by user engagement (clicks in Navboost / Chrome data or recent indexing freshness), which elevates that source page into the high-quality link index tier. A link sitting on an unclicked, zero-traffic page remains unactivated.
2. Mathematical Dilution of topicalityWeight
Inside the RepositoryWebrefOutlinkInfo module of the leaked API documentation, Google defines the attribute topicalityWeight as: “The topicality_weight for each link with this target URL. The weight of how relevant each link is to the topic for the target URL”.
Intuitively, many SEOs assume that adding more internal links to a page always improves its ranking. The mechanism of topicalityWeight proves the opposite: because Google evaluates the relevance distribution of links pointing to a target URL, adding low-relevance sitewide or boilerplate links dilutes the average topicalityWeight.
To illustrate how an averaging distribution operates mathematically, suppose a page has two highly relevant editorial links with weights of 0.82 and 0.71 (an average topicalityWeight of 0.765), and a site owner adds 10 unrelated links each carrying a low score of 0.15; the average weight drops by 67.1% to 0.252. This causes three parallel penalties: the relative ratio of supplementary context drops due to denominator expansion, the semantic vector centre-of-gravity is neutralised (weakening the match between the page’s WebRef topicality and the query’s QRef topicality), and supplementary anchor explanations are physically discarded due to anchor storage slot monopoly. Reverting noisy internal anchors typically restores rankings within 1 to 2 weeks.
3. Contextual Windows, the 16-Word Limit, and anchorMismatchDemotion
Google does not merely read the anchor text inside the <a> tag; it extracts the surrounding prose before and after the link using Anchors.Anchor.fullLeftContext, Anchors.Anchor.fullRightContext, Anchors.Anchor.context, and Anchors.Anchor.context2, and checks whether the source page is a homepage via Anchors.AnchorSource.homePageInfo. Furthermore, empirical testing documented in the Hobo Web Internal Links SEO Checklist demonstrates two critical thresholds:
- Single-Anchor Influence: Where an internal anchor text pointing to a page is the only mention of a target keyword phrase on the site, it takes only ONE internal anchor text link to provide a noticeable influence in specific keyword phrase rankings and relevance (T*).
- The 16-Word Anchor Ceiling: In
Google.co.uktesting (including across 301 redirects), any keyword signal within a 16-word limit in anchor text is consistently detected, whereas nothing above the 16-word threshold is detected.
Conversely, if inbound anchor text is topically irrelevant to the destination page, Google applies CompressedQualitySignals.anchorMismatchDemotion (alongside pageMismatchTaggedAnchors), actively demoting pages that attempt to manipulate rankings with mismatched links. Spikes in manipulative commercial anchors are flagged by phraseAnchorSpamPenalty (LINK_SPAM_PHRASE_SPIKE), penguinPenalty, and BadBackLinks (which disables improvanchor scoring). As Google’s internal link labelling disclosure explained in the Hobo Web Link Building Strategy:
“So disavow is again, basically, just a label internally. It’s applied on the links and anchors. And then you can see that, as well. Basically, you could have like a link from, I don’t know, WhiteHouse.gov, and it has labels Penguin RT, footer and disavow.” – Google Search Quality Disclosure
Deep Dive into Pillar C: Clicks, Dwell Time, and Navboost
In Trial Exhibit UPX0004 (“Life of a Click (user-interaction)”), Google engineer Dr. Eric Lehman summarised the Three Pillars of Ranking (Body, Anchors, and User-interactions) and explained how Google uses human reactions to sustain its ranking induction loop:
“We look at people. If a document gets a positive reaction, we figure it is good. If the reaction is negative, it is probably bad. Grossly simplified, this is the source of Google’s magic… In designing user experiences, SERVING the user is NOT ENOUGH. We have to design interactions that also allow us to LEARN from users. Because that is how we serve the next person, keep the induction rolling, and sustain the illusion that we understand.” – Dr. Eric Lehman, Google Search Engineer, Trial Exhibit UPX0004
Lehman’s presentation also exposed Google’s strict internal secrecy policy regarding click signals: “Do not discuss the use of clicks in search, except on a need-to-know basis with people who understand not to talk about this topic externally”.
As detailed in the Hobo Web Navboost Guide, the user engagement data that provides the foundational Clicks (C) signal for T* (dwell time before bouncing back to the SERP) also powers Navboost, which Pandu Nayak confirmed dates to ~2005 and is “one of the important signals that we have”. Operating on a 13-month rolling window of memorised click data (processed via the internal Craps module and stored in the Glue super query log), Navboost slices user interactions by device type (mobile vs. desktop) and locality. It distinguishes between:
badClicks: Immediate bounces / pogo-sticking back to the SERP, signalling dissatisfaction.goodClicks: Visits where the user dwells on the destination page for a meaningful duration.lastLongestClicks: The final result a user clicks and dwells on without returning to the SERP, signalling that their search journey ended successfully.
To prevent manipulation, Google stores squashed (logarithmically normalised) and unsquashed click counts side-by-side, applies IP-based priors, and filters thin data on distinct users. Whilst Navboost refines blue-link web rankings, its sister system Glue tracks multi-touch gestures (clicks, hovers, scrolls, swipes) across all rich SERP features to feed the Tetris / Tangram whole-page layout engine, supported by Instant Glue (operating on a 24-hour window with ~10-minute latency for fresh queries). Pages that bait clicks with misleading titles or frustrate users with heavy ad clutter (clutterScore) or mobile pop-ups (violatesMobileInterstitialPolicy) accumulate badClicks and suffer navDemotion or serpDemotion.
Clicks vs. Quality: Why Authority (Q*) Overrides Popularity
One critical revelation from the trial was Google’s caution against using raw click metrics as a proxy for quality. An internal evaluation found that “a large number of clicks on a link does not necessarily mean that the page is of high quality” (regmedia.co.uk), and Pandu Nayak noted that page quality tends to be “anti-correlated” with pure click-through rates in some cases – improving result quality in tests sometimes led to fewer clicks, as users often chase sensationalist clickbait even when higher-quality information is available. Discussing an internal slide during his testimony, Dr. Eric Lehman warned:
Q: “How does that relate to the question of user data or user interaction data?” A: “So the chart is a little bit complex, but what it’s illustrating is one of the problems with using click data in connection with ranking search results. It’s a very strong observation that people tend to click on lower-quality, less-authoritative content than we would like to show on our search engine. Our goal is to show – when someone issues a query, to give them information that’s relevant and from authoritative, reputable sources. People tend not to click on those so much. So if we’re guided too much by clicks, our results would be of a lower quality than we’re targeting.” – Dr. Eric Lehman, Google Search Quality Testimony, U.S. v. Google LLC
Google’s ranking engineers therefore treat raw click volume with scepticism when evaluating authority. Whilst Google uses nuanced engagement signals (such as dwell time for Navboost and T*) to refine relevance, it relies on query-independent authority signals – specifically PageRank (PageRank-NearestSeeds, measuring link distance from trusted seed sites, per Exhibit PXR0356 at -744) and the site-wide Quality Score (Q*) (siteAuthority, scamness, unauthoritativeScore, and pandaDemotion) – to keep search results genuinely trustworthy. As HJ Kim observed in Trial Exhibit PXR0356 regarding why Q* is essential against content farms and generative AI spam: “Nowadays, people still complain about the quality and AI makes it worse”.
Expressed through the analytical macro heuristic (final_score = α · topicality + β · quality + γ · navboost), Google uses nuanced click signals to refine relevance, whilst relying on Q* and E-E-A-T as an algorithmic ceiling to enforce trust and authority.
Download Your Free eBook
This article was an excerpt from my SEO book – Strategic SEO 2025 and first published here on Hobo Web on 18 July 2025. I have since expanded it to what you read today.
Academic Bibliography and Primary Source Registry
1. Primary Court Exhibits, Official Documentation & Industry Analyses
- Google Search Central (2022-2026). A Guide to Google Search Ranking Systems & Introducing Our New Guide to Google Search Ranking Systems. Google Developers. Available at: https://developers.google.com/search/docs/appearance/ranking-systems-guide and https://developers.google.com/search/blog/2022/11/introducing-guide-to-ranking-systems.
- Goodwin, D. (2023-2025). How Google Search Ranking Works (Pandu Nayak Testimony) & The ABCs of Google Ranking Signals. Search Engine Land. Available at: https://searchengineland.com/google-abc-ranking-signals-455360, https://searchengineland.com/how-google-search-ranking-works-pandu-nayak-435395, and https://searchengineland.com/google-search-ranking-documents-434141.
- Kim, H.-J. & U.S. Department of Justice (2025). Interview with Google Engineer H.J. Kim Notes (February 18, 2025). Trial Exhibit PXR0356 (
USDOJ-GOOG-00195742), United States et al. v. Google LLC, Case No. 1:20-cv-03010-APM. Available at: https://www.justice.gov/atr/media/1398871/dl. - Lazuk, E. (2025). What Has Pandu Nayak Said About Page-Level Rankings? (Analysis of DOJ Trial Transcript & Exhibit PXR0357). Available at: https://ethanlazuk.com/blog/what-has-pandu-nayak-said-about-page-level-rankings-a-tales-from-the-query-post/.
- Lehman, E., Lehman, K., & Nayak, P. (2023-2024). United States v. Google LLC Trial Transcripts & Proposed Findings of Fact. Available at: https://regmedia.co.uk/2024/02/26/us_v_google.pdf and https://www.justice.gov/d9/2023-11/417508.pdf.
- Mehta, A. P. (2025). Memorandum Opinion on Remedies (September 2, 2025). Document 1436, United States et al. v. Google LLC, U.S. District Court for the District of Columbia, Case No. 1:20-cv-03010-APM. Available at: https://law.justia.com/cases/federal/district-courts/district-of-columbia/dcdce/1:2020cv03010/223205/1436/.
- Montti, R. (2024-2025). What Are Google’s Core Topicality Systems? & Google Discusses Core Topicality Systems (Elizabeth Tucker & Liz Reid). Search Engine Journal. Available at: https://www.searchenginejournal.com/what-are-googles-core-topicality-systems/523537/ and https://www.searchenginejournal.com/core-topicality-systems/521046/.
- Saito, R. (2025). Analysis of Topicality (T*), ABC Signals, topicalityWeight, and WebRef Entity Annotations (topicalityE2). Note Technical Series. Available at: https://note.com/roberto_saito/n/n572a6bf23a06?hl=en.
- Słowik, S. (2025). NavBoost: Court-Confirmed User Interaction Signals and Content Warehouse Schema Analysis. Available at: https://www.szymonslowik.com/navboost/.
- Turrado, N. (2025). How Google Defines and Calculates Quality (Q*) and Topicality (T*) in Search Ranking. Natzir.com. Available at: https://natzir.com/en/seo/google-quality-topicality-ranking/.
2. Companion Hobo Web Architecture Series
- Hobo Topical Authority Frameworks (2025-2026). Guide to Topical Authority and Semantic SEO Strategy. Referenced via: https://www.hobo-web.co.uk/topical-authority/.
- Anderson, S. (2025a). Strategic SEO 2025 (Version 2): Deconstructing Google’s Ranking Architecture. Hobo Web (Updated October 12, 2025). Available at: https://www.hobo-web.co.uk/strategic-seo-2025/.
- Anderson, S. (2025b). DOJ v. Google Disclosure: The Popularity Signal (P* Score). Hobo Web (Published September 14, 2025). Available at: https://www.hobo-web.co.uk/doj-v-google-disclosure-the-popularity-signal-p-score/.
- Anderson, S. (2025c). What is Google PageRank? PageRank-NearestSeeds and the Quality Score. Hobo Web (Published October 30, 2025). Available at: https://www.hobo-web.co.uk/google-pr-update/ and https://www.hobo-web.co.uk/is-google-pagerank-still-relevant-in-2026/.
- Anderson, S. (2026a). How Google Works: The Architecture of Q*, P*, Topicality (T*), and Navboost. Hobo Web (Updated October 1, 2026). Available at: https://www.hobo-web.co.uk/how-google-works/.
- Anderson, S. (2026b). The Q* Metric: Google’s Site-Level Quality Score. Hobo Web (Updated January 20, 2026). Available at: https://www.hobo-web.co.uk/the-q-metric-google-quality-score/.
- Anderson, S. (2026c). Technical SEO: The New Canon of Truth and the CompositeDoc Architecture. Hobo Web (Updated September 29, 2026). Available at: https://www.hobo-web.co.uk/technical-seo/, https://www.hobo-web.co.uk/compositedoc/, and https://www.hobo-web.co.uk/perdocdata/.
- Anderson, S. (2026d). Navboost: How Google Uses Large-Scale User Interaction Data to Rank Websites. Hobo Web. Available at: https://www.hobo-web.co.uk/navboost-how-google-uses-large-scale-user-interaction-data-to-rank-websites/.
- Anderson, S. (2026e). “We Fake It”: Google’s Illusion of Understanding Content (Exhibit UPX0203 & UPX0004 Analysis). Hobo Web. Available at: https://www.hobo-web.co.uk/we-fake-it-googles-illusion-of-understanding-content/.
- Anderson, S. (2026f). Topical Authority: Semantic Understanding, siteFocusScore, and siteRadius. Hobo Web. Available at: https://www.hobo-web.co.uk/topical-authority/.
- Anderson, S. (2026g). Super Topicality: Creating an Algorithmic Gravity Well for Authority. Hobo Web. Available at: https://www.hobo-web.co.uk/super-topicality-creating-an-algorithmic-gravity-well-for-authority/.
- Anderson, S. (2026h). Internal Links SEO Checklist & The Definitive Guide to Link Building After the Google Content Warehouse Leak. Hobo Web. Available at: https://www.hobo-web.co.uk/internal-links-seo-checklist/ and https://www.hobo-web.co.uk/the-definitive-guide-to-linkbuilding-after-the-google-content-warehouse-leak/.
- Anderson, S. (2026i). What is E-E-A-T in SEO, Contextual SEO, On-Page SEO, and The ContentEffort Attribute. Hobo Web. Available at: https://www.hobo-web.co.uk/what-is-e-e-a-t-in-seo/, https://www.hobo-web.co.uk/eeat/, https://www.hobo-web.co.uk/contextual-seo/, https://www.hobo-web.co.uk/on-page-seo/, and https://www.hobo-web.co.uk/the-contenteffort-attribute-the-helpful-content-system-and-e-e-a-t-is-gemini-behind-the-hcu/.

Disclaimer: Note that this is not official advice from Google. It is an opinion. It is SEO theory based on my own primary analysis of Google patents and code leaks. Any article (like this) dealing with the Google Content Data Warehouse leak requires a lot of logical inference when putting together the framework for SEOs, as I have done with this article. I urge you to double-check my work and use critical thinking when applying anything about the leaks to your site. My aim with these articles is essentially to confirm that Google does, as it claims, try to identify trusted sites to rank in its index. The aim is to irrefutably confirm white hat SEO has purpose in 2026 – and that purpose is to build high-quality websites.
The Google API leak exposes data structures and protocol buffers, not the live executing C++/Elixir code or algorithmic weights. Take for example siteQualityStddev; While the attribute exists, no one outside Google knows its exact mathematical multiplier in real-time query retrieval. Treat most leaked parameters similarly. Some parameters listed may be inactive or superseded in production. Modern search models (like RankEmbedBERT) don’t use fixed C++ formulas at all. They rely on neural network weights trained on millions of interactions. You can’t “decode” a neural network’s decision-making just by looking at its input schema. Also note I only discuss sales, editorial and marketing-related attributes. I do not explore attributes tied to Abuse, Sensitive or Inappropriate classifiers in documents in any of my writings (and will not). (My thesis builds on this: “There was a leak of Google documents which named certain components of Google’s ranking system, but the documents don’t go into specifics of the curves and thresholds“ – The most direct and significant acknowledgement of this is from within the context of the DOJ’s antitrust litigation based on a call dated February 18, 2025, with Google Engineer HJ Kim). I view the leaks as an architectural map of possibilities, not a cheat code. I suggest you do the same. Feedback and corrections welcome.
Disclosure: I use generative AI when specifically writing about my own experiences, ideas, stories, concepts, tools, tool documentation or research. My tool of choice for this process is Google Gemini Pro 2.5 Deep Research and later models. I have over 25 years of experience writing about accessible website development and SEO (search engine optimisation). All content was conceived, edited, and verified as correct by me, Shaun Anderson, aka Hobo (and is under constant development). Corrections and contributions welcome. See my AI policy.