# E-E-A-T Mechanics — machine-readable reference

Evidence-disciplined mapping of Google ranking systems to the E-E-A-T frame. Raw source for use in LLMs (Claude Code, Claude Skills, agents, RAG). German version: /llms.txt

© Thomas Wawra · Senior SEO Manager · wetter.com — a Funke Digital company · https://www.wetter.com/impressum/

Version: v1 · 2026 · Norm: QRG 2025-09-11 · 25 systems · 38 fields

Evidence codes: [A] sworn/DOJ · [B] leak · [O] official · [C] interpretation

Reading rule: [A] and [B] are documented, [C] is interpretation and must be labeled as such when quoted. Page Quality and Needs Met are judged separately, never averaged.

## Ranking systems

### 1. NavBoost / CRAPS

- Function class: Behavior
- E-E-A-T: E=indirect · Exp=indirect · A=primary · T=primary
- Axis: Usage (Search)
- Reach: Doc
- Calibration source: Behavior
- Evidence code: A/B
- Leak fields: goodClicks, badClicks, lastLongestClicks
- Feeds into: Tangram
- Fed by: IP-Prior / Squashing / voterToken

**What it measures:** NavBoost watches what users do after a Google search: do they click on a weather page and stay — or jump straight back to Google? The last, longest click of a session is the strongest signal because it shows the need was genuinely met. Real user behavior becomes a re-ranking signal, without Google having to read the content directly.

**Derived from:** Anonymized, aggregated click logs from Google search results — per query and per document. Google observes patterns across many users, not the behavior of individuals.

**Metric detail:** goodClicks/badClicks = clicks rated as satisfying or disappointing; lastLongestClicks = the last, longest click of a session — the strongest satisfaction signal, because it marks the need that was met. [B] fields, [C] thresholds.

**Evidence:** [B] field names; [A] confirmed in DOJ/sworn material as a central signal.

**Why this attribution:** Authority and Trust are primary: when users click a page and don't immediately bounce back, it proves the page is perceived as a trustworthy, recognized source. Expertise and Experience are indirect — NavBoost acts as a 'lie detector': pages that are factually weak get avoided, even when that's not visible from the outside.

**Strategic consequence:** NavBoost cannot be manipulated directly — it reflects genuine user satisfaction. For wetter.com this means: titles and snippets in search results must deliver on their promise. A page about general climate served for 'storm Berlin tomorrow' risks immediate bounces — and a worse ranking as a result. The best optimization is a page that genuinely answers the user's need.

### 2. Glue / Instant Glue

- Function class: Behavior
- E-E-A-T: E=indirect · Exp=indirect · A=primary · T=primary
- Axis: Usage (Search)
- Reach: Doc
- Calibration source: Behavior
- Evidence code: B
- Leak fields: Feature-Klicks/Hovers (wie NavBoost)
- Feeds into: Tangram

**What it measures:** The counterpart to NavBoost for special Google elements like carousels, 'People also ask' boxes, or news features. Glue measures whether users interact with these elements — tapping a weather carousel, expanding a forecast in a PAA box, or arriving at wetter.com via a news feature. Instant Glue is the real-time variant for fast-moving events.

**Derived from:** Hover and click interactions with SERP features (carousels, PAA boxes, news features) on Google search result pages.

**Metric detail:** Measures interactions with SERP features (hovers, clicks on carousel/PAA) rather than web clicks; Instant Glue = short-term real-time aggregation of the same signals. [B/C]

**Evidence:** [B] described as NavBoost's counterpart for features.

**Why this attribution:** Works by the same logic as NavBoost: authority and trust are confirmed through lived demand. When users reach wetter.com via SERP features and stay, it counts exactly like a regular web click.

**Strategic consequence:** Wetter.com should prepare content for Google feature inclusion: structured data (Schema.org), clear forecast widgets, timely weather reports with news value. Appearing in carousels and PAA boxes and generating clicks there earns the same ranking signals as organic web clicks.

### 3. Tangram

- Function class: Behavior
- E-E-A-T: E=none · Exp=none · A=indirect · T=indirect
- Axis: Usage (Search)
- Reach: Doc
- Calibration source: —
- Evidence code: B
- Leak fields: –
- Fed by: NavBoost / CRAPS, Glue / Instant Glue

**What it measures:** Tangram evaluates nothing itself — it is the assembly layer that takes the ratings from NavBoost and Glue and builds the final search results page from them. Think of it as a conductor: it decides which elements appear where on the results page, but it doesn't judge whether the content is good.

**Derived from:** The outputs of NavBoost and Glue as input — Tangram combines these behavioral ratings into a SERP layout, making no judgments of its own.

**Metric detail:** GlueResponse = the behavioral input delivered by Glue, which Tangram assembles into the final SERP layout; Tangram itself is the composition layer, not a measured value. [B/C]

**Evidence:** [B] field names.

**Why this attribution:** Tangram has no direct E-E-A-T bearing because it doesn't judge. Authority and Trust appear only indirectly because Tangram arranges the results of systems that do measure those dimensions.

**Strategic consequence:** There is no direct lever at Tangram. Whoever optimizes NavBoost and Glue through genuine user satisfaction automatically benefits — Tangram then shows the content in the right place in the right order.

### 4. Chrome-Systeme

- Function class: Behavior
- E-E-A-T: E=none · Exp=none · A=indirect · T=primary
- Axis: Usage (Browser)
- Reach: Site
- Calibration source: Behavior
- Evidence code: B
- Leak fields: chromeInTotal, uniqueChromeViews

**What it measures:** How often wetter.com is accessed in the Chrome browser overall — regardless of whether someone arrived via Google search or typed the URL directly. High direct visits show that users trust the site enough to navigate there without asking Google. It measures brand popularity and willingness to return.

**Derived from:** Aggregated, anonymized browser usage data from Chrome — total views and unique visitors at site level, not page level.

**Metric detail:** chromeInTotal = total Chrome views of a site; uniqueChromeViews = unique viewers. Together a site-wide, topic-agnostic usage/popularity measure from browser data. [B fields, C weighting]

**Evidence:** [B] field names.

**Why this attribution:** Trust is primary because direct browser visits are a strong trust statement — you go straight to the source without asking Google. Authority is indirect. No expertise or experience bearing: Chrome data is topic-agnostic and measures brand recognition, not subject-matter competence.

**Strategic consequence:** This metric cannot be improved through page optimization alone — it arises from genuine brand recognition and return-visit habits. For wetter.com, app usage, newsletters, browser bookmarks and direct visits strengthen the Chrome signal far more than any SEO work on individual pages.

### 5. IP-Prior / Squashing / voterToken

- Function class: Behavior
- E-E-A-T: E=none · Exp=none · A=none · T=indirect
- Axis: Usage (Search)
- Reach: Site
- Calibration source: Pattern
- Evidence code: B
- Leak fields: unscaledIpPriorBadFraction, voterToken
- Feeds into: NavBoost / CRAPS

**What it measures:** This system protects the integrity of click data. It detects when click patterns look unnatural — for example when many clicks come from the same IP addresses or are artificially directed at a page. Such suspicious clicks are filtered out ('damped') before they can influence ranking.

**Derived from:** IP-address-based trust priors and a 'ballot' system (voterToken) that identifies individual click units to prevent double-counting.

**Metric detail:** unscaledIpPriorBadFraction = share of suspicious clicks per IP range before scaling (manipulation damping); voterToken = a unit that identifies a click 'ballot' to make repeat voting harder. [B fields, C mechanics]

**Evidence:** [B] field names.

**Why this attribution:** Only Trust (indirect): the system makes no statements about content, it protects the reliability of measurement data. It is the quality guardian for all behavioral signals.

**Strategic consequence:** No direct action needed for wetter.com. Important to know: clickbait campaigns or purchased click packages remain ineffective — they are filtered out before touching the ranking. Genuine user satisfaction is the only sustainable strategy.

### 6. Q* (Saldo)

- Function class: Quality & Prediction
- E-E-A-T: E=indirect · Exp=indirect · A=primary · T=primary
- Axis: Standing
- Reach: Site
- Calibration source: Rater (IS)
- Evidence code: B
- Leak fields: siteAuthority + lowQuality + nsr (Aggregat)
- Feeds into: SegIndexer / Tiers
- Fed by: NSR (+ Fallback-Vererbung), hostAge-Sandbox

**What it measures:** Q* is a summary quality score for the entire website — a 'balance' made up of the site's authority (siteAuthority), a low-quality flag (lowQuality), and the overall NSR value. This score changes slowly: internal Google documents describe it as 'largely static' — a sluggish foundation, not a day-to-day signal.

**Derived from:** A combination of siteAuthority and lowQuality (both derived from the NSR system) with the overall NSR value. Q* consumes NSR, it doesn't compute it.

**Metric detail:** Not a single field: siteAuthority (authority) + lowQuality (low-quality flag) + NSR converge into the Q* balance; siteAuthority and lowQuality both stem from quality_nsr and are applied in Qstar. [B]

**Evidence:** [B] siteAuthority 'converted from quality_nsr ... applied in Qstar'.

**Why this attribution:** Authority and Trust are primary: Q* reflects the long-term reputation of the site — is it perceived as a reliable, recognized source? Expertise and Experience feed in indirectly because the underlying content signals are factored into Q*.

**Strategic consequence:** Q* does not react to individual page improvements — it requires consistent quality over months and years. For wetter.com this means: long-term content quality beats short-term SEO tricks. Pages with persistently low quality scores drag down the entire site's Q* value.

### 7. NSR (+ Fallback-Vererbung)

- Function class: Quality & Prediction
- E-E-A-T: E=none · Exp=indirect · A=primary · T=indirect
- Axis: Standing
- Reach: Site
- Calibration source: mixed
- Evidence code: B/C
- Leak fields: predictedDefaultNsr, nsrConfidence
- Feeds into: Q* (Saldo)
- Fed by: chard (+ YMYL/Hoax), contentEffort, tofu / keto / Rhubarb / Subchunks, OriginalContentScore, Topic-Embeddings (Fokus/Radius)

**What it measures:** NSR (Normalized Site Rank) is the real quality core — a normalized quality and authority value computed for individual 'chunks' (subsections) of a website. New sections initially inherit the average of their neighboring chunks ('fallback inheritance'). The historicized signal (predictedDefaultNsr) shows development over time — the trajectory matters, not today's single value.

**Derived from:** Four combined data vectors: (1) link structure from the web, (2) click and satisfaction behavior, (3) thematic focus vectors (mathematical representations of topic space, called embeddings), (4) aggregated content quality from chard, OCS and contentEffort.

**Metric detail:** predictedDefaultNsr = historicized baseline quality score (VersionedFloatSignal -> the trajectory/consistency is the signal, not the daily value); nsrConfidence = Google's confidence in its own NSR judgment (deprecated). [B fields, C interpretation]

**Evidence:** [B] fields + conversion to siteAuthority/lowQuality; [C] vector weighting.

**Why this attribution:** Authority is primary — NSR is the machine core of authority assessment. Expertise and Trust are built in indirectly. Experience has no bearing: first-hand material from individual articles averages out when aggregated to site level.

**Strategic consequence:** For wetter.com the long-term trajectory matters, not today's value. New topic sections start with the house average — weak neighboring content pulls them down. Actively pruning thin pages raises the average of the whole section and with it the NSR.

### 8. chard (+ YMYL/Hoax)

- Function class: Quality & Prediction
- E-E-A-T: E=none · Exp=primary · A=none · T=indirect
- Axis: Form quality
- Reach: Doc→Site
- Calibration source: Rater (IS)
- Evidence code: B
- Leak fields: chardScores, chardVariance
- Feeds into: NSR (+ Fallback-Vererbung)
- Fed by: IS-Eichung (Rater/QRG)

**What it measures:** chard rates the content quality of individual pages — calibrated against the judgments of human quality raters that Google trains according to strict guidelines (QRG). A variance score shows how confident the model is in its verdict. chard is especially sharp on YMYL content (Your Money or Your Life — topics with real life consequences such as health, safety, severe weather) and on suspected misinformation.

**Derived from:** A machine content classifier working with chardScore and chardVariance. It is trained to judge the way a trained human rater would. Its output feeds into the NSR content vector.

**Metric detail:** chardScore = content-quality value per document; chardVariance = its dispersion/uncertainty (high variance = an uncertain judgment about the page). [B fields, C meaning of the variance]

**Evidence:** [B] fields; [A/C] rater calibration backed by a documented chain.

**Why this attribution:** Expertise is primary — chard directly measures whether content is factually grounded and correct, the way an expert would assess it. Trust is contained indirectly.

**Strategic consequence:** Severe weather warnings and storm alerts on wetter.com fall in the YMYL category — people make real decisions (cancel trips, clear cellars) based on this content. Such pages must be accurate, current and clearly written. Errors or misleading information can not only harm users but also permanently damage the chard score.

### 9. contentEffort

- Function class: Quality & Prediction
- E-E-A-T: E=indirect · Exp=primary · A=none · T=indirect
- Axis: Form quality
- Reach: Doc→Site
- Calibration source: Rater (IS)
- Evidence code: B
- Leak fields: contentEffort (LLM-Aufwand)
- Feeds into: NSR (+ Fallback-Vererbung)
- Fed by: IS-Eichung (Rater/QRG)

**What it measures:** A language model estimates how much editorial effort and depth went into a page. The system distinguishes between shallow, machine-generated content and pages with recognizable research, structure and substance. It is not a hard measurement but a model estimate — one that increasingly reliably separates thin AI content from genuine editorial engagement.

**Derived from:** A single AI-estimated value distilled from the page's text content. The model was trained to recognize effort and depth. Its output feeds into NSR.

**Metric detail:** A single LLM-estimated value for creation effort/depth — not a raw measurement but a model estimate; separates substance from thin AI output. [B field, C scale]

**Evidence:** [B] field.

**Why this attribution:** Expertise is primary — high creation effort is a sign of subject-matter engagement. Experience is indirect because effortful content often also contains first-hand material. Trust is indirect.

**Strategic consequence:** For editors at wetter.com: automatically generated forecast pages without editorial added value are at a disadvantage compared to pages with context and explanation. A storm report that explains what this means for everyday travelers earns more contentEffort points than one that only lists temperatures and wind speeds.

### 10. tofu / keto / Rhubarb / Subchunks

- Function class: Quality & Prediction
- E-E-A-T: E=none · Exp=indirect · A=none · T=indirect
- Axis: Form quality
- Reach: Doc→Site
- Calibration source: Rater (IS)
- Evidence code: B/C
- Leak fields: tofu, keto, Rhubarb, Subchunk-Deltas
- Feeds into: NSR (+ Fallback-Vererbung)
- Fed by: IS-Eichung (Rater/QRG)

**What it measures:** These codenames stand for refinement signals that fine-tune the quality estimate at a very granular level — individual subsections (subchunks) of a page. Their exact function is not fully documented in the available material. They are not a standalone main system but refine chard and NSR at the section level.

**Derived from:** Internal codenames from the leak material; the exact input data is not fully documented. The weakest evidence in the entire matrix.

**Metric detail:** Codenames for refinement/delta signals at the subchunk level; subchunk deltas = changes in the quality estimate between sub-sections. Function largely conjectured. [B existence, C mechanics]

**Evidence:** [B] existence of the codenames; [C] precise function — the weakest evidence in the matrix.

**Why this attribution:** Expertise and Trust are indirect and deliberately rated conservatively because evidence for the exact mechanics is lacking. These signals are fine-tuning, not main levers.

**Strategic consequence:** No lever of its own. Whoever improves chard and NSR through good content quality automatically improves these derived signals too. Communicate internally: the codenames are documented but their exact effect remains interpretation — don't make promises based on speculation.

### 11. OriginalContentScore

- Function class: Quality & Prediction
- E-E-A-T: E=primary · Exp=indirect · A=none · T=none
- Axis: Form quality
- Reach: Doc→Site
- Calibration source: Rater (IS)
- Evidence code: B/C
- Leak fields: OriginalContentScore
- Feeds into: NSR (+ Fallback-Vererbung)
- Fed by: IS-Eichung (Rater/QRG)

**What it measures:** This score measures how much first-hand material a piece of content contains — information found only there, that an AI could not reconstruct from other sources. The higher the score, the more original the content. It is the closest machine equivalent to the E-E-A-T dimension 'Experience' (lived experience and direct observation).

**Derived from:** An AI-based score derived from the text content, measuring how much first-hand material is present. It feeds into the NSR content vector.

**Metric detail:** A value for the originality/first-hand-material degree of content: the higher, the less reconstructable elsewhere (the closest machine proxy for Experience). [B field, C Experience reading]

**Evidence:** [B] field; [C] Experience reading.

**Why this attribution:** Experience is primary — this score is the only content signal that can directly measure 'own experience'. Expertise is indirect. Important: when aggregated to site level (NSR), Experience 'evaporates' — a single original article gets lost in the average of many pages.

**Strategic consequence:** For wetter.com: own measurement data, meteorological assessments from the in-house sensor network, direct observations — all of this increases the OriginalContentScore. An editor writing 'our station at Hamburg-Fuhlsbüttel is currently measuring 87 km/h wind gusts' creates more original value than someone copying the same number from another service. Flag your own data visibly.

### 12. IS-Eichung (Rater/QRG)

- Function class: Quality & Prediction
- E-E-A-T: E=primary · Exp=primary · A=primary · T=primary
- Axis: Norm (overarching)
- Reach: —
- Calibration source: Source
- Evidence code: A
- Leak fields: IS-Score, siteQualityStddev
- Feeds into: chard (+ YMYL/Hoax), contentEffort, tofu / keto / Rhubarb / Subchunks, OriginalContentScore, Panda / BabyPanda

**What it measures:** Not a live signal used daily in ranking — but the foundation on which all quality models are calibrated. The IS-Score (Information Satisfaction) summarizes how well a page meets a user's information need, rated by trained human raters. The QRG (Quality Rater Guidelines) are the public rulebook by which these raters judge.

**Derived from:** Judgments from trained human raters following the QRG — especially 'Needs Met' (does the page serve the search intent?) and 'Page Quality' (is the page itself high quality?). siteQualityStddev measures the consistency of quality across the whole site.

**Metric detail:** IS-Score = the Information Satisfaction score distilled from rater judgments (the calibration norm itself); siteQualityStddev = standard deviation of quality across a site (consistency). [A IS, B/C Stddev]

**Evidence:** [A] the hardest evidence (DOJ/sworn).

**Why this attribution:** This norm defines what E-E-A-T means — it is the source, not a player. All four dimensions are fully affected because the QRG explicitly names Experience, Expertise, Authority and Trust as rating dimensions.

**Strategic consequence:** The QRG are publicly available and should be required reading for wetter.com's editorial and product teams. What a rater would judge as 'high quality' per the QRG is the goal — not what tools measure as 'optimized'. The QRG distinguish strictly between 'Page Quality' (the quality of the page itself) and 'Needs Met' (does the page match the search query).

### 13. Topic-Embeddings (Fokus/Radius)

- Function class: Quality & Prediction
- E-E-A-T: E=none · Exp=indirect · A=indirect · T=none
- Axis: Geometry
- Reach: Site
- Calibration source: Language
- Evidence code: B
- Leak fields: siteEmbeddings, siteFocusScore, siteRadius
- Feeds into: NSR (+ Fallback-Vererbung)

**What it measures:** Google computes a kind of 'topic map' from the entire text content of a website (embedding — a mathematical vector representation of the topic space). siteFocusScore measures how sharply the site's topic field is defined; siteRadius measures how widely it scatters thematically. A small radius means: a tightly focused site. The system measures thematic consistency, not subject-matter competence.

**Derived from:** Self-supervised from the site's language use — from the actual text of all pages, without human labels. The resulting vectors feed into NSR as a focus vector.

**Metric detail:** siteEmbeddings = vector representation of the site's topic space; siteFocusScore = how sharply outlined the topic field is; siteRadius = how widely the site scatters in topic space (small = focused). [B fields]

**Evidence:** [B] fields.

**Why this attribution:** Expertise and Authority are indirect ('flanking'): a focused weather site is more likely to be perceived as a weather expert than one that also offers recipes and travel reports alongside weather. But focus alone doesn't substitute for quality — it is a multiplier.

**Strategic consequence:** For wetter.com, a clear thematic focus (weather, climate data, meteorology) feeds the focus score positively. Content far outside the thematic core can raise the radius. For international expansions like pogoda24: plan thematic coherence from day one.

### 14. QualityBoost-Demotions

- Function class: Penalty & Spam
- E-E-A-T: E=none · Exp=none · A=none · T=primary
- Axis: Penalty ledger
- Reach: Doc
- Calibration source: Pattern
- Evidence code: B
- Leak fields: navDemotion, anchorMismatch, serpDemotion

**What it measures:** A penalty system for concrete quality and navigation deficiencies: navDemotion penalizes poor page guidance (unclear navigation, misleading structure); anchorMismatch penalizes links where the link text doesn't match the destination (e.g. 'more info' instead of 'Germany rain radar'); serpDemotion is derived from negative user behavior in search results — users click, immediately bounce back, click on competitors.

**Derived from:** Three separate deficiency signals: navigation problems on the page, inconsistency between link text and link destination, poor click behavior in search results.

**Metric detail:** navDemotion = a deduction for poor user guidance/navigation; anchorMismatch = anchor text does not match the target; serpDemotion = a deduction derived from SERP behavior. [B fields, C triggers]

**Evidence:** [B] fields.

**Why this attribution:** Only Trust is primary — penalty deductions are trust bookkeeping: each of these signals shows that a page breaks its implicit promise (good navigation, honest links, relevant content).

**Strategic consequence:** For editorial and UX at wetter.com: internal links should concretely describe where they lead ('Germany rain radar' rather than 'click here'). Pages that are often clicked in search results and cause immediate bounces are actively penalized — regularly reviewing these pages is worthwhile.

### 15. clutterScore / scamness / unauth.

- Function class: Penalty & Spam
- E-E-A-T: E=none · Exp=none · A=indirect · T=primary
- Axis: Penalty ledger
- Reach: Doc
- Calibration source: mixed
- Evidence code: B
- Leak fields: clutterScore, scamness, unauthoritativeScore

**What it measures:** Three quality deductions in one: clutterScore measures layout overload (too many ads, pop-ups, distracting elements that disrupt the information flow); scamness measures fraud proximity (patterns that resemble disreputable sites); unauthoritativeScore measures missing credibility (no author details, no recognizable institution behind the page).

**Derived from:** Machine pattern recognition from layout, text and structural features of the page.

**Metric detail:** clutterScore = a measure of layout overload/distraction; scamness = scam/fraud proximity; unauthoritativeScore = lack of authoritativeness. Higher values = a stronger trust deduction. [B fields, C scale]

**Evidence:** [B] fields.

**Why this attribution:** Trust is primary — all three signals are direct trust violations. Authority is indirect because unauthoritativeScore has a connection to authority.

**Strategic consequence:** For wetter.com: advertising and promotional elements should not interrupt the information flow — especially on severe weather pages where users need fast clarity. 'Answer first' as a principle: the relevant weather information appears at the top, not behind an ad banner. Visibly naming authors and meteorological sources strengthens the authoritativeness score.

### 16. Panda / BabyPanda

- Function class: Penalty & Spam
- E-E-A-T: E=none · Exp=indirect · A=none · T=primary
- Axis: Penalty ledger
- Reach: Site
- Calibration source: Rater (IS)
- Evidence code: B/O
- Leak fields: Panda, BabyPandaV2
- Fed by: IS-Eichung (Rater/QRG)

**What it measures:** Panda is a site-wide demotion system for pages with thin or low-quality content — applying to the entire website, not just individual pages. When too large a share of a site is rated as 'thin content' (content-poor pages), all pages of the site rank worse. BabyPandaV2 is a newer, more finely calibrated variant of the same principle.

**Derived from:** A demotion system calibrated to rater judgments — similar to chard it judges content quality, but based on site-wide patterns rather than individual pages.

**Metric detail:** Site-wide thin/low-quality demotion; BabyPandaV2 = a newer/lighter variant. Not a single-field score but a demotion system oriented to rater judgments. [B/O]

**Evidence:** [B] fields; [O] Panda historically communicated.

**Why this attribution:** Trust is primary (a site with lots of thin content is seen as unreliable), Expertise is indirect (thin content signals a lack of subject-matter competence to raters).

**Strategic consequence:** For wetter.com with thousands of automatically generated location pages, Panda is a real risk. Pages showing only weather data without context or added value can drag down the site average. Active pruning (removing or improving weak pages) is more important than constantly adding new ones.

### 17. Anchor-Spam (Penguin-Erbe)

- Function class: Penalty & Spam
- E-E-A-T: E=none · Exp=none · A=indirect · T=primary
- Axis: Anchor
- Reach: Doc
- Calibration source: Pattern
- Evidence code: B
- Leak fields: phraseAnchorSpamPenalty, IsAnchorBayesSpam

**What it measures:** This system detects manipulative link patterns in anchor text (the clickable text of hyperlinks). When a page is linked to exclusively with exact keyword texts ('Berlin weather', 'free weather forecast'), it looks unnatural and like deliberate SEO manipulation. Penguin is the historic Google update that addressed this problem — the logic lives on in today's system.

**Derived from:** phraseAnchorSpamPenalty (penalty score for over-optimized anchors) and IsAnchorBayesSpam (a classifier that gives a yes/no verdict on whether anchor spam is present).

**Metric detail:** phraseAnchorSpamPenalty = a penalty for over-optimized anchor phrases; IsAnchorBayesSpam = a Bayes-classifier flag (yes/no) for anchor spam. [B fields]

**Evidence:** [B] fields.

**Why this attribution:** Trust is primary — anchor manipulation is a direct breach of trust toward Google and users. Authority is indirect because the link profile influences a page's authority.

**Strategic consequence:** For the link-building team at wetter.com: natural, descriptive link texts are better than keyword-optimized ones. When asking external partners for links, don't prescribe ready-made anchor texts. Varied internal link texts ('Hamburg rain radar', 'North Sea storm warning') instead of always 'weather forecast' reduces the risk.

### 18. SpamBrain

- Function class: Penalty & Spam
- E-E-A-T: E=none · Exp=none · A=none · T=primary
- Axis: Penalty ledger
- Reach: Site
- Calibration source: Pattern
- Evidence code: B/O
- Leak fields: SpamBrain (scaled content abuse)

**What it measures:** Google's AI-based spam detection system. The aspect most relevant to wetter.com is 'scaled content abuse' — the mass production of content following the same pattern every time, without genuine editorial added value. Typical example: thousands of location pages that just plug variables (city name, coordinates, temperature data) into a template, without any human editorial quality check.

**Derived from:** A machine learning system that detects spam patterns in content. It delivers a binary spam verdict, not a quality score.

**Metric detail:** An ML system against spam; the named aspect 'scaled content abuse' = mass-produced thin content. It delivers a spam verdict, not a substance score. [B field, O system]

**Evidence:** [B] field; [O] officially communicated as a spam system.

**Why this attribution:** Only Trust — SpamBrain is pure trust protection: it protects search results from content that tries to deceive the system.

**Strategic consequence:** For wetter.com with many auto-generated location pages (and for expansions like pogoda24): pure template pages without editorial added value are a SpamBrain risk. Real added value per page — local peculiarities, climate history, own measurement data — protects against being classified as 'scaled content abuse'. Scaling content production must be flanked by quality assurance.

### 19. hostAge-Sandbox

- Function class: Penalty & Spam
- E-E-A-T: E=none · Exp=none · A=none · T=indirect
- Axis: Standing
- Reach: Site
- Calibration source: Pattern
- Evidence code: B
- Leak fields: hostAge
- Feeds into: Q* (Saldo)

**What it measures:** New domains and new hosts initially receive less trust credit from Google — not because they are penalized, but because there is no behavioral history yet from which Google could draw conclusions. This 'sandbox effect' describes passive damping, not an active penalty. As the host ages and builds a positive usage history, the damping fades.

**Derived from:** The age of the host (hostAge) — how long the domain or host has been known and active.

**Metric detail:** hostAge = age of the host; young hosts receive less trust credit — passive damping, not an active deduction. [B field, C sandbox effect]

**Evidence:** [B] field.

**Why this attribution:** Only Trust (indirect) — the missing trust credit is the only E-E-A-T connection. No bearing on Expertise, Experience or Authority.

**Strategic consequence:** Domain migrations (e.g. pogoda24 as a separate domain) or new sub-domain launches must account for an initial ramp-up period. Strategically: set up new domains early to start the clock — even if the full content launch follows later. Patience is the only strategy.

### 20. SegIndexer / Tiers

- Function class: Index & Infrastructure
- E-E-A-T: E=none · Exp=none · A=indirect · T=none
- Axis: Index economy
- Reach: Doc
- Calibration source: mixed
- Evidence code: B
- Leak fields: scaledSelectionTierRank (0-32767)
- Fed by: Q* (Saldo)

**What it measures:** Google doesn't index all pages equally quickly or cheaply. The SegIndexer decides which 'tier' (level) of the index a page lands in — from expensive, lightning-fast servers (for important pages) to slower, cheaper storage (for rarely needed pages). A high tier rank means: the page is quickly available when queries arrive.

**Derived from:** scaledSelectionTierRank — a rank value between 0 and 32,767 that reflects a mix of demand (how often is this page searched?), authority and storage economics.

**Metric detail:** scaledSelectionTierRank = a rank value for index-tier selection; 0-32767 is the technical 16-bit maximum (a finely resolved rank scale, not a quality value in itself). [B field, C tier logic]

**Evidence:** [B] field; [B/C] tier architecture.

**Why this attribution:** Authority is indirect — index-worthiness correlates with authority but does not judge content quality. The system is a 'gate' before the actual ranking: pages not in the fast tier appear more slowly in search results.

**Strategic consequence:** For wetter.com with thousands of location pages: weak pages occupy tier capacity that more important pages need. Active content pruning frees up capacity and improves indexing efficiency for the pages that matter — such as current severe weather warnings that must land in the fast tier.

### 21. Crawl-Budget

- Function class: Index & Infrastructure
- E-E-A-T: E=none · Exp=none · A=indirect · T=none
- Axis: Index economy
- Reach: Site
- Calibration source: Behavior
- Evidence code: O
- Leak fields: Crawl-Kapazität x Nachfrage

**What it measures:** How often and how regularly Google visits and updates a website's pages (crawling — Google 'reading in' the content). The budget is determined by two factors: capacity (how fast and error-free does the server respond?) and demand (how important does Google find these URLs?). For a dynamic weather site this is highly relevant: if Google doesn't crawl a page in time, the weather data in search results is outdated.

**Derived from:** Server response times and error rates (capacity side), plus popularity and freshness data (demand side).

**Metric detail:** Not a leak field but an official concept: capacity (what the server can handle — response time, error rate) × demand (how much Google wants to fetch the URLs — popularity, change frequency). [O]

**Evidence:** [O] officially documented by Google.

**Why this attribution:** Authority is indirect — crawl budget is a demand correlate and the earliest gate in the entire chain: pages that aren't crawled aren't indexed and can't rank.

**Strategic consequence:** For wetter.com with many thousands of URLs, crawl budget is a real bottleneck. Technical measures such as fast server response times, clean sitemap structures and the removal of unimportant or duplicate URLs free up budget — for pages that genuinely need it: daily severe weather warnings, real-time radar and forecast pages.

### 22. Mustang / SuperRoot etc.

- Function class: Index & Infrastructure
- E-E-A-T: E=none · Exp=none · A=none · T=none
- Axis: Infrastructure
- Reach: —
- Calibration source: —
- Evidence code: B
- Leak fields: Mustang, Ascorer, SuperRoot, Twiddler (inkl. Host-Crowding)

**What it measures:** Mustang, Ascorer, SuperRoot and Twiddler form Google's technical scoring infrastructure — they calculate and coordinate the ranking but do not themselves judge content quality. Twiddler also contains 'host-crowding': a rule that prevents a single domain from dominating too many search results at once. These systems are the 'negative entry' of the matrix — there is nothing to optimize here.

**Derived from:** The outputs of all other rating systems — these systems coordinate, they measure nothing themselves.

**Metric detail:** Mustang = scoring/serving infrastructure; Ascorer = the scorer; SuperRoot = the central coordinator; Twiddler = a re-ranking layer (incl. non-evaluative host-crowding/diversity). No measured values. [B names, B/C roles]

**Evidence:** [B] names; [B/C] roles.

**Why this attribution:** No E-E-A-T bearing on all four dimensions — pure infrastructure, no judgment.

**Strategic consequence:** No direct action needed. Only Twiddler's host-crowding is practically relevant: if wetter.com has too many similar pages for the same query (e.g. weather for Berlin-Mitte, Berlin-Prenzlauer Berg, Berlin-Kreuzberg as separate pages), Google may limit the number of results shown to 2-3 per domain. Consolidating overly similar pages helps here — not as a quality issue, but as a display rule.

### 23. RankBrain / DeepRank / FastSearch

- Function class: Index & Infrastructure
- E-E-A-T: E=none · Exp=indirect · A=indirect · T=indirect
- Axis: AI citation
- Reach: Doc
- Calibration source: Behavior
- Evidence code: B/O
- Leak fields: RankBrain, DeepRank, RankEmbedBERT

**What it measures:** These systems measure the semantic fit between a search query and a page — whether the meaning matches, not just the keywords. RankBrain was Google's first AI system for this purpose; DeepRank is a deeper language understanding layer; FastSearch accelerates this estimate for faster responses. The key point: these systems measure relevance, not content quality.

**Derived from:** Learned mathematical representations (embeddings) of search queries and documents, plus behavioral calibration (which search results did users actually rate as satisfying?).

**Metric detail:** RankBrain = an ML system for (novel) queries via embeddings; DeepRank = deeper language understanding (BERT-related); RankEmbedBERT = the BERT-based embedding field. They measure fit, not quality. [B fields, O RankBrain/BERT]

**Evidence:** [B] fields; [O] RankBrain/BERT officially communicated.

**Why this attribution:** Expertise, Authority and Trust are each indirect — these systems flank all E-E-A-T dimensions through better meaning understanding, but don't measure them directly. Experience has no bearing: meaning similarity doesn't touch first-hand material.

**Strategic consequence:** For wetter.com: keyword stuffing (cramming in as many search terms as possible) no longer works — RankBrain recognizes whether the meaning is right. Writing text that genuinely answers the information need behind a query performs better. Also, RankBrain bridges to AI citation: pages that are semantically well-positioned have better chances of appearing in AI overviews (AIO).

### 24. FreshnessTwiddler / RealTimeBoost

- Function class: Freshness
- E-E-A-T: E=none · Exp=none · A=none · T=indirect
- Axis: Freshness
- Reach: Doc
- Calibration source: Pattern
- Evidence code: B
- Leak fields: FreshnessTwiddler, RealTimeBoost
- Fed by: Datums-Triangulation

**What it measures:** For certain queries Google expects fresh content — the concept is called 'Query Deserves Freshness' (QDF). FreshnessTwiddler gives current pages a temporary ranking boost. RealTimeBoost is the sharper variant for real-time events like severe weather or breaking news. Being current means ranking higher — but only if the query actually demands freshness.

**Derived from:** A combination of the document's freshness state (when was it last updated?) and the query's freshness demand (are many people currently searching with a time reference for this topic?).

**Metric detail:** FreshnessTwiddler = a freshness-dependent re-ranking boost (Query Deserves Freshness); RealTimeBoost = a sharper, short-lived boost for real-time/event spikes. [B fields, O/C QDF]

**Evidence:** [B] fields; [O/C] the QDF concept.

**Why this attribution:** Trust is indirect — an outdated severe weather warning is a breach of trust: someone relies on wrong information. Otherwise this system is E-E-A-T-neutral: freshness is not a dimension of expertise or authority.

**Strategic consequence:** For wetter.com freshness is central — especially for severe weather warnings, storm track forecasts and real-time radar. But the freshness boost only works if pages are crawled and indexed in time (cf. crawl budget, system 21). Editorial processes for fast updates during weather events must be synchronized with the technical infrastructure.

### 25. Datums-Triangulation

- Function class: Freshness
- E-E-A-T: E=none · Exp=none · A=none · T=primary
- Axis: Freshness
- Reach: Doc
- Calibration source: Pattern
- Evidence code: B/O
- Leak fields: bylineDate, syntacticDate, semanticDate
- Feeds into: FreshnessTwiddler / RealTimeBoost

**What it measures:** Google checks a page's date by comparing three different signals: (1) the visible date in the article (bylineDate — easy to fake), (2) the date from URL, markup or timestamp (syntacticDate — also easy to manipulate), (3) the date inferred from the text content (semanticDate — can only be changed through genuine content updates). When the three don't agree, Google detects a 'date lie'.

**Derived from:** Three independent date signals with different costs to fake — the system uses the difference between easy-to-fake and hard-to-fake signals to verify date honesty.

**Metric detail:** bylineDate = a visibly stated date (cheaply faked); syntacticDate = extracted from URL/markup/timestamp (cheaply faked); semanticDate = inferred from the content (only fakeable through genuine updating). The cross-check exposes date lies. [B fields, O date determination]

**Evidence:** [B] fields; [O] officially communicated for date determination.

**Why this attribution:** Trust is primary and the only primary signal in the entire freshness class — date deception is a direct breach of trust: disguising an old document as 'new' by changing the date is detected.

**Strategic consequence:** For wetter.com and especially for seasonal content (spring weather, summer tips, winter warnings): a date should only be updated when the content has genuinely been updated. Simply changing the date in the CMS without content changes — a widespread SEO practice — is detected by semanticDate and damages the trust score. Content must be substantively revised.

## Metric catalog (leak fields)

| Field | Definition | System | Class | Provenance | Evidence |
| --- | --- | --- | --- | --- | --- |
| `goodClicks / badClicks` | Clicks of a query×page rated as satisfied or disappointed. | NavBoost | Behavior | Behavior | [B] |
| `lastLongestClicks` | Last, longest click of a session — the strongest satisfaction signal. | NavBoost | Behavior | Behavior | [B] |
| `GlueResponse` | Behavioral input on SERP features that Tangram assembles for display. | Glue / Tangram | Behavior | Behavior | [B/C] |
| `chromeInTotal / uniqueChromeViews` | Total Chrome views or unique viewers = site-wide attachment. | Chrome | Behavior | Behavior | [B] |
| `unscaledIpPriorBadFraction` | Share of suspicious clicks per IP range (manipulation damping). | IP-Prior | Behavior | Pattern | [B/C] |
| `voterToken` | Identifies a click 'ballot', making repeat voting harder. | IP-Prior | Behavior | Pattern | [C] |
| `siteAuthority` | Authority distilled from quality_nsr, applied in Q*. | Q* / NSR | Quality & Prediction | Composite | [B] |
| `lowQuality` | Low-quality flag, converted from quality_nsr.NsrData. | Q* | Quality & Prediction | Composite | [B] |
| `predictedDefaultNsr` | Historicized baseline quality score (VersionedFloatSignal → the trajectory counts). | NSR | Quality & Prediction | Composite | [B] |
| `nsrConfidence` | Confidence in its own NSR judgment (deprecated). | NSR | Quality & Prediction | Composite | [B] |
| `chardScore` | Content-quality value per document, calibrated to rater judgments. | chard | Quality & Prediction | Rater | [B] |
| `chardVariance` | Dispersion/uncertainty of the chardScore (high = an uncertain judgment). | chard | Quality & Prediction | Rater | [B/C] |
| `contentEffort` | LLM-estimated creation effort and depth of a page. | contentEffort | Quality & Prediction | Rater | [B] |
| `OriginalContentScore` | Originality/first-hand-material degree — the closest machine Experience proxy. | OCS | Quality & Prediction | Rater | [B/C] |
| `tofu / keto / Rhubarb` | Codenames for refinement/delta signals at the subchunk level. | tofu-Familie | Quality & Prediction | Rater | [B/C] |
| `siteEmbeddings` | Vector representation of the site's topic space. | Topic-Embeddings | Quality & Prediction | Language | [B] |
| `siteFocusScore` | How sharply outlined the site's topic field is. | Topic-Embeddings | Quality & Prediction | Language | [B] |
| `siteRadius` | How widely the site scatters in topic space (small = focused). | Topic-Embeddings | Quality & Prediction | Language | [B] |
| `IS-Score` | Information Satisfaction from rater judgments — the calibration norm. | IS-Eichung | Quality & Prediction | Rater | [A] |
| `siteQualityStddev` | Standard deviation of quality across a site (consistency). | IS-Eichung | Quality & Prediction | Rater | [B/C] |
| `navDemotion` | Deduction for poor user guidance/navigation. | QualityBoost | Penalty & Spam | Pattern | [B] |
| `anchorMismatch` | Anchor text does not match the link target. | QualityBoost | Penalty & Spam | Link | [B] |
| `serpDemotion` | A deduction derived from SERP behavior. | QualityBoost | Penalty & Spam | Behavior | [B] |
| `clutterScore` | A measure of layout overload and distraction. | clutterScore | Penalty & Spam | Pattern | [B] |
| `scamness` | Scam/fraud proximity of a page. | clutterScore | Penalty & Spam | Pattern | [B] |
| `unauthoritativeScore` | Lack of authoritativeness. | clutterScore | Penalty & Spam | Pattern | [B] |
| `BabyPandaV2` | Site-wide thin/low-quality demotion (Panda heritage). | Panda | Penalty & Spam | Rater | [B/O] |
| `phraseAnchorSpamPenalty` | Penalty for over-optimized anchor phrases. | Anchor-Spam | Penalty & Spam | Link | [B] |
| `IsAnchorBayesSpam` | Bayes-classifier flag for anchor spam (yes/no). | Anchor-Spam | Penalty & Spam | Link | [B] |
| `hostAge` | Host age; young hosts receive less trust credit. | hostAge | Penalty & Spam | Pattern | [B/C] |
| `scaledSelectionTierRank` | Rank for index-tier selection (0–32767 = 16-bit maximum). | SegIndexer | Index & Infrastructure | Composite | [B/C] |
| `crawl capacity × demand` | What the server can handle × how much Google wants to fetch the URLs. | Crawl-Budget | Index & Infrastructure | Behavior | [O] |
| `RankEmbedBERT` | BERT-based embedding for deeper query understanding. | RankBrain / DeepRank | Index & Infrastructure | Language | [B/O] |
| `Mustang / Ascorer / SuperRoot` | Scoring/serving infrastructure and compositor — evaluates nothing. | Mustang | Index & Infrastructure | Infrastructure | [B/C] |
| `bylineDate` | A visibly stated date — cheaply faked. | Datums-Triangulation | Freshness | Pattern | [B] |
| `syntacticDate` | Extracted from URL/markup/timestamp — cheaply faked. | Datums-Triangulation | Freshness | Pattern | [B] |
| `semanticDate` | Inferred from the content — only fakeable through genuine updating. | Datums-Triangulation | Freshness | Pattern | [B] |
| `FreshnessTwiddler / RealTimeBoost` | Freshness boost (QDF) and short-lived real-time spikes. | FreshnessTwiddler | Freshness | Pattern | [B] |

## Best practices → named systems

### Original information, own research or analysis; substantial, complete coverage of the topic that goes beyond the obvious — not mere summaries or rewrites.

- Guidance: Helpful Content: Content & Quality
- Leak fields: OriginalContentScore [B], contentEffort [B], chardScore [B] · density: high
- Rater anchor: primary — chard, OriginalContentScore and contentEffort are directly calibrated to rater judgments.
- Officially named systems:
  - Original content systems [O] — Ensures first-hand material is surfaced; does not reward scraping or copying. (https://developers.google.com/search/docs/appearance/ranking-systems-guide)
  - Helpful content system (part of Core Ranking) [O] — Evaluates people-first helpfulness site-wide. (https://developers.google.com/search/docs/appearance/ranking-systems-guide)
  - Panda (integrated into Core Ranking) [O] — Favored high-quality, original content. (https://developers.google.com/search/docs/appearance/ranking-systems-guide)

**Comment:** Strongest cluster: substantial first-hand content feeds multiple official systems at once.

### Content for people, not primarily for search engines; no mass or automated production without added value; no content that exists only to capture traffic.

- Guidance: Helpful Content: People-first / Avoid search-engine-first
- Leak fields: Panda / BabyPandaV2 [B], niedriger contentEffort (Risikomarker) [B] · density: medium-high
- Rater anchor: primary — The Helpful content system and Panda are calibrated to rater quality judgments.
- Officially named systems:
  - Helpful content system [O] — Demotes search-engine-first content. (https://developers.google.com/search/docs/appearance/ranking-systems-guide)
  - Spam detection systems (incl. SpamBrain) [O] — Addresses scaled content abuse among other issues. (https://developers.google.com/search/docs/appearance/ranking-systems-guide)

**Comment:** The penalty side of the substance cluster: the same property viewed from the demotion angle.

### The content satisfies what someone is actually searching for and leaves the feeling of being well served — clear search intent met.

- Guidance: Helpful Content: Search intent & good experience
- Leak fields: NavBoost / Glue (goodClicks, lastLongestClicks) [B] · density: high
- Rater anchor: indirect — Neural matching/RankBrain/BERT are algorithmic; raters confirm indirectly via the Needs Met rating.
- Officially named systems:
  - Neural matching [O] — Understands meaning behind query and content. (https://developers.google.com/search/docs/appearance/ranking-systems-guide)
  - RankBrain [O] — Connects words with concepts. (https://developers.google.com/search/docs/appearance/ranking-systems-guide)
  - BERT [O] — Understands word combinations and intent. (https://developers.google.com/search/docs/appearance/ranking-systems-guide)
  - Passage ranking system [O] — Finds relevant individual passages. (https://developers.google.com/search/docs/appearance/ranking-systems-guide)

**Comment:** The official side is query/meaning understanding; behavioral confirmation (NavBoost) is partly outside the text itself.

### Cleanly produced: free of errors, not carelessly made, without intrusive advertising, usable on mobile.

- Guidance: Helpful Content: Presentation & production
- Leak fields: clutterScore [B], scamness [B] · density: medium
- Rater anchor: none — Page Experience relies on technical signals — no direct rater involvement.
- Officially named systems:
  - Page experience system [O] — Mobile-friendliness, HTTPS, Safe Browsing, no intrusive interstitials. (https://developers.google.com/search/docs/appearance/ranking-systems-guide)

**Comment:** Directly addressable; provenance pattern.

### Recognizable expertise in the topic area; the site/author has a traceable background; others recommend or cite the source.

- Guidance: Helpful Content: Expertise + E-E-A-T
- Leak fields: NSR / siteAuthority [B], Topic-Embeddings (siteFocusScore) [B], IS-Score (Information Satisfaction) [A] · density: medium
- Rater anchor: primary — Reliable information systems and IS-Score are built on rater judgments.
- Officially named systems:
  - Reliable information systems [O] — Elevates authoritative pages, demotes low-quality ones, rewards quality journalism. (https://developers.google.com/search/docs/appearance/ranking-systems-guide)
  - Link analysis systems & PageRank [O] — Reputation/endorsement through the link structure. (https://developers.google.com/search/docs/appearance/ranking-systems-guide)

**Comment:** Partly only earnable over time and through external attribution, not by writing alone.

### Clear who created the content and why; the responsible party/author is identifiable and trustworthy; for YMYL: accuracy and reliability as the core E-E-A-T component.

- Guidance: Helpful Content: Who/How/Why + Trust
- Leak fields: unauthoritativeScore [B], Trust-Demotions (QualityBoost-Familie) [B] · density: low
- Rater anchor: primary — Strongest rater dependency: Google checks trust/accountability primarily through rater judgment — no algorithmic field replaces it.
- Officially named systems:
  - Reliable information systems [O] — The only broad official system; demotes unreliable content, elevates authoritative. (https://developers.google.com/search/docs/appearance/ranking-systems-guide)

**Comment:** KEY FINDING: most thinly backed by named official systems. Google checks accountability primarily through rater judgment — no checkable field. No marker strategy works here; genuine substance counts most.

### For time-sensitive topics, content is current and maintained/updated as needed.

- Guidance: Helpful Content (implicit) + QRG Freshness
- Leak fields: FreshnessTwiddler / RealTimeBoost [B], semanticDate (teuer fälschbar) [B] · density: medium
- Rater anchor: none — Freshness systems work algorithmically on date signals — no rater calibration.
- Officially named systems:
  - Freshness systems [O] — Surfaces fresher content where recency is expected (QDF). (https://developers.google.com/search/docs/appearance/ranking-systems-guide)

**Comment:** Strategically central for wetter.com: time-critical forecasts and news.

### No manipulative practices: no keyword stuffing, no link schemes, no misleading techniques.

- Guidance: Spam Policies + Search Essentials
- Leak fields: phraseAnchorSpamPenalty / IsAnchorBayesSpam [B] · density: high
- Rater anchor: none — SpamBrain and Penguin are ML/pattern-based — no rater involvement.
- Officially named systems:
  - Spam detection systems (SpamBrain) [O] — Detects spam patterns algorithmically. (https://developers.google.com/search/docs/appearance/ranking-systems-guide)
  - Penguin (integrated into Core Ranking) [O] — Demotes spammy link building. (https://developers.google.com/search/docs/appearance/ranking-systems-guide)
  - Exact match domain system [O] — Prevents excessive credit for keyword-match domains. (https://developers.google.com/search/docs/appearance/ranking-systems-guide)

**Comment:** Well backed; avoidance cluster (what NOT to do).