The E-E-A-T mechanics, well-sourced and medium rare
An evidence-disciplined reference of the 25 documented evaluation systems from the Content-Warehouse leak and DOJ material — their metrics and their contribution to Experience, Expertise, Authoritativeness, Trust.
The short version
E-E-A-T is not a ranking signal and not a score in the code. It is the checklist that Google's human quality raters use, and part of the ranking models is calibrated against it. The checklist is the Quality Rater Guidelines. This reference separates what Google measures, what it supports indirectly, and what you can only comply with.
~16,000
external Search Quality Raters worldwide
Google employs ~16,000 external raters who continuously evaluate search results according to the QRG. Quality, trust and relevance require human judgment that algorithms alone cannot capture. The raters represent real users across languages and cultures. Their assessments calibrate the models that then rank billions of pages.
What Google defines as spam and which practices lead to penalties.
Page Quality
Page Quality asks how well the page fulfills its purpose: substance, creation effort, reliability. E-E-A-T sits here. It is a property of the page itself and does not depend on the query.
Needs Met
Needs Met asks how well the page serves the concrete search intent. It depends on the query. You cannot read it from the page alone, only in relation to the query.
↳ Never an average: Page Quality and Needs Met are judged separately.
Five roles in ranking
Every ranking system does one of five jobs — they are grouped by that job here.
BehaviorUser interaction — clicks and satisfaction as a re-ranking signal.n=5
Quality & PredictionContent quality and its model-based prediction.n=8
Penalty & SpamDeductions and manipulation defense.n=6
Index & InfrastructureProcessing, selection and serving — gates, not judgment.n=4
FreshnessRecency and its dating.n=2
§B
Where the systems pay in
Each row shows which E-E-A-T dimensions a system feeds. Filled (●) means it measures or verifies the dimension directly. Outline (◐) means it contributes indirectly. Chip color shows the role. Click a system to open its detail.
IS calibration (Rater / QRG) — the norm that defines all four dimensions · ●●●● · [A]
This is not a signal used at ranking time. It produces the target values the content models are trained against. It stands above the columns, not in one.
E
Experience
Self-experienced?
● 1 · ◐ 4
● primary
◐ indirect
E
Expertise
Expert-grounded?
● 2 · ◐ 9
● primary
◐ indirect
A
Authority
Recognized source?
● 4 · ◐ 8
● primary
◐ indirect
T
Trust
Reliable, honest?
● 10 · ◐ 9
● primary
◐ indirect
Foundation — Mustang / SuperRoot / Twiddler: scoring and serving infrastructure, evaluates nothing · ––––. Not optimizable, pays into no dimension.
§C
What each system derives its judgment from
Signal source decides how you can influence a dimension and how much you can move it. It cuts across the role.
How much you can move it ↓ ① Rater (steerable through substance) › ② Behavior (only via genuine satisfaction) › ③ Link (external attribution) › ④ Pattern (compliance only)
1Rater-based
Calibrated to human quality judgments (IS / QRG). The only provenance that addresses substance work directly.
Calibrated to lived user demand (clicks, satisfaction). Not directly buildable — only earned through genuine satisfaction.
NavBoostGlueChromeRankBrainCrawl-Budget
3Link-based
From the link graph / external attribution. Barely a system of its own — flows primarily as a vector into NSR / siteAuthority. Penguin is the abuse side.
Anchor-Spam (Penguin)Link vector → NSR
4Pattern-based
Rule-, heuristic- and classifier-based (spam, dates, demotions). Not to be built up but to be complied with and avoided.
Learned from language use itself (embeddings). Measures geometry / focus, not capability — a multiplier, not a substitute.
Topic-Embeddings
Composite & Infrastructure
Bundles several sources (NSR: links + behavior + focus + content) or judges nothing at all (Mustang, Tangram).
NSR (mixed)clutterScoreSegIndexerMustangTangram
The almost empty Link band is the finding: in the documented material, links are barely a standalone evaluating system. They flow as a vector into the NSR composite. Dashed means not a standalone evaluating system (vector, composite part or infrastructure).
§D
Field catalog
This catalog lists every documented leak field with its name (monospace), definition, source badge and signal source. You can search and filter it.
⌕
38/38 fields
Role
Signal source
Source
Behavior6 fields
goodClicks / badClicksB
Clicks for a combination of search query and page, rated as satisfied or disappointed.
Behavior
lastLongestClicksB
The last and longest click of a session. It is the strongest satisfaction signal.
Behavior
GlueResponseB/C
Behavioral data on SERP features (extra elements on the results page) that Tangram assembles for display.
Behavior
chromeInTotal / uniqueChromeViewsB
Total Chrome views or unique viewers. Both measure site-wide attachment.
Behavior
unscaledIpPriorBadFractionB/C
Share of suspicious clicks per IP range. It dampens manipulation.
Pattern
voterTokenC
Identifies a click 'ballot' and so makes repeat voting harder.
Pattern
Quality & Prediction14 fields
siteAuthorityB
Authority distilled from quality_nsr. It is applied in Q*.
Composite
lowQualityB
Flag for low quality, converted from quality_nsr.NsrData.
Composite
predictedDefaultNsrB
Baseline quality score, stored with its history (VersionedFloatSignal), so the trajectory counts.
Composite
nsrConfidenceB
Confidence in its own NSR judgment (deprecated, meaning marked as obsolete).
Composite
chardScoreB
Content-quality value per document, calibrated against the judgments of human raters.
Rater
chardVarianceB/C
Spread, meaning uncertainty, of the chardScore. A high value means an uncertain judgment.
Rater
contentEffortB
Creation effort and depth of a page, as estimated by an LLM (large language model).
Rater
OriginalContentScoreB/C
Degree of originality and first-hand material. It is the measure that comes closest to Experience for a machine (a proxy).
Rater
tofu / keto / RhubarbB/C
Codenames for refinement and delta signals at the subchunk level.
Rater
siteEmbeddingsB
The site's topic space, expressed as a vector (a list of numbers).
Language
siteFocusScoreB
How sharply outlined the site's topic field is.
Language
siteRadiusB
How widely the site spreads across topic space. A small value means focused.
Language
IS-ScoreA
Information Satisfaction, taken from the judgments of human raters. It serves as the calibration norm.
Rater
siteQualityStddevB/C
Standard deviation of quality across a site. It measures consistency.
Rater
Penalty & Spam10 fields
navDemotionB
Deduction for poor user guidance and navigation.
Pattern
anchorMismatchB
Anchor text does not match the link target.
Link
serpDemotionB
A deduction derived from behavior on the search results page (SERP).
Behavior
clutterScoreB
A measure of layout overload and distraction.
Pattern
scamnessB
How close a page is to scams or fraud.
Pattern
unauthoritativeScoreB
Lack of authoritativeness.
Pattern
BabyPandaV2B/O
Site-wide demotion for thin or low-quality content, inherited from Panda.
Rater
phraseAnchorSpamPenaltyB
Penalty for over-optimized anchor phrases.
Link
IsAnchorBayesSpamB
Yes/no flag from a Bayes classifier for anchor spam.
Link
hostAgeB/C
Age of the host. Young hosts receive less trust credit.
Pattern
Index & Infrastructure4 fields
scaledSelectionTierRankB/C
Rank used to select the index tier. Values run from 0 to 32767, the 16-bit maximum.
Composite
crawl capacity × demandO
The product of two factors: what the server can handle and how much Google wants to fetch the URLs.
Behavior
RankEmbedBERTB/O
BERT-based embedding (a numeric representation of text) for deeper query understanding.
Language
Mustang / Ascorer / SuperRootB/C
Infrastructure for scoring and serving, and a compositor. It evaluates nothing.
Infrastructure
Freshness4 fields
bylineDateB
A visibly stated date. It is cheap to fake.
Pattern
syntacticDateB
Extracted from the URL, markup or timestamp. That is cheap to fake.
Pattern
semanticDateB
Inferred from the content. It can only be faked through genuine updating.
Pattern
FreshnessTwiddler / RealTimeBoostB
Freshness boost (QDF, Query Deserves Freshness) and short-lived real-time spikes.
Pattern
§E
System detail
All 25 systems with the matrix columns. Mechanics shows the technical internals; Plain text foregrounds the strategic consequence.
Google's official Helpful Content best practices, mapped to the officially named ranking systems. The mapping itself is our assessment [C]. Both ends are documented by Google [O].
Mapping = our assessment [C] · Both ends documented by Google [O] · Leak fields [B]/[A] in optional deep-dive
Original information, own research or analysis; substantial, complete coverage of the topic that goes beyond the obvious — not mere summaries or rewrites.
Official backing
High
Rater involvement●Direct
chard, OriginalContentScore and contentEffort are calibrated directly against the judgments of human raters, the people who review search results.
The best-backed group of practices: substantial first-hand content serves several official systems at once.
Technical deep-dive
Leak fields — not officially confirmed by Google.
OriginalContentScore[B]
contentEffort[B]
chardScore[B]
Content for people, not primarily for search engines; no mass or automated production without added value; no content that exists only to capture traffic.
Official backing
Medium
Rater involvement●Direct
The Helpful content system and Panda are calibrated against quality ratings by human raters.
Addresses, among other issues, scaled content abuse (content mass-produced mainly to manipulate rankings).
The penalty side of the substance group: the same property, seen as demotion, meaning a lower ranking.
Technical deep-dive
Leak fields — not officially confirmed by Google.
Panda / BabyPandaV2[B]
niedriger contentEffort (Risikomarker)[B]
The content satisfies what someone is actually searching for and leaves the feeling of being well served — clear search intent met.
Official backing
High
Rater involvement◐Indirect
Neural matching, RankBrain and BERT are algorithmic. Raters confirm indirectly, through the Needs Met rating (how well a result meets what the searcher needed).
The official systems here understand queries and meaning. Whether users were satisfied is confirmed by their behavior (NavBoost, the system that evaluates clicks from search), which lies partly outside the text itself.
Technical deep-dive
Leak fields — not officially confirmed by Google.
NavBoost / Glue (goodClicks, lastLongestClicks)[B]A stand-in measured from user behavior for Needs Met (how well a result meets the searcher's need). Text alone does not produce it. It is earned through genuine satisfaction.
Cleanly produced: free of errors, not carelessly made, without intrusive advertising, usable on mobile.
Official backing
Medium
Rater involvement–None
Page Experience relies on technical signals — no raters are directly involved.
Reputation and endorsement, read from which sites link to which.
Partly, this can only be earned over time and through what others say about you, not by writing alone.
Technical deep-dive
Leak fields — not officially confirmed by Google.
NSR / siteAuthority[B]
Topic-Embeddings (siteFocusScore)[B]
IS-Score (Information Satisfaction)[A]
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.
Official backing
Low
Rater involvement●Direct
Strongest dependency on human raters: Google checks trust and accountability mainly through their judgment. No algorithmic field replaces it.
The only broad official system. It ranks unreliable content lower and authoritative content higher.
KEY FINDING: this practice is backed by the fewest named official systems. Google checks accountability mainly through the judgment of human raters, and there is no checkable field. A strategy based on adding markers does not work here. Genuine substance counts most.
Technical deep-dive
Leak fields — not officially confirmed by Google.
unauthoritativeScore[B]
Trust-Demotions (QualityBoost)[B]
For time-sensitive topics, content is current and maintained/updated as needed.
Official backing
Medium
Rater involvement–None
Freshness systems work algorithmically on date signals. They are not calibrated against human raters.
Prevents excessive credit for domains whose name matches a keyword.
Well backed. This group is about avoidance: what NOT to do.
Technical deep-dive
Leak fields — not officially confirmed by Google.
phraseAnchorSpamPenalty / IsAnchorBayesSpam[B]
Context: two systems without a best-practice equivalent
Site diversity system [O]: max. ~2 results per domain in top results — a display rule, not a quality judgment. Consequence: consolidate to one strong URL per search intent.
Product reviews system [O]: rewards well-researched, original reviews (relevant only for sites publishing reviews).
Almost every official best practice has mechanical counterparts — often several. But the density is uneven: where many systems underlie (substance, spam avoidance, intent), the guidance is hard-backed. Where few do (accountability/trust), Google relies on rater judgment and learned representations — no marker strategy works there, only genuine substance. Lead with the guidance; verify with the systems — never the reverse.
§G
System dependency chains
Select a system to see what feeds it (upstream) and what it feeds into (downstream). The rater connection shows whether Quality Rater judgments inform this system directly, indirectly, or conceptually.