How Google Scores Your Content: The Five Quality Models Behind Every Ranking

The Honest Hedge

Every analysis has a threshold where certainty ends and inference begins. Here's where that line falls for this research:

---

---

The Source: Google's QualityNsrPQData Module

In March 2024, an automated bot accidentally pushed over 2,500 pages of Google's internal Content Warehouse API documentation to a public GitHub repository. SEO researchers Mike King and Rand Fishkin brought the leak to public attention, revealing more than 14,000 attributes used in Google's ranking systems.

Among the most significant revelations was a module called QualityNsrPQData — a data structure that encapsulates page-level quality signals feeding directly into Google's core ranking systems. Inside this module sit five quality scoring models, each with a food-related codename, each evaluating your content from a different angle.

Google later acknowledged the leak's authenticity while cautioning against "inaccurate assumptions" based on "out of context, outdated, or incomplete information." That's Google's way of saying: yes, these documents are real, but the system is more complex than what you can see here.

---

Google's Quality Scoring Ensemble: Five Models, One Verdict

Google doesn't run one quality check. The API leak reveals five quality-scoring models, each measuring a different dimension of quality:

Model	What It Measures	Scope
Chard	Content quality baseline	Site + Page level
Tofu	Initial quality for new URLs	Page level (bootstrap)
Keto	Versioned quality prediction	Site + Page level
Rhubarb	Page quality vs. site quality delta	Page-relative-to-site
contentEffort	LLM-judged authoring effort	Page level

The co-presence of five quality attributes in a single module is consistent with ensemble architecture — the same approach used in modern ML systems where multiple models contribute to a prediction, and the aggregate is more reliable than any individual model. If these models do function as an ensemble, disagreement between them would reduce confidence while agreement would increase it.

---

Google's Chard Quality Scorer: The Content Quality Baseline

Chard is the baseline quality classifier — the model most consistent with the original Panda algorithm's design. It evaluates content quality using content-based signals (not behavioral or link-based), generating a score that describes how inherently good or bad a page's content is.

What makes Chard distinctive is its variant system. The API leak reveals multiple Chard variants, each specialized for different detection tasks:

• Standard Chard — General content quality assessment
• Hoax detection variant — Identifies misinformation and factually unreliable content
• Translated content variant — Detects machine-translated or poorly translated pages
• YMYL variant — Applies heightened scrutiny to Your Money or Your Life content (health, finance, safety)

Chard's dual-level operation (page AND site) means it generates two signals: how good this specific page is, and how good the entire site is. The site-level signal creates a ceiling. If your site's aggregate Chard score is low, individual pages face a headwind — unless Rhubarb rescues them (more on that below).

Chard's Patent Connection to Google Panda

Chard is likely the production implementation of the phrase-model quality prediction described in Patent US9767157B2 (Predicting Site Quality), where n-gram frequency patterns are mapped to baseline quality scores. The patent, co-invented by Navneet Panda himself, describes precisely this kind of content-based quality predictor.

---

Google's Tofu Quality Model: Trust on First Use

Every new page has a cold start problem. NavBoost needs click data. Chard needs enough content to analyze. Behavioral signals take weeks to accumulate. So how does Google decide where a brand-new page should rank?

Tofu — likely Trust on First Use — appears to be a predictive quality score assigned to new URLs before they have any history, giving them a starting position in search results.

Based on its bootstrap function, Tofu likely evaluates signals such as:

• Site-wide quality signals — The parent domain's existing quality scores (inheriting from Chard's site-level assessment)
• Technical health indicators — Page speed, mobile-friendliness, proper markup
• Topical alignment — Whether the new page's content is consistent with the site's established topical focus
• Outbound link profile — What the page links to as an indicator of its quality neighborhood

☝ The API reveals tofu as a single numeric attribute — these input dimensions are inferred from its bootstrap function, not from the API documentation itself.

The implication is clear: new pages on established, high-quality domains get a better starting position than new pages on fresh domains. This isn't a conspiracy — it's a sensible bootstrapping strategy. If a domain has consistently produced quality content, its next page is likely quality too.

---

Google's Keto Quality Model: The Versioned Predictor

Keto is Google's most actively versioned quality model, described in the API documentation as a VersionedFloatSignal. That data type is significant — it means Google can deploy new versions of Keto independently, and the versioning structure is consistent with A/B testing different model iterations.

While Chard appears to use a phrase-model approach (rooted in the 2013 Panda patent), Keto's architecture is unknown. The VersionedFloatSignal type tells us the output format, not the model behind it — though the versioning pattern is consistent with modern ML deployment practices.

What the versioning tells us:

• Continuous improvement — Google isn't running one static quality model. Keto gets updated, and each version is tracked separately
• A/B testing capability — Different Keto versions can run on different traffic segments, allowing Google to validate improvements before full rollout
• The quality bar moves — Content that scored well under Keto v1 might score differently under Keto v3. The target is moving

---

Google's Rhubarb Quality Model: The Page-Site Delta Scorer

Rhubarb answers the question every content creator on a mediocre domain has asked: "Can my individual page overcome my site's reputation?"

The answer appears to be yes — and Rhubarb is the strongest candidate mechanism.

• Positive Rhubarb → Page is better than the site average. The page can rise above its domain's quality ceiling
• Zero Rhubarb → Page is consistent with its site. Standard site-level signals apply
• Negative Rhubarb → Page is worse than the site average. Even a strong site can't save a weak page

This has profound implications for content strategy. You don't need to fix every page on your site before your best content can rank. You need to make your best content demonstrably better than your site average. Rhubarb will notice.

---

Google's contentEffort Quality Model: The LLM-Based Quality Judge

This is the newest and most remarkable signal in the ensemble. Google is using a large language model to evaluate whether content shows genuine human effort — and the signal feeds directly into ranking.

Based on its alignment with Google's Quality Rater Guidelines, contentEffort likely evaluates dimensions such as:

Dimension	What It Looks For	Low-Effort Signal
Utility	Does the content solve the user's problem?	Generic advice, surface-level coverage
Expertise	Does the author demonstrate domain knowledge?	Wikipedia-style summaries, no depth
Originality	Does the content add something competitors don't?	Rehashed talking points, template content
Trustworthiness	Are claims sourced? Is authorship clear?	Unsourced assertions, anonymous content

☝ The API reveals contentEffort as a single attribute — these evaluation dimensions are inferred from alignment with Google's Quality Rater Guidelines, not from the API documentation itself.

The contentEffort Connection to Google's E-E-A-T Framework

contentEffort is the strongest candidate for a technical implementation of E-E-A-T's "Experience" dimension. When Google's Quality Rater Guidelines say "Does the content demonstrate first-hand experience?" — contentEffort is likely the system that answers that question at scale, algorithmically.

E-E-A-T Dimension	Technical Signal	How It's Measured
Experience	contentEffort	LLM evaluates first-hand knowledge, original examples
Expertise	isAuthor + relevanceScore	Author detection, topic relevance
Authoritativeness	isPublisher + normalizedTopicality	Publisher detection, topical focus
Trustworthiness	nsr + scamness + spambrainLavcScore	Site quality, scam detection

☝ These E-E-A-T-to-API mappings reflect attribute function alignment — Google has not published an official mapping between E-E-A-T dimensions and specific API attributes.

---

How Google's Quality Scoring Ensemble Feeds the Three Pandas

The five quality models don't produce a ranking directly. They feed into Google's quality firewall — the Three Pandas:

Variant	Scope	Demotion Range
Panda	Site-level	0–1023
BabyPanda	Page-level	0–1023
BabyPanda V2	Page-level (updated model)	0–1023

Each Panda variant produces a granular demotion score from 0 to 1023. This isn't binary pass/fail — it's a sliding scale. A page might receive a BabyPanda demotion of 150 out of 1023, meaning it's not severely penalized but is slightly held back compared to higher-quality competitors.

Key signals that likely contribute to the Panda demotion calculation:

• contentEffort — from the ensemble
• page2vecLq — neural similarity to known low-quality pages
• clutterScore — ad density and layout quality
• spambrainLavcScore — AI-generated content fingerprints
• racterScores — additional auto-content detection

---

Patent Connection: Google's Panda Patent US9767157B2

While the five quality models are API leak revelations, the theoretical foundation traces back to a specific patent: US9767157B2 — Predicting Site Quality, co-invented by Navneet Panda and Yun Zhou.

The patent describes a phrase-model architecture for predicting site quality:

1. Collect n-grams (2-grams through 5-grams) across all pages on a site
2. Calculate relative frequency — how often each phrase appears relative to the site's total content
3. Map frequencies to quality scores — using baseline scores from previously-evaluated sites
4. Aggregate — compute a weighted average across all phrases to produce a site quality score
5. Smooth — apply linear interpolation to prevent outlier phrases from dominating the prediction

Read the full patent analysis: Patent US9767157B2: Predicting Site Quality — The Panda Patent

---

How to Score Well Across All Five Google Quality Models

Model	What Drives High Scores	Common Mistakes
Chard	Original, comprehensive content with professional language patterns	Template content, thin pages pulling the site average down
Tofu	Publishing on an established, quality domain with clear topical focus	Launching on a fresh domain with no quality history
Keto	Continuously improving content — the model evaluates current state	Leaving content stale and outdated
Rhubarb	Making individual pages significantly better than your site average	Making every page the same quality — no standouts
contentEffort	Original data, case studies, methodology, first-hand experience	Generic AI-generated content with no unique contribution

Citation Network

Source Evidence

Unlike the patent-based analyses in this series, the Quality Scoring Ensemble draws from API leak evidence rather than granted patents. The primary source is the QualityNsrPQData module within the leaked Content Warehouse API, corroborated by Google's acknowledgment of the leak's authenticity.

Patent Foundation

• US9767157B2 — Panda (Site Quality) — The phrase-model architecture in this patent is the theoretical backbone for Chard. It describes the n-gram frequency analysis that Chard likely implements. The patent predates the API leak by over a decade.

Related Articles on This Site

• US8661029B1 — NavBoost (CRAPS) — NavBoost feeds click data alongside these quality signals. Quality scoring gates entry; NavBoost reorders what remains. Together they are the two halves of Google's quality-and-satisfaction loop. See also: How NavBoost Really Works.
• US10235423B2 — Entity Scoring — Entity scoring determines ranking position; quality scoring determines whether demotion applies. The two systems operate in parallel — who ranks is determined by entity metrics; whether they stay ranked is determined by quality signals.
• US11409748B1 — Passage Ranking — Passage ranking operates at the sub-page level while quality scoring evaluates at the page level. contentEffort likely examines the same heading structures that passage ranking uses as mathematical vectors.
• US7346839B2 (Historical Data) — Historical Data's temporal link signals feed into the same quality aggregation pipeline. The velocity and freshness checks from this 2003 patent contribute to the overall quality picture that Chard, Keto, and the ensemble evaluate.
• US7603350B1 (Entity Trust) — Entity Trust evaluates whether the people behind content are themselves trusted. Quality scoring evaluates whether the content itself is quality. Together they answer two questions: is this content good, and should we believe the person who wrote it?

Quality Scoring: What Doesn't Matter as Much as SEOs Think

The nature of this system is ancient: good content is recognizable. Humans can tell the difference between a carefully researched article and a template. Between original investigation and rehashed summaries. Between writing that comes from experience and writing that comes from autocomplete.

The ensemble is Google's attempt to mechanize that recognition. Five different models looking at the same page, each asking a different version of the same question: is this genuinely good?

The flavor will change. Keto will get new versions. contentEffort's LLM will improve. Chard's YMYL variant will evolve. New models might join the ensemble. Old ones might be retired.

But the nature — evaluating content from multiple independent angles and using consensus to determine quality — is permanent. It's how peer review works. It's how editorial boards work. It's how any serious quality evaluation system works. Google is doing at machine scale what humans have done for centuries: getting multiple independent opinions and trusting the consensus.

The practical takeaway is brutally simple. Create content that five different quality evaluators, each looking at different aspects, would all call "good." That's your north star. Not one metric. Not one optimization. Five independent judges, all agreeing.

That's a high bar. It's supposed to be.

---

Frequently Asked Questions