Unhelpful Content: The Patterns That Trigger Suppression

Every major quality system in Google's leaked API and patent corpus works by identifying negative signals. The Panda patent (US9031929) scores content against known-quality baselines. Low-quality pages depress a ratio. The system does not reward quality; it punishes its absence. The N-gram quality patent (US9767157) builds phrase models from known-quality sites, then flags pages whose distributions deviate. It detects formulaic content, not original content.

NavBoost (US8595225) feeds behavioral signals as negative quality indicators. The signal taxonomy from the API leak contains goodClicks, badClicks, lastLongestClicks, and unicornClicks. Bad clicks trigger demotions. Good clicks inform ranking but produce no "helpfulness" score. SpamBrain assigns document-level spam probability from 0 to 1. The entire spam infrastructure is a classifier for bad. There is no "SpamFree" bonus.

The contentEffort attribute, an LLM-based effort estimation for article pages, is the closest thing to a positive quality scorer. But its name reveals its framing: it measures effort (the absence of laziness), not helpfulness. The practical implication is straightforward. "Be more helpful" is circular. The priority is eliminating the specific patterns each classifier detects. A page that avoids all unhelpfulness signals ranks by default because it survives the quality gauntlet.

The CompressedQualitySignals module in the leaked API contains 11 named demotion types and exactly one promotion type. This ratio is not an accident. It reflects a system architecturally weighted toward identifying what is wrong.

The sole promotion type, authorityPromotion, is the exception that proves the rule. Authority is the one dimension where Google explicitly boosts rather than demotes. Every other quality dimension operates through suppression. This has a direct consequence for how practitioners should allocate effort: the return on removing unhelpfulness signals is higher than the return on chasing a positive "helpfulness" score that the system does not compute.

The Panda patent computes site quality as a ratio: navigational queries (people searching for the site by name) divided by informational queries (queries the site appears for). Every low-quality page indexed from a domain inflates the denominator without contributing to the numerator. The Q* synthesis scores sites 0 to 1. Below 0.4 disqualifies from rich results.

Detection is site-level, not page-level. A single excellent page on a site with thousands of thin pages inherits the site's suppressed quality score. This is why programmatic SEO builds fail when they inflate the index with low-value pages, and why content pruning consistently improves metrics for remaining pages. The trigger is the ratio, not any single page.

Common sources of dilution: CMS-generated tag and category archives, faceted navigation generating thousands of near-duplicate URLs, out-of-stock product pages still indexed, and placeholder or "coming soon" pages. Index tiering means these pages may never reach the base index, but they still count against the site-level ratio.

One e-commerce site identified 600,000 pages that received zero clicks in 12 months and no-indexed them. The result: a 30% increase in clicks and impressions, double the keywords ranking in the top 3 four months later, and the highest signups and revenue in the company's history. Only about 1% of the original pages remained indexed. The ratio is the mechanism. Removing the denominator is often faster than growing the numerator.

Patent US9767157 builds phrase models (2-gram through 5-gram) from known-quality sites and scores unknown pages against this distribution. The classifier measures relative frequency: how many pages on the site contain each phrase, divided by total pages. If 8,000 of 10,000 pages share 70% identical boilerplate with only entity substitution, the n-gram model flags the distribution.

The canonical failure pattern is shallow variable substitution in programmatic pages: "Find the best [CITY] [SERVICE] near you" repeated across thousands of URLs with only the city name changed. This is not a hypothetical. It describes the majority of local SEO landing page builds that the March 2024 core update suppressed.

Machine-generated text, whether from LLMs or template engines, tends to produce repetitive structural patterns that this classifier catches. The n-gram model does not care whether a human or a model wrote the text. It measures the output distribution.

NavBoost processes click data over a rolling 13-month window, segmented by geography and topic. Two demotion types in CompressedQualitySignals are directly tied to user behavior. serpDemotion fires when users consistently return to the SERP quickly after clicking a result. Google publicly denies pogo-sticking as a signal. The API leak confirms it. navDemotion fires when users click through but cannot find what they need, navigate poorly, or abandon the page.

The triggers are structural. Content that does not match the query intent causes the user to bounce. Misleading titles or meta descriptions that overpromise cause the user to bounce. Poor page UX (intrusive interstitials, aggressive ad placement, broken navigation) causes the user to bounce. Pages that technically answer the query but force unnecessary friction (pagination walls, login gates for free content, excessive scroll to reach the answer) cause the user to bounce.

The API leak also confirmed chromeInTotal as a signal source, contradicting Google's public statements. Chrome usage data feeds engagement quality assessment beyond what SERP clicks alone capture.

Copia ("abundance") monitors content velocity: the ratio of URLs generated during specific periods against substantive articles produced. Firefly synthesizes Copia velocity signals, QualityNsrPQData quality signals, and NavBoost user dissatisfaction signals into high-confidence abuse determinations. The system is method-agnostic. It targets the pattern, not the production method.

Google's March 2024 "scaled content abuse" policy replaced the older "spammy automatically-generated content" policy. The new framing is explicit: "many pages are generated for the primary purpose of manipulating search rankings and not helping users," whether with automation, people, or a combination. A content farm staffed by humans triggers the same classifier as one powered by GPT-4. What matters is the ratio of daily clicks to daily good clicks (clicks where users did not return to the SERP).

The scale of the problem these classifiers address: as of September 2025, 17.3% of the top 20 Google results contain AI-generated content, up from 2.3% in 2019 (Originality.ai, 500 keywords sampled). The percentage peaked at 19.6% in July 2025. Copia and Firefly do not detect AI authorship. They detect the behavioral and velocity patterns that mass production creates regardless of production method.

The contentEffort attribute is an LLM-based effort estimation for article pages. A model (likely Gemini-class) assesses depth of knowledge and original research. This is the algorithmic operationalization of "did someone actually put work into this?"

Low scores correlate with reformulated content from other sources without original analysis, absence of unique images or embedded tools, shallow treatment of topics that competitors cover in depth, and generic advice without specific examples or measurements. High scores correlate with unique data points, in-depth linguistic complexity, original media, and evidence of first-hand experience.

contentEffort is the closest algorithmic proxy for the "Experience" dimension of E-E-A-T. Content with genuine first-hand experience scores higher because it contains details that require effort to produce: specific measurements, dated observations, original photographs, test results. These details are expensive to fabricate, which is precisely what makes them reliable signals.

Two independent scoring mechanisms target unoriginal content. OriginalContentScore (0-512, 0-127 for short content) measures page-level uniqueness. CopycatScore (in BlogPerDocData) uses shingling, overlapping text chunks plus fingerprinting, to detect content copied from other sources without exact duplication.

The triggers are predictable. Aggregated content that reformats publicly available data without adding analysis scores low. Articles that paraphrase competitors without original perspective score low. Syndicated content without canonical attribution scores low. The threshold is not "is any of this information available elsewhere?" Nearly all information is. The threshold is "does this page provide perspective, analysis, or data not available elsewhere in substantially similar form?"

The clutterScore attribute in QualityNsrPQData penalizes pages with distracting elements that degrade the content consumption experience. Aggressive ad placement (especially above-fold, interstitial, or mid-content), pop-ups that interrupt reading, excessive sidebar widgets competing with main content, and auto-playing media all contribute.

Lily Ray's analysis of HCU-hit sites found strong correlation between penalties and two specific patterns: excessive ads and broken navigation. Cookie consent banners and GDPR notices are legally required, but their implementation quality matters. A full-screen overlay that obscures content for five seconds while the user searches for a dismiss button is a different signal than a subtle banner at the bottom of the viewport.

Clutter creates a compounding problem. It degrades user engagement (feeding Pattern 3, NavBoost demotions), it reduces the content-to-noise ratio (feeding Pattern 1, Panda ratio), and it signals low editorial standards (feeding Pattern 5, contentEffort). Patterns do not fire in isolation.

E-E-A-T does not function as a scoring weight where more is better on a linear scale. It functions as a gate. Content either passes the threshold or it does not. Below the threshold, other quality factors are irrelevant because the content is filtered before LLM re-ranking.

The specific absence patterns that fail the gate: no named authors or author bios, no evidence of real staff or organizational expertise, AI-generated author personas (fake headshots, fabricated credentials), no first-hand experience markers (original photos, specific test results, dated observations), and YMYL content without credentials or institutional backing.

The HCU hit pattern (no authors, no bios, no evidence of staff) is a checklist of absences, not a failure to reach a high bar. Sites that added real author bios and experience markers saw the earliest recoveries. The fix is not "demonstrate more E-E-A-T." The fix is "stop omitting the signals the gate checks for."

When a site publishes content outside its core topic, it does not just fail to build authority in the new area. It actively dilutes authority in the core area. Google evaluates topical chunks independently via NsrChunks. Off-topic content drags down the chunk score. The siteFocusScore attribute measures how concentrated a site's content is around its core topic.

The "content for traffic" strategy, publishing on high-volume keywords unrelated to the site's expertise, is the canonical trigger. HubSpot's traffic losses in core updates were attributed partly to tangential content. IBM, Progressive, and DoorDash saw organic traffic gains after pruning off-topic content. The mechanism is the same as Pattern 1 (Panda ratio) but operates at the topical authority level rather than the content quality level.

The strongest content on the wrong page type for the query intent still fails. Google classifies query intent granularly using CommercialScore as a filter and asteroidBeltIntents for multi-label intent. The system evaluates whether the page type matches what the user is looking for.

Blog posts ranking for transactional queries (user wants to buy, page wants to inform). Product pages appearing for informational queries (user wants to learn, page wants to sell). Listicles serving navigational intent (user wants a specific site, page offers alternatives). A 3,000-word guide when the user needs a quick reference table. All of these are mismatches that the system detects and penalizes through retrieval-time filtering.

Of 40 pages that lost rankings in core updates, 11 had drifted from their original intent match. 7 of those 11 recovered within 60 days of realignment. Intent mismatch is one of the faster patterns to fix because it operates at retrieval time, not at the pre-retrieval stage where site-level signals take months to recalculate.

Content that exists but cannot be efficiently consumed. This is distinct from clutter (Pattern 7): clutter is noise around the content, structural inaccessibility is the content itself being poorly organized. Semantic units exceeding 180 words cause comprehension to drop in scanning studies. Long-form content without headings, lists, or tables becomes a wall of text that neither humans nor crawlers can navigate to specific sections.

The connection to AI citation is direct. Pages with structured formats (tables, FAQ patterns, clear heading hierarchies) are cited at 2.3x the rate of unstructured pages. Sequential headings produce 2.8x citation lift. Sentences exceeding 17 words in extractable passages exceed the extraction ceiling for AI systems. 57% of user attention goes to above-fold content. Critical information buried below the fold is both a user experience failure and a retrieval accessibility failure.

The API leak includes unauthoritativeScore in CompressedQualitySignals. Limited documentation exists on what feeds this score, but its presence as a named attribute suggests an independent assessment of authority separate from the other quality signals.

New domains without entity recognition signals, sites lacking external corroboration (no mentions on trusted third-party sources), and content on topics where the site has no demonstrated track record are the likely triggers. This connects to the hostAge sandbox (new domains face indexing friction) and the entity stacking evidence that Google needs 30-50 unique trust signals before it considers a brand a real entity. A site that has not built these signals starts with an authority deficit that content quality alone cannot overcome.

This pattern has the lowest measurement confidence of any in the catalog. The attribute exists. Its inputs are poorly documented. But its architectural position in CompressedQualitySignals, alongside pandaDemotion and siteAuthority, suggests it operates as a pre-retrieval gate.

SpamBrain includes a dedicated keywordStuffingScore, a 7-bit integer (0-127). This is a legacy pattern. Most practitioners stopped worrying about keyword density a decade ago. But the dedicated scoring attribute in the leaked API confirms it remains actively monitored, and the 127-point granularity suggests more than a simple binary check.

The modern version of keyword stuffing is subtler than the 2005 variety. It shows up in heading tags that repeat the primary keyword in every H2, alt text that uses the same phrase on every image, meta descriptions that read like keyword lists, and hidden text or off-screen blocks. The threshold is not about a specific density percentage. It is about unnatural phrase distributions that the n-gram model (Pattern 2) and the dedicated keywordStuffingScore flag independently.

These 13 patterns do not all fire at once. They trigger at different stages of the ranking pipeline. This matters for remediation sequencing: a failure at an earlier stage blocks the impact of fixes at later stages.

The binding constraint is the highest-severity pattern at the earliest pipeline stage. A site with a Panda ratio problem (Pattern 1, pre-retrieval) will not benefit from fixing intent mismatch (Pattern 10, retrieval-time) until the ratio is addressed. This sequencing logic is the same principle behind the clinical diagnostic framework: test gates in order, stop at the first binding constraint.

Of approximately 400 sites tracked by Glenn Gabe hit by the September 2023 Helpful Content Update, only about 88 showed 20%+ traffic lifts by August 2024. That is a 22% recovery rate after more than 18 months. The HCU classifier was folded into the March 2024 core update, which took 45 days to complete (the longest core update on record). Recovery timelines average 17 months for domain-migration-class problems.

Sector-specific data makes the pattern sharper. Of 671 travel publishers analyzed, 32% lost more than 90% of their organic traffic from the September 2023 HCU. The December 2025 core update hit affiliate sites hardest (71% affected), followed by health and YMYL content (67%) and e-commerce (52%). Recovery timelines for YMYL sites stretched to 6-12 months even after remediation.

Recovery is slow because no single pattern is "the helpful content classifier." The classifier is the composite effect of all these systems. A site suppressed by the HCU likely triggered multiple patterns simultaneously: thin content ratio (Pattern 1) combined with low effort (Pattern 5) combined with E-E-A-T absence (Pattern 8) combined with clutter (Pattern 7). Fixing one pattern while three remain does not clear the composite threshold.

The actionable takeaway is not "be more helpful." It is: map which patterns are active on your site, sequence remediation by pipeline stage (pre-retrieval first), and address multiple patterns simultaneously rather than serially. The gauntlet tests 13 things. Passing 12 of them is not enough if the 13th is a gate.