The Mt. AI Effect: Why AI-Generated Content Surges Then Crashes in Google

You publish 200 AI-generated articles. Traffic spikes. Impressions climb. For a few weeks, you feel like you've cracked the code. Then, somewhere around the three-month mark, the graph reverses - sharply, irrecoverably, and across every search surface you thought you'd conquered. SEO practitioners have started calling this pattern "Mt. AI." It's what happens when sites scaling heavily with AI-generated content initially surge in Google, but then drop heavily as Google's systems adjust. The term was coined by Barry Schwartz at Search Engine Roundtable, and it has quickly become shorthand for a phenomenon now backed by multiple large-scale experiments and real-world case studies. Understanding why this mountain forms - and why it always collapses - is no longer optional for anyone with a content budget. The data is not ambiguous. After grading 42,000 blog posts with an AI detector, Semrush found that content classified as fully human-written outperformed content classified as AI-generated across all top 10 positions, with the gap most striking at position 1, where pages have an 80.5% probability of being human-written compared with just 10% for AI-generated. Yet paradoxically, 86.5% of top-ranking pages contain some AI-generated content, making AI in content creation the norm, not the exception. The distinction between these two facts explains everything about the Mt. AI effect - and what separates sustainable AI-assisted content from the kind that craters.

The Anatomy of the Surge: Why AI Content Ranks Fast (and Looks Promising)

Google indexes new content eagerly. It has to. When a fresh page appears that matches real queries, Google tests it. In the Search Engine Land/SE Ranking 16-month experiment across 20 brand-new websites, about 71% of new AI-generated pages were indexed within the first 36 days, generating over 122,000 impressions and 244 clicks, with 80% of sites ranking for at least 100 keywords each.

That initial velocity is intoxicating. Overall, these results showed AI-generated content can gain traction quickly - even without backlinks, editorial input, or additional SEO work. In the short term, content alone was enough to get indexed and appear in search. When you publish at scale against low-competition long-tail queries, the math looks irresistible: volume multiplied by easy wins equals hockey-stick impressions charts. Here's the mechanism behind the early surge. Google's crawling and indexing system is separate from its quality evaluation system. Indexing is relatively fast and somewhat generous - Google is still willing to crawl and index AI-generated content in most cases. Quality assessment, however, operates on a longer cycle. The initial indexing may have been a grace period. Google's systems need time to evaluate content at scale. You can index 900,000 pages relatively quickly. Evaluating whether those pages actually deserve to rank takes longer.

This temporal gap between indexing and evaluation creates the illusion that raw AI output works. Teams see the climb and invest more. They scale the publishing cadence. The mountain rises.

The Crash: Three Months and the Cliff Edge

The SE Ranking experiment tracked what happened next. Around February 3, 2025, roughly three months after publication, the experiment hit a turning point. Only 3% of pages remained in the top 100, down from 28% in the first month. In practical terms, the content remained indexed but rarely appeared where users could see it.

That 28-to-3 percent collapse is the defining signature of Mt. AI. And it doesn't recover. The experiment ran for over a year to see if rankings would recover. For the most part, they didn't. After the drop around the three-month mark, visibility remained extremely low for the rest of the experiment.

The question worth sitting with: why does the crash happen at all if Google doesn't penalize AI content per se? Google does not penalize content simply for being created with AI. However, it does penalize low-quality, unhelpful, and spammy content, regardless of how it's produced. In 2025, Google's updated guidelines reaffirm this principle. The crash isn't about detection. It's about what the content lacks.

The Missing Signals That Trigger the Fall

The AI-generated articles lacked many signals Google uses to assess quality and trust: Authority - no backlinks or external validation, meaning new domains struggle to compete with established sites. Expertise and credibility - no authors, credentials, or real-world expertise. Content differentiation - much of the content resembled what already exists, lacking unique insights. Site structure - no internal linking, topical organization, or clear hierarchy.

Every one of those gaps compounds. An AI article without an author byline carries no expertise signal. Without backlinks, it has no authority. Without original data or practitioner perspective, it offers no information gain - the very thing Google's systems are increasingly tuned to reward.

The Authority Floor Effect

One of the most underappreciated dynamics of scaled AI content is its domain-level impact. Publishing generic content doesn't just fail to rank - it actively degrades the ranking potential of everything else on your site.

Each undifferentiated article adds to the low-quality signal without contributing E-E-A-T markers to offset it. The authority floor begins to decline. New content published to the domain inherits the degraded baseline signal. Rankings for existing high-quality pages begin to soften - not because those pages changed, but because the domain-level signal that previously supported them has weakened.

This creates a destructive feedback loop. The strategy that appeared to offer compounding organic growth is producing compounding organic decline. Most teams respond by publishing more to compensate, which accelerates the collapse. The ranking decline that follows sustained volume AI content production is not a temporary algorithmic adjustment. It is a structural condition that requires structural correction to reverse. Recovery for a domain that has accumulated significant helpful content penalty signals is measured in months, not weeks.

Grokipedia: The Mt. AI Case Study in Real Time

No example illustrates the Mt. AI effect more dramatically than Grokipedia, xAI's AI-generated encyclopedia. In November 2025, Grokipedia received 19 clicks from Google Search for the entire month. By January 2026, that number was 3.2 million clicks per month - from 19 to 3,200,000 in roughly 60 days. The site was ranking for 6 million keywords with over 900,000 indexed pages. Then it fell. The illusion of invincibility shattered abruptly on February 6, 2026. Observations filtering through industry channels, notably flagged by Glenn Gabe, confirmed Grokipedia had begun a significant, precipitous plunge in Google rankings.

What made the Grokipedia collapse particularly instructive was the cross-platform contagion. Grokipedia wasn't just losing Google organic rankings. It was losing visibility across Google organic search, Google AI Overviews, Google AI Mode, and ChatGPT - where it had previously been cited as a source. All of them, at the same time. This is an important detail. If you've been thinking about AI and LLM visibility as a separate channel from Google organic, this event suggests they're more connected than we thought.

By February 10, Wikipedia was ranking above Grokipedia for the search query "Grokipedia." Google ranked a competitor above Grokipedia for its own brand name. That's about as definitive a signal as Google sends.

The case didn't end there. The site's visibility continued to decline with the March 2026 broad core update, reflecting similar trends across various AI search platforms. Peec AI's analysis found that 36% of brands included in the success stories of a popular AI content generation tool have a typical Mount AI visibility/traffic trend in Google. This is not an edge case. It is a pattern.

What Google Actually Targets: Scaled Content Abuse, Not AI Itself

Getting the mechanism right matters. Google's official position is clear. According to Google Search Central, the focus is on creating "helpful, reliable, people-first content." The documentation explicitly states that automation, including AI, is acceptable when used to create helpful content rather than to manipulate search rankings. Google's algorithms evaluate whether content serves users, not whether a human typed every word.

The actual penalty target is specific: scaled content abuse. Scaled content abuse occurs when someone generates large amounts of content primarily to manipulate search rankings rather than help users. The defining characteristic is mass production with no meaningful added value. Google added "scaled content abuse" as a specific spam category in early 2025.

According to Search Engine Land's coverage, the policy targets specific behaviors. Notably, "uses AI" is not on the list. The penalty targets behavior, not tools. A company using AI to produce well-researched, thoroughly edited articles that genuinely help its audience is just doing content production. A company publishing 100 AI-generated pages that all follow the same template with no original information is engaging in abuse. The enforcement has been severe. The February 2026 core update sent Semrush Sensor to 9.4 as mass AI content sites saw 40-60% traffic drops. Before that, in March 2024, 1,446 websites had manual actions applied to them. The analysis found 100% of the websites had some posts that were AI-generated, and 50% of the sites had 90-100% of their posts as AI-generated.

The brands that lost 40–55% of organic traffic through those algorithm cycles had things in common: over-reliance on AI-generated content at scale, weak E-E-A-T signals, and content strategies designed for rankings rather than readers.

Information Gain: The Metric That Explains the Crash

If scaled content abuse is the what, information gain is the why. Google filed a patent in June 2022 for an "information gain score" that uses the amount of unique information in content as a ranking factor, favoring pages that add fresh perspectives over those that simply repeat existing information.

The concept is structurally incompatible with generic AI output. The information gain score suggests a new algorithm element targeting AI-generated content and new content farms. Content might be demoted if it lacks uniqueness, even if it consists of different words arranged differently. AI language models are trained on the corpus. Their default output is, by definition, a synthesis of what already exists. Without human intervention - original data, practitioner experience, unique analysis - the output has zero information gain.

Now that AI can compile and synthesize comprehensive coverage from ten articles in seconds, "comprehensive" is no longer the differentiator - it's the baseline. The information gain theory has evolved from patent filing to practical necessity. Which means every piece of content now demands a more honest question: "Does this need to exist?"

This reframes the entire AI content debate. The problem isn't that AI wrote the sentences. The problem is that AI-only content adds nothing new to the information ecosystem. Google's systems - whether through the information gain scoring mechanism, user engagement signals, or quality rater feedback - eventually identify this and respond accordingly. Practitioners creating information gain do so through specific, replicable methods: original data from surveys or polls, expert insights from industry professionals, unique perspectives on topics, bonus resources like checklists or tutorials, and real-world examples that help readers understand practical applications.

The Quality Rater Reality: How Google's Human Evaluators See AI Content

Google's 16,000+ quality raters are not abstract. They directly shape algorithmic direction. Google's Quality Rater Guidelines were updated in early 2025 to address AI content specifically. Raters are instructed to evaluate AI content on the same quality dimensions as human content. The guidelines rate AI content as "Lowest" quality only when it "lacks human oversight and review." Content that has been edited, fact-checked, and improved by humans doesn't receive this rating regardless of its origin.

The January 2025 update went further. One of the biggest highlights was the definition and guidance around AI-generated content. In the guidelines, Google refers to generative AI as a "useful tool," but also highlights that bad actors could potentially misuse its capabilities.

Raters now look for specific AI fingerprints. Google's January 2025 Quality Rater Guidelines update introduced AI "fingerprint" flags - content containing phrases like "As an AI, I don't have opinions" is rated lower without verified human review. But the evaluation extends far beyond mechanical detection. Google doesn't rely on AI detectors. Instead, it evaluates quality signals such as originality, depth, factual accuracy, and user engagement.

This human evaluation loop feeds directly into algorithm training. If your AI-generated article is full of quality red flags, a human rater will likely mark it as low-quality. That feedback then helps the algorithms get even better at spotting and devaluing similar content across the web. The system is continuously tightening. Content that passed muster in 2024 may not pass in 2026.

What Sustainable AI-Assisted Content Actually Looks Like

The path forward isn't avoiding AI. It's using AI in a way that produces content Google's systems reward. The distinction between content that climbs Mt. AI and crashes versus content that sustains its rankings comes down to four structural differences.

Human Editorial Architecture, Not Just Editing

At Semrush, AI supports how they produce content, but it only gets them part of the way. For every step in which they use AI - ideation, outlining, drafting - they have a human reviewing the output. Every article involves at least three people (a strategist, a writer, and an editor), often up to five. And all of them are subject matter experts or work very closely with them. They also have a deep understanding of their tools and how to use them to address real problems their customers face.

This isn't a "light editing pass." It's an editorial system where AI accelerates the mechanical work while humans supply the judgment, experience, and original insight that creates information gain.

Original Research and Proprietary Data

Content that gets cited in AI Overviews tends to contain something unavailable elsewhere. Clients who transitioned to "cite-worthy" content - focused on case studies and original data - gained visibility despite the increased presence of AI Overviews. Original data acts as a moat. An AI model cannot fabricate your proprietary survey results. A language model cannot replicate your customer case study. These elements function as information gain by definition.

E-E-A-T as Architecture, Not Cosmetics

Bolting an author bio onto an AI-generated article doesn't create expertise signals. Adding author bios and credentials to your website is necessary but not sufficient. E-E-A-T requires genuine demonstration of expertise over time, not cosmetic additions. Companies add a headshot and 50-word bio, then wonder why nothing changes. AI systems evaluate expertise signals across the entire internet, not just what you say about yourself on your own website.

Real E-E-A-T signals include verifiable author credentials linked across platforms, consistent topical authority demonstrated through a content hub architecture, external citations from reputable industry sources, and transparent editorial processes. E-E-A-T determines eligibility, while SEO and other optimization efforts determine selection within the eligible content. Google's guidelines explicitly state that "Trust is the most important member of the E-E-A-T family."

Topical Segmentation Over Volume

A 2025 study of 300 B2B SaaS websites found that companies segmenting by industry increased Top 10 Google rankings by 43.4% on average. Companies without segmentation saw rankings decline by 37.6%. The segmented sites achieved 15.7X higher organic traffic growth. Narrow, audience-specific content generates information gain because industry-specific advice cannot be replicated by generic AI articles. Instead of "The Ultimate Guide to X" for everyone, write "X for Fintech Startups" or "X for Healthcare Platforms."

The Cross-Platform Cascade: Why Mt. AI Now Means Mt. Everywhere

Perhaps the most consequential development in the Grokipedia saga - and the broader Mt. AI pattern - is the cross-platform visibility linkage. The Grokipedia case shows how when you lose rankings in Google, you don't just lose Google traffic but the impact carries over to AI search engines too. This happens because ChatGPT uses Google Search during its grounding process. Perplexity crawls the web using similar quality signals. AI Overviews draw from Google's core ranking systems.

Google's systems can often adjust and the site can plummet in rankings. And when you drop in Google, you can often drop in AI Search as well. One penalty, one algorithmic reassessment, and you become invisible everywhere simultaneously. This transforms the risk calculus entirely. The old argument for scaled AI content was that even if Google caught on, you'd have built presence across LLMs. The evidence now suggests the opposite: losing Google credibility cascades into losing LLM citations.

AI Overviews now appear in 82% of B2B technology searches, a significant increase from 36% in 2025. The #1 organic ranking page sees a 34.5% drop in CTR when an AI Overview is present. But being cited within that AI Overview changes the equation entirely - being cited within an AI Overview increases brand clicks by 35%. The content that earns those citations is precisely the kind that doesn't climb Mt. AI in the first place: original, expert-informed, and structurally trustworthy. --- The Mt. AI effect is not a bug in Google's system. It is the system working as intended - testing new content against established quality signals and demoting what fails to earn its position over time. The initial surge is simply the evaluation window. What comes after is the verdict. For practitioners, the takeaway is blunt: AI is a production tool, not a strategy. The results of the 16-month experiment don't mean AI content is useless. They show AI alone isn't enough to drive lasting impact. Early traffic and impressions may look promising, but without a clear SEO strategy and human guidance, those gains will likely fade within a few months. The teams that thrive will be the ones that use AI to accelerate the creation of content that is genuinely worth ranking - content that carries original insight, demonstrated expertise, and verified trustworthiness. The mountain is always a mirage. The plateau, earned through real expertise, is what lasts.