AI content Google penalty: what the policy actually says

TL;DR: Google does not have an AI content penalty. It has a spam policy against “scaled content abuse,” and that policy applies to AI-written pages, human-ghostwritten pages, and spun content the same way. The March 2024 core update deindexed roughly 1.7-1.8% of monitored sites, and studies found the deindexed sites were disproportionately mass-produced with AI, not merely AI-assisted. Ahrefs analyzed 331,000 pages and found AI content ranks fine when it clears the same quality bar human content has to clear. The failure mode isn’t the model. It’s publishing volume without editorial judgment.

What Google’s policy actually says
The March 2024 core update: what actually got hit
Scaled content abuse, defined
What the ranking data shows about AI content
The 7 failure modes that actually trigger action
Manual action vs algorithmic demotion
How to publish AI content without tripping the policy
AI content and AI Overviews / AI Mode
Comparison: what actually gets penalized vs what doesn’t
Frequently asked questions
Key takeaways

What Google’s policy actually says

Google published its position on AI-generated content directly, and it hasn’t reversed course since. The February 2023 Search Central post states it plainly: “Our focus on the quality of content, rather than how content is produced, is a useful guide that has helped us deliver reliable, high quality results to users for years” (Google Search Central). That sentence is the whole policy. Production method is not a ranking factor. Quality is.

The same page ties quality back to E-E-A-T: “Google’s ranking systems aim to reward original, high-quality content that demonstrates qualities of what we call E-E-A-T: expertise, experience, authoritativeness, and trustworthiness” (Google Search Central). AI, human, or a mix, the bar is the same four letters.

Google’s fundamentals documentation on generative AI content repeats this without softening it: automation has been part of publishing for decades (think templated financial reports, sports scores, weather updates), and using it well was never against the rules (Google Search Central). What changes the outcome is intent. Content built to manipulate rankings, at scale, without adding value, breaks the rules whether a person or a model typed it.

On May 15, 2026, Google formally confirmed that its spam policies apply to generative responses inside Search itself, including AI Overviews and AI Mode (ppc.land). That update didn’t introduce new rules. It closed a gap: the same standards that governed a blue link now explicitly govern an AI-generated answer box, so gaming one surface with junk content risks both.

The March 2024 core update: what actually got hit

If you want to know what Google actually penalizes, look at what it actually removed. The March 2024 core update, paired with new spam policies, deindexed roughly 837 of 49,345 monitored sites (1.7%) in one tracking study, and a separate 79,000-site study found 1,446 deindexed (1.8%) (Search Engine Journal). Those 837 sites had been pulling in more than 20.7 million organic visits a month combined before the manual action hit, and the estimated monthly display-ad revenue wiped out across them was $446,552 (Search Engine Journal).

The AI-content connection is the part most summaries skip: a follow-up study of the deindexed sites found 100% showed signs of AI-generated content, and half had 90-100% of their posts AI-generated (Search Engine Journal). That’s a strong correlation. It is not proof that AI use alone caused the deindexing; it’s evidence that AI-at-scale-without-editing was the dominant business model on the sites Google decided were manipulating rankings. A separate 671-site travel publisher analysis found 32% (213 sites) lost more than 90% of their organic traffic after the same update (Search Engine Journal).

Zoom out past the deindexed sites and the damage was broader: a six-month study of niche sites found 80% lost some traffic, 50% lost more than half, and 20% lost all of it (Search Engine Journal). Some of that is the routine churn every core update causes. Some of it is sites that had been riding thin AI content into rankings finally getting re-evaluated.

Scaled content abuse, defined

The policy doing the actual work here is called scaled content abuse, and it predates the AI panic. Google’s spam policy documentation defines it as generating “many pages for the primary purpose of manipulating Search rankings and not helping users,” regardless of how those pages were produced, by automation, humans, or a combination of both (Google Search Central).

Read that definition again. It names three production methods explicitly: automation, human labor, and hybrid workflows. A content farm running 200 freelancers through a templated brief violates this policy exactly as much as a content farm running 200 AI prompts through a templated brief, if both are unoriginal, low-value, and built to rank rather than help. Google has enforced this against human-written spam for over a decade. AI just made the volume achievable for smaller teams, which is why enforcement got louder, not because the rule changed.

The March 2026 spam update, which specifically targeted mass-produced content detection through Google’s SpamBrain system, completed in under 20 hours, the fastest spam update in Google’s documented history (ppc.land). Speed like that means the detection is largely automated pattern-matching on output characteristics (templated structure, thin differentiation across pages, near-duplicate phrasing patterns) rather than a slow manual review queue.

flowchart TD
    A[Content published] --> B{Does it add value<br/>beyond existing pages?}
    B -->|No| C{Was it published<br/>at scale with<br/>ranking as primary goal?}
    C -->|Yes| D[Scaled content abuse risk]
    C -->|No, just low quality| E[Algorithmic demotion risk]
    B -->|Yes| F{Passes E-E-A-T signals:<br/>expertise, experience,<br/>trust markers?}
    F -->|No| G[Ranks poorly, no penalty needed]
    F -->|Yes| H[Eligible to rank on merit,<br/>regardless of AI use]
    D --> I[Manual action:<br/>deindex or demote]
    E --> J[Lower rankings,<br/>no manual flag]

What the ranking data shows about AI content

Policy statements are one thing. What actually ranks is the harder evidence, and multiple independent studies now agree on the shape of it even if they disagree on the exact percentages.

Ahrefs analyzed 331,000 pages and published results under the headline “Google doesn’t punish AI content, it punishes bad content” (Ahrefs). That’s a vendor with a direct incentive to tell SEOs the truth, since their tool is what those SEOs use to check the data themselves.

A separate Ahrefs-sourced analysis of 600,000 pages found 86.5% of top-ranking pages show some AI involvement, reported by eMarketer under the headline “Google doesn’t penalize AI content” (eMarketer). Read that carefully: “some AI involvement” is a low bar that likely captures grammar tools, outline assistance, and light editing alongside fully generated drafts, so it measures AI-touched content, not AI-authored content.

Originality.ai runs an ongoing tracker and found that as of September 2025, 17.31% of top-20 results are AI-generated by their detection threshold, peaking at 19.56% in July 2025 before declining slightly (originality.ai). Their data also shows 81.9% of top results are mixed content, not purely AI or purely human (originality.ai).

Where the studies converge is on the No. 1 spot specifically. A Semrush analysis of 42,000 blog posts found human-written content holds position one 80% of the time versus 9% for purely AI-generated pages, a roughly 8x gap reported by Search Engine Land (Search Engine Land). That’s the number that should reshape your strategy more than any policy quote: AI content gets into the results, but pure, unedited AI output rarely wins the top spot against content with a real editorial pass on it.

The 7 failure modes that actually trigger action

Every documented case of AI content getting hit traces back to one of these, not to AI use itself.

Publishing volume that outpaces editorial review. If your team can’t read every page before it ships, you can’t verify it clears E-E-A-T, and Google’s pattern detection notices the resulting uniformity.
Templated structure with swapped keywords. “Best [City] plumbers” times 400 cities, each page identical except the city name, is the textbook scaled content abuse example Google names directly (Google Search Central).
Zero first-hand experience on experience-dependent topics. Product reviews, “I tried this” guides, and local recommendations need actual use of the thing. AI can’t have used the thing.
No named author, no credentials, no edit history. E-E-A-T signals are checkable claims, not vibes. A byline with nothing behind it doesn’t pass and neither does an anonymous “Team” byline on YMYL content.
Duplicate or near-duplicate phrasing across your own site. If ten of your pages answer the same question in barely reworded language, that’s the “many pages, little differentiation” pattern the spam policy targets.
Publishing on YMYL topics (health, finance, legal, safety) without subject-matter review. The quality bar rises specifically where mistakes cause real harm, and Google’s evaluators are trained to weight that.
Buying or scraping your way to content instead of reporting it. If the underlying facts were never verified by anyone at your organization, no amount of prompt engineering fixes that at the ranking-signal level.

None of these require an AI detector to find. They’re visible from the page and from the site’s publishing pattern, which is exactly why Google’s enforcement doesn’t need to detect “AI” at all. It detects the pattern AI makes easy to produce at scale.

This matters because AI detection tools themselves are not part of Google’s public ranking documentation. Google has never confirmed it runs your content through a standalone “AI or not” classifier as a ranking input. What it evaluates, by its own account, are quality signals: originality, depth, factual accuracy, and the E-E-A-T markers described above. A page can read as obviously AI-generated to a human and still rank fine if it clears those signals, and a page can be entirely human-typed and still get caught by the scaled content abuse policy if it’s templated, thin, and published in bulk. Treating “will an AI detector flag this” as your quality bar is solving the wrong problem. Treating “does this add something the top 10 results don’t already say” as your quality bar solves the one Google actually checks.

Why chasing an AI detector score is a wasted editorial cycle

Teams that add an AI-detector pass to their publishing workflow are usually optimizing for a signal Google doesn’t use. The useful version of that review step isn’t “does this score as human,” it’s “would a subject-matter expert sign their name to this claim.” That reframing changes what the editor actually looks for: unverified statistics, generic advice that could apply to any competitor’s page, and sections that restate the question instead of answering it. Those are the failure modes that show up in low-ranking pages regardless of how the draft was produced, and they’re the ones a detector score can’t catch because a detector reads style, not substance.

Manual action vs algorithmic demotion

These get conflated constantly and they’re not the same enforcement mechanism.

A manual action is a human reviewer at Google flagging your site, visible in Search Console under Manual Actions, usually triggered by a spam report or a targeted sweep. It’s binary: your pages get removed or demoted until you fix the issue and file a reconsideration request. The deindexed sites from the March 2024 update mostly fall in this bucket.

An algorithmic demotion is your site simply ranking lower because a core update re-weighted quality signals and your content no longer clears the bar. Nothing shows in Search Console. There’s no reconsideration request. The 80% of niche sites in the six-month study that “lost some traffic” without full deindexing are almost certainly experiencing this, not manual action (Search Engine Journal).

If your traffic dropped and you have no manual action in Search Console, you’re not being penalized for AI content. You’re being re-ranked against a quality bar your content used to clear and no longer does, which is a distinct problem worth investigating with a structured traffic-drop diagnosis rather than an AI-panic response like ripping down every AI-assisted page you’ve published.

How to publish AI content without tripping the policy

The practical fix isn’t “write everything by hand again.” It’s closing the gap between AI output and the E-E-A-T bar before publishing.

Set a review gate every page must pass, no exceptions for volume. If a human with subject knowledge hasn’t read it, don’t ship it, no matter how many pages are in the queue.
Add first-hand detail an AI can’t invent. A screenshot from your own account, a specific number from your own test, a named client outcome. This is the fastest way to separate your page from the ten AI-generated competitors ranking near you.
Byline real people with real credentials, linked to a bio page, not a generic “editorial team.”
Control output length and depth instead of letting a model pad to a target word count. Thin AI content padded to “look complete” is more detectable, and less useful, than a shorter page that actually answers the query. This is the exact problem Vrid.ai’s word-count control addresses: you set the depth the topic actually needs instead of forcing every article to an arbitrary length, so pages don’t get bloated with filler to hit a number.
Cap your publishing rate to what your review team can actually clear. If you can review 20 pages a week, publish 20, not 80 with a lighter pass.
Differentiate every page in a cluster. If two pages could be merged without losing information, merge them. Redundant pages are the clearest scaled-content signal you can send.
Treat YMYL content as a different review tier. Slower, with named subject-matter review, regardless of how the draft was produced.

None of this is AI-specific advice dressed up. It’s the same discipline that kept human-written content farms out of rankings for the fifteen years before generative AI existed. The tool changed. The bar didn’t.

AI content and AI Overviews / AI Mode

The May 2026 policy confirmation that spam rules cover AI Overviews and AI Mode responses matters for a separate reason: it means gaming AI-generated answer surfaces with manipulative content now carries the same risk as gaming classic rankings (ppc.land). If you were hoping thin AI content might slip into an AI Overview citation even after it stopped ranking traditionally, that gap is closed. The content still has to clear the quality bar to be cited, not just to rank.

This connects directly to how share of answer works as a metric: getting cited inside an AI-generated response requires the same underlying signals as ranking a blue link, sourced facts, clear structure, and enough authority that the AI system trusts your page as an input. Scaled, unoriginal content doesn’t clear that bar any more than it clears traditional ranking.

Comparison: what actually gets penalized vs what doesn’t

Content pattern	Penalized?	Why
AI-drafted, human-edited, fact-checked article with a named author	✗ Not penalized	Passes E-E-A-T regardless of drafting tool (Google Search Central)
Hundreds of templated city/service pages with swapped keywords	✓ Penalized	Textbook scaled content abuse (Google Search Central)
Human-written spun spam at high volume	✓ Penalized	Same policy applies regardless of production method (Google Search Central)
AI content citing real sources, adding original data or a first-hand test	✗ Not penalized	Original, verifiable value, tool-agnostic evaluation
AI content on YMYL topics with no subject-matter review	✓ High risk	Quality bar is stricter on health, finance, legal, safety
Short, direct AI-assisted answer that fully resolves the query	✗ Not penalized	Quality, not length, is the signal
Duplicate AI pages published faster than any human reviews them	✓ Penalized	Scale plus zero differentiation is the trigger, not the model
AI-generated product descriptions written to spec with real product data	✗ Not penalized	Factual, useful, no ranking-manipulation intent

Frequently asked questions

Does Google have a specific AI content penalty?

No. Google has a general spam policy against scaled content abuse that applies to AI-generated, human-written, and hybrid content equally when it’s mass-produced without adding value. There’s no separate rule that treats AI-authored text worse than human-authored text of the same quality (Google Search Central).

Will using ChatGPT or Claude to write my blog posts get my site penalized?

Not by itself. What gets sites penalized is publishing high volumes of unedited, undifferentiated AI output built primarily to rank rather than to help readers. A single AI-assisted, human-reviewed post carries no inherent risk.

What percentage of top-ranking Google results contain AI content?

Estimates vary by detection method. Originality.ai’s tracker put it at 17.31% of top-20 results as of September 2025, while an Ahrefs-based analysis found 86.5% of top pages show at least some AI involvement, a much looser threshold that likely includes light AI editing assistance (originality.ai, eMarketer).

Is it true that human content ranks #1 far more often than AI content?

Yes, based on the data available. A Semrush analysis of 42,000 blog posts found human-written content holds position one 80% of the time versus 9% for purely AI-generated content, an 8x gap (Search Engine Land). Mixed AI-plus-human content performs closer to human-only content than pure AI.

What is scaled content abuse, exactly?

It’s Google’s spam policy against generating many pages primarily to manipulate rankings rather than help users, regardless of whether automation, humans, or both produced the content (Google Search Central). Volume plus low differentiation plus ranking intent is the trigger, not any single production method.

How many sites got deindexed in the March 2024 update?

Tracking studies found 837 of 49,345 monitored sites (1.7%) deindexed in one dataset, and 1,446 of 79,000 (1.8%) in another (Search Engine Journal). That’s a small percentage of the total web but a meaningful volume of organic traffic, over 20.7 million monthly visits across the deindexed sites in the first study.

Did those deindexed sites actually use AI content?

A follow-up study found 100% of the deindexed sites showed signs of AI-generated content, with half having 90-100% of posts AI-generated (Search Engine Journal). That’s a strong correlation between AI-at-scale publishing and the sites Google flagged, though it doesn’t prove AI use alone was the cause versus the quality and volume of what was published.

What’s the difference between a manual action and an algorithmic ranking drop?

A manual action is a human Google reviewer flagging your site, visible in Search Console, requiring a fix and reconsideration request to reverse. An algorithmic demotion is your rankings dropping because a core update re-weighted quality signals, with no notification and no appeal process, just better content to compete against.

If my traffic dropped and I use AI content, was I penalized for the AI?

Check Search Console for a manual action first. If there isn’t one, you’re facing an algorithmic re-ranking, which is about content quality relative to competitors, not a flag against AI use specifically. Diagnosing a traffic drop with no announced update walks through separating these causes.

Does Google’s policy apply to AI Overviews and AI Mode, or only classic search results?

As of May 15, 2026, Google confirmed its spam policies formally apply to AI Overviews and AI Mode responses as well as classic rankings (ppc.land). Content built to manipulate one surface risks losing eligibility on both.

Can AI content pass E-E-A-T signals?

Not on its own. E-E-A-T requires expertise, experience, authoritativeness, and trustworthiness, and an AI model has no first-hand experience with anything. Passing E-E-A-T means adding human verification, real credentials, and first-hand detail on top of an AI draft, not publishing the draft as-is.

Is YMYL content held to a higher standard for AI use?

Yes. Health, finance, legal, and safety topics carry a higher quality bar because mistakes cause real harm, so unreviewed AI content in these categories carries more risk than in low-stakes topics like recipe variations or entertainment trivia.

What triggered the March 2026 spam update?

It targeted mass-produced content detection specifically, using Google’s SpamBrain system, and completed in under 20 hours, the fastest spam update in Google’s documented history (ppc.land). The speed suggests largely automated pattern detection on output characteristics rather than manual review.

Should I stop using AI to write content entirely?

No, based on the evidence. AI content that clears the same quality and originality bar as good human content ranks fine, as Ahrefs’s 331,000-page study found (Ahrefs). The risk is in unreviewed volume, not the tool.

Does adding a human editing pass actually change ranking outcomes?

The data suggests yes, indirectly. Mixed AI-and-human content performs closer to fully human content than to fully AI content in ranking studies, and originality.ai found 81.9% of current top results are mixed content rather than purely one or the other (originality.ai).

What does “primary purpose of manipulating rankings” actually mean in practice?

It means the page exists mainly to capture search traffic through keyword coverage rather than to answer a real question in a way that couldn’t be found elsewhere as well. A city-by-city page with identical content except the city name fails this test even if every sentence is factually accurate.

Can I use AI for first drafts and still be safe if editors rewrite heavily?

Yes. Google’s policy is explicit that production method doesn’t matter; a heavily edited AI first draft is functionally the same as human-written content by the time it publishes, assuming the edit adds real verification and value rather than surface-level rewording.

Is there a word count or publishing frequency that’s “safe”?

No fixed number exists in Google’s documentation. The risk factor is review capacity relative to publishing volume, not a specific word count or posts-per-week threshold. A team that can rigorously review 5 posts a week is safer than one publishing 50 posts a week with a light pass.

What should I check first if I suspect an AI content penalty?

Search Console Manual Actions report, first. If it’s empty, your issue is algorithmic quality re-ranking, not a policy violation, and the fix is improving content depth and E-E-A-T signals rather than removing AI-assisted pages.

Does Google detect AI content directly, or detect the patterns AI makes common?

Google has never confirmed a standalone “AI detector” in its ranking systems. The documented spam policies target patterns, scale without differentiation, templated structure, thin value, that AI makes cheap to produce, not the presence of AI-generated text itself.

Key takeaways

Google’s stated policy is production-method-agnostic: quality and originality determine rankings, not whether AI wrote the draft (Google Search Central).
The real enforcement mechanism is the scaled content abuse policy, which applies equally to AI, human, and hybrid content published at scale without adding value.
The March 2024 core update deindexed 1.7-1.8% of monitored sites, and a follow-up study found 100% of those sites showed AI-generation signs, evidence of correlation with unreviewed volume, not proof AI use alone caused it.
Human-written content still wins the No. 1 ranking position roughly 8x more often than purely AI-generated content, per a 42,000-post Semrush analysis.
Manual actions (visible in Search Console) and algorithmic demotions (no notification) are different problems with different fixes; check which one you’re actually facing before you change your content strategy.
The fix is editorial review, first-hand detail, real bylines, and publishing volume capped to review capacity, the same discipline that predates AI by over a decade.

If you’re building an AI-assisted content workflow and want the drafting step to produce right-sized pages instead of padded filler, Vrid.ai’s word-count control lets you set the depth a topic needs rather than forcing every article to hit an arbitrary target, which is one less reason for a page to look scaled and thin.

AI content penalty: what Google's policy really says