Google's Scaled Content Abuse Policy: What It Really Bans

Google wrote a spam policy that everyone calls "the AI content policy," and it never once makes AI the offence. The actual sentence, live on Google's spam policies page as of its 15 May 2026 update, ends like this: unoriginal content that provides little to no value to users, "no matter how it's created."

Five words, and they carry the whole rule.

The stakes are not theoretical. In May 2026, SEO consultant Lily Ray published an analysis of more than 220 websites identified as customers of AI content-creation platforms. Of those sites, 54% had lost 30% or more of their peak organic traffic, 39% lost 50% or more, and 22% lost 75% or more. Her title for the pattern: "it works, until it doesn't."

So both things are true at once. Google does not care that a machine helped write your post. Google cares enormously about what happens when a site starts producing pages faster than it can make any of them worth reading. This post reads the policy closely — the actual words, the enforcement mechanics, the 2026 changes almost every guide on the topic missed — and turns it into publishing rules you can actually follow.

Read the sentence, not the headlines

Here is Google's definition, verbatim, from the Search spam policies documentation:

"Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users. This abusive practice is typically focused on creating large amounts of unoriginal content that provides little to no value to users, no matter how it's created."

Two conditions, both required. Many pages. Primary purpose of manipulating rankings rather than helping users. One page of thin content is a quality problem, not scaled content abuse. A thousand pages of genuinely useful documentation is not scaled content abuse either. The policy is aimed at the intersection.

Now the five examples Google gives. Read each one to the end, because the ending is where the rule lives:

1. "Using generative AI tools or other similar tools to generate many pages without adding value for users"
2. "Scraping feeds, search results, or other content to generate many pages (including through automated transformations like synonymizing, translating, or other obfuscation techniques), where little value is provided to users"
3. "Stitching or combining content from different web pages without adding value"
4. "Creating multiple sites with the intent of hiding the scaled nature of the content"
5. "Creating many pages where the content makes little or no sense to a reader but contains search keywords"

Every single one carries a value qualifier or an intent qualifier. Example 1 is the one people quote as proof that Google bans AI writing — but the sentence does not stop at "generate many pages." It stops at "without adding value for users." Strip that clause and you have a different policy, one Google conspicuously declined to write.

It declined on purpose, and said so at launch. When the policy arrived alongside the March 2024 core update, Google described the target as "producing content at scale to boost search ranking — whether automation, humans or a combination are involved." Danny Sullivan, Google's Search Liaison, put it even more plainly when the quality rater guidelines were updated: "It doesn't matter how you're doing scaled content, whether it's AI, automation, or human beings. It's going to be an issue."

The same pattern shows up in the quality rater guidelines themselves. The January 2025 update tells raters to apply the "Lowest" rating when "all or almost all of the MC on the page (including text, images, audio, videos, etc) is copied, paraphrased, embedded, auto or AI generated, or reposted from other sources with little to no effort, little to no originality, and little to no added value." The AI clause never stands alone. It is always bundled with effort, originality and added value.

There is a useful sanity check here. In April 2026, Google announced a brand-new spam policy targeting back button hijacking, enforceable from 15 June 2026, with a named mechanic and a compliance date. That is what it looks like when Google decides to ban a specific behaviour. It has written policies like that repeatedly. It has never written one against "content written by a machine."

What the policy does not ban

The misreadings are expensive, so here are the things people avoid unnecessarily.

AI authorship. Google's own generative-AI optimisation guide — published 15 May 2026 and last updated 10 July 2026 — describes the risk as creating content variations where "doing so primarily to manipulate rankings or generative AI responses in Google Search violates Google's scaled content abuse spam policy." The trigger is the primary purpose. The tooling is not mentioned as a factor.

Automated translation. This one has an on-record clarification. In June 2025, Google told SEO consultant Glenn Gabe: "Our policies do not strictly define content that has been translated by AI as spam. Our scaled content abuse policy mentions automated transformations, including translations, as part of the overall warning against creating large amounts of unoriginal content that provides little to no value to users." Translating your genuinely useful documentation for a new market is not the thing being described.

Long posts. Google's helpful content guidance, last updated 10 December 2025, lists this among its warning signs: "Are you writing to a particular word count because you've heard or read that Google has a preferred word count? (No, we don't.)" We publish long posts here and it is worth saying out loud: length is a consequence of covering a topic properly, never a target to hit.

Publishing on a schedule. Volume alone is not the offence — but the reason for the volume can be. The same guidance flags: "Are you adding a lot of new content or removing a lot of older content primarily because you believe it will help your search rankings overall by somehow making your site seem 'fresh?' (No, it won't)" Publishing weekly because you have something weekly worth saying is fine. Publishing weekly to look busy to a crawler is the thing being warned about.

A specific page count. There isn't one. Google says "many pages" and has never quantified it, which frustrates people looking for a safe number. But a threshold would be a target, and a target would be gamed. Enforcement is pattern-based: volume, originality, value and intent considered together, at the level of a site rather than a single URL.

The evidence: AI content ranks, scaled AI content decays

Two datasets look like they contradict each other. They don't — they are photographing different things.

The snapshot comes from Ahrefs, which ran an AI detector across 600,000 pages drawn from the top 20 results for 100,000 random keywords. Findings: 13.5% were classified pure human, 4.6% pure AI, and 81.9% a mix — meaning 86.5% of top-ranking pages contained some AI-generated content. The correlation between a page's AI share and its ranking position was 0.011, which the authors describe as "effectively zero." Their conclusion: "Google probably doesn't care how you made the content. It simply cares whether searchers find it helpful."

A separate Ahrefs study of 900,000 newly created English-language pages in April 2025 found 74.2% contained some AI content. Originality.ai's longitudinal sampling of 500 informational keywords, meanwhile, found AI-flagged pages made up 17.31% of top-20 results in September 2025, up from 2.27% in February 2019. Different detectors and different populations — but the gap between "74% of new pages" and "17% of top results" is worth naming rather than smoothing over.

The time-lapse is Ray's 220-site study, and it sees something no snapshot can. The recurring shape she documents: rapid page growth over six to twelve months, a traffic peak within three to six months of maximum content production, then a steep decline over the following year that frequently lands below the original baseline. Her highest-risk templates are recognisable to anyone who has browsed a content-marketing site lately — comparison pages at scale, "what is X" glossaries, "best X for Y" listicles, competitor-alternatives pages, programmatic location scaling, FAQ farms.

Put the two together and the picture resolves. AI-assisted pages rank perfectly well. Sites that scale AI pages as a strategy tend to spike and then fall. Cross-sectional data cannot see the fall, because sites in the middle of one are still ranking on the day you measure them.

One caveat cuts against every percentage above, including the ones that support this argument: AI detectors are unreliable. Stanford HAI tested seven detectors against TOEFL essays written by non-native English speakers and found 61.22% were flagged as AI-generated. Ahrefs says the same of its own tooling: "No AI content detector is perfect… They can be incredibly accurate, but they always carry the risk of false positives." Google has never said it uses a detector as a ranking input, so "will a detector flag this?" is the wrong question to optimise for.

Three cases with nothing in common but volume

Case one: AI, actioned. In 2023, marketer Jake Ward exported a competitor's sitemap and used a GPT-4-based tool to generate 1,800 articles in hours for a business-planning software client, publicly reporting 3.6 million visits and 13,000 first-page keywords. Then the traffic went from roughly 610,000 visits in early October 2023 to about 190,000 by mid-December — near the site's early-2023 baseline of around 200,000. The offence was not that a model wrote the words. It was mapping another site's page inventory and regenerating it. A room full of freelancers doing exactly that would violate exactly the same policy.

Case two: humans, same outcome. HubSpot's content operation is professionally staffed, well resourced and human-written. Its organic traffic still fell from about 13.5M in November 2024 to 8.6M in December 2024 on Semrush data. Aleyda Solis's independent analysis tracked the blog subdomain from 3.94M monthly visits in January 2024 to 1.1M that December — roughly a 72% decline — with the whole domain down 23.2% across the year. The losses concentrated in high-volume, low-relevance pieces like "The 100+ Most Famous Quotes of All Time" and "How to Type the Shrug Emoji." Solis's read: "Google is doing a better job at ranking content that showcases real subject matter expertise." No AI involved, and the same shape.

Case three: AI at enormous scale, untouched. Gabe has documented Reddit ranking millions of AI-translated URLs — 2.3M in France, 2.4M in Spain — with no observed manual action, and Gizmodo indexing roughly 7,000 AI-translation-labelled articles on its Spanish subdomain with rising visibility. Those translations sit on top of genuine human discussion and reporting, and they serve readers who could not otherwise access it.

What the three have in common is not the tooling — it varies in every direction. What varies with the outcome is whether the pages added something a reader wanted. It is also worth being honest about the third case: enforcement is discretionary and appears sensitive to site-level trust. "The policy technically permits this" is a thinner shield for a two-year-old site than for Reddit.

How enforcement actually works

Google enforces "both through automated systems and, as needed, human review that can result in a manual action," and notes that sites violating policies "may rank lower in results or not appear in results at all." Most of what happens is algorithmic and silent. You will not get a notification.

When a manual action does land, the relevant one is filed under "Major spam problems", and its Search Console wording reads: "The site appears to use aggressive spam techniques such as scaled content abuse, cloaking, and/or other repeated or egregious violations of Google's spam policies." Recovery means fixing the site and filing a reconsideration request "with examples of bad content that you removed and good content that you added."

Two nuances that matter more than they get credit for.

First, there is a separate, softer manual action for "Thin content with little or no added value" — shallow pages, thin affiliate content, scraped content, doorway pages. A small blog publishing shallow posts can be hit by this without ever reaching anything Google would call "scaled." Volume is not your only exposure.

Second, seeing "Major spam problems" does not automatically mean your writing is the problem. Gabe documented a ~20,000-URL site that received a sitewide version of that action where the root cause was not content at all: a non-canonical www homepage had been hijacked and redirected to a gambling site via a dangling DNS record, producing nearly 12,000 clicks a day on gambling queries. Once fixed, the reconsideration request was approved in about a day. Diagnose before you rewrite.

One suggestive strand, clearly labelled as speculation: third-party analysis of the leaked Google Content Warehouse API documentation describes a site-level structure with fields like numOfUrlsByPeriods (new pages found across successive 30-day windows) sitting beside quality and engagement counts — a shape consistent with detecting scaled abuse from the ratio of pages produced to good pages over time. That is analysis of a leak, not Google guidance. Treat it as a plausible mental model, not a fact.

The 2026 wrinkle: the policy now covers AI answers

Most guides to scaled content abuse on the web today quote a version of the surrounding policy that no longer exists.

On 15 May 2026, Google rewrote the opening definition of spam. It now reads: "In the context of Google Search, spam refers to techniques used to deceive users or manipulate our Search systems into featuring content prominently, such as attempting to manipulate Search systems into ranking content highly or attempting to manipulate generative AI responses in Google Search." The previous wording ended at "ranking content highly."

That clause extends every named spam policy — scaled content abuse included — to AI Overviews and AI Mode. The surface you can be demoted from just got bigger.

Google's generative-AI optimisation guide, published the same day, is refreshingly boring about what to do next. Asked whether SEO still matters, it answers: "In short, yes! The best practices for SEO continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems." The same guide says you don't need machine-readable AI text files, don't need to break content into tiny pieces, and don't need special schema.org markup. Its content advice is one line: "Don't just recycle what others on the internet have already said, or could easily be produced by a generative AI model."

Meanwhile the map between rankings and citations is loosening. Ahrefs' March 2026 study across 863,000 keyword SERPs and 4 million AI Overview URLs found only 37.9% of AI Overview citations now come from the top 10, down from roughly 76% in a July 2025 study. So the durable play is the unglamorous one: be the page that had something worth citing.

Practical tactics for increasing your share of AI Overview citations — what signals to surface and how to structure content for generative models — are covered in How to Get Cited by ChatGPT and AI Overviews.

One correction while we're here. The March 2026 and June 2026 spam updates were both described by Google only as normal spam updates rolling out for all languages and locations. Neither announcement named scaled content abuse, despite claims to the contrary. Where you see specific drop figures attached to them, credit the analyst who measured them.

What a safe automated blog actually looks like

Google's helpful content guidance doubles as the best free pre-publish checklist available. Six questions from it, reframed as things to check before anything ships:

• Does the post provide original information, reporting, research, or analysis — or is it mainly summarising what others said?
• Does it offer insightful analysis beyond the obvious?
• If it draws on other sources, does it avoid simply copying or rewriting them and add substantial value?
• Does it contain any easily-verified factual errors?
• Does it look produced well, rather than sloppy or hastily assembled?
• Would a reader need to search again afterwards to get better information elsewhere?

Google also asks, in the same document, whether "the use of automation, including AI-generation, [is] self-evident to visitors through disclosures or in other ways." Disclosure is framed as a trust signal, not a legal obligation. Most of this post was drafted by an agent, fact-checked against its sources, and reviewed by a human before it went live — which seems like the kind of thing worth stating rather than hiding.

Now the honest objection, because skipping it would be its own kind of dishonesty. Orbit Media's twelfth annual blogger survey — 808 respondents, data collected August 2025 — found that the roughly one in ten marketers who use AI to write complete articles are the least likely of any AI-using group to report strong results. The overall benchmark for "strong results" is 21%. Bloggers who use AI for ideas came in at 23%. Orbit's summary of the winning pattern: "not too much AI, not too little."

That finding is the objection to this entire product category, and the difference it points at is where the human sits, not whether a model was involved.

Two more numbers from the same survey complicate the tidy "publish less" advice. Posts of 2,000+ words reported strong results at 39% versus the 21% benchmark, and people publishing multiple times weekly reported 37%. Frequency and depth both correlate positively with results. So the rule is not "publish less." It is publish nothing you wouldn't defend — and then keep doing that on a schedule you can actually sustain. The average post in that survey took 3 hours 25 minutes to write, which is precisely why most calendars quietly collapse into either silence or filler.

For a practical publishing schedule that balances depth, cadence and budget, see how often you should publish blog posts in 2026.

If you're weighing depth versus cadence, here’s what a 3,000-word AI blog post costs to generate in 2026 to help you budget and choose the right size and frequency.

This is the shape Magic Share is built around. Every draft is held until a human approves it — nothing publishes on its own — and no draft is saved until every cited URL resolves, with a verification report attached to the draft. That matters for the checklist above, because "does the content have any easily-verified factual errors?" and "is there clear sourcing?" are the two questions that are genuinely tedious to answer by hand at any cadence, and the two a machine can answer exactly.

The cadence controls are the same idea expressed as product design. A topic ledger that remembers everything written or planned is a direct answer to "are you producing lots of content on many different topics in hopes that some of it might perform well?" A banned-topics list keeps a site inside its own expertise rather than drifting toward shrug-emoji territory. And a publish schedule spaces posts out deliberately — weekly, not hourly. A garden tended year-round beats a field seeded all at once, which is almost exactly the shape of Ray's traffic curves.

For a practical blueprint of those controls — topic ledgers, verification reports, cadence gating and human-in-the-loop approval — see how to build an AI content agent: 6 parts beyond the prompt.

To see a concrete, small-site example of overlapping topics stealing traffic and how to fix it, read Keyword Cannibalization on a Small Blog: The Real Cost.

Frequently asked questions

Does Google penalise AI-generated content? Not for being AI-generated. The scaled content abuse policy applies to unoriginal, low-value pages produced at volume to manipulate rankings, "no matter how it's created." Ahrefs found 86.5% of top-20 pages contain some AI content and effectively zero correlation (0.011) between AI share and position. What draws enforcement is volume without value, not authorship.

How many pages count as "scaled"? Google has never published a number and almost certainly won't, since a published threshold would immediately become a target. Enforcement looks at patterns — volume relative to originality, value and intent, assessed at site level. Note that the separate "thin content with little or no added value" manual action can apply to small sites that never approach scale.

Do I have to disclose that AI helped write a post? There is no requirement. Google's helpful content guidance asks whether automation is "self-evident to visitors through disclosures or in other ways" and frames it as a trust signal. A short, plain note about how a post was researched and checked usually costs nothing and reads as confidence.

Does the scaled content abuse policy apply to AI Overviews? Yes, since 15 May 2026. Google's revised definition of spam now covers "attempting to manipulate generative AI responses in Google Search," which extends every named spam policy — scaled content abuse included — to AI Overviews and AI Mode.

If you're wondering whether a robots-style llms.txt can influence AI Overviews and your site's citations, see how llms.txt affects SEO and AI Overview citations.

The short version

Read the policy as written and it stops being mysterious. It bans many pages, unoriginal, low value, published mainly to capture rankings. It does not ban the tool you used, a long post, a weekly schedule, or a translation that helps someone read your work. The three cautionary tales above have wildly different tooling and identical arithmetic underneath: pages produced faster than value was added.

Which makes the safe path for an automated blog fairly plain. Fewer posts than your tooling could physically produce. Real sources, actually checked. A human deciding what ships. A record of what you've already said, so you never publish the same thing twice with different words. And a cadence you can hold for a year.

That is the whole product thesis here, and it is also just good publishing. If you'd like a blog that grows that way — researched, fact-checked, and waiting for your approval each morning — start free, or watch this blog fill up with posts written exactly the way this one was.