Google's scaled content abuse policy: the AI pipeline that survives

Scaled content abuse is a Google spam policy that targets pages mass-produced to win rankings rather than to help anyone, whether a person wrote them, a model wrote them, or the two split the work. The policy is method-agnostic on purpose. Google rewrote it in March 2024 because the old wording chased automation, and automation had stopped being the tell.

So the question your content pipeline has to answer is not whether a model wrote the page. It's whether the page would exist if search did not.

What the policy actually says

The official definition in Google's spam policies is narrow and worth reading literally: pages "generated for the primary purpose of manipulating search rankings and not helping users". Two conditions, joined by an and. Volume alone is not the violation. Volume in service of ranking, with nothing underneath it for a reader, is.

The documented examples are specific. Using generative AI to produce many pages without adding value. Scraping feeds or search results into pages. Stitching content from other pages without adding anything. Spinning up extra sites to hide how scaled the operation is. Publishing pages that carry keywords but make little sense.

Notice what is absent from that list. There is no page-per-month threshold, no word count, no AI-detection score. Google's separate guidance on generative AI content makes the same point from the other side: the ranking systems reward original, helpful content that shows expertise, experience, authoritativeness and trust, and how it was produced is not the deciding factor.

What changed in 2026

Two things moved this year, and both tighten the frame around automated publishing.

On 15 May 2026, Google extended the spam policies to cover attempts to manipulate generative AI responses in Search, not just the ten blue links. Every tactic the policy already named, scaled content abuse included, now applies when the target is an AI Overview or an AI Mode answer. If your pipeline exists to get cited by an AI engine, it is inside the policy's scope. That is worth saying plainly, because GEO tactics have been marketed for two years as though they sat outside spam enforcement.

Then, on 24 June 2026, Google shipped the June 2026 spam update. It ran globally, across all languages, on SpamBrain, and finished rolling out on 26 June, about two days end to end. Google confirmed the update did not target link spam and did not target site reputation abuse. Ruling those two out points the update at content-level violations, and scaled content abuse is the most aggressively policed of those since 2024.

Two days is the number to sit with. A demotion that arrives over 48 hours gives you no window to react mid-rollout. Whatever your pipeline is doing when an update lands is what gets judged.

Why "add a human pass" is not the fix

The common reflex after a spam update is to bolt an editor onto the end of the pipeline. Someone skims each draft, fixes the tone, ships it. The output is now human-touched, so the thinking goes, and therefore safe.

It isn't, because provenance was never the thing being measured. Ahrefs looked at 600,000 pages and found a correlation of 0.011 between AI-generated content and ranking penalties. That is noise. And it has to be noise, because 74.2% of new pages contain some AI-written text, with only 2.5% written purely by a model. A signal that demoted AI presence would demote three quarters of the new web.

A polish pass changes how a page reads. It does not change why the page exists. If the answer to "why does this URL exist" is still "because the keyword had volume", the editor added cost and no defence.

What actually works

Five constraints, in rough order of how much they protect you.

Cap the output at what you can defend

Pick the number of pages you could personally justify to a reviewer, one by one, and publish that number. Not the number your tooling can produce. Those are different figures, and the gap between them is where the risk lives. A pipeline that can write forty articles a week and publishes four is a pipeline making an editorial choice. One that publishes forty is making a bet on detection.

Give every page a job that predates the keyword

Work from a backlog you wrote before you saw the search data, where each entry names a question you have an actual answer to. Search data then tells you which of those to write first. Invert the order, and you are letting keyword volume decide what you believe, which is the exact behaviour the policy describes.

Make research a gate, not a garnish

Every claim with a number, a date, a version or a regulation in it gets verified against a primary source before the draft is written, and the topic gets dropped when the sources come back thin. This is the step that separates a page carrying information from a page carrying prose about a topic. It is also the step that automation skips first, because it is the slowest.

Edit for the claim, not the prose

Reading a draft for tone catches nothing that matters here. Read it for the assertions: which are load-bearing, which are sourced, which are true only in a sentence that sounds nice. Cut the third category. A model will generate a plausible statistic without malice, and plausible statistics are the fastest route to a page that helps nobody.

Refresh what you publish, or take it down

Anything old enough to be wrong is a page that no longer helps a user, and it counts against you whether a model wrote it or you did. Pick a window, revisit what falls outside it, and delete what cannot be brought current. Publishing is an obligation you take on per URL, and it does not end at the publish button.

What this looks like in practice

We run one of these pipelines on this blog, so the constraints above are ours and not hypothetical. It publishes one article per run. It picks from a backlog written months ahead, cross-referenced against Search Console for which topic to pull forward. It verifies claims against primary sources before drafting, and it drops the topic when the research is thin. It ships in English, Italian and Spanish. Anything past 90 days gets revisited and updated or retired.

That has produced 92 articles. A pipeline optimising for volume would have produced several thousand in the same period, and it would have been the wrong bet. The cost of the constraints is real: we publish maybe a tenth of what the tooling could emit. The benefit is that every URL has a reason to exist that we could defend out loud, which is the only test the policy actually applies.

The uncomfortable part is that none of this is verifiable from outside. You cannot look at a page and see whether a backlog or a keyword tool chose it. Google can only infer it, imperfectly, from the pages themselves. Which means the discipline is genuinely voluntary, and the sites that got hit in June were, for the most part, sites that took the other option.

Sources

• Spam policies for Google web search
• Google Search's guidance on generative AI content
• Google Search: new updates to address spam and low-quality results (March 2024)
• Google completes its June 2026 spam update rollout
• AI-generated content does not hurt your Google rankings (600,000 pages analysed)
• 74% of new webpages include AI content (study of 900k pages)