On September 24, 2026, at 6:15 PM Beijing time, Google quietly began rolling out its fourth confirmed spam update of the year. As of today it's still running, with a scheduled completion around October 8. If you've opened Search Console in the past week and watched your traffic curve drift downward, this is almost certainly why.
On September 24, 2026, at 6:15 PM Beijing time, Google quietly began rolling out its fourth confirmed spam update of the year. As of today it's still running, with completion expected around October 8. If you've opened Search Console in the past week and watched your traffic curve drift downward, this is almost certainly why.
But the update itself isn't what I want to talk about. The more important development is how Google catches spam content now — it has shifted from reviewing pages one by one to following the trail and taking down entire networks at once.
How Google Detects Mass AI Content Today
According to research Google has published, the company runs a system called S-CTS (Scalable Cluster Termination System). The mechanics are straightforward:
First, it converts massive numbers of pages into vectors using Sentence-BERT and maps them into a semantic space. Articles produced from the same prompt and generation pipeline cluster together in this space — even when the wording differs.
Second, it analyzes connections between accounts and domains: publishing cadence, template structure, even overlapping infrastructure. If you run 20 sites using the same tooling and the same rhythm, the system sees them as one cluster.
Third, once enough accounts inside a cluster show the same templated narrative and abnormal publishing patterns, the entire cluster gets actioned — not just individual pages.
This explains a phenomenon many site owners have reported recently: it's not a few articles losing positions, it's batches of pages, or entire domains, vanishing at once.
More importantly, the system adapts quickly. When spammers switch models or prompts, the classifiers can be retrained in short order. The window for dodging detection by changing tactics keeps shrinking.
Who the September Update Hit Hardest
Based on cases aggregated across the SEO community over the past week, three groups have suffered most:
Pages with heavily templated structure. Formulaic titles, formulaic paragraphs, even the closing CTAs identical every time.
Programmatically mass-produced content. The classic examples are city and location pages — one article regenerated hundreds of times with a different place name swapped in. Product category guide pages replicated at scale on e-commerce sites fall into the same bucket.
Fully automated AI publishing with zero human involvement. Multiple practitioners have observed a pattern (an observation, not an official Google statement): among sites using AI, the ones that kept a human review step in their pipeline weathered this round noticeably better.
Interestingly, Google still hasn't acknowledged that this update specifically targets AI content. The official line is unchanged: it goes after "content produced at scale primarily to manipulate rankings," regardless of the tool used.
That's true — but mass-producing ranking-focused content with AI is now so cheap that enforcement naturally concentrates there.
What About Baidu?
Baidu takes a different route but heads in the same direction. Its algorithm system is organized by problem type: Hurricane handles site networks and aggressive scraping, Gale targets misleading aggregation pages, Drizzle covers low-quality B2B pages and advertorials, Blue Sky deals with directory trading.
For ordinary content creators, Hurricane and Gale are the easiest to trip — the batch site-network model, and pages that promise one thing in the title but deliver only aggregated filler with no real answer.
One data point is worth noting: Baidu now filters roughly 120 million abnormal links per day, and the weight of "link naturalness" in its scoring continues to rise. Old tactics like bulk-buying links and mass reciprocal linking are effectively self-incrimination now.
How a Legitimate Site Stays Alive
Based on the signals from this update, here's what we advise the GEO clients we work with at SilkGeo:
Don't touch the site-network model. It's the most direct target of cluster detection. One operator running a swarm of near-identical, interlinked sites is the textbook definition of a "generation cluster" in S-CTS terms. Better to build one site deep than spread ten sites thin.
De-template your content structure. This doesn't mean dropping H2s and H3s — it means not writing every single piece in the same "what / why / how / summary" four-part pattern. Real humans sometimes open with a story, sometimes lead with the conclusion, sometimes spend an entire piece arguing against a popular view. Structural variety across your articles is itself a strong human signal.
Publish like a human. A brand-new site dropping 30 posts in one day and then going silent for a week is more suspicious than any content signature. A steady, sustainable cadence — even one post a day — is far safer than pulse-style flooding.
Every piece needs a "human anchor." A specific experience, real numbers, your own judgment, even a mistake you made. These aren't just quality signals; in semantic space they naturally keep your content from clumping with everyone else's — because they're genuinely unreplicable.
Don't make panic edits while the rollout runs. Both Google and veteran practitioners stress the same point: mass changes during a rollout destroy your ability to diagnose. Pull data from September 24 onward, see which page types and directories are dropping, identify the pattern, then act. Fix the cause, not the symptoms.
The Bigger Picture
There's a related story unfolding in parallel: the Penske Media antitrust case against Google (Penske owns Rolling Stone, Variety and more). During a recent hearing, the presiding judge directly questioned the nature of Google's AI Overviews, describing them as something built "on the backs of publishers."
No ruling has been issued yet, but read that case alongside S-CTS and they point to the same trend: the rules of the AI-era content ecosystem are being rewritten. On one end, AI is consuming content at scale and serving answers directly on the results page. On the other, search engine tolerance for low-quality mass production is approaching zero.
The space in between belongs to content that offers genuine, unique value.
That's not actually a bad outcome. When the flood recedes, the naked swimmers leave — and the people doing serious work become easier to see. Whether by humans or by AI.