Google S-CTS y S-BERT: qué pueden aprender los equipos SEO

Google published a research paper that some SEOs are reading as Google can now detect AI content at scale. That is almost right. But the more useful lesson is sharper than that. The paper is not proof that every AI assisted article is about to get punished in Google search. It is a warning about systems that look like synthetic content clusters. Welcome to SEO for Francisco. I'm Francisco. I have more than 15 years in SEO, including more than 7 years as SEO lead at Shopify, more than 4 years at Pornhub, and hands-on work with multinationals through my agency Growing Search. In this video, I'm going to show you what Google's SCTS paper actually says. Where the embedding model fits, and the five checks I would use before scaling any AI assisted content program. Here's the first important distinction. SCTS stands for scalable cluster termination system. That name matters. It is not called one paragraph detector. It is not called single blog post classifier. It is a cluster system. Google Research describes a deployed defense system for a major online video platform. The system looks for coordinated groups of accounts with a high concentration of adversarial synthetic content. That means the system is looking beyond one piece of text. It is looking at patterns, content account and behavior signals, synthetic artifacts, and the relationship between those signals. Now, why should SEOs care if the paper is about a video platform and not a confirmed Google search ranking system? Because the mental model is portable. If you publish AI assisted content at scale, the risky question is no longer just whether one page sounds AI written. The better question is, if a platform looked at the entire publishing system, would it see human judgment, or would it see a synthetic cluster? That is the part most SEOs miss. The model name is BERT. In plain English, embeddings help compare meaning and structure. So, two articles can use different words, but still have the same footprint, the same opening, the same advice order, the same examples, the same generic conclusion, the same thin comparison table. That is not just an AI detector problem. That is an editorial quality problem. So, here is the practical audit. Check one. Source evidence. Before a page is drafted, does it have primary sources, screenshots, test data, customer language, or first-party observations? If the answer is no, the model is probably going to fill the page with generic probability. Check two. Unique page purpose. Would this URL still deserve to exist if search volume were zero? If the only reason for the page is that there is a keyword, that is a bad sign. Check three. Template risk. Group your pages by production process, not just by topic. If 20 pages share the same semantic path, same H2 structure, same table format, and same conclusion, you may have a cluster problem, even if every page passes a basic grammar check. Check four. Expert review. Someone has to remove weak statements, include caveats, reject pages that should not publish, and add experience that a prompt cannot invent. If the prompt template makes every important decision, you are not scaling content quality. You are scaling sameness. Check five, publication velocity. If output volume rises faster than source review and expert review, the risk goes up. That is where many AI content operations get into trouble. They increase publishing speed, but they do not increase evidence, editing, or accountability. Now, here is what this paper does not prove. It does not prove that SCT is used as a ranking system in Google Search. It does not prove that this system is part of Google's public spam detection systems. It does not prove that AI-assisted writing is automatically spam. And it does not prove that a specific ranking drop was caused by this research. The main takeaway is different. Google's anti-spam research is moving towards systems-level pattern detection. So, your SEO quality process also needs a systems-level view. Do not only ask, "Is this one article readable?" Ask, "What pattern does this publishing operation create?" A healthy AI-assisted workflow starts with a source pack. Then it creates a draft. Then it gets expert edits. Then it checks internal links and related context. Then it publishes only when the page has a reason to exist. A risky workflow starts with a keyword list, sends it through one template, adds a feature image, and publishes the same kind of page again and again. That is the difference. AI assistance is not the core problem. The core problem is making the same kind of page again and again at scale. If you want the full checklist, I published the complete breakdown on SEOfrancisco.com. I also included the source links, the paper caveats, and the cluster risk audit table. And I want to know how you are handling this. Are you auditing AI content page by page, or are you auditing the whole content cluster? Leave that in the comments. If this helped, like the video, subscribe to SEOfrancisco, and I will see you in the next one.