Google Can Detect Your AI Content. Here's What Happens Next.

0:00Google's John Mueller liked this LinkedIn post and has been liking other posts just like it. John Mueller is basically the face of Google search. He's Google search advocate, answers a lot of SEO questions, and it's kind of shocking that he liked this LinkedIn post. This is from Christine who said, "I said it a couple months ago and I'm going to say it again. Google is going to come out with an AI content killing algorithm sometime in the next 6 to 8 months. These spam updates are testing it. Why? Because regurgitated AI content destroys their models and right now the internet is something like 50% to 60% AI

0:41slop. Google doesn't care how you create content is completely untrue. I wish they would stop saying that because they have said that all AI content out of the box is considered low quality. That means they do care. And she makes a good point with the August spam update. The August spam update killed so many AI sites all over the place. You saw people saying I had this AI content site and it was destroyed this August. It's the overwhelming belief that Google's August

1:12spam update focused on SEO AI content. It was even covered in Search Engine Journal. Reports indicate Google spam update focused on SEO AI content. So we have to revisit this. How easily can Google detect AI content? How cheap is it for them to do so? And are they actually going to do anything about it? So even though Google is not going to get the keys to cheaply detect AI text from other AI companies, Article 50 of the EU's AI Act, which is being adopted globally by all the major AI companies, requires, as I understand it, third parties like Google to get a reasonably practical way to detect AI generated

1:52text with costs kept in mind. So it actually might become cheaper for Google to detect AI-generated text. Personally, I don't think it will become that much cheaper. Google has already had AI text detection through SynthID, which they developed years ago and have been iterating on since. And it's also to be determined how much watermarking will change output because all the major AI companies have committed to watermarking their text for AI detection because of Article 50 of the EU AI Act. But, I don't think large language models are going to worsen their text just for watermarking. No way. And I will still put forth that this might still be a

2:35marketing move and nothing is actually going to change in terms of output, which is what I said in the previous episode on this, episode 1,140 of the show. Claude is watermarking everything it writes, Google doesn't have the key. But, after seeing this LinkedIn post, which John Mueller has liked, I think it's important to ask ourselves, let's suppose that Google can much more cheaply detect AI text. Again, I think it's unlikely, but it's still an important question to ask, what now? Is Google going to take action and remove all AI text from their index, show no more AI text to searchers, not train on any AI text? What now if this is the

3:18case? So, I'm going to share this really great blog post from Dan Petrovic. Dan is a legend in SEO. I'm sharing his stuff so much on the podcast. He's been doing the coolest SEO experiments for decades, and I love sharing his stuff. He wrote this post, "Relax, Google won't penalize all AI-generated content." So, I'm going to cherry-pick my favorite lines from this. He says, "Google evaluates content based on substance and user behavior signals rather than penalizing pages simply because AI generation is detected." Just because Google can detect AI generated content does not mean it will automatically penalize it. One, Google is heavily invested in the commercialization of

4:01Gemini. They're not going to undermine its integration into virtually all of their products. Two, bad content predates LLMs and humans have generated plenty of it already. Google leads with state-of-the-art content quality evaluation algorithms and have more user behavior data than any other company on the planet. And that is certainly true. Not just with Google Search, but also with YouTube. Google has all the behavior data. And three, AI detection is scalable and computationally cheap, but represents a separate layer to content quality evaluation and spam detection. So, Dan makes the argument in this post that AI text detection was already very cheap. So, we continue with this post. Dan says, "Low-quality filler

4:45existed decades before large language models, spanning content farms, link networks, and mechanical article spinners. An LLM on its own possesses no more agency, purpose, or creative ambition than a spell checker, an autocomplete utility, or a kitchen toaster. It functions entirely as a mechanical tool guided by a human operator. Search engines evaluate documents through a multi-stage quality pipeline where initial detection is merely a preliminary signal, passing the content to content substance classification and downstream user behavior signals to reward high-utility pages regardless of their origin. Yes,

5:26detection is computationally cheap. Identifying machine-generated text does not require large deep learning models or extensive compute clusters. We demonstrate that distinguishing model outputs can be accomplished using parameter-free gzip compression distance, which applies normalized compression distance alongside K-nearest neighbors classification. Because text generated by specific models compress together in predictable ways, attribution and origin identification are light on computational resources. In search architectures, origin detection serves as an initial classification signal rather than a trigger for exclusion. Search engines evaluate multiple downstream heuristics before determining document placement, ensuring that machine-assisted documents with high utility remain discoverable. Search systems increasingly differentiate actionable information from filler using specialized semantic classifiers. And this next part is super important. These

6:30text evaluations operate alongside downstream user behavior signals. Search engines assess post-click interactions, session durations, and navigation paths to evaluate whether a page resolves reader queries. Perceptions of content quality remain inherently subjective. David Kwaed, friend of the podcast, probably the guest who's been on the most, is always saying this. Text that appears redundant or uninspired to one audience may provide practical utility to another, creating risks of collateral

7:01damage during automated sweeps. Addressing varied quality preferences over the long term requires transparent labeling and user-controlled filtering rather than blanket origin suppression. Origin suppression meaning just excluding all AI content. Of course, whenever there are classifiers in the mix, there will be casualties, but it won't be as doom and gloom as many fear. So, that's all from this excellent post from Dan Petrovic, but we still have to revisit the fact that John Mueller liked this LinkedIn post saying "Regurgitated

7:35AI content destroys Google's models. Google doesn't care how you create content is completely untrue." So, what I'm going to do is I'm going to end this episode with three possibilities and then what I'm doing with AI content and AI text. Possibility one is AI content performs in search, so it's not excluded in SEO, but it does get removed from training data. And this will hurt companies that are using AI to write most of their content, but at least if they have customers sharing them if these companies have customers sharing them with without AI, then that will help them. There will be all this user-generated content about the

8:14companies that's authentic that is actually used in training data. And with option one, if AI text is not used in training data, that means that large language models have to do more searches to get their information because they don't have as much in their training data. And I actually think this is entirely possible. You might see that AI content is removed from training data in order to make the models better. And as a result, the large language models are doing more searches. Option two is AI content performs in search and knowingly gets trained on, but AI content receives a less weight. That one is possible, too. And option three is AI content

8:55performs in search and gets trained on depending on user signals for where the content came from and some categorization like YMYL informational AI text would be completely ignored, but commercial non-YMYL text would be used. This one could be true for Google, but other AI companies don't have all the behavior signals that Google has. So, I would not bank on this one. It really could be just something like number one is what happens. Does that mean you shouldn't use AI at all to make your SEO

9:26content? Well, for us, we are using AI. We're using AI as a writing assistant, not to write everything. AI content that is 100% generated from start to finish sucks. We are using AI to help articulate and to ideate or even to summarize with heavy heavy editing. But when you just give a short prompt and you say make this and you just put it out, you don't really think about it, that content is going to suck. It's going to suck, it's going to be unoriginal, it's going to be fluffy, it's probably not going to give the searchers exactly what they want. If you're trying to use it to sell, it

10:00might do a worse job of selling because it's actually ignoring selling points or re phrasing things in a way that don't sound as confident or don't put your most strong suits forward. But using it again as an assistant to help articulate and to ideate, that's great. And with heavy editing, it's not going to read as much like AI. However, that doesn't mean that Google won't take all the text on a website or huge chunks of text on a website, multiple pages to see because it's easier to detect AI when you have more text. So then they can see okay, does this website actually use a lot of

10:34AI? But for me, that's okay because with high sales intent prompts where people are actually looking to convert, nine out of 10 times, the LLM is going to do a search anyway. And the more you are showing up in search because I don't believe that AI text will be excluded from search. It's going to be a signal that is used if there's tons of bad behavior signals. You want to show up in search. And when you show up in search and get customers and do a good job satisfying them, those customers are

11:03going to put out user-generated content. And you're going to appear on Reddit, you're going to appear on Trustpilot, you're going to appear on all these places. Customers are writing about you using their own language. That's being trained on more or that's receiving more weight. I don't think any of this is a reason to stop using AI text to do your SEO, especially high-intent SEO where you are trying to convert. But you should absolutely be reviewing, making sure that your text satisfies search intent, trying to satisfy search intent right away, heavily editing, using it to ideate, not not just having it write everything from one short prompt. Pretty

11:39much what we've been saying on the show for the last 1,000 episodes. All right, if you want to learn how to do very high intent SEO where you are targeting people who are looking to buy something, use something, make a discovery call, and this is SEO that you can use with AI. This These are strategies and methods that you can use with AI. I have

11:57a course on this, compactkeywords.com. 13 and 1/2 hours, written sections, Google Docs, Google Sheets, templates, visual templates, page templates, everything is in this course. If you haven't checked it out, you're going to love it. This is episode 1,153 of the podcast, 1,153 days in a row doing this show. If you watched this on YouTube, thank you so much for watching. If you listened on Spotify or Apple Podcasts, thank you so much for listening. And I will talk to

12:23you again tomorrow. Bye now.