Claude's Watermarks Just Broke SEO transcript

Anthropic just started watermarking every single word that Claude writes, invisible, embedded in the text itself, and it rides along when you paste it onto your website. So, if you've been publishing AI content, there's now a permanent mark on it saying where it came from. It's on everything, every single model launched since August 2nd across the API, Claude Code, Cowork, all of it, worldwide. It survives copy and paste.

It survives editing. And Anthropic says a detector is coming. Plus, they're working on retrofitting the older models. So, follow that chain out. That mark is on your pages. Anthropic is going to ship the detector, and now Google knows exactly which pages on the internet were written by AI down to the exact paragraph. That's the fear, not the watermark itself, but what a search engine does once it can sort the entire index into human and machine.

Now, if that flips into a demotion, every AI published page you got turns into a liability overnight. And there's no real way to walk this back at scale. That's the version currently circulating right now, and every link in it sounds reasonable. But none of this makes sense until you understand what the watermark actually is. Because almost everyone reporting on it got this part wrong. The watermark is not metadata. There's no hidden file attached to your text.

There's no invisible characters buried in it. The watermark is the word choices themselves. When Claude writes, it's choosing each word from a set of options. The watermark nudges those choices in a pattern. Every individual word looks completely normal. But across a few hundred words, the pattern of choices adds up to a signature. This is why it survives copy and paste. You're still of course copying the words, and the words are the watermark.

It's also why heavy rewriting kills it, but light paraphrasing usually won't. You have to change enough of the word choices so that the pattern falls apart. Now, we have to hold onto to because this matters for something later. There's one more thing that most people are getting wrong here. A detected mark does not prove that Claude wrote it, okay? This is Anthropic's own language in their own documents. If you wrote an entire article yourself by hand, you paste it in, and you ask Claude to clean up the grammar, the output will still likely carry that mark.

It's the same if you ask it to translate or summarize. All it tells you then is that the text passed through Claude at some point, which is a much narrower claim than a lot of people are making. So, that's what the mark is. It's real, it's durable, and it's a lot dumber than most of the coverage suggests. But, the version circulating right now breaks in three separate places. I'm going to show you. First, Anthropic ships a detector, and it gets into Google's hands.

So, okay, this one clearly plausible. Anthropic has said it intends to publish the details and allow third parties to check it. So, I'm obviously not going to stand here and tell you that detection is impossible. It is likely coming. This makes the second link the important one because the whole chain assumes that once Google can detect AI content, Google will do something about it. And honestly, that assumption has never been tested with text.

Text watermarking is like 10 days old. Nobody has data on it. Nobody can have data on it. But, it has been tested somewhere else for 3 years in public, and almost one in this industry has really bothered to look. So, before I show that to you, I want you to look at what Google has actually said about AI-generated content because their documented position is not silence. They wrote it down. Their guidance says appropriate use of AI or automation is not against their guidelines, and that however content is produced, you should be aiming for original, high-quality, people-first work.

This is Google's own blog post from 2023 and it's still up. Now, I'm not going to sit here and tell you that you should trust Google's public statements. Google has spent years telling this industry that clicks were too noisy to use in ranking. Then, their VP of Search testified under oath that Navboost is one of their important signals and later a former engineer testified that Google was told internally not to admit this publicly because SEOs would exploit it.

This is the same story with the helpful content update. Google said it would reward people-first content. Then, Danny Sullivan sat in a room with publishers it had crushed and told them there was nothing wrong with their content. Google's systems just weren't surfacing it. Those people paid their own way to that meeting. Okay, so a Google blog post proves basically nothing on its own. What I want is behavior, years of it, with the detection already working.

This brings me back to what's been running in public since 2023, images. Google has a watermarking system called SynthID. It's how Google marks the images, the video, and audio its models produce. And it's been live since 2023. It shows up inside Google Search, in Lens, in Circle to Search, where it'll tell you whether a picture was AI-generated. Now, every major lab does some version of this. It's been the industry standard for years.

So, if Google were ever going to rank against AI content, images are where this would have happened first. The mark exists, the detection works, it's been sitting there for over 3 years. Now, let me show this to you. This is a page that my Core 30 agent designed for local SEO built. And here's an image that was AI-generated, start to finish. It's a good image. Google Gemini generated it. You'd scroll right past this if you visited this website and never really think twice.

I mean, maybe it looks a little weird that he's taking a wrench to this part of the water heater, but it's not a super weird image. Now, let me show you what this looks like on a free detector. This is just one random detector I found and it says 99% AI generated. This is off the pixels alone in like 2 seconds. There's no login, there's no key, there's no special access. Anybody can run this detection including Google and honestly Google can do it better because Google's own model made the image and Google's own watermarks are in its output.

So, this is not a close call. Every layer of detection that exists catches it. Now, here's what happened to these pages. Let me show you some rank maps. This is from people running that very same Core 30 agent before and after. AI written content, AI generated images published to their sites. Here's Here's this one, May 27th to June 16th. So, this is roughly 3 weeks of a before and after all based on content written by the Core 30 agent.

Here's another one. Uh this May 2nd to June 25th. So, this is almost 2 months. Uh I can show another page here. Uh we have This is July 16th to July 25th. So, this is 9 days, right? Insane progress for 9 days. We can show another one here. This is the before, June 29th. Here's the after, July 24th. A little bit less than a month. Absolutely crazy results that people are getting with the Core 30 agent. And remember, this is with AI generated images, AI generated text, 3 years of working free public detection and it has never once been what decides whether a page ranks.

That's the second link gone. Google having the capability and Google using it against you have never been the same thing and these are the receipts to prove it. Which leaves the third link. And this is the one that's going to cost you money if you get it wrong. Because somebody is going to tell you that you need to strip the watermark out anyway just to be safe. Now, these methods exist. There's published research on this.

Documented academic papers. You can run text through a paraphraser. You can translate it into another language and back. You can do character level attacks that swap letters for look-alikes from other alphabets. And there's newer work on more targeted removal. Some of it actually needs access to the model's internals, which obviously you're never going to get through a hosted API, but most of it doesn't. Most of it you can just run this afternoon on content if you wanted to.

But they all cost something. Remember how the watermark works, that the signal lives in the word choices themselves. That's why paraphrases and round trip translation defeat it. But keep in mind, you can't use Claude or OpenAI or Gemini to do the paraphrasing or to do the round trip translation. This has to be done by hand. OpenAI, Gemini are signing on to their own watermarks. This is part of a European Union law. So, we're not deleting a tag.

We're not turning every word on the page. Doing this, especially by hand, the writing will flatten, errors will creep in that no one proofreads for, and you've ended up rewriting a page that was working. And the character swaps are the easiest but the worst of the bunch because of how this works under the hood. Remember that Google doesn't match keywords, it resolves entities. It reads the city Molden and connects it to everything we know about that city.

It reads your business name and connects it to your Google Business Profile. But none of it happens visually. Of course, Google's algorithm doesn't have eyes. Text gets broken into tokens before anything looks at it. And if you've replaced the Latin symbol A with a Cyrillic character that looks identical, it's a still a completely different character underneath. So, a swapped word, it's not a misspelling that something can work out from context.

It doesn't resolve to anything at all. Do that to a service name, a category, a city, the business name, and the page stops making the connection it was built to make. On top of that, mixed script text is trivially detectable, and Google has run this as a spam signal for years. You'd be breaking the entities on your page. You'd be handing Google something it flags on site. You'd be degrading real pages to hide from a consequence that doesn't exist.

Okay, so look, things change. If there's ever evidence that Providence affects ranking, I'll talk about it, and I'll say so plainly. I have enough clients under management that are ranking with AI-written content and AI-generated images that we would see it immediately at my agency. But right now, the evidence runs the other way, and it's been running that way for 3 years. If you want to see the system that produced these rank maps, I put together a full walk-through of how the Court 38 Agent works, what it writes, and what the pages look like.

You can check out that video here.