Claude Now Watermarks Its Writing. What That Means for Your Site.

A headline went around this month: AI watermarks just broke SEO. If you have used Claude to help write a service page, a blog post, or a Google Business Profile update, that headline is built to worry you.

Here is the short version. Anthropic did start watermarking Claude's text. That part is real. Most of what got stacked on top of it is not. No search engine has announced a penalty. The mark does not prove who wrote anything. And it is weakest on exactly the kind of short, factual copy a local business publishes.

This is what actually changed, taken from Anthropic's own explanation, and what it means for a business in Central Illinois.

What actually changed

Starting August 2, 2026, new Claude models generate text carrying a machine-readable watermark. Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, alongside roughly 190 other signatories. The requirement traces back to the EU AI Act.

Anthropic turned it on worldwide, not just in Europe. Its stated reason is practical: the company does not yet have a durable way to limit the feature by region. The mark covers text from supported models across Claude, the API, and its other products. Older models are being added over the coming months.

This is also not a Claude-only story. Other major model providers signed the same code and are implementing their own marks. If you use a different assistant, expect something similar.

What the watermark actually is

A language model writes one word at a time. At each step it picks from a short list of reasonable next words. Take "the weather today was cold and." The next word could be "overcast" or "grey." Either one works, and the meaning barely moves. Normally that tie is settled by a random number.

Watermarking changes where that randomness comes from. Instead of an arbitrary number, the model uses a secret key plus the preceding words to settle the tie. The words still look random. But someone holding the key can check the sequence and estimate whether it matches the pattern the key would produce.

That is the whole mechanism. In Anthropic's words: "Nothing is added to the text and there are no hidden characters." There is no invisible symbol to strip out, no tag in your file, no tracker. The method is a version of SynthID-Text, published by Google DeepMind in Nature in 2024.

It also carries nothing about you. Anthropic says the mark cannot be traced to a person, an organization, or a conversation. It does not slow the model down and does not cost extra.

Three claims that do not hold up

Most of the panic comes from three ideas that the source material does not support.

First: that a detected mark proves an AI wrote the page. It does not. Anthropic is direct about this. The watermark "cannot distinguish 'Claude wrote this' from 'Claude heavily edited this.'" It also cannot tell you the text was written by some other AI, and the absence of a mark does not mean a human wrote it.

Second: that there are hidden characters in your text you need to clean out. There are not. The pattern lives in which ordinary words got chosen, not in anything added to the file.

Third: that Google will demote watermarked pages. Nobody has said that. Not Google, not Anthropic. Google's published guidance on AI content has said the opposite for years: its systems aim at "rewarding high-quality content, however it is produced," and "appropriate use of AI or automation is not against our guidelines." What is against the guidelines is using automation to generate content mainly to manipulate rankings. That was already true before August.

Where the mark is weakest is where local copy lives

This is the part the scary version leaves out. The watermark needs room to work, and a local business rarely gives it any.

It needs length. Anthropic says detection "doesn't work well on small samples," because a short passage holds too few word choices to read a pattern from. A Google Business Profile post runs a couple hundred characters. There is almost nothing there to detect.

It needs choices. The mark is thinner on factual writing, because facts do not leave room to pick a different word. Anthropic's example is a sentence naming Isaac Newton's Principia, where only one word is correct. Your hours, your prices, your service area, and your phone number work the same way.

And it barely survives real editing on your side. When a model only cleans up grammar on something you wrote, Anthropic notes that "nearly all the words are the person's, there's very little (if anything) for the watermark to attach to."

One more thing worth knowing: there is no public detector yet. Anthropic says a detection API is coming and that it is still working out the details. Right now, nobody outside the company can check a page.

What has not changed at all

Picture two plumbing pages in the same town. One lists real prices, the neighborhoods served, photos of finished work, and what happens on the first visit. The other says the company is committed to quality service for all your plumbing needs.

The first page wins. It wins whether a machine helped draft it or a person typed every letter. It won last year, and it will win next year. A watermark does not enter into it.

The pages that should worry an owner are the ones that were published fast and never touched again. Six city pages that swap only the city name. A services list that could belong to any company in the state. Those pages were a problem before this announcement. The announcement did not create them, and scrubbing a watermark would not fix them.

What to do this week

Six practical moves, in order.

1. Do not rewrite good pages to remove a mark. Anthropic says only a rewrite replacing every word clears it, which means rewriting the page from scratch. That is real cost against a penalty no one has announced.
2. Pull up the pages you published quickly and never revisited. Read them as a customer. If a page does not say anything a competitor could not also say, that is the page to fix.
3. Add the things only you can add: your prices, your service area, photos of actual work, the real steps of the job, the names of the towns you drive to.
4. Keep one person accountable for every page before it goes live. Someone who checks the phone number, the hours, the claims, and whether the page answers the question it was written for.
5. If you work in a regulated field, or a client contract requires disclosure of AI assistance, go read that contract. That is a disclosure obligation, not a ranking question, and it deserves a straight answer.
6. When the detection API ships, treat any result as a probability, not a verdict. A mark means a model was likely involved somewhere. It does not tell you who wrote what.

The part worth remembering

A watermark reports that a machine may have touched some text. It says nothing about whether the text is accurate, useful, or worth showing to someone searching for a plumber in Morton.

The question was never whether a machine helped write the page. It is whether the page is worth the click.

That was the standard in July. It is the standard now. The tools changed. The bar did not.

If you want a second set of eyes on the pages you are worried about, Find What's Costing You. I will tell you which ones are actually weak and which ones are fine as they are.