Claude watermarks its text now. What it means for your site

Anthropic started hiding a mark inside the text Claude writes. The mark is a pattern in which words the model picks, and it survives a copy and paste into your website.

TechCrunch reported the plan on August 11. Anthropic published its technical explainer on August 14, and TechCrunch followed up the next day with more detail. Search Engine Journal and Search Engine Land covered it all week.

If you draft blog posts with Claude, that text now carries a fingerprint. I write with these tools every day, so I read the whole explainer.

How the watermark works

Anthropic changed how Claude flips its coins. When the model writes, several words often fit the same slot, like overcast or grey.

It still picks one at random. The company swapped out where that randomness comes from, so a secret key plus the words already on the page steer the pick.

Do that a few hundred times and the choices form a pattern. Someone with the key can read it. You can't, and neither can your reader.

I had this wrong at first. I pictured an invisible character wedged between letters, the kind of junk you strip out with find and replace.

Nothing gets added at all. The words themselves are the code, which is kind of crazy when you sit with it.

Their help doc says the mark covers all generated text across Anthropic's products, including the API and Claude Code. It travels with the text when someone pastes it somewhere else.

Will a watermark hurt your Google rankings?

Nobody has said, and honestly that's the whole answer right now. Search Engine Journal made the same point on August 14. No one has stated whether marked text gets treated any differently once Google crawls it.

Google signed the same EU code of practice that pushed Anthropic here. Google already runs SynthID checks on images, video, and audio. For text, there is no public checker.

Google's published line on AI content hasn't moved. It rewards pages that help the person reading them, whatever tool produced the draft. I'd plan around that instead of a rumor.

What the mark can and can't prove

A detection says Claude probably touched the text. It doesn't say who wrote your page, and it doesn't say how much of the page came from the model.

Four limits worth knowing:

• It needs length. Search Engine Journal reports the mark may not show up on text under roughly 200 tokens, so your titles and meta descriptions are basically invisible to it.

• It thins out on factual writing and on code, where the model has fewer optional words to play with.

• False positives run around 1 in 1,000, per that same writeup.

• Anthropic says light editing probably leaves the mark intact, while a rewrite that replaces every word wipes it.

Researchers have already stripped watermarks like this with better than 80% success for under $50, Search Engine Journal reported. Treat detection as a signal and nothing stronger.

What this changes for your blog

Rewrite the draft in your own words, which is what I'd tell you if none of this had happened.

Picture the homeowner searching "furnace making a clicking noise" at 9pm. She lands on your post. She reads two paragraphs of advice she already saw on four other sites, then she's gone.

She left because the page told her nothing she couldn't get anywhere else. No watermark involved in that.

Use the model for the boring parts, like outlines and a rambling first pass you'll clean up later.

Then put in what only you have. Tell them about the call you took last Tuesday, or what you charge for a Saturday visit.

Half the houses on that block run the same 1970s furnace, and you're the one who knows it.

That takes reps. It's the part you can't do in a day, and it's most of what we spend six months and more on for clients.

I still don't know whether Google will ever read these marks. I also don't know whether clients start running your posts through that detection API before they pay the invoice.

Draft with whatever helps you start. Publish the version you'd be willing to say out loud to the person with the clicking furnace.

Sources used

Anthropic, How Claude's text watermarking works: https://www.anthropic.com/news/claude-text-watermark

Anthropic Help Center, How Claude marks AI-generated content: https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content

TechCrunch, Anthropic shares more details about how Claude's new watermarks will work: https://techcrunch.com/2026/08/15/anthropic-shares-more-details-about-how-claudes-new-watermarks-will-work/

TechCrunch, Anthropic says it will watermark text generated by its AI models: https://techcrunch.com/2026/08/11/anthropic-says-it-will-watermark-text-generated-by-its-ai-models/

Search Engine Journal, What A Claude Watermark Can & Can't Tell You About Authorship: https://www.searchenginejournal.com/what-a-claude-watermark-can-cant-tell-you-about-authorship/585754/

Search Engine Journal, Anthropic Reveals What The Watermark Is And How It Can Be Defeated: https://www.searchenginejournal.com/anthropic-reveals-what-the-watermark-is-and-how-it-can-be-defeated/585930/

Search Engine Land (homepage scan, August 17, 2026): https://searchengineland.com

Hacker News front page (August 17, 2026): https://news.ycombinator.com