As of this week, every text Claude writes carries an invisible watermark. You can't see it, it survives copy-paste, and nobody knows yet how much editing it takes to get rid of it.
On August 11, Anthropic announced that text from all new Claude models will carry a machine-readable watermark — worldwide, not just in the EU. It applies across Claude, the API, Claude Code, Claude Cowork and Claude Tag. For files like images, Anthropic uses the open C2PA standard. For plain text, the mark is embedded in the text itself.
Claude isn't alone here. Google has been marking Gemini output with SynthID for a while, and OpenAI is following. The direction is clear: unmarked AI text is becoming the exception.
How the watermark works
Language has slack. Many words can be swapped without changing the meaning:
- begin ↔ start
- quick ↔ fast
- help ↔ assist
That slack is exactly what the watermark uses. Wherever several words fit equally well, Claude doesn't pick randomly — it picks according to a secret key. A single word reveals nothing. But across hundreds of words, a pattern emerges that a detector can prove conclusively, while a reader never notices.
Three properties matter:
- It lives in the text itself, not in metadata. Copy, paste, change formats — the pattern travels with it.
- Light paraphrasing isn't reliably enough. Anthropic itself says the watermark may persist through editing. How much editing it would take is an open question.
- It only proves that AI generated the text. Not who the author is, and not how much human work went in. A heavily edited AI draft can still carry the pattern even though the substance is yours.
Why now: the EU AI Act
The timing is no coincidence. Since August 2, 2026, Article 50 of the EU AI Act applies: providers of generative AI must mark their outputs as artificially generated in a machine-readable way — text included. Systems already on the market have until December 2, 2026. Penalties reach €15 million or 3% of worldwide turnover.
Anthropic could have shipped this for the EU only, but is rolling it out globally. That's the real story: a European rule is currently setting the world standard for AI text.
For you as a company, a second part of Article 50 matters. If you publish AI text that informs the public on matters of public interest, you must disclose it. But there is a decisive exception: text that has undergone substantive human review, with a person holding editorial responsibility, is exempt. Remember that wording — it comes back below.
What this means for SEO
Until now, AI detection was guesswork. Detectors were often wrong, and Google could only estimate from quality signals whether a text came out of a machine. Watermarks turn the estimate into proof: the EU Code of Practice explicitly expects providers to offer detection mechanisms so third parties can verify whether content is AI-generated. A third party can also be a search engine.
Early data shows where this leads. An August 2026 analysis by First Page Sage (1,682 pieces of content, 139 B2B websites) found:
- Watermarked content ranked at an average of position 11 on Google versus position 6 for human content. That's the difference between page one and page two.
- The study didn't fully control for quality — but the direction is unmistakable.
For context: Google still says that using AI is not against its guidelines. What gets punished isn't the tool, it's low-effort content. The difference: Google used to have to guess. Now it can measure. Raw, unedited AI output becomes a measurable signal.
What this means for GEO
For visibility in AI answers — GEO — there are two effects, and the second one is bigger.
Effect 1: citation rates. In the same analysis, watermarked content was cited about half as often in AI answer surfaces (Google AI Overview, ChatGPT, Claude): 7% versus 12%. AI systems already appear to prefer sources that don't look like AI.
Effect 2: training data. AI companies have a strong self-interest in not training on their own models' output — it degrades the next model generation. Until now, AI content was hard to spot in training data. With watermarks, filtering it out becomes trivial.
Concretely: if your website consists mostly of raw AI text, it can be dropped entirely from the training data of future models. The next model then simply doesn't know your brand. No mentions, no recommendations, no citations. It's the quietest effect of all — and the one with the biggest long-term damage.
What to do now
The answer is not to throw AI out of your content process. The answer is a role change: AI as the tool, not the author.
- Substantive editing instead of copy-paste publishing. Exactly the wording of the EU exemption: human review, editorial responsibility. What you should do for quality anyway is now also the regulatorily safe side.
- Put your own substance in. Your data, project experience, examples, an opinion. None of that is in any model — and it's the part AI answers most like to cite.
- Make editorial responsibility visible. A named author, a review process, a last-updated date. Good for readers, good for E-E-A-T, good for Article 50.
- Measure how AI reads you. Which AI bots are reading your website, where you get cited and where you don't. Without measurement, every GEO discussion is gut feeling.
Producing content with AI has never been easier. Becoming visible with it has never been harder. If you want to know where your website stands in AI answers and how to get recommended there: that's exactly what I work on with B2B companies as a GEO agency in Munich.