In August, Anthropic confirmed that Claude now applies an invisible watermark to everything it produces: text, images and code. The trigger is a European Union law requiring AI-generated content to be labelled, which Anthropic has signed up to. Google is doing the same across Gemini, while OpenAI hasn’t confirmed its plans, though it will likely follow.
The detail that matters for marketers is how far the mark travels. It persists even when Claude only enhances content rather than writing it from scratch, so a paragraph you drafted yourself and ran through Claude for a tidy-up can still carry the watermark. Abongile Silonga, our search engine optimisation (SEO) expert, doesn’t see a crisis. The watermark isn’t the thing to worry about, Abongile says. The quality of the final page is.
How AI text watermarking works
A watermark isn’t a hidden signal tucked into the text. It’s a pattern in the words the AI picks. Each time an AI model writes, it chooses every word from a shortlist of good options, and a secret key quietly tilts those choices in a set direction. You won’t see a thing, but a detector holding the same key can spot the pattern. Copy and paste the text and the pattern comes along with it.
| Common assumption | What actually happens |
| Watermarks add hidden characters | The signal sits in word choice, so retyping doesn’t remove it |
| Anyone can check for one | Detection needs the provider’s secret key, and Claude’s detection tool is in private preview for approved organisations |
| A few edits wipe it out | Light edits weaken it, while heavy rewriting or translation breaks it |
| It proves a human didn’t write the piece | It can’t tell whether AI wrote the whole piece or a person rewrote an AI draft |
How Google Could Treat Watermarked Content
Nobody knows yet. Google hasn’t said whether it will use watermarks as a ranking signal, and a detection tool from Anthropic and Google is planned, which would make the marks far easier to find. Our view is that the risk is lower than the headlines suggest. Google has never banned content for involving AI. It rewards pages that show experience, expertise, authoritativeness and trustworthiness (E-E-A-T), and an expert-led article with a light AI polish still ticks those boxes. A generic page that repeats what’s already online was always going to struggle, watermark or not.
It’s also worth knowing that the watermark can be weakened by editing the content heavily or running it through a different model. That makes an unmarked page poor evidence of anything, because the publishers with the most reason to strip the mark are the ones mass-producing thin content.
The problem with labelling light AI edits
The fairer criticism is about scale. If proofreading triggers the same watermark as a fully generated article, original human work gets lumped in with machine output and risks losing value. Abongile has no issue with experts using AI to express what they know, since not every specialist is a copywriter. The line is clear, though: AI can improve how the expertise reads, but it can’t supply the expertise itself.
How Watermarking Affects Content and Ads
The practical impact depends on what the content is and who stands behind it. Paid media is already ahead of search here: Google Ads and Meta both prompt advertisers to declare AI-generated creative themselves, so disclosure is becoming a normal part of publishing.
| Content type | Watermark risk | What protects it |
| Expert-led articles with AI proofreading | Low | Named author, first-hand insight |
| Generic pages drafted entirely by AI | Higher, mainly because of quality | Original input and a clear point of view |
| Ad creative made or edited with AI | Disclosure required | Self-declaring in Google Ads and Meta |
The wider industry should expect a scramble. Detection tools will arrive, people will test how much editing defeats them, and the cycle of gaming search will repeat. Brands that invest in genuine expertise won’t need to join in.
How to Keep AI-Assisted Content Competitive
Treat AI as a tool, not as the source of expertise. Keep your subject matter experts involved from first idea to final edit, and make that expertise visible on the page. Four habits will keep you covered whatever Google decides:
- Put a named, credible author on every article, with a short bio that shows their experience.
- Add first-hand detail that no generic source holds, such as project results, client questions or original data.
- Use AI for structure, grammar and clarity, then challenge anything that reads as generic.
- Check accuracy before publishing, and declare AI-generated ad creative where the platform asks.
The watermark is a label. Usefulness is still what decides who ranks.
The last word
If you’re unsure where AI fits into your content workflow, or you’d like a second opinion on whether your pages show real expertise, get in touch with Herdl. We’re always happy to help you build content that earns its place in search.