--- url: 'https://seoafterai.com/news/claudes-text-watermark-and-google-rankings-what-is-known/' title: 'Claude’s text watermark and Google rankings: what is known' author: name: Nadia Sorensen url: 'https://seoafterai.com/author/n-sorensen/' date: '2026-09-28T13:00:14+00:00' modified: '2026-09-28T13:00:14+00:00' type: post summary: 'Anthropic now marks text that Claude writes. Within days, posts on LinkedIn and X said Google would stop ranking such pages. Nothing published by Google or Anthropic supports that claim.' categories: - SEO News tags: - Gemini - Spam policies image: 'https://seoafterai.com/wp-content/uploads/2026/09/cover-250-dc662abf.png' published: true --- # Claude’s text watermark and Google rankings: what is known Anthropic now marks text that Claude writes. Within days, posts on LinkedIn and X said Google would stop ranking such pages. Nothing published by Google or Anthropic supports that claim. ## Table of contents - [What this is about](#what-this-is-about) - [The facts](#the-facts) - [What we can and cannot say yet](#what-we-can-and-cannot-say-yet) - [What to watch](#what-to-watch) - [FAQ](#faq) **Key takeaways** - Anthropic’s mark shows only that Claude was likely involved at some point. By Anthropic’s own account, it cannot separate a Claude draft from a human draft that Claude edited. - The mark exists to meet the EU Code of Practice on Transparency of AI-Generated Content, signed in July 2026. Google Search is not a party to it. - Google DeepMind has had a similar method, SynthID-Text, since 2024. Google has never said it uses it for ranking. - The measured traffic risk sits elsewhere. In Lily Ray’s review of 200+ templated sites, more than half lost 30% or more of peak organic traffic. ## What this is about On August 14, 2026, Anthropic described how it watermarks Claude’s text. Claims soon spread that Google would demote pages written with Claude. This piece sets out what the documents say, what has been measured, and what is still guesswork. It concerns anyone who publishes AI-assisted content. ## The facts **How the mark works.** Anthropic’s August 14 post explains the method. When the model picks between two equally good words, a random number normally settles the choice. With watermarking, a secret key settles it instead. No characters are added, and speed and cost stay the same. The key can later detect the pattern of choices. **What it can and cannot tell.** Per Anthropic, detection means Claude was “likely involved”. It does not mean a human was absent. Short texts are hard to detect, and confidence grows with length. Where a single exact answer exists, as in code or in a plain fact, there is little room for the mark. **Why it exists.** Anthropic and about 190 other signatories joined the EU Code of Practice in July 2026. It requires marking AI-generated text. Anthropic applies the mark worldwide, since it cannot yet limit it by region. Models released before August 2 get it over the coming months. **Google’s side.** Google DeepMind published SynthID-Text in *Nature* in 2024 and has used it on Gemini since. Google Search Central’s guidance of February 8, 2023 says Search rewards quality, however the content was made. Google has announced nothing in response to the EU code. **Quality.** Google DeepMind tested a watermarked model on part of live Gemini traffic. User ratings showed no statistically significant difference. Human raters comparing answers side by side also saw none. **The risk that has been measured.** Seer Interactive, in a recent post by a team of nine, points to Lily Ray’s study of sites that scaled content by copying a template: | Share of 200+ templated sites | Loss of peak organic traffic | | --- | --- | | More than half | 30% or more | | Nearly 40% | 50% or more | | One in five | 75% or more | Seer’s post gives no dates for these losses. ## What we can and cannot say yet We can say that no public document links the mark to rankings. Google has held a similar key for its own model for two years, and its stated policy has not changed. We cannot say what Google’s systems do internally. Nobody outside Google can. A claim of “Claude pages are being demoted” would need a before-and-after comparison on matched pages. So far no one has published one. Detectors are a separate problem. Seer’s Alex Ramadan warns that false positives will flag Claude where it was never used. A Stanford team (Liang and co-authors, *Patterns*, 2023) tested seven detectors. They labelled more than half of TOEFL essays by non-native English writers as AI-written, and almost never did so for native writers. A detector score is a weak basis for accusing a brand of anything. We ran a detector on our own articles and [wrote up what we found](https://seoafterai.com/analytics-data/we-ran-the-detector-on-our-own-12-articles-nine-of-them-ai-written/). ## What to watch Three signals would change this picture: 1. A statement from Google Search Central that names watermarks or the EU code. Today there is none. 2. A published test comparing traffic for matched pages with and without the mark over several weeks. 3. The pattern after the next core or spam update. If templated sites fall again and mixed human–AI sites hold, the mark is beside the point. Our notes on [the August spam update](https://seoafterai.com/news/what-the-community-is-reporting-about-googles-august-spam-update/) show how thin early reports usually are. For now the practical step is plain: check your templated pages first, not your AI vendor. ## FAQ Will Google stop ranking content written with Claude? There is no evidence of that. Google has said nothing about the mark, and its guidance judges quality, not the tool. Its own model has carried a similar mark since 2024. Does a light Claude edit on a human draft carry the mark? Often very little. The mark attaches only to words Claude chooses. If most words stay human, there may be too few for detection, especially in a short text.