How AI watermarks content

Every piece of text you generate with ChatGPT, Gemini or Claude can now carry a watermark you can't see. The three leading LLM providers have each landed on a slightly different answer to the same question: how do you watermark AI-generated content?

Google Gemini: SynthID watermarking

Google got there first. SynthID has watermarked Gemini text since 2024, and Google open-sourced the text version that same year. It works by nudging the probability scores the model assigns to each candidate word. Readers see nothing unusual; the pattern only shows up statistically, across many word choices.

The detection side has grown fast. On 7 October 2026 Google opened its SynthID Detector portal to everyone, and verification is now built into Search, the Gemini app and Chrome, handling over a million checks a day. Business users get an AI Content Detection API on Google Cloud.

OpenAI ChatGPT: textGrain watermarking

OpenAI held back the longest on text but on 5 October it announced textGrain, its own statistical watermark on word choice. It becomes the default for eligible ChatGPT and Codex output (in the EU first), on all plans.

OpenAI is unusually candid about the limits. Its own tests show that swapping one in four words drops detection to about 17%. The detector is restricted to approved researchers at launch. Images are handled separately, with C2PA Content Credentials and, since May, Google's SynthID.

Anthropic Claude: global watermarking by default

Anthropic took the broadest route. After signing the Code, it confirmed in August that Claude models launched from 2 August watermark their text at launch, with older models being updated. The mark applies worldwide, not just in the EU, across the Claude apps, the API, Claude Code and cloud platforms like AWS and Google Cloud. Users can't switch it off.

The technique resembles SynthID: Claude uses the many low-stakes word choices in a response to leave a pattern. Generated files get signed C2PA provenance metadata instead. A public detection API is still in development, and Anthropic stresses that a mark is a signal, not proof.

Note: Other large models such as Meta's Muse and DeepSeek's model also apply watermarking to AI-generated content in a similar way.

What watermarking AI content means for SEO and GEO

For brands publishing AI-written content, the honest answer is: nobody outside Google knows yet how much watermarks will count. Google's position remains that AI content isn't against its guidelines, and a watermark is a transparency measure, not a penalty. But the infrastructure now exists.

Google Search already runs SynthID checks, and answer engines that pick sources for AI Overviews or ChatGPT search could easily weigh provenance as one more quality signal. The bigger risk is not the mark itself but what it exposes: unedited, prompt-and-paste pages are exactly the content that keeps a clean watermark, and exactly the content search engines already struggle to reward.

Heavily edited text with original data, real expertise and a named author loses most of the signal and gains what actually ranks and gets cited. The practical move is the same one good content teams were already making: use AI for drafts and research, then rewrite enough that the final piece is genuinely yours, which is exactly what Graded Text is built to help with.

Sources: European Commission Code of Practice (Freevacy) · SynthID Detector rollout (Unite.AI) · OpenAI textGrain (VKTR) · OpenAI EU watermark (BleepingComputer) · Anthropic watermarking (Computing) · How Claude watermarks text (BleepingComputer) · Claude watermarking scope (Indian Printer Publisher)