AI watermarks are readable in Google Search now. Here's what they do and don't do to your content

Two things became true this year that are constantly confused with each other. Google can now read provenance signals — SynthID watermarks and C2PA Content Credentials — directly in Search and Chrome. And Google's September spam update began rolling out on 24 September, with a rollout window of up to two weeks. Plenty of people are about to connect those facts and conclude that AI watermarks are costing them rankings. They are not. But provenance does now change something real about how you produce and publish brand media, and most content teams are quietly destroying the signal in their image pipeline without knowing it.

What actually shipped, and when

The correction: provenance is labelling, not ranking

Google reads provenance marks in order to tell a user how a piece of media was made. That is the entire function. There is no documented mechanism by which the presence of a SynthID or C2PA signal reduces a page's ranking, and the strongest available evidence against it is simply time: Google has been applying SynthID-text to Gemini's own output for around two years and has not used it as a ranking input.

What Google's spam systems act against is scaled content abuse — pages produced at volume that add little or nothing and exist to manipulate rankings. The distinguishing feature of sites caught by that policy is not the tool used to write them. It is the absence of anything useful on the page, produced fast, at scale, against a keyword list. A watermark is a fingerprint on the output; the policy is about what the output is for.

This matters for a practical reason beyond reassurance. If you spend the next fortnight trying to strip watermarks from your text, you will have spent effort on a non-problem while the actual cause of any traffic change — a two-week spam update rollout, thin pages, a crawler configuration, seasonality — goes undiagnosed. The honest version of this advice has not changed in three years: the question Google is answering is whether the page is useful, original and trustworthy, and our guide to E-E-A-T for business websites covers what that means in practice.

Signal by asset type: who embeds, who reads, what it triggers

The word "watermark" is doing too much work in most coverage. Four different things are in play and they behave differently.

Two conclusions follow. First, the text watermark that caused the most alarm is the one with the least practical consequence: it is not read by Search, and it degrades under exactly the editing any competent content process performs anyway. Second, the signals that do have visible consequences are on media — images, video, audio — which is precisely the category most marketing teams never think about in provenance terms.

Where your stack destroys the signal

This is the section that should change what somebody does tomorrow. C2PA Content Credentials are signed metadata attached to a file. Almost every step of a normal publishing pipeline is capable of discarding metadata, and most do it by default in the name of file size. If you care about being able to demonstrate that your own product photography is real, this is the work.

1. CMS upload. Many platforms re-encode on upload and generate derivatives. Check whether the original with its manifest is retained anywhere, or whether only stripped derivatives survive.
2. Image optimisation and format conversion. Automatic WebP and AVIF conversion is the single most common point of loss. The converted file is a new file, and unless the tool is provenance-aware, the credentials do not come with it.
3. CDN transforms. On-the-fly resizing, cropping and recompression produce new derivatives. The same logic applies.
4. Build-time pipelines. Image plugins in static site generators and bundlers routinely strip all metadata as an optimisation, and nobody reviews the default.
5. Social and marketplace re-encoding. Platforms re-encode uploads for their own delivery. Assume credentials do not survive unless the platform states otherwise.
6. Screenshots and re-exports. A screenshot discards metadata entirely. This is also why watermarks that survive visual transformation and metadata that carries richer context are complementary rather than redundant.

The practical protocol is short: keep signed originals in your asset library rather than only web derivatives; decide deliberately whether public derivatives should carry credentials; verify one real page after any change to your image pipeline instead of trusting the configuration; and treat this as part of the same maintenance surface as the rest of your delivery stack. Guidance around the Google rollout says the same thing more bluntly: attach Content Credentials to original media, then confirm your CMS or CDN is not stripping the signed metadata on upload.

A disclosure policy you can actually apply

Provenance visibility raises a question that is separate from compliance: what do you tell your audience? Here is a defensible position by asset use, with the obligations separated from the judgement calls.

The line that holds up across all six rows: disclose when the medium itself could mislead, and stay accountable when the substance is what matters. That is the same standard our notes on platform rules on low-effort AI content and on AI avatars and brand trust arrive at from different directions, and it is the one that survives the next round of platform policy changes because it is not tied to any specific one.

Frequently asked questions

Does an AI watermark hurt your Google rankings?

No. Provenance signals such as SynthID and C2PA Content Credentials are used to label how media was created, not to rank pages. Google has applied SynthID-text to Gemini's own text output for around two years without using it as a ranking input. What Google's spam systems act on is scaled content abuse, meaning pages produced at volume that add little value and exist to manipulate rankings, which is a judgement about the content rather than about the tool that produced it.

What does Google actually do with SynthID and C2PA signals?

It reads them to answer whether a piece of media is AI-generated or AI-edited, and surfaces that answer to users. Following announcements at Google I/O in May 2026, C2PA verification and SynthID detection run natively in Google Search and Chrome, and users can ask whether an image is AI-generated through Lens, AI Mode and Circle to Search. Where an image carries a C2PA manifest indicating AI generation or AI-assisted editing, that is treated as strong evidence for applying an AI label.

Does Claude's text watermark affect SEO?

There is no evidence that it does. Anthropic announced on 11 August 2026 that all new Claude models embed an invisible machine-readable watermark in generated text, primarily in response to marking requirements such as Article 50 of the EU AI Act, which took effect on 2 August 2026. Text watermarks are not read by Google Search as a ranking factor, and they degrade under the ordinary editing, restructuring and merging that any real editorial process performs.

Why do Content Credentials disappear from images on my website?

Because Content Credentials are signed metadata attached to a file, and most publishing pipelines discard metadata to save bytes. The common loss points are re-encoding on CMS upload, automatic WebP or AVIF conversion, on-the-fly CDN resizing and cropping, build-time image plugins that strip metadata by default, and re-encoding by social platforms. A screenshot removes it entirely. Keep signed originals in your asset library and verify a live page after any change to the image pipeline.

Do I have to disclose that I used AI to write an article?

There is no search-engine requirement to disclose AI assistance in writing, and Google's position is about whether a page is useful, original and trustworthy rather than how it was typed. Disclose the author instead: a named person who researched, edited and is accountable for the piece. Disclosure becomes important where the medium itself could mislead, such as synthetic voice, an AI presenter, or imagery a customer would rely on to judge a product.

Should brands add C2PA Content Credentials to their own photography?

It is worth doing for assets whose authenticity carries commercial weight, such as product photography, documentary and case-study imagery, and anything that might be disputed later. The value is being able to demonstrate provenance rather than any ranking benefit. The prerequisite is a pipeline that preserves the credentials from capture through to the file your site actually serves, which is where most implementations currently fail.

The takeaway

Provenance became real infrastructure this year: marking is effectively mandatory for model providers, the major image and text generators all embed something, and Google reads those signals in Search and Chrome. None of that is a ranking mechanism, and treating it as one will send you chasing a problem that does not exist while a two-week spam update rollout and your own thin pages go unexamined. The work that does pay is unglamorous: know which of your assets carry credentials, find the step in your pipeline that strips them, keep signed originals, and hold a disclosure position based on whether the medium could mislead rather than on which tool was open at the time.