Your AI content is cooked

Want to automate the grunt work?

Meet Agent A, the AI teammate with full access to your Ahrefs data. From fixing keyword cannibalization to shipping technical reports straight into Notion or Google Docs, it does the work you’d rather not.

A few days ago, Anthropic published a support article titled "How Claude marks AI-generated content." Oh boy, AI content creators are being cooked

Starting August 2nd 2026, every new Claude model will watermark what it writes. You won’t even be able to catch it.

The mark is woven into the text itself, and it travels with the words when someone copies and pastes them somewhere else. It survives a fair amount of editing too, though not all of it.

For images and files, Claude attaches signed C2PA metadata - it’s the same open provenance standard the rest of the industry is converging on.

Google's doing something similar with SynthID. OpenAI is using cryptographic metadata - Meta, Microsoft, Black Forest Labs- they're all in on this now.

The reason is the EU AI Act - it came to the real world on August 2nd, 2026, and it requires AI companies to mark generated content in ways that can be detected later.

AI content was already getting easier to spot. Not X but Y, Not only X but also Y, the rhythm, the same three-sentence-paragraph structure everywhere. People were already more and more capable of smelling it.

Now there's a technical layer sitting underneath the writing style and it is directly embedded in it.

There are two takes I keep seeing, and I think they're both a little off.

One camp thinks Google will now detect AI content and kill it.

The other thinks watermarks don't matter at all because nobody can actually read them and Google has said for a long time now that AI content isn't penalized just for being AI.

The second camp is closer to correct, but only technically. Google's official position, going back to their 2023 guidance, is that they reward quality content however it's produced.

Using automation with the primary purpose of manipulating rankings violates their spam policies. Using AI to help you write something genuinely useful does not. That's still the stated policy.

Google has never confirmed they use watermarks as a ranking signal, and there's no evidence they're scanning for SynthID or Claude's watermark and docking pages for it.

Anthropic themselves are careful about this too. Their own documentation says a detected mark means content "may have been processed by Claude,".

What is real, though, is a study First Page Sage published literally 2 days ago. They found watermarked content ranking about five positions lower on average than non-watermarked content, often the difference between page one and page two.

On AI answer engines, unwatermarked sources got cited almost twice as often. That's a real gap, and I don't want to wave it away.

Human-written or heavily human-edited content tends to get more editorial care regardless of whether it's watermarked. More fact-checking, better structure, more actual original thinking.

The ranking gap might have nothing to do with the mark itself and everything to do with what the mark tends to correlate with, which is content that skipped the editing step.

For the last two years, the working theory for a lot of AI-assisted content was simple: generate it, lightly skim or edit it, publish it, and unless it read really badly, nobody would know for sure.

I hated it because it’s not true and it actually worked like shit - not what I want to do for long-term SEO.

Readers noticed AI content, editors already knew. Search quality raters are not clueless. But there was enough plausible deniability in the system that it was easy to convince yourself the risk was low.

Watermarking AI content makes it easier to detect. That’s it.

An actual signal in the word choices that survives copy and paste is a very different thing than a stylistic tell a sharp reader might or might not catch.

I think this is the actual story, more than any ranking algorithm. Substack partnering with Pangram to flag AI-generated newsletters. C2PA metadata becoming an industry default on images. Every major model now shipping some form of provenance signal. None of these individually change how Google ranks a page.

Together they change the assumption that AI Content is undetectable that was the foundation of many content strategies. AI content is traceable by default now.

So what do you do with this, practically, if you create content for your business or your clients?

You don’t need to throw out Claude - it’s still good for research, pulling out arguments, etc. Don’t use it as an easy “click to generate content” button. Write your own words. Yeah, I know scary shit.

It’s not only because Google could look for a watermark but because the entire internet is looking for them - it’s getting easier to tell the difference between what was AI-supported and AI-generated.

EEAT is still with us in SEO, and it didn’t change. Your original data, expertise and POV matter a ton, and you can AI-generate it.

Watermarking removes the shortcut where you could publish unedited AI output and reasonably assume nobody would ever be able to prove it. Don’t be cringe.

For technical SEO specifically, I'd treat this as a prompt to actually audit your content workflow rather than a reason to panic about rankings.

Good content was never at risk. If you write from your own experience, in your own voice. Just blabbering on a piece of paper but in your own way, you will do well. Just structure it well for SEO, you know.

The internet is catching up to finding thin, generic, poopy AI-generated content that bored readers weren’t even reading.

It’s time to stop pretending you write quality content - go and actually write it.

Reply to this email with “Content” and your domain.

I am selecting ONE site to do a free SEO analysis including content structure. Let me show you how to do content better.