For most of my career I was anti-AI. Not curious about it, not cautiously testing it. Against it. Somewhere over the last year that flipped completely, and I now build systems around AI most weeks – custom GPTs, structured briefing workflows, a full content pipeline that starts with research and ends with a published article. I’m still the one doing SEO. AI just does more of the mechanical part than it used to.
That flip happened slowly enough that I didn’t notice it until Anthropic announced, in August 2026, that Claude would start embedding invisible watermarks in every piece of text it generates. Google’s Gemini has done something similar with images for a while – the small badge in the corner, the metadata underneath. Instagram rolled out its own “Made with AI” label across organic posts and ads earlier this year. None of this is new in spirit. What’s new is that it’s arriving for writing specifically, and writing is the one output most SEOs and marketers still don’t want anyone to know AI touched.
I had that conversation with a colleague recently. The reservation wasn’t really about Google. It was about being replaced – about handing over the first draft to a machine and wondering what’s left for the writer to do. I think that fear is proportional to how bad your system is. If your process is a bare ChatGPT window and a prompt, the fear is justified. If it’s something you’ve actually built, the watermark doesn’t change anything, because you were never trying to hide the AI. You were trying to make sure what it produced was worth publishing.
What the watermarks actually do – and what they don’t
Worth being precise here, because most of what’s circulating about this is wrong.
Anthropic’s watermark is a version of Google DeepMind’s SynthID-Text method. It doesn’t add hidden characters or extra tokens to the text. It changes which word Claude picks when there are multiple equally good options – the difference between “overcast” and “grey” in a sentence where either works. Do that enough times across a long passage and you get a statistical pattern that’s undetectable to a reader but checkable against Anthropic’s key. Short passages carry almost no signal. Factual sentences carry almost none either, because there’s rarely a choice to nudge – “Isaac Newton’s most famous work was called Principia” doesn’t have an equally good alternative word to swap in.
The watermark can’t identify you, your organization, or your specific chat. It can only answer one question: was Claude likely involved in producing this text, and how much of it did Claude actually write versus lightly edit. Anthropic is implementing this because it signed, along with roughly 190 other organizations, the EU’s Code of Practice on Transparency of AI-Generated Content, which took effect on 2 August 2026 under the EU AI Act.
Here’s the part most people miss: OpenAI signed the same code of practice and has not shipped text watermarking in ChatGPT. It ships SynthID for images and audio. For text, it’s held back, citing concerns about false positives and the fact that watermarks can be stripped by a heavy enough rewrite. So the premise that “all the AI tools are now watermarking your writing” isn’t accurate. One major provider is. The others are watching.
None of this changes whether your content ranks. It changes whether someone with the right key could, in principle, check.
AI as a junior writer, not an author
I’ve started describing a well-built AI tool to colleagues as a junior content writer. Not because it’s simple, but because that’s the correct working relationship. A junior writer gives you a genuinely useful first draft. Sometimes it’s sharp. Sometimes it gets facts wrong, or misses the point of the brief entirely. Either way, nobody hands a junior writer’s copy straight to a client. It goes through an editor first.
The mistake most people make isn’t distrust of AI. It’s the opposite – they think “editing” means a quick read-through for typos, when it needs to be closer to rewriting. A junior writer’s draft gets restructured, fact-checked, and rewritten in places before it goes anywhere near a client. Treat an AI draft with the same discipline you’d give a junior writer’s copy, and it works. Treat it like something written by a senior practitioner who doesn’t need supervision, and it falls over – not because the AI is bad, but because you skipped the part of the job that was always yours.
That editorial layer is not optional and it is not fast. It’s where the actual SEO happens.
Why a bare chat window fails
Open a fresh ChatGPT or Claude session and ask it to write you an SEO article. It’ll produce something. It might even read fine on the surface. But it knows nothing about the client, the site, the audience, or the competitive position – it’s pattern-matching against general training data, not against anything true about the business you’re writing for.
I’ve watched this fail in a specific, repeatable way. A brief goes into a bare chat with no site context, no fact-checking layer, and no guardrails. The output can’t be grounded in anything real because there’s nothing real feeding it. It reads generically because it is generic. And there’s a second problem that gets less attention: if memory is switched on, whatever tone, client context, or instruction got used in a previous conversation in that same chat can bleed into the next one. Anyone else touching that chat – a colleague, a different brief, a different client – can quietly shift the voice and content of work that has nothing to do with them. That’s not a minor inconvenience. It’s an unsafe way to produce anything you’re putting a client’s name on.
Once a proper system replaces the bare chat, the difference in the first draft is immediate. It’s more accurate. More on point. Genuinely usable as a starting point rather than something you’d have to gut and rewrite. And that changes what you spend your time doing next – instead of fixing a draft that never had a foundation, you’re improving one that already stood on solid ground.
The system that actually works
Here’s what I actually built, because “have a good process” is a slogan until you can see the mechanics.
I keep every client, and my own site, in an isolated Claude or ChatGPT project rather than one shared chat. Each project holds a fixed set of source files: who the client is, how they write, their products and audience, and what they will not say. For deeper context I run a Screaming Frog crawl of the site, pull the full page copy through a custom export, then run it through Gemini for semantic embedding – which surfaces how pages actually relate to each other, not just what a sitemap says they should. That becomes the site’s content map: what the business offers, what it doesn’t, and where the internal linking opportunities actually sit.
Before any of that goes into a project, I run it through a structured interview – a long prompt built specifically to extract a client’s business identity, audience, competitive position, voice, and hard nos, one question at a time, refusing vague answers. If someone describes their tone as “professional but approachable,” the process pushes back: give me a sentence from your content that proves it, or write one that sounds right. That interview alone runs to roughly 40-plus questions and takes the better part of ninety minutes to work through properly. It’s tedious. It’s also the difference between a system that produces something usable and one that produces something that sounds like every other SEO article written this year.
The writing itself runs on a similar principle. I built what I call a ghostwriter skill, grounded in my own experience, case studies, and the words I do and don’t use. It doesn’t write on its own. It interviews me in two stages. The first round asks why I’m writing about this topic now, what the intent is, and what I actually want to say – and it won’t move forward until I’ve given it something substantial. Once I approve the resulting outline, it interviews me again, this time pulling out my actual opinions, results, and experience on the specific subject: have you seen this happen with a client, what’s the anecdote, what would you tell someone who tried this and it didn’t work. I generally answer that second round by dictating – raw, unedited, in my own words – because that’s what the system is built to work from. Without a brief already prepared, the whole process runs to eight stages. With one, it’s four. Either way, nothing gets published until I’ve supplied something a competitor’s writer, or a model trained on the current search results, genuinely couldn’t produce on its own.G
Skip those steps and the output is slop – technically fine, structurally sound, and indistinguishable from everything else already ranking for the same query. Run them properly and the first draft is something I can actually build on.
Slower is the point
The industry’s current obsession is speed. Every tool pitch is some version of “generate a full article in twenty minutes.” I think that’s the wrong goal entirely.
If the twenty-minute article is genuinely as good as one that took three days, fine – use the twenty-minute version. But that’s rarely the comparison actually being made. What’s usually being compared is a twenty-minute AI draft against a mediocre three-day manual one, and the AI wins by default because the bar was low. The real question is what happens when you take the time AI saves you on the mechanical part of writing and reinvest it into the parts that actually differentiate the piece – sourcing better data, finding the angle nobody else has covered, going back and rewriting the sections that don’t earn their place.
That’s the trade I make deliberately. The system produces a genuinely usable first draft in a day, sometimes less. I then spend the next two days on it – not fixing what’s broken, but making it better than it needed to be. More research. Sharper angles. A second and third editing pass. The total time doesn’t necessarily go down. What changes is where that time gets spent – less of it wasted wrestling a blank page or a broken draft into shape, more of it spent on the work that actually makes a piece worth reading.
I’ll be direct about why I built this for myself specifically. I’m not a natural writer. I can talk for an hour about a topic and never once feel stuck, but sitting down to write the same thing from scratch is a different skill, and it’s not one I have. Without this system, I wouldn’t produce content on my own site at all – not because I lack the thinking, but because turning that thinking into finished prose was always the bottleneck. The tool didn’t replace my thinking. It removed the bottleneck between having something to say and actually saying it.
What this means if you write for a living
If you’re a content writer reading this and wondering whether AI is coming for your job, here’s the honest mechanic, not the reassurance.
The work is shifting from drafting to editorial judgment. Knowing what’s true, what’s missing, what sounds wrong, and what a reader actually needs – none of that goes away when AI writes the first pass. If anything it becomes the entire job, because the drafting used to eat the time that judgment now needs. That’s a different skill from sentence-level writing, not a lesser one. The writers who struggle with this shift are usually the ones who were relying on the drafting itself to prove their value. The ones who’ll do fine are the ones whose value was always in knowing what good looks like – they just used to spend most of their day producing it word by word instead of shaping it.
Google’s own position, going back to a February 2023 blog post that still holds, is that using automation to generate content purely to manipulate rankings violates its spam policies – but that not all automation is spam. The March 2026 core update leaned harder into this than anything before it, and sites publishing large volumes of AI-generated pages with no editorial layer saw traffic drops of 50 to 80 percent, in some cases losing their entire blog catalogue overnight. None of that is really about AI. It’s about content with little effort, little originality, and little value behind it – which was always going to get punished eventually, AI or not.
So the question was never whether to use AI to write. It’s whether you’ve built something around it that makes the output worth someone’s time to read. Most people haven’t, because building it is slower than not bothering. That’s exactly why it’s worth doing.
FAQ
Does ChatGPT watermark text the same way Claude does? No. Anthropic began embedding invisible watermarks in Claude’s text output in August 2026, using a method based on Google DeepMind’s SynthID-Text. OpenAI signed the same EU Code of Practice but has not deployed text watermarking in ChatGPT, citing concerns about false positives and the ease of stripping a watermark through heavy rewriting. OpenAI does use SynthID for images and audio generated through ChatGPT.
Can someone remove an AI watermark by editing the text? Light editing usually won’t strip it completely, but a thorough enough rewrite – where nearly every word choice changes – will remove most or all of the signal. At that point it’s fair to ask whether the text is still meaningfully AI-generated at all, since the watermark only tracks word choices Claude actually made.
Is AI-generated content bad for SEO? Not inherently. Ahrefs’ analysis of 600,000 pages found 86.5% of top-ranking pages contained some AI input, and the correlation between the percentage of AI-generated content and ranking position was statistically negligible. What gets penalized is thin, low-effort content, whether a person or a model wrote it – and Google’s March 2026 core update specifically targeted large volumes of AI content published without editorial review.
What’s the difference between using AI to write and using AI to edit? Using AI to write means it produces the first draft, which then needs the same scrutiny you’d give a junior writer’s copy – fact-checking, restructuring, and a genuine rewrite pass where needed. Using AI to edit means a human wrote the original and AI is proofreading or lightly polishing it. The two produce very different watermark signals, and very different levels of risk if you skip the review step.
How do I know if my content sounds like AI wrote it, even after editing? Look for uniform sentence length, hedging without purpose (“may,” “might,” “could potentially”), vague attribution (“experts say”), and generic openings that could apply to any business in the category. If a paragraph could run on a competitor’s site with the names swapped, it hasn’t earned its place yet.
Do I need to disclose that I used AI to write my content? Google doesn’t require a disclosure for AI-assisted content the way Instagram now requires a “Made with AI” label on certain posts. What matters for search is whether the content is genuinely useful, not how it was produced. That said, transparency tends to build more trust than it costs – hiding AI involvement only becomes a liability if the content itself isn’t good enough to stand on its own.
