Straight answer: there’s no magic percentage, but from tracking my own site and client sites for the best part of a year, pages that keep at least 30 percent pure human input, meaning rewritten openings, real examples, and facts you checked yourself, hold up in rankings and reader trust. Pages built almost entirely from an unedited AI draft tend to plateau or quietly slide within a few months. The ratio matters less than what that human third contains.
Why I started counting the percentage in the first place
I run this blog and I write client strategy documents, so I was using AI drafts for both long before most people admitted to it. Early last year I got lazy on a client’s blog. Six posts went up that were roughly 90 percent AI-generated, tidied for typos, nothing more. Traffic on those posts dropped 22 percent over six weeks compared to the posts before them. Not a Google penalty, nothing dramatic, just a slow bleed. Average time on page fell from around two minutes forty to under a minute. People were landing, skimming, and leaving.
Worth reading next: How Do You Test Your Website SEO? A Step-By-Step Process That Works in.
I rewrote all six. Same topics, same keywords, but I added named examples, specific numbers, and opening paragraphs that sounded like me rather than a generic consultant. Ten weeks later those six posts were outperforming the originals by 15 percent. Nothing else changed. That gap is where the real answer to this question lives.
The number Google cares about (it’s not a percentage)
Google has said publicly, more than once, that it doesn’t penalise content for being AI-assisted. What it penalises is content built to game rankings rather than help a reader, whatever wrote it. That distinction matters because it means the “how much AI is allowed” question is the wrong shape entirely. You can have a site that’s 80 percent AI-drafted and rank well, and a site that’s 100 percent human-typed and rank badly, because it’s thin, generic, and says nothing a reader couldn’t get from ten other sites saying the exact same thing.
I keep a local page for Dallas with the workflows that go first there: AI consultant in Dallas.
This is the part a lot of business owners find uncomfortable: the risk was never really the AI. The risk is that AI made it easy to produce content that sounds like everyone else’s content, and Google, and readers, notice sameness fast. I wrote about this in why small business websites all sound the same now, and the honest bit in that piece still holds: most of that sameness predates ChatGPT. AI just made the copying quicker.
The uncomfortable bit nobody wants to say out loud
Here’s the thing that annoys me about most posts on this topic. They tell you AI content is fine “as long as you edit it,” then never say what editing means, or admit that most people’s edit pass is a spell check and a swap of one adjective. A edited AI draft and a lightly skimmed AI draft look identical to a reader for about four seconds, and then they don’t. The lightly skimmed one has no specific numbers, no named client, no year that’s current, no sentence that could only have come from someone who’d done the thing. It reads fine. It just doesn’t stick.
The other uncomfortable truth: plenty of “100% human-written” content on small business blogs is worse than a well-edited AI draft, because it was written fast by someone who doesn’t enjoy writing and it shows. Purity of authorship isn’t the goal. Usefulness is. I’d rather publish a post that started as an AI outline and ended up 40 percent rewritten with real detail than a post typed entirely by a tired founder at 11pm that says nothing specific at all.
My actual rule, step by step
This is the process I now use on every piece that goes on this site, and the one I hand to clients when they ask me the same question you’re asking now:
- Let AI do the first structural draft. Research, outline, a rough pass. This can be close to 100 percent AI and it’s fine, nobody reads your outline.
- Rewrite the opening two paragraphs entirely yourself. This is where a reader decides whether to trust you, and it’s the part AI is worst at because it doesn’t know your actual week.
- Add one real example per major section: a client, a number, a date, a mistake you made. My rule is one specific detail for every 300 words minimum.
- Cut every sentence that starts with “in today’s fast-paced world,” “it’s important to note,” or “when it comes to.” These are AI’s verbal tics and readers now spot them instantly.
- Read the whole thing aloud. If it sounds like a brochure rather than a person, rewrite the paragraph, don’t just swap synonyms.
- Check the ending. AI-generated conclusions restate the intro. Replace it with an actual next step or opinion.
Follow those six steps and you can start from an 80 or 90 percent AI draft and end with something that reads as fully yours, because by the time you’re done, most of the sentences a reader remembers were rewritten by you.
What the AI detector tools won’t tell you
I get asked constantly whether tools like GPTZero or Originality.ai should be the test. They shouldn’t. I’ve run my own hand-typed paragraphs through them and had perfectly human writing flagged as 90 percent AI-generated, simply because it was clean and well-structured. These tools guess at pattern, not truth, and Google itself doesn’t use them to judge your site. Chasing a low AI-detection score is a waste of an afternoon. Chasing specific, checkable, useful content is not.
I learned this the hard way running an automation experiment I wrote up honestly in Week 17 Flop, where I let AI handle an entire outbound sequence hands-free. It read fine sentence by sentence and completely failed at the job, because nothing in it sounded like it came from a person who’d sent a cold email and got a reply. The same failure shows up in website copy. Fluent is not the same as useful.
Where this shows up across a whole website, not just the blog
The percentage question gets harder once you look past blog posts. Your homepage, your services pages, your about page, all get the same AI-draft-then-edit treatment from most businesses now, and it’s why so many small business sites read like the same template with the logo changed. A cold email has the exact same problem, which is why I wrote a separate piece on how to write a cold email that gets a reply: the AI-drafted version is always polite, always correctly formatted, and always ignored, because it has nothing specific to the person receiving it.
Your about page is the worst offender. I see AI-generated “founder stories” constantly, and they all share the same three beats: passion discovered early, gap noticed in the market, mission to help people like you. None of it is false, all of it is forgettable. That page should be closer to 90 percent your own words, because it’s the one page where sameness costs you the most trust.
A quick way to audit your own site this week
Pull up your last ten published pages. For each one, ask three questions: is there a number that couldn’t have come from a generic prompt, is there a name of a real person or client, and could a reader tell this was written by someone at your specific business rather than any business in your sector. If a page fails all three, that’s your 90-plus percent AI page with no edit pass, and it’s the one worth rewriting first, whatever traffic it currently gets, because it’s the page most likely to start slipping once a reader compares it to a competitor’s page that passed all three tests.
I do a version of this audit every few months and report what’s changing in the AI landscape in my weekly AI news roundups, partly because the tools change so fast that last year’s “safe” workflow needs checking again this year.
So what’s the actual number
If you want one figure to hold onto: aim for at least 30 percent of your final published word count to be sentences that only you could have written, meaning specific facts, named examples, opinions with a reason behind them. That’s not a Google requirement, it’s a reader-trust requirement, and in my own testing it’s the point where content stops reading as filler and starts reading as evidence you know what you’re talking about. Below that, on any topic with real competition, you’re gambling that nobody notices the sameness. Increasingly, they do.
Frequently asked questions
Will Google penalise my website for using AI-written content?
No, not for the fact of it being AI-written. Google has said repeatedly that its systems assess helpfulness and quality regardless of how content was produced. What gets penalised is thin, unhelpful, or duplicate-feeling content, which AI makes easier to produce at volume but doesn’t cause on its own.
What percentage of my website should be human written?
There’s no official figure, but based on tracking my own pages, keeping at least 30 percent of the final draft as human, meaning rewritten intros, real examples, and specific numbers, is where content stops plateauing and starts holding or growing its rankings.
Do I need to tell visitors my content used AI?
There’s no legal requirement in the UK or US to disclose AI-assisted drafting for standard website or blog content. The exception is anything making medical, financial, or legal claims, or content in regulated sectors, where disclosure and accuracy standards are stricter regardless of who or what wrote it.
How do I know if AI content is already hurting my traffic?
Check time on page and scroll depth for pages published in the last year. A sharp drop against your older, more manually written pages is the clearest early signal. If several recent pages show visitors leaving within 30 to 40 seconds, that’s usually a sign the draft never got the human rewrite pass it needed.
