Two numbers, from the same research, that sound like they can't both be true. Roughly half of new articles published on the internet are now AI-generated. And 86% of the articles that actually rank on Google are written by humans.
Both come from Graphite, which analyzed tens of thousands of web pages from the Common Crawl archive and ran them through AI detection. The publishing side is what you'd expect by now: machine-written articles have climbed to about half of everything new that goes up. The distribution side is the surprise. When Graphite checked what search actually surfaces, AI-generated articles took just 14% of Google rankings — and when they do rank, they rank lower. The AI assistants are even harsher on their own kind: 82% of the articles ChatGPT and Perplexity cite in their answers are human-written.
So the flood happened. And almost none of it is reaching anyone.
Publishing got free. Distribution didn't.
Here's the mechanism worth understanding. AI collapsed the cost of producing an article to nearly zero, so the volume of content exploded. But the number of spots on a search results page didn't change. The number of citations in an AI answer didn't change. The number of minutes your customer spends reading didn't change. Supply doubled; shelf space didn't. Which means the filters between what's published and what's seen — Google, the AI assistants, and ultimately the reader — got twice as aggressive, overnight.
And what those filters select against isn't AI. It's genericness. The pages getting filtered out are the ones that restate what a hundred other pages already say — no original data, no first-hand experience, no point of view. It just happens that unattended AI produces exactly that, at industrial scale.
The line isn't human vs. machine. It's grounded vs. generic — and generic is now half the internet.
What this kills, and what it rewards
The first casualty is the volume play. If your content plan for the fall is "use AI to publish five posts a week," you now know exactly where those posts go: into the 50% that the filters exist to remove. You'd be spending real hours adding to a pile that Google and ChatGPT are actively working to ignore. More output of the same quality is no longer a strategy. It's the thing the strategy has to beat.
What gets rewarded is the material only you have. Your actual numbers — what the campaign cost, what it returned. What your customers said, in their words. The mistake you made and what it taught you. The opinion you'll defend. None of that exists in any training data, which is precisely why the filters treat it as signal. A machine can't fake having run your business.
The practical version, in three moves
Feed before you ask. The difference between AI content that gets filtered and AI content that works is what goes in before the writing starts. Give the model your sales calls, your reviews, your real results, your actual stance — then let it draft. Same tool, opposite outcome, because now it's writing something that exists nowhere else.
Publish fewer, denser pieces. One article with a real number in it will outperform five that could have been written about any business in your category. If a competitor could publish your post unchanged, it isn't done.
Put the specific thing in the headline. The filters — and the readers — are skimming for evidence that a page contains something real. A concrete claim up front is how you signal it.
An obvious note, from an obvious messenger: this post was written by AI. It also cites its sources, contains actual research numbers, and takes a position — because someone pointed the tool at something real. That's the whole difference. The half of the internet getting ignored isn't ignored for being machine-made. It's ignored for being about nothing.