Machines Now Write the Content. Distributing It Is Still on You

Producing content is no longer an advantage: machines write roughly every second new article on the web, and everyone can do it. Distribution, though — landing in AI-search answers — is still on the human: among the articles Google puts in position 1, only 7% are machine. The winner isn't whoever produced more, but whoever got chosen. And that, unlike volume, is measurable.

Machines already write every second new article on the web. Looks like the human got squeezed out. But AI search still surfaces the human — and here's why.

How much machine content is really out there

Let's start with scale. No scare stories — measurements.

Graphite took random articles from the open web archive and ran them through detectors. Fresh data, through March 2026: in Q4 2025 machines wrote 50.9% of new web articles, in Q1 2026 — 49.9%. Roughly half. And it has been sitting at half for more than a year — a plateau since May 2024.

For contrast: in January 2020 it was 2.2%. Machines crossed the 50% line back in November 2024 and have simply held it since. So “AI is taking over the internet” isn't last week's news. The break happened a year and a half ago, and then it froze.

Now, about method — this matters. The latest version averages three detectors at once — Pangram, GPTZero, Copyleaks — not one. And still: these are detector estimates, not a signature on the text. A detector catches style, not authorship. Keep that in mind for the whole piece.

Next, the nuance the scare stories always hide. “Contains AI” and “written by AI” are two different things. Ahrefs checked 900,000 new pages (April 2025): 74% contain traces of AI text, but purely machine-written — only 2.5%. Everything else is a hybrid. Human and machine, four hands. That's why the “share of AI” numbers jump around so much: it depends on what you count.

And it's not just English. An academic paper from AWS (ACL 2024): 57% of sentences on the multilingual web have been machine-translated into three or more languages at once, and the more languages they were spread across, the worse the quality. Russian is in the sample. So part of the Russian-language web is machine exhaust too, not a live human.

It reached science. Stanford and Kobak, two separate studies: traces of AI editing show up in 13.5–17.5% of scientific abstracts. The tell is the vocabulary — the beloved “delve”, “intricate”. Careful: these are “traces of editing”, not “a robot wrote the whole thing”. But the trend is clear.

And there are outright farms. NewsGuard counts sites that churn out AI news on a conveyor. As of June 23, 2026 — 3,749 such sites in 16 languages. A year ago it was 1,121. It started with 49 in May 2023. It grows by hundreds a month.

The takeaway is simple. Content pours in from everywhere, in every language. Producing text today costs nothing. Which means the value in the act of producing is also — zero.

And here's where it gets interesting.

Distribution is still on the human

Half is produced. You'd expect half the search results to be machine too, right? Nope.

Same Graphite, June 2025, 31,493 search queries. Look at what Google actually puts in results. Among ranked articles, 14% are machine. The other 86% — humans. And in position 1, only 7% are machine. Human content takes the top spot eight times more often. And it's no coincidence: the result is statistically significant, odds of chance below one in a million.

Same story with AI bots. In ChatGPT and Perplexity citations, machine content is 18% each. Graphite's own words about these AI articles: they “largely do not appear in Google and ChatGPT”. Produced a mountain, and one sixth made it into the answer.

Separately — Google AI Overviews, the AI answers right inside search. Originality.AI checked 29,000 queries (October 2025): 10.4% of citations are machine, 74% humans. And a curious detail: 52% of AI Overviews citations come from beyond the first hundred ordinary results. So you can land in an AI answer without being in Google's top. That's a separate game, and few people play it.

Now let me kill the obvious objection. “If machine content doesn't rank, does that mean writing with a machine is harmful?” No. Remember the Ahrefs number: 74% of pages already contain AI text, and it's the hybrids that win in results. Machine wrote it, a human finished and distributed it. Google says plainly that it judges by usefulness, not by method of production. What loses isn't machine text as such. What loses is bare mass-production — generated and dumped, with no human on top.

There's the break. A mountain of machine content barely turns into visibility. Everyone learned to produce and buried everything. And what gets cited is still the human and the original. The advantage moved from the point of “producing” to the point of “being chosen”. And “being chosen” is, for now, human work.

The scarcity shifted. Not the text. Distribution.

But before we talk about what to do with this, let's clear out the trash. Because there's no less lying around the topic of AI content than there is AI content itself.

Half the numbers about AI are slop themselves

You've surely seen the figure “90% of the internet will be AI by 2026”. Pretty, viral, everyone loves to repost it. And it's not a measurement. It's a forecast. And originally — about something else entirely.

The first person to name the figure was Victor Riparbelli, CEO of Synthesia. The company makes synthetic video — so, a directly interested party. And he said it about video: up to 90% of video content could become synthetic in three to five years. Then Nina Schick's book relayed the figure. And then Europol, in its report, quietly widened “video” to “all online content by 2026”. One more step — and now it's pinned on “experts” and on Gartner.

See it? One vendor's forecast about video, inflated to “the whole internet”. Nobody measured this number. It simply has no author who actually counted it — only a chain where each hand added a little of its own.

And now — a live example, straight from under this article. While preparing the material, I went looking for the freshest number. First thing I found was something pretty: “per MIT and the Oxford Internet Institute — 64% of all content is AI, 8.3 billion AI articles”. Solid. I went to the primary source — and there is none. No such joint study, no such numbers. Just a link to an authority that never existed. That's exactly how it works: slop dressed up as science. A pretty number, a solid name, “all by the documents, of course” — and nothing underneath.

And since I'm on honesty. Even the real measurements I lean on here are detector estimates, not gospel. Plus the two main sources, Graphite and Originality, are commercial firms — the “humans win” framing suits them a little. That's why I also hold on to academia — AWS, Stanford — as an independent counterweight. Not “trust me”, but checkable.

In this noise, trust itself became the scarce thing. And whoever AI search trusts in that mess — gets the floor.

Why the original is now worth more

And this is no longer about marketing. It's about how AI itself is built.

There's a study in Nature (2024, Shumailov's group). If you train an AI model on its own synthetic output, over and over, you get “model collapse” — it loses diversity and quality, and the damage is irreversible. Careful: this is about training on synthetic data, not “the internet will break”. Let's not inflate it. But the consequence matters.

The more machine text there is on the web, the more expensive clean human data becomes for the AI labs themselves. The original and the primary source get literally more valuable, as a resource. Not in the “appreciate authors” sense, but the direct one — people will pay for live text.

And the platforms are moving the same way. Google spelled it out in its rules: it fights “scaled content abuse” — mass-stamping pages to rank. Doesn't matter whether a human or a machine wrote them. Judged by usefulness, not by method. Same thing again.

The market, the tech, and the platforms all move in one direction. From volume — to the value of the original.

What this means for your business

Let's boil it down to four steps.

One. Machines already write half the web. Producing content costs nothing and sets no one apart. Hand that part to the machine — there it's your friend.

Two. But distribution — landing in AI-search answers — is still on the human. 7% machine in position 1, 18% in bot citations. The hybrids win, where a human finished the text and distributed it.

Three. So the winner isn't whoever produced more, but whoever AI chose to cite. The scarcity isn't the text — it's distribution. And that, for now, is human work.

Four, and the main one. Citability, unlike “volume”, can be measured. With a score, before and after. And you can move it.

I tested it on myself. I rebuilt my own site for AI search and pushed citability from 25 to 64 in two weeks. Honestly: 64 is the ceiling for now, the 70–80 come as authority builds up, and that's not instant. But the shift is real and measurable, not “ran a set of activities”.

Where you start. The first thing worth checking — whether you land in AI answers at all, and what it says about you. From there it's clear what to fix.