Graphite's analysis of Common Crawl puts the share of newly published English articles that are primarily AI-generated at about parity with human-written ones — a share that climbed from near zero before ChatGPT to roughly 50% by late 2024, then plateaued.
2026-08-21 · by Moe Ameen
The SEO firm Graphite published research measuring how much of the open web is now written by AI, and the widely cited number is stark: the share of newly published English-language articles that are primarily AI-generated has risen to roughly parity with human-written ones. In Graphite's first analysis, released in October 2025, machine-written articles crossed the 50% line for the first time around November 2024, edging past human output — then the trend flattened rather than continuing to climb.
The change tracks almost exactly to ChatGPT, which OpenAI launched on November 30, 2022. Graphite reported that articles published before the launch registered only a low, false-positive-level rate of AI detection; within about a year of ChatGPT's release the AI share had jumped to roughly 39%, and it kept rising until it hit parity in late 2024. A follow-up analysis extending the data into early 2026 used three separate AI detectors — Pangram, Copyleaks, and GPTZero — averaged together across roughly 55,000 Common Crawl articles, and landed on a more conservative figure: AI-written content peaked around 50.9% before settling near 49.9% in the first quarter of 2026, about three percentage points below the single-detector estimate. Both studies classified an article as AI-generated when more than half its text read as machine-written.
The methodology matters, and Graphite is candid about its limits. The samples are drawn from Common Crawl, a broad crawl of the public web; AI detectors are probabilistic, not perfect, which is why the follow-up averaged three of them to cut error. The takeaway everyone repeats — "half the internet is AI" — is really "about half of newly published articles in this sample are primarily AI-written," which is a narrower and more defensible claim.
The finding that actually matters for anyone publishing is the one that got less attention. Graphite noted that these AI-generated articles largely do not show up in Google search results or in ChatGPT's answers. In other words, the web filled up with AI text, but that text mostly isn't the content winning distribution. Volume surged; visibility didn't follow it.
The instinct this study tends to trigger — "everyone else is mass-producing AI articles, so I need to out-produce them" — is exactly the losing move. Graphite's data says the web is already saturated with AI text and that most of it wins no search traffic. Winning more of the same commodity harder is not a strategy. The opening the numbers actually reveal is that AI text is where the flood is, and video, carousels, avatar content, and native social posts are not — those are the formats where a distinct, on-brand asset still cuts through. That is the half of the map [Kompozy](/) is built for. From a single source, it generates the content the study didn't count: captioned [Persona Shorts](/glossary/persona-shorts) narrated by your AI Influencer persona, brand-exact [Carousel Posts](/glossary/output-buckets) rendered through [HyperFrames](/glossary/hyperframes), quote graphics, photo posts, plus the text formats — but as one voice across all of them, not another anonymous article.
The second half of the finding — that AI content mostly isn't the content getting distributed — is a distribution problem, and it's the one Kompozy is designed to solve. Generating a post is the cheap part now; getting it in front of an audience is the constraint. Kompozy's per-post review gate and [Autopilot](/glossary/autopilot) schedule and publish that whole set across the eight social platforms (Instagram, Facebook, TikTok, YouTube, LinkedIn, X, Pinterest, Threads) plus blog and email — where your actual audience is, instead of into the article graveyard the study measured. The lesson from Graphite isn't "stop using AI." It's "stop shipping the undifferentiated thing the whole web is already drowning in." Use AI to produce a recognizable, multi-format body of work under one governed voice, and route it to owned and social channels — reach you control rather than search luck you don't.
Graphite's analysis of Common Crawl articles found that primarily AI-generated content is at roughly parity with human-written content among newly published English-language articles — its follow-up study put the Q1 2026 figure near 49.9%. That is best read as "about half of new articles in this sample are primarily AI-written," not "half of everything on the internet." An article was counted as AI-generated when more than half its text read as machine-written.
Graphite's first study found the share of primarily AI-generated articles crossed 50% around November 2024, edging past human-written output for the first time, then plateaued rather than continuing to climb. Before ChatGPT launched on November 30, 2022, AI detection on published articles registered only a low, false-positive-level rate.
Largely not, according to Graphite. The firm noted that despite AI content's prevalence, these articles mostly do not appear in Google search results or ChatGPT answers. So the web filled with AI text, but that text is not what's winning distribution — volume and visibility came apart. That is the practical warning for anyone publishing at scale.
Compete on differentiation, not volume. In a web where the median new article reads machine-made, the durable advantages are a recognizable voice, an owned persona, and formats that resist one-prompt reproduction — short-form video, carousels, avatar content. A content engine like Kompozy generates those on-brand formats from one source and publishes across the eight social platforms plus blog and email, so reach comes from channels you control rather than from search luck.