Half the new web is written by AI. It wins 14% of Google.

Pew's Data Labs team pulled 490,000 English-language web pages from the Common Crawl archive, spanning January 2021 to July 2026, and ran them through the open Pangram detection model. Two numbers lead the report. 10% of all sampled pages show significant signs of AI authorship, and among pages published after ChatGPT's November 2022 launch the figure passes one in three. The full report is public, and TechCrunch covered it the same day.

Pew is careful about the limits, and so are we. Detection models misclassify individual documents. A single em dash proves nothing about a single page. The claim holds at the level of half a million pages, where the statistical fingerprints of model prose separate cleanly from human writing. Read it as a directionally correct census, and treat any one page's score as a guess.

The part of the report nobody else led with is the trend data on specific writing patterns. Between January 2023 and July 2026, across the whole sample of web text, em dashes went from 5.79 to 11.19 per 10,000 words. Oxford commas rose 63%, from 34.04 to 55.51. AI-typical vocabulary, the delve-and-showcase word list, more than doubled from 11.94 to 26.02. And negative parallelism, the "this isn't X, it's Y" construction, nearly tripled from 0.87 to 2.36.

Treat that trend data as a public, quantified checklist of what makes bought content read machine-made. Your readers pattern-match on these without knowing their names. So do the reviewers at any publication you pitch, and, increasingly, the classifiers that decide what surfaces in AI search. A content operation that ships drafts carrying these fingerprints is publishing its own production process.

The AI share is not evenly spread. Commercial .com domains show significant AI-authorship signs at roughly 10%, about ten times the rate of .edu and .gov pages, which sit near 1%. Nonprofit .org domains land at 4.6%.

That distribution is a trust map. The domains with editorial gatekeeping barely moved, and the domains where publishing is cheap and unsupervised absorbed almost all of the flood. Every buyer of content now publishes into the crowded end of that map. The scarce signals, a verifiable number, a source someone can open, a position someone will defend, are worth more there than they were in 2022.

Pew counts pages. Graphite, an SEO research firm that runs its own detection studies, counts articles, and its numbers frame the same event from the supply side: AI-generated articles hit 35.9% of new articles within twelve months of ChatGPT's launch, crossed 50.9% in late 2025, and have sat flat near half for five straight quarters. The growth stopped. Whatever economics were pushing raw AI publishing found their ceiling more than a year ago.

Why the ceiling? Graphite's 2025 companion study of 31,493 keywords answers it with the demand side. In Google's results, 86% of ranking articles are human-written. In ChatGPT and Perplexity citations, 82% are. Only 7% of number-one Google results are AI-generated, and the AI-written articles that do appear cluster after position ten, where nobody clicks. The machines writing half the web are, for the most part, writing pages that no search engine and no answer engine ever shows anyone.

A 2026 SEO study from Rankability that scored 487 competitive commercial results found the same shape: 83% of top results read as human-written. Three separate measurements with different methodologies land on the same shape. The flood is real, and the flood is losing.

Google's published position is that it rewards quality regardless of how content is produced, and penalizes scaled content whose primary purpose is manipulating rankings. Believe the second half of that sentence more than the first. What the ranking systems reward has converged on exactly what raw model output lacks: information that exists nowhere else, backed by a claim with a source under it.

Raw AI drafts fail that test in a detectable way. Pew's tell data shows the fingerprints are getting more common, which means the average unedited draft is getting easier to classify. And the same statistical signals a research lab can count at scale, a ranking system can weight at scale. Nobody needs a courtroom-grade detector to downrank a page. They need a probability, and Pew just published how legible that probability is.

Half the web conceding the quality bar means the bar is cheaper to clear than it has ever been. A company that publishes twenty pages a year with real numbers, named sources and an argument someone owns is now competing against a field that is one-third statistical filler. The .edu and .gov numbers show what scarcity looks like: the trusted end of the web is barely 1% machine-written, and trust concentrates where the flood is not.

AI still belongs in that operation. The drafting cost collapse is real and you should take it. What the data rules out is the version where the model's output ships unread. Read Graphite's own caveat closely: it did not evaluate AI-assisted content with heavy human editing, and it names that unmeasured route as possibly an effective strategy. So nobody has measured edited drafts failing the way raw ones do, and the mechanism points the other way, because what a detector counts is the raw output's fingerprints and an editor's whole job is to replace them with judgment. Your team's judgment is the thing the measurement rewards.

This is the pipeline we sell, so weigh our bias accordingly. When we build content production systems for a studio or a marketing team, the model does the assembly and a person does the arguing: the workflow drafts, then routes every piece through an editor who owns the claims before anything ships. We wire that gate in as part of a custom AI agent build, because the gate is the product. The machine gets you to a reviewable draft in minutes. It never gets to publish.

Publish raw model output at volume and you are competing for the 14% of ranking slots that AI content actually wins, mostly below position ten, with pages that carry fingerprints growing more legible every quarter. Rankability's study documents the terminal version: a site running 100% AI-detected content that was deindexed after a Google update, and recovered only after a human rewrite.

The cost is not only rankings. A third of the post-ChatGPT web reading as machine-made has trained your buyers to bounce on the pattern. The em dash cadence and the paragraph that summarizes itself now work as a label, and the label says nobody here checked this. Once a prospect files your blog under that label, no individual page gets a second read.

And the near-term risk compounds, because AI search is repricing visibility right now. The same week Pew published, Axios reported that Reddit's share of ChatGPT Search citations collapsed by 86% in about a week as OpenAI reweighted its sources. Citation flows in answer engines move that fast, with no notice. Content that earns citations on trust signals survives reweighting. Content that got lifted by one engine's temporary source mix does not.

Run the audit Pew just handed you. Count the tells in your last ten published pieces per 10,000 words and compare against the averages above; if your blog tracks the 2026 curve, your production process is visible to anyone who looks. Mapping that gap, across content and every other function, is half of what our 10X audit exists to do. Then fix the pipeline rather than the prose. Every AI-drafted piece gets a named human owner who edits the argument, not the commas. Give each figure a primary source a reader can open, and make each piece carry one thing a model cannot generate, whether that is your own data or a worked example from your own operation.

Measure the demand side while you are at it. Track which pages earn citations in ChatGPT, Perplexity and AI Overviews, and treat a citation as the new ranking. It is the same discipline as running evals on a production agent: measure the output, or you are guessing. The ROI numbers on AI investments follow measurement discipline everywhere else we have looked, and content is no different.

If you would rather buy the pipeline than build it, that is the trade an AI content creation agency exists for, and the same one our own AI automation agency sells, ours included: a fixed-price build that wires drafting, editing gates and source checks into one system your team runs. Whoever builds it, hold them to the same bar this post is held to. Ask where the human edit happens, and walk if the answer is nowhere.