DropSlop research · The Slop Index: AI writes more of the web every year
new English web pages are now written by AI.
Pages of 150 words or more with a publish date, in the Common Crawl archive; latest archive snapshot, .
Key findings
01 · The rise
It has never fallen back: past one in ten by , past one in five by , and above both lines ever since.
The first snapshot made wholly after ChatGPT, , found ; the latest, , . The snapshots before ChatGPT, from a time before chatbots were in wide use, run from to around the baseline: that is the noise in this measure.
Check your own text
Paste a paragraph and see which phrases read machine-made. Free, no account. To read a whole site page by page, run a site audit.
02 · Year by year, by page type Exploratory
Written by AI, by page type and year published (2023 to 2026). The span each panel’s change covers is under its name.
03 · Big sites and the long tail
04 · Page types
Blog posts are the most AI-written kind of page: of new ones.
Pages say what they are in their own markup. Each type is measured against its own pages from before ChatGPT, so a type the detector reads harshly is not held against it. Blog posts are clearly ahead of articles, the next type (). At the other end: , .
05 · Domain endings Exploratory
Country endings
Endings with no country
06 · Telltale words
Extra uses per million words, pages published since ChatGPT against pages before it, for words and phrases often named as AI tells. Only terms used at least times before ChatGPT, on at least sites, are listed; how many times as common is shown only where that figure is stable. Words that also name things (dives, landscaping) are left out.
07 · A check on the dates Exploratory
Both ways of dating a page show the rise, but one rises about twice as far.
Method
How we did it
The method
We scored pages published since ChatGPT and before it with DropSlop’s detector. On pages from before ChatGPT, when chatbots were not yet in wide use, it flagged ; on AI-written pages made for the study it caught . “Written by AI” takes out the first and puts back the misses. It is an estimate.
Limits
- About wholly AI-written pages. The detector misses most pages that mix human and AI writing, so the real share of AI involvement is higher.
- Catch rates come from new, unedited text. We generated the AI-written test pages with five 2026 models and did not edit them, so text from older models, or edited text, may be caught at a different rate.
- Pages with a date only. Only pages that carry a publish date can be placed before or after ChatGPT, and pages that declare their date in markup rise further (section 07).
- Pages, not people. The detector can be wrong about any one page; these are shares of many.
- Not the same sites over time.
Data: s01-deep.json · s01-deep.csv
Fair useUse these findings freely with a link back to this page.