deslopify.me

← Research

Research

How much of the web is AI-written now?

The published estimates span 17% (Google's top SERP results) to 74% (all new web pages, broadly AI-touched). The numbers disagree because every measurement comes from an AI detector and the detectors disagree. Every study still shows the AI share rising.

The headline numbers

Nov 2024

Graphite (single detector, 43,000 URLs)

AI articles cross 50%, first overtake of human on the open web

Q1 2026

Graphite (3-detector ensemble, 55,400 URLs)

~50.4% of indexed articles are primarily AI; plateaued 5 quarters

Apr 2025

Ahrefs, 900,000 new pages

74.2% AI-touched · only 2.5% pure AI · 71.7% hybrid

2025

Detector-vendor study, LinkedIn long-form

>50% of posts likely AI

Sept 2025

Detector-vendor study, Google top-20 SERP

~17% AI in top-ranking results (down from ~20% July)

Mar 2026

NewsGuard, AI content farms

3,006 unreliable AI news sites, 16 languages (49 in May 2023)

What to actually believe

The Ahrefs number is the most defensible

Ahrefs (April 2025) analyzed 900,000 newly indexed web pages, one per domain, English-only, using their internal bot_or_not classifier. Their headline finding doesn't collapse to a single number on purpose: 74.2% of new pages contain detectable AI content; only 2.5% are fully AI; 71.7% are human-AI blend. That tracks how the web actually works in 2025 and 2026. Most new pages have AI somewhere in the pipeline. Only 2.5% are raw model output pasted to a page.

Graphite's 50% headline is real but framed for SEO marketing

Graphite's Q1 2026 update used a three-detector ensemble of leading AI classifiers on 55,400 CommonCrawl URLs. They report 50.4% AI vs 49.9% human on indexed articles, with the share plateaued near 50% for five quarters. The methodology is stronger than their May 2025 study, which used a single detector. But Graphite is an SEO agency selling "topical authority" services, so "the AI takeover" framing is also a marketing hook. The 50% number is plausible. Leading every headline with "AI takeover" is a sales decision.

Volume isn't visibility

Search Engine Land ran a 16-month experiment publishing pure-AI pages and watching how they ranked. Only ~3% of pure AI pages stayed in Google's top 100 by month 3, rising back to ~20% after the August 2025 spam-update reshuffle. The takeaway: AI content dominates new-page volume but loses badly on the pages that actually get traffic. A separate detector-vendor Google SERP analysis puts AI at 17.3% of top-20 results in September 2025, consistent with that gap.

NewsGuard tracks the worst of it

NewsGuard's AI content farm tracker counts sites, not articles: 3,006 unreliable AI news sites across 16 languages as of March 2026, up from 49 in May 2023. These are sites publishing AI-generated news with no editorial oversight and no disclosure. Useful for trend direction; not a share-of-web number.

Why the numbers disagree

Every estimate above ends up running text through an AI detector. The detectors disagree about the same text, especially when the text has been edited at all. See why detectors fail. That has three consequences:

  • "Pure AI" numbers are unstable. A page lightly edited by a human will read as AI to some detectors and human to others. Studies that force a binary classification understate hybrid; studies that allow a fractional "AI-touched" label (like Ahrefs) get bigger numbers and lose the "pure" share.
  • Vendor incentives shape framing. Graphite (50%) sells topical-authority services; a higher AI-pollution share helps that pitch. Ahrefs (74%) launched its own AI detector during the study window. Detector vendors publishing usage stats are, by definition, selling detectors. Read every "X% of the web" press release with the incentive structure in mind.
  • Sampling matters. CommonCrawl-based studies over-represent indexed pages. NewsGuard counts whole-site judgments. Google-SERP studies sample only the top of search rankings, which are heavily filtered.

What that means for a writer

The competitive landscape for the kind of writing you do is heavily AI-touched, but the writing that ranks and gets read is still mostly human or heavily edited. That's consistent with Zhang & Gosline 2023: AI marketing copy works for short product ads. For anything longer, readers reward writing that sounds like a specific person wrote it — which is also what survives Google's spam updates. The Slop Score is for that second category. If you're competing on volume against the AI-touched 74%, aim for the shape of the other 26%: a lower stylistic register score and claims specific enough to be wrong.

The sources