AI-Generated Content Statistics (2026)

What this graphic shows

The AI Gradient in Figure 1 stacks Google's first page one position at a time. In Ahrefs' June 2026 study, the share of ranking pages that are at least 80% AI-generated climbs from 8.4% at position #1 to 11.7% at #10: a slope, not a wall. The supporting cards show what that means for indexing, how fast AI-written text is spreading across the web, and the writing habits that give it away.

The data arrives as the rules change. Since August 2, 2026, Article 50 of the EU AI Act requires providers of generative AI systems to mark their output in a machine-readable way, and Anthropic has started watermarking text written by Claude worldwide. Ahrefs' Ryan Law used these figures to argue the watermark is unlikely to become a ranking signal.

How much AI content ranks in Google?

5.3% of pages in Google's top three positions are 100% AI-generated, and 9.0% are at least 80% AI, according to Ahrefs' AI content detector. Most top results are still mostly human: 54.7% of top-3 pages have less than 20% AI content, and pages under 50% AI hold 82.2% of top-3 rankings.

AI content gets rarer as you climb the results page, but only slightly. The average page at position #1 scores 27.1% AI against 30.9% at #10. A separate Ahrefs study of 600,000 pages in 2025 found a correlation of 0.011 between AI content share and ranking position, which Ahrefs called "effectively zero".

Does Google penalize AI-generated content?

Not outright: 40.35% of very-high-AI pages were indexed, against 49.28% of low-AI pages, a gap of about 9 percentage points. If Google were filtering AI text out of its index, that gap would be far wider.

AI use does track with weaker results. Ahrefs found low- and moderate-AI pages earned 2–3× the impressions of high- or very-high-AI pages, with no sudden drop-off over time. Ahrefs reads this as Google punishing low-quality content, which often overlaps with AI content, rather than AI itself. Google's own guidance, published in February 2023, is headlined "Rewarding high-quality content, however it is produced" and treats AI content made mainly to manipulate rankings as spam.

How much of the internet is AI-generated?

About 10% of English webpages in a July 2026 sample show significant signs of AI authorship, according to Pew Research Center, and over one-third of pages published since ChatGPT's release do. Pew ran the Open Pangram detector over 490,000 pages sampled from Common Crawl since 2021.

Commercial sites lead. In 2026, 9.35% of .com pages showed signs of AI authorship, against 4.59% of .org pages and about 1% of .edu (1.03%) and .gov (0.76%) pages. Newer pages run higher still: Ahrefs found 74.2% of 900,000 pages first crawled in April 2025 contained some AI text, though only 2.5% were pure AI.

What are the signs of AI-written text?

"Negative parallelism" ("it's not just X, it's Y") nearly tripled on recent webpages, from 0.87 to 2.36 uses per 10,000 words between early 2023 and 2026, Pew found. AI-favored words such as "delve", "interplay" and "testament" more than doubled, em dashes roughly doubled and Oxford commas rose 63%. None of these proves AI authorship on its own, since people use them too.

Does Claude watermark its text?

Yes. Anthropic says supported Claude models weave an invisible watermark into generated text, and models launched on or after August 2, 2026 support it at launch. The marking applies across the Claude API, Claude apps and Claude Code wherever Claude is offered, worldwide. Files get signed C2PA Content Credentials instead.

A detected mark only means text "may have been processed by Claude", so a human draft that Claude only proofread can carry it. Anthropic notes the mark can disappear when text is heavily edited, paraphrased, translated, mixed with other writing or very short. Text-watermark detection is in private preview for eligible organizations such as regulators, researchers and media.

Will other AI companies watermark their output?

Yes: 95 organizations had signed the provider section of the EU's Code of Practice on Transparency of AI-generated Content when we checked on November 13, 2026, including Anthropic, Google, Meta, Microsoft, Mistral and OpenAI. Another 192 signed the deployer section, which covers labeling. The European Commission put the total at about 190 organizations by the end of July 2026.

Google already does it. In its 2024 Nature paper, Google DeepMind reported a live test of SynthID-Text across about 20 million watermarked and unwatermarked Gemini responses: thumbs-up rates differed by 0.01% and thumbs-down rates by 0.02%, both statistically insignificant. The method now watermarks Gemini responses.

How do AI text watermarks work?

The model nudges its word choices toward a secret, randomly chosen set of "green" tokens. Kirchenbauer et al. (2023) described the approach as promoting green tokens during sampling so the signal is invisible to readers but detectable "from a short span of tokens" with a statistical test. Any single word looks normal; across a few hundred words the bias becomes measurable by anyone holding the key.

The limits are real. Google DeepMind says SynthID's confidence scores "can be greatly reduced" when text is thoroughly rewritten or translated, and the method is less effective on factual answers, where there is less room to vary wording.

Methodology & caveats

Every AI-content share in this graphic is a detector estimate. Ahrefs used its own detector and Pew used Open Pangram; they are different models with different thresholds, so their percentages are not directly comparable, and neither can confirm who actually wrote a page. Ahrefs only classifies pages of 350 words or more.

Periods differ by study. Ahrefs' ranking and indexing data is from June 2026 (indexing used Search Console impressions since January 2026); its new-pages study is April 2025 and its 600,000-page correlation study is 2025. Pew's trend uses six-month averages of Common Crawl samples labeled by period start, and its "10%" and "over one-third" figures refer to a July 2026 snapshot. Ahrefs' 5.3% (100% AI) is part of the 9.0% (≥80% AI), not in addition to it; the Ahrefs article that led to this graphic describes the 9% as "another" 9%.

Derived figures are marked in card notes: the 8.93-point indexation gap, the 3.8-point average rise, the unitemized ~9.8-point slice of Ahrefs' 2025 mixed pages, and the Fig. 7 multiples, which divide Pew's 2026 value by its H1 2023 value. EU signatory counts come from the Commission's live list and will change; the provider (Section 1) and deployer (Section 2) lists overlap, so they are not added together.