Pew: 10% of the Web Is Written With AI — Authorship Detection Is Not GEO Detection

Three-line summary

• AI writing is close to becoming the default on the web. Pew Research analyzed 490,000 webpages and found signs of AI authorship on 10% of pages in the July 2026 crawl, rising to over one-third among pages whose publication date falls after ChatGPT's release.
• "Who wrote this" and "what was this built to win" are separate questions. Pew's detector identifies the writer. GEO detection research identifies the intent behind the optimization. Being flagged by one says nothing about the other.
• Scrubbing surface signals is not the answer. Em dashes and Oxford commas appear in human writing too, and neither detector determines whether an AI answer will cite your page.

What Pew actually measured

On August 20, 2026, Pew Research Center Data Labs published "How Much of the Internet Is Written With AI?". This is not a survey — it is a large-scale analysis of a web archive.

Component	Detail
Sample	49 Common Crawl snapshots (2021-01 to 2026-07) x 10,000 pages each = 490,000 English pages
Detector	Open Pangram (editlens_Llama-3.2-3B), scoring 0 (human) to 1 (AI)
Threshold	0.2 and above counted as meaningful signs of AI authorship
Cross-validation	62,370 pages checked against Pangram commercial model 3.3 → 96% agreement, Cohen's kappa 0.61
Coverage	Publicly accessible pages only — paywalled and login-gated content is underrepresented

A threshold of 0.2 is deliberately low, and Pew spells out what that produces: a page a person wrote and then cleaned up with an AI tool lands in the same bucket as one the model produced from scratch.

AI detection models are probabilistic tools, and individual page classifications should not be treated as definitive. — Pew Research Center, methodology (2026-08-20)

10% of the web, and over a third of recent pages

All figures below come from the July 2026 crawl.

Segment	Share with AI authorship signs
All pages	10%
Pages with a post-ChatGPT publication date	Over one-third
.com domains	~10%
.org domains	4.6%
.edu domains	~1%
.gov domains	~1%

The gap between the two headline numbers is a timing effect. The sample contains a large volume of pages from before 2021, which could not have been written with these tools at all. Read together, they say that the share is far higher among content being produced now.

One constraint matters here. Pew could parse a publication date for only 10-15% of pages, so "over one-third" describes dated content, not the web at large.

The domain split is easier to read. A .com page is about twice as likely to carry these signs as a .org page, and roughly ten times as likely as .edu or .gov. AI writing is concentrated where content is produced commercially and at volume — which also means most of the pages you compete against are being made the same way.

The style markers have already shifted

Separately from the classification scores, Pew counted how often specific language features appear per 10,000 words (2023 → 2026).

Signal	2023	2026	Change
Em dashes (—)	5.79	11.19	~2x
Oxford commas	34.04	55.51	+63%
AI-typical vocabulary ("delve", "interplay")	11.94	26.02	~2x
Negative parallelism ("it's not X, it's Y")	0.87	2.36	~2.7x

The tempting conclusion is to strip those constructions out. Pew closes that door itself: taken one at a time, none of these features can identify a specific document, and human writers use all of them. The shift only means something across hundreds of thousands of pages.

Authorship detection and GEO detection answer different questions

This is the distinction that gets lost most often in practice. Two studies landed within days of each other, but they measure different things.

	Pew (Open Pangram)	GEO-Flag
Question	Who wrote this page	What this page was built to win
Target	Traces of AI generation or AI editing	Strategic intervention aimed at AI citation
Prevalence	10% of all pages / over a third of recent ones	8.9% of pages / 16.4% of pages modified in 2026
Text only polished with AI	Counted (threshold 0.2)	Trained to be excluded

The same page can read as "AI-written" under Pew's method and "not optimized" under GEO-Flag, and the reverse holds too — a page a person wrote by hand, with mechanical optimization patterns layered on top, falls into the second case.

The figures and the practical line on the optimization side are covered in GEO Content Is Now Detectable.

So what does this mean for citations?

There is a question neither study answers: does a page flagged as AI-written become less likely to be cited in an AI answer? No direct measurement of that exists in either paper. What is established stops here.

• A substantial share of the web is AI-written (Pew)
• That is technically detectable (both studies)
• No search engine or AI platform has announced a penalty based on it

On citation probability itself, other research has produced answers. A survey reviewing 45 GEO studies reported cases where rewriting that manipulated form without adding substance actually reduced exposure (-9% / -16% / -6%). That result points at whether the content genuinely answers the question, not at what wrote it.

In other words, the writing tool is not the deciding variable. Whether the draft came from a model or from a person, what remains is whether the page carries the information an answer needs, in a form that can be extracted.

What to check now

1. Rewording to dodge detection is not a priority. Removing em dashes neither evades a classifier nor earns a citation. With a 0.2 threshold and kappa of 0.61, any single page verdict is probabilistic to begin with.
2. Review depth matters more than whether AI wrote the draft. The fact that AI-edited text is classified alongside AI-generated text also means the reverse: whatever produced the draft, quality diverges once a person verifies the facts, attributes the sources, and organizes the structure.
3. Understand the competitive baseline numerically. At roughly 10% of .com pages and over a third of recent content, an average AI draft does not differentiate. What does: first-party data, your own measurements, and specific cases.
4. Check how your page actually reads first. Independent of tooling, the precondition for citation is whether an AI system can use your page while composing an answer.

Frequently asked questions

Will using AI to write get me penalized in search?

No penalty policy has been announced. The Pew study measures prevalence, and no search engine or AI platform has stated that it demotes AI-written content across the board. That said, now that detection has been demonstrated at scale, it is reasonable to treat "platforms may use this as one quality signal" as a long-term risk.

If a person writes it and AI polishes it, does that count as AI-written?

Under Pew's method, yes. The 0.2 threshold is set to capture mixed human-and-AI writing broadly, so a polished draft and a fully generated one land in the same category. GEO-Flag's detector was trained the other way — AI polishing alone does not raise its score. The answer depends on which detector you mean.

Can I avoid detection by removing em dashes and words like "delve"?

Not recommended. Pew states that these features cannot identify a specific document and that human writers use them. Removing them offers no guarantee the classification changes, and even if it did, citation probability is a separate matter. Time spent adjusting style is better spent strengthening the content — that helps on both fronts.

On a small budget, if AI drafting has to continue, what should come first?

Shifting resources from drafting to review is the more efficient move: verifying quantitative claims against primary sources, placing the answer sentence near the top, and adding your own data or cases. All three act directly on citation likelihood regardless of what produced the draft.

Is the 10% figure reliable?

The sample and validation are solid — 490,000 pages across 49 crawls, with 96% agreement against a separate commercial model. But Cohen's kappa of 0.61 is substantial agreement, not perfect agreement, and about 1% of pre-ChatGPT documents were false positives. Treat it as a reliable trend at the aggregate level and as unsuitable evidence for judging any individual page.

Do the same rates apply to non-English content?

Unknown. Pew's sample is limited to English pages. No comparable large-scale public study of AI authorship in other languages is available, and applying English-web figures directly to another language is not supported by evidence.

Wrapping up

Pew's analysis of 490,000 pages turned an impression — "AI is writing the web" — into a number: 10% overall, over a third of recent content. That number does not translate into a citation penalty. Authorship detection and optimization detection answer different questions, and neither one determines whether an AI answer will cite you.

One standard survives. Whatever produced it, does the page hold information that genuinely answers the question, in a form that can be extracted? RanketAI's free page structure diagnostic shows how your page reads right now.

Related reading

• GEO Content Is Now Detectable — 8.9% of Web Pages Already Optimized — the detection side that targets optimization intent, and where the practical line sits.
• Review of the 45-Study Critical GEO Survey — the backfire figures for form-only rewriting.
• What Is GEO (Generative Engine Optimization)? — verified optimization methods and the conditions under which they work.
• 9 AI Search Myths vs 15M Data Points — the record of shortcut strategies failing under measurement.