One in Four Web Pages Is Now AI Text. Where It Grew Tells You What to Do

What 5.1 million pages show, and why the average hides the story

The count comes from Adam Visokay and a team spanning the University of California, Berkeley, the University of Washington and several other departments, in a paper posted on 27 October 2026. They took 5,118,347 pages from 505,974 domains, across 71 Common Crawl snapshots running from January 2019 to August 2026 — every language, the full link graph, sampled by two-stage stratified cluster sampling and then weighted back to the crawlable population.

Measuring text is not like counting boxes, so the method matters. They ran several AI detectors against labelled benchmarks, kept the two that performed best, and then passed the raw flags through a Bayesian model that corrects for each detector's false positives and false negatives. The headline is not a detector's output. It is an estimate with the detector's error removed.

That estimate: in August 2026, 24% of crawlable pages carried text generated or substantially edited by an AI model, with a 95% credible interval of 22.9% to 24.4%. The trend behind it is steeper than the level. The average annual increase was 0.2 percentage points from January 2019 to February 2023, 4.6 points a year from February 2023 to January 2025, and 8.1 points a year from January 2025 to August 2026. The change in the rate of change is itself significant: 3.6 points per year, interval 3.3 to 3.9.

A second team reached a smaller number with a narrower sample. Pew Research Center sampled 490,000 English-language pages from the same archive and found 10% of a random July 2026 sample showing signs of AI authorship. Filter down to pages published after ChatGPT's release and the same sample gives more than a third. 10%, 24% and 35% are not in conflict. They are the same fact measured on different slices, and knowing which slice is yours is the next section.

Where the text actually sits, and why your pages are in that tier

The paper sorts every domain into six tiers by PageRank — how many other well-linked domains point at it. The top tier is 0.001% of domains and 2.7% of pages: WordPress.org, Wix, Vimeo, the Internet Archive, the European Union, the IRS, Harvard, the Financial Times. The bottom two tiers are 99% of all domains and 70.5% of all pages. The authors list what lives there: a Swedish rabbit breeder, a Louisville dog-walking service, a German hair salon, a French roofing contractor, a pizza restaurant. If you run a small business site, that sentence is about you.

Prevalence by tier in August 2026 ran from 11.6% in the most central tier to 27.5% in the least central. Every tier rose, but the growth concentrated at the edge: about four-fifths of the total increase came from domains on the periphery of the link graph. Of the 21.7-point rise between December 2021 and August 2026, 15.7 points came from rising prevalence inside the bottom two tiers and only 1.4 points from those tiers taking up a bigger share of pages. The flood is not a wave washing over the whole web. It is filling the corners where nobody is checking credentials.

Domain endings tell the same story more bluntly. In the latest snapshot, 29.5% of .com pages contained AI text, against 20.4% of .net, 14.6% of .org and roughly 6% of .edu and .gov. Since December 2021, prevalence in .edu and .gov rose by 4.9 and 4.0 points; on .com it rose by 27.1. Pew's independent sample puts .com at about ten times the rate of .edu and .gov. English pages are the other concentration: 32.8% against 17.0% for pages with no English detected, a ratio that has widened every year since 2023.

Read those numbers from the outside and they look alarming. Read them from a business owner's chair and they are clarifying. The volume of generated text is highest exactly where there is no reputation to protect and no customer to keep. That is not a description of a serious local business with a name, a phone number and a van.

"Google doesn't punish AI content" is true, and narrower than it sounds

That summary circulates in every SEO agency post, and the source is Google itself. In a February 2023 post signed by Danny Sullivan and Chris Nelson, Google wrote: "Using AI doesn't give content any special gains. It's just content." Its systems, the post says, reward original, high-quality work regardless of how it was produced, and automation has generated useful pages for years — scores, weather, transcripts.

The same document draws a line you can trip. Using automation "with the primary purpose of manipulating ranking in search results is a violation of our spam policies". The policy that covers it is scaled content abuse, defined as many pages generated mainly to manipulate rankings rather than help users, "no matter how it's created". Its examples name generative AI directly. And it ends with an instruction most owners never read: if you are hosting that content, exclude it from Search.

Google's own guidance to site owners also sends you to the Search Quality Rater Guidelines, sections 4.6.5 and 4.6.6, which cover scaled content abuse and main content "created with little to no effort, little to no originality, and little to no added value". Those raters do not set your ranking directly. They are how Google grades whether its own systems work, and pages in that band are graded at the bottom.

Bing states the same foundation from the other side. Its webmaster guidelines say Bing and Copilot rest on the same crawling, indexing and ranking base as classic search, so the practices that earn citations in AI answers are the practices that earn indexing — plus a new entry on the list: content written to manipulate the language model, which can cost you visibility or the whole index.

The 86/14 gap: publishing is not the same as being visible

Now measure the other end of the pipe. Graphite, a search-analytics firm that publishes its methodology, collected the first two pages of Google for 31,493 keywords across 10 categories in June 2025 and classified every article. 86% of the articles ranking were written by people; 14% were AI-generated. The same study pulled citations from ChatGPT and Perplexity: 82% human, 18% AI. Only 7% of the articles sitting in position one were AI-generated. When it matched keywords that had both kinds of page competing, the human page ranked higher — statistically significant at p < 0.000001. Their detector's false-positive rate was 4.2% and its false-negative rate 0.6%, both stated in the methodology.

Set that against how much AI text is published. The same firm's broader study found AI-written articles reached 49.6% of sampled published articles in the first quarter of 2025, 50.9% in the fourth quarter, and 49.9% in the first quarter of 2026 — roughly half, and flat for five quarters. Their own explanation for the plateau is the number above: practitioners found that purely generated articles do not perform, and stopped scaling them. An independent study using the Internet Archive, arXiv:2604.26965, put AI-generated or AI-assisted sites at about 35% of newly published websites by mid-2025, up from zero before ChatGPT, and found that the rise tracks falling semantic diversity and rising positive sentiment. It found no significant evidence of lower factual accuracy, which is worth saying out loud: the panic is ahead of the data on that one.

Half of what gets published is generated. One in seven of what ranks is. Everything between those two numbers is the filter that decides whether a small business website has traffic or an archive. That filter is not a detection tool — it is whether anyone finds the page worth linking to, quoting or calling.

The commodity test, and the arithmetic behind it

Google's guidance for its own AI features names the thing that separates a page that survives from a page that does not: non-commodity content. Its example is worth keeping. "7 Tips for First-Time Homebuyers" is commodity — common knowledge that could have come from anyone. "Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line" is not, because it carries what the writer actually saw. Same topic, same buyer. One of them could have been generated by any model on earth.

Here is what that means in money, with the assumptions stated so you can swap in your own. Say you run a 14-person heating and plumbing firm and pay an agency $450 a month for four articles: 48 pages a year, $5,400. Assume the pages bring 1,900 clicks across the year and two quote requests, which is a realistic shape for generic content in a trades category. That is $2,700 per enquiry.

Now the other page. You spend two Saturday mornings writing "What we charge to replace a 12-year-old boiler, and the three jobs where we tell you to repair instead". You put in your real standing charges, your two-hour callout window, the warranty terms you actually honour and a photo of a boiler you pulled out last month. Call your own time $800 for those two mornings. Assume the page earns 40 clicks a week and converts at 3%, which is not generous for a page naming prices — that is roughly 60 enquiries a year, at $13 each, and it keeps earning after you stop touching it. Even if both estimates are half wrong, the gap between $2,700 and $26 does not close by producing more of the first page.

Notice what the second page contains. Not better prose, and not a better prompt. A price, a warranty, a photograph of a specific job, and a rule for when the customer should do the opposite of what you sell. No model can invent those, because they are not on the internet — they are in your invoicing system and your head. That is the asset the flood cannot copy, and it is the only reason the next section exists.

A Monday-morning audit you can finish in a week

None of this is worth anything until it touches your own pages. Here is the sequence, in the order it makes sense to do it.

1. Open Search Console, go to Performance, set the range to 12 months and switch to the Pages tab. Export it. You want impressions, clicks and average position per URL, not a site-wide total.
2. Mark every page that already contains something only your business could supply: a price, a lead time, a warranty term, a photograph of your own job, a named customer, a mistake you made and fixed. These are keepers. Do not touch them yet.
3. For every page left over, ask the blunt question: could a random model produce this page from the top ten results for that query? If yes, it is commodity, whatever tool wrote it — including you, in 2019.
4. Commodity pages that are bringing enquiries stay. They are doing a job. Leave them alone and come back to the list next quarter.
5. Commodity pages with traffic but no enquiries get rewritten first — one a week, not all at once. Add the numbers you have. This is where an afternoon of your time has the highest return, because the reach already exists.
6. Commodity pages with no traffic in twelve months get noindex or deleted rather than rewritten. Google's own spam policy says that if you are hosting content that adds nothing, exclude it from Search.
7. Where a reader would reasonably wonder how a page was made, add one honest line about it. Google's guidance calls those disclosures useful, and it says putting an AI system in the author byline is probably not the right way to do it.
8. Put a person's name and a date on the pages where a byline is expected. Google treats those as trust signals, and so do the assistants pulling citations.

When to do nothing, and when to stop paying for content

Most of the audience for this article should not rewrite anything. If your site is a dozen pages that answer real questions your customers ask, and those pages bring calls, the 24% is a fact about other people's websites. The number is a population estimate, not a verdict on your URL.

Three more cases where the right move is to wait. First, if your workflow is genuinely "a model drafts, a person checks every claim and adds their own numbers", you are outside the studies above: Graphite measured fully generated articles and says outright that it did not evaluate AI-assisted content with heavy human editing. Second, some categories carry far less generated text than others — in their sample the fewest high-confidence AI articles were in Food at 4.5%, News at 4.7% and Travel at 6.5%, against 14.1% in Productivity and 13.3% in Crypto. Third, do not touch a page that is earning clicks just because it looks thin. Reach is the scarce thing; tidiness is not.

And do not chase AI-detector scores. The detectors in these papers are research instruments with known error rates, not a ranking input, and Google's published position is about quality rather than production. Writing worse on purpose to look human is the only guaranteed way to spend a day and gain nothing.

The decision that does survive eighteen months is about what you buy. If you are invoiced per article, you are paying for the one input that just became free. A supplier producing four commodity pages a month at $450 is selling volume in a market where volume is worthless and specificity is scarce. Renegotiate the shape of the work: fewer pages, more of your time in them, priced on whether enquiries move. That is a change to a process you already have, and it does not depend on what any model costs next year.

One page that says what you charge, what you refuse to do and what went wrong last time will outlast a hundred that say what everyone already knows.

Frequently asked questions

Is AI-written content against Google's rules?

Not by itself. Google's published position is that its systems reward quality rather than caring how a page was made, and it says automation has produced useful pages for years. What crosses the line is scaled content abuse: generating many pages for the primary purpose of manipulating rankings instead of helping readers. The distinguishing test is intent and volume, not which tool held the keyboard.

Will a detector flag my pages and drop my ranking?

Google has not published AI detectors as a ranking signal, and the tools used in these studies are research instruments. Their error rates are real — 4.2% false positives in the Graphite study, and Pew states plainly that detectors misclassify pages in both directions. The pattern that costs you visibility is not "a detector guessed AI". It is a page that adds nothing, which any quality system can judge without needing to guess who wrote it.

How much of the web is AI-generated, in one number?

24% of crawlable pages in August 2026, with a credible interval of 22.9% to 24.4%, from a weighted sample of 5.1 million pages. The number moves a lot with the slice you take: 10% of a random English-language sample, over a third of pages published after ChatGPT launched, and about 50% of newly published articles in a separate sample. Quote the slice, not the headline.

Should I delete the AI-written posts I already published?

Only the ones with no traffic and nothing of yours inside them. A page earning clicks stays, whatever wrote it. For the dead weight, Google's spam policy tells you what to do with content that adds nothing: exclude it from Search. That is a noindex or a deletion, and it is faster than rewriting a page nobody has ever visited. Keep the ones that carry your prices, your photographs and your own jobs, and improve those instead.

Do I have to say that AI helped write a page?

Google calls AI and automation disclosures useful where a reader would reasonably ask how the content was created, and says giving AI the author byline is probably not the way to do it. Name the person who checked the page. That covers the disclosure question and the byline question in the same line, and it is the detail the assistants pulling citations tend to look for.