The Internet After the Internet: How Much AI Content Is Already Online and What Happens When Machines Write for Machines

The Internet After the Internet: How Much AI Content Is Already Online and What Happens When Machines Write for Machines

Rating
—/5 0 ratings
Follow us on Google News
Perspectives15 min read

AI content on the internet is no longer an exception that can be easily separated from the rest of the web. In August 2026, Pew Research Center published one of the broadest estimates yet: in a random sample of 10,000 English-language pages collected in July 2026, about 10% showed substantial signs of AI authorship or editing. When considering only pages published after ChatGPT launched, the share rises to roughly 35%.

For more on artificial intelligence, see our article on AI’s energy consumption reaching the scale of a city: what happens when artificial intelligence meets electricity limits, which offers related context.

This does not mean that “35% of the entire internet is written by AI”. The web contains billions of older pages, and AI detectors are probabilistic tools rather than perfect tests of authorship. But the data point to a structural change: a growing share of the new web is generated, rewritten or assisted by models.

At the same time, the internet is no longer read only by people. Cloudflare says bots and automated agents already generate more than half of the web requests observed on its network, while AI crawler traffic is used for search, model training and carrying out tasks on users’ behalf. The web is therefore beginning to take on an unusual structure: machines write pages, other machines read them, and further models use the result to answer people without necessarily sending them to the original page.

How much AI content is actually on the internet?

The accurate answer is that nobody can count it perfectly. There is no central register of pages created with AI, and an article may be entirely human-made, entirely generated, or any combination of the two: a human outline, assisted research, AI rewriting, automated proofreading or full generation.

Pew Research Center analysed 490,000 pages drawn from 49 Common Crawl snapshots between January 2021 and July 2026. The researchers used an open-weight detector and checked some of the results with a separate commercial model.

In the July 2026 snapshot, around one in 10 pages in the sample showed substantial signs of AI authorship or editing. But after filtering out older pages and looking only at content dated after 30 November 2022, the researchers found such signals in approximately 35% of pages.

A separate study by researchers at Imperial College London, the Internet Archive and Stanford University, published as a preprint in April 2026, reached a similar figure: roughly 35% of websites newly published by mid-2025 were classified as AI-generated or AI-assisted.

The convergence of the two estimates is notable, but it should not be treated as a perfect measurement. AI detection can produce both false positives and false negatives, and web archives represent some kinds of pages better than others.

The internet is not yet “mostly AI”

This distinction matters because the phrase “dead internet” is often used alongside claims that most online content is already produced by bots. The available data in 2026 do not demonstrate that.

In Pew’s sample, 90% of web pages analysed in July 2026 did not exceed the threshold set for substantial signs of AI authorship. The 35% figure applies to a subset: pages with publication dates after ChatGPT’s launch.

The more accurate picture, then, is not that “the internet has died and been replaced by AI”, but that the new internet is changing quickly. If the trend continues, the share of synthetic material could become much larger in the years ahead.

The web’s language is also changing

Pew found that certain linguistic features frequently associated with generative models became more common after 2022. The frequency of em dashes approximately doubled, use of the Oxford comma rose by 63%, and certain words and constructions favoured by AI models appear far more often than before.

None of these signs, on its own, proves that a text was written by AI. People used these punctuation marks and words long before ChatGPT. But across hundreds of thousands of pages, the statistical shift suggests that generative tools are leaving an imprint on the web’s collective style.

An even more interesting effect emerges at the semantic level. The study by Dolezal and colleagues found a negative correlation between the growth of AI-generated content and semantic diversity: pages tend to cluster within a narrower range of phrasing and ideas. The researchers did not, however, find sufficient statistical evidence that this phenomenon had already reduced the factual accuracy or stylistic diversity of the internet as a whole.

When bots become a larger part of the web audience

The change is not occurring only among authors. Readers are changing too. Cloudflare describes the internet of 2026 as an environment in which bots and automated agents account for more than half of web requests observed on the company’s infrastructure.

That does not mean that “more than half of visitors are AI”. One machine can make thousands of requests in the time it takes a person to load one page. Request metrics and the number of people are different things.

Still, the distribution shows why the web’s design is starting to adapt to non-human readers. Cloudflare already separates AI traffic into three purposes: Search, Agent and Training. Search indexes information for search services; Agent carries out actions on a user’s behalf; and Training collects content to train or fine-tune models.

In 2026, nearly half of AI crawler traffic measured by Cloudflare was for training

Cloudflare Radar data from autumn 2026 show that, among traffic attributed to AI crawlers, around 48–49% of requests declared training as their purpose, while another roughly 39–41% combined training and search. The values vary from one period to another, but the direction is clear: a very large proportion of automated access to pages is not primarily intended to send a reader to the site.

This breaks a long-standing economic convention of the internet. Traditional search engines indexed a page and, in return, could send users to the publisher. The publisher bore the cost of producing the content, while a user visit could generate advertising, subscription revenue or other income.

An AI system can read a page, synthesise an answer and present the information directly in its own interface. A site may receive a citation, but the user no longer always has a reason to open the source.

Hundreds or thousands of crawls for one referral

Cloudflare measures this imbalance with its crawl-to-refer ratio: the number of HTML pages requested by a crawler for every visit a platform sends to a site.

During a recent period in autumn 2026, Cloudflare Radar showed approximately 447 crawls for one referral in OpenAI’s case, around 488:1 for Anthropic and roughly 3,500:1 for Perplexity. Google was much closer to the traditional model, at about 5:1.

These figures must be interpreted carefully. Cloudflare notes that native applications may not send the Referer header, meaning that some AI-originated visits may not be attributed to the platform. Even so, the difference in scale is large enough to explain the tension between publishers and AI companies.

Google Search is also starting to answer before sending users to sites

The shift is not limited to chatbots. Google has integrated AI Overviews and AI Mode into Search, meaning that users can receive a summary before the conventional list of links.

A Pew study of the behaviour of 900 US adults in March 2025 found that when an AI summary appeared, users opened a traditional result in about 8% of visits, compared with 15% when no AI summary appeared. Links cited directly in the AI summary were clicked in only about 1% of visits.

The data are from 2025, and behaviour may evolve as the product changes. But they illustrate the economic mechanism: the more complete the answer in an intermediary’s interface, the fewer reasons a user may have to visit the original source.

In 2026, Google began reporting AI-feature visibility separately

The change became important enough for Google Search Console to introduce reports dedicated to visibility in generative features in June 2026; on 31 August, the company expanded this information globally.

Publishers can now see which pages appear in AI Overviews, AI Mode and other generative experiences. In September, Google also added reporting for multimodal search.

This change says something important: an article’s success no longer means only ranking from 1 to 10 in a list of results. It can also mean appearing as a source in an AI-generated answer, even if no click follows.

SEO is not dying, but it is no longer the only game

Terms such as AEO — Answer Engine Optimization — and GEO — Generative Engine Optimization — emerged in 2026. The idea is to create content that systems generating answers can readily find, understand and cite.

Google nevertheless rejects the notion that there is a separate magic formula for AI. In official guidance published in May 2026, the company says core SEO practices remain relevant and that generative features are connected to its existing ranking and indexing systems.

Google’s message to publishers is essentially unchanged: valuable, unique content that is difficult to replace. That becomes even more important on a web where thousands of sites can quickly generate near-identical articles on the same subject.

Google does not ban AI content, but it does ban low-value mass production

One of the most widespread misconceptions is that Google automatically penalises any text written with AI. Its official documentation says otherwise. Generative AI can be used for research, structuring and production, provided the result meets quality standards.

The problem arises when large volumes of pages are generated chiefly to manipulate rankings, without original information or value for readers. Google calls this practice scaled content abuse.

That means the distinction between a premium site and an AI content factory will not be the tool used, but the editorial process: primary sources, fact-checking, selection, expertise, context, relevant images and a conclusion that readers will not find copied across another 100 pages.

What happens when AI increasingly learns from AI?

There is a deeper technical problem as well. Language models are trained on enormous quantities of text. If the share of synthetic data in future training sets rises too far, models may end up consuming their own approximations of reality.

The phenomenon is known as model collapse. In recursive training scenarios, one generation of a model produces data for the next, and errors and simplifications can be amplified. Rare distributions tend to disappear, diversity may decline, and the model can become increasingly confident in a narrow version of the world.

A study published in npj Artificial Intelligence in July 2026 tested ways of reducing this effect. The authors showed that loss functions which reduce the weighting of tokens generated with excessive confidence can delay degradation and allow a model to tolerate more than twice as much synthetic content before collapse appeared in their tests.

That shows model collapse is not an inevitable sentence. Clean data, filtering, provenance and training strategies can matter enormously. But the issue becomes more important as the web contains more model-generated text.

Original human content may become more valuable because it is rarer

Paradoxically, the explosion of AI content may increase the value of authentic material. An original statement, interview, directly obtained document, analysis based on proprietary data, photograph taken on location or specialist’s experience cannot be generated from nothing.

Models can summarise such information once it exists, but someone has to produce the information in the first place. If the web becomes too reliant on summaries of other summaries, first-hand sources become a scarce resource.

This may be where serious publications hold an advantage: not competing with AI on generation speed, but producing what AI does not have before publication — data, verification, access, selection and context.

The internet may divide into two layers

One plausible scenario is the emergence of a two-tier web. The first remains visible to people: pages, apps, video and publications. The second becomes a layer optimised for agents: structured data, APIs, feeds, cryptographic bot identity and formats more efficient than traditional HTML.

Cloudflare and GoDaddy already speak of an “agentic web” in which sites must be able to distinguish legitimate agents from unknown automated traffic. Cloudflare offers separate controls for search, agents and training and, from September 2026, changed default settings for some sites so monetised pages could block training and agents more aggressively.

In other words, robots.txt — a file that has existed for decades — is starting to be supplemented by much more granular rules: who can read, for what purpose, and whether there is a form of attribution or compensation.

A new economy: content may be licensed directly to AI

As referral traffic declines, publishers are seeking other economic models. AI companies have begun signing licensing agreements with media organisations and data providers, while web infrastructure is beginning to develop mechanisms for measuring how bots use content.

In 2026, Cloudflare launched tools allowing sites to see more clearly which AI company accesses their pages and how much human traffic it sends back. The company has even discussed automatic compensation models for access to content.

If this direction matures, the internet could move from “free crawling in exchange for traffic” to a market in which some sources charge for access, licensing or compensation for systematic use by models.

The authenticity question: how do we know who created something?

When text, images, voice and video can be generated quickly, the question “is it real?” is no longer enough. A synthetic image can accurately illustrate a concept, while a real photograph can be used to support a false claim.

The internet will increasingly need provenance: who published the material, when it was created, whether it was modified and which sources support its claims. Digital signatures and provenance standards may become as important to media as HTTPS became for the secure transmission of data.

AI detectors are not sufficient as the sole mechanism. Pew explicitly warns that a detector can misclassify an individual page. It is more useful when applied statistically to large collections, rather than as an absolute verdict on an author.

For publications, speed is no longer the decisive advantage

A model can produce a generic article on almost any topic in seconds. If a publication competes only on volume and speed, it has lost the competition from the outset.

The advantage lies in information that is difficult to commoditise: verifying a primary source, linking it to local data, expertise, comparing studies, identifying limitations and explaining consequences for readers.

Google makes the same point in guidance for its generative features: valuable, unique, non-commodity content remains fundamental. On a web where anyone can publish 1,000 articles a day, volume stops being a signal of value.

InfoBuz has also examined technologies that could change the world by 2030. But AI’s transformation of the web differs from many other technologies: it changes not only the devices we use, but the very information raw material that the next generation of systems will read and learn from.

At the same time, the InfoBuz article on how online life takes up more and more of our attention takes on a new dimension: it is not only the time spent online that is changing, but also the share of interactions in which the other party is no longer necessarily a person.

What the internet could look like in 2030

  • Fewer traditional searches: a larger share of questions will be answered directly in AI Mode, chatbots or agents.
  • More AI-written pages: the cost of producing text, images and video will continue to decline.
  • More machine-to-machine traffic: agents will access sites, compare products and carry out actions on users’ behalf.
  • Differentiated access: sites will separately decide what they permit for search, training and agents.
  • More valuable premium sources: original data, journals, research and human expertise may become more important for grounding.
  • More important provenance: readers and models will need better signals about content origins.

The internet is not dying. Its economic and information contract is being rewritten

“Dead Internet Theory” is too simplistic for what is happening. Data from 2026 do not show a web where people have disappeared and everything is artificial. They do, however, show an internet in which 10% of pages in a broad sample, and approximately 35% of new content dated after ChatGPT’s launch, show signs of AI authorship.

At the same time, automated systems account for a huge share of web requests, while AI interfaces are starting to answer users without always directing them to the source. This combination may alter the model on which the open web was built: people publish, engines index, and users visit.

The next web may work differently: people and AI create content together, agents read it, models synthesise it, and original sources must demonstrate why they deserve to be accessed, cited and funded.

The real question is not whether AI will fill the internet. That process has already begun. The question is whether mechanisms can be built so that original, verified and valuable information remains economically viable in a system capable of copying and summarising almost instantly.

Frequently asked questions about AI content on the internet

How much of the internet is written by AI? Pew Research Center estimates that around 10% of a random sample of English-language pages from July 2026 showed substantial signs of AI authorship. For pages dated after ChatGPT’s launch, the share was approximately 35%.

Does that mean 35% of the entire internet is AI? No. The figure refers to pages published after ChatGPT’s launch that had detectable publication dates. The web includes a great deal of older content.

Does Google penalise AI-written articles? Not automatically. Google says the problem is mass-produced content with no value for users, whether it was generated by AI or by other methods.

What is model collapse? It is degradation that can occur when successive generations of models are trained largely on synthetic data produced by other models, progressively losing diversity or rare information.

Does AI send traffic to the sites it reads? Sometimes, but the ratio differs sharply between platforms. Cloudflare observes hundreds or thousands of crawls for one visit sent to the source for some services.

Does SEO still make sense in the AI era? Yes. Google explicitly says that core SEO practices remain relevant for AI Overviews and AI Mode, and Search Console reports visibility in generative experiences separately.

Sources and transparency

Etichete: AEOAI OverviewsAI-generated contentCloudflareAI contentGoogle SearchArtificial intelligenceInternet
Share this article
COMUNITATEA INFOBUZ

Discuția cititorilor

Ai o opinie sau o întrebare despre acest articol? Spune-ne ce crezi.

0 comentarii

Fii primul care deschide discuția.

Lasă un comentariu

Email address este vizibilă doar administratorilor și nu va fi publicată. Câmpurile marcate cu * sunt obligatorii.

Te rugăm să păstrezi un limbaj civilizat și să discuți la subiect. Comentariile spam sunt filtrate automat.