THE LOCKE LEDGER.
Independent thinking about the work behind the work

Myth-Busting

Bots Send About Half of Web Requests. People Haven't Left.

The theory spread from a 2021 forum post. Bot counts from the security companies Thales and Cloudflare, and a Pew survey of 5,023 American adults, now measure parts of what it claimed.

Bots make about half of web requests by two companies' counts, but the claim that people have left the internet does not hold up. The doubt behind the theory is real, and a business website has to answer it.

On January 5, 2021, an anonymous user on Agora Road, a small online forum, opened a thread titled "Dead Internet Theory: Most of the Internet is Fake." The post said that around 2016 or early 2017 the message board 4chan had filled with posts by "someone or something," and it went further: the United States government, the author wrote, was running an "artificial intelligence powered gaslighting of the entire world population."

The author wrote that the theory had first been drafted on another message board with other anonymous users. No evidence for a government plot appeared in the post, and none has appeared since. The phrase spread anyway, and it is now a common search, while some of the numbers published since then look like what the theory predicted. Thales, the technology group that owns the bot protection company Imperva, reported in April 2026 that bots made up 53 percent of web traffic in 2025.

The dead internet theory is wrong in its literal form, because the people did not leave. The doubt it describes is real and has been measured. In a Pew Research Center survey of 5,023 American adults in June 2025, 76 percent said it was extremely or very important to be able to tell whether pictures, videos and text were made by AI or by people. In the same survey, 53 percent said they were not too or not at all confident they could tell. A small business website will meet some of those doubtful visitors.

What is the dead internet theory?

The original version had two parts. One was an observation that a growing share of what people see online is produced by software. The other was an explanation: that the change was coordinated, deliberate and aimed at controlling the public.

Over time the phrase lost most of its second half. A survey paper posted to arXiv, the open research archive, in January 2025 by Prathamesh Muzumdar and five coauthors describes the theory as the idea that much of the internet, social media in particular, is dominated by bots, AI generated content and corporate interests. Its abstract, which is the only part of the paper reviewed for this article, argues that platforms built to reward engagement have pushed meaningful human connection aside. That argument rests on incentives, and the abstract names no hidden hand running the web.

For a business, the questions are narrower than either version: how much of the activity a website sees is automated, how much of what people read is machine written, and what a visitor needs before believing a real company is on the other end.

How much of the internet is bots?

Two security companies publish widely cited answers, and they measure different things.

Thales's 2026 Bad Bot Report, based on full year 2025 activity, puts bots at 53 percent of internet traffic, up from 51 percent the year before, with bad bots at 40 percent and human traffic at 47 percent. The report says it combines findings from Thales's threat research and security analyst teams. Its summary page says Thales blocked 17.2 trillion bot requests in 2025 but does not explain how the traffic shares were sampled. The numbers come from a company that sells bot protection, measuring with its own teams.

Cloudflare's 2025 Radar Year in Review, covering January 1 to December 2, 2025, counts requests for web pages, which it calls HTML requests, and splits them three ways. It reports that bots other than AI crawlers began the year responsible for half of those page requests, seven points more than people. By December 2, human traffic accounted for 47 percent and those bots for 44 percent. AI bots averaged 4.2 percent of page requests over the year, and Googlebot, the crawler that builds Google's search index, accounted for 4.5 percent on its own.

Neither number means half of the people online are fake. A bot request is a machine asking for a page. Many of those machines are search engines indexing a site so it can be found, uptime monitors checking that it loads, or price trackers and AI crawlers copying its text. Some are attacks. Reading those shares as counts of fake people goes beyond what either report supports.

Google Analytics, a free tool many small business websites use to count visitors, already filters out part of this traffic. Its help center says traffic from known bots and spiders is excluded automatically, using Google's own research and the International Spiders and Bots List maintained by the Interactive Advertising Bureau, and that the filter cannot be turned off. The page does not say what happens to bots Google does not recognize.

Is the dead internet theory true?

The traffic numbers are only one test. The theory's stronger claim is about content: that most of what people read is written by machines.

A widely repeated forecast on that point came from Europol, the European Union's police agency. Its Innovation Lab's 2022 report on deepfakes said experts estimated that as much as 90 percent of online content might be synthetically generated by 2026, with a footnote to a 2020 book on deepfakes by Nina Schick. Futurism and other outlets spread the line. The copy Europol hosts today, a file dated January 31, 2024, keeps the paragraph on synthetic media but no longer contains that sentence. The report offered no measurement of its own behind the figure, and 2026 has now passed.

A measured estimate exists for one slice of the web. Graphite, a search marketing firm, published a study on May 15, 2026 that ran 55,400 English language articles from Common Crawl, a public archive of the web, through three AI detection tools. By the firm's count, the share of new articles that were primarily AI generated reached about half in early 2025 and then stopped climbing, holding near 50 percent through the first quarter of 2026, the last period studied, when it was 49.9 percent. Graphite's own tests found each detector wrongly flagged roughly 1.4 to 1.8 percent of articles published before ChatGPT's release, and the study did not check how well the tools classify AI drafts that a person then edited. It covers pages marked up as articles, not reviews, comments or social posts.

LinkedIn, the professional network, offers a closer look at one platform. Originality.ai, a company that sells an AI detection tool, ran 8,795 LinkedIn posts of 100 words or more, published between January 2018 and October 2024, through its detector and estimated that 54 percent were AI generated, The Register reported in November 2024. The company also found that the average word count of long posts had risen 107 percent since ChatGPT launched. In a separate sample of 5,000 public long posts gathered in July 2026, Originality.ai classified 81.2 percent as likely AI. The two samples were drawn in different ways, so together they do not show a trend. LinkedIn told The Register in 2024 that it does not track whether generative AI was used to write or edit posts.

Engagement on the platform also fell in 2025. Metricool, a social media scheduling company, counted 23 percent fewer LinkedIn impressions and 14 percent fewer interactions than the year before across the accounts in its yearly study, and its December 2025 release put that down to more posting and more competition for attention. No source here shows that AI written posts caused the drop. LinkedIn has since changed how it handles such posts. In May 2026 the company said content that appears to be generated by AI and lacks a clear perspective is less likely to be shown beyond the author's immediate network.

None of those counts measures how many people are online, so none of them shows that people have left.

Where a small business meets the bots

For a small business, automation shows up in three places.

The first is the contact form. Automated scripts submit forms with fake names and junk addresses, and when a form submission counts as a conversion in an ad account, those fills can teach the bidding system to find more of them. The guide to Google Ads spam leads covers the form checks Google recommends and the cost of each.

The second is reviews. The Federal Trade Commission's 2024 rule on reviews and testimonials names reviews by people who do not exist, including AI generated fake reviews, among the practices it bars; the guide to getting more Google reviews within the rules covers what the rule allows. A business on the receiving end of fake reviews has a separate problem, and the guide to removing a fake Google review walks through Google's process.

The third is the business's own content. Google's spam policies define scaled content abuse as many pages "generated for the primary purpose of manipulating search rankings and not helping users," and say the policy applies no matter how the pages are created. The account of Google's September 2026 spam update explains what that has meant for small business sites.

Showing a real person is there

Google's guidance for site owners asks for proof of the kind that answers the doubt Pew measured. Its page on helpful, reliable content asks whether pages carry a byline where a reader would expect one and whether the byline leads to more about the author. It says the company "strongly encourage[s] adding accurate authorship information." It also calls fabricated creator profiles, including AI generated headshots, made up names and false credentials, a form of deception.

A small business website can answer the doubt with things a stranger can check: a phone number that reaches a person during the posted hours rather than a voicemail box nobody clears, a street address that matches the Google Business Profile and the invoices, photos of people who work there, and case studies and testimonials from real customers who agreed to be named. Each of these is a fact a visitor or a search engine can confirm against another source, which is the idea behind making business information easier to verify.

Each also has a cost. Photographing real staff takes a shoot and goes out of date when people leave. Publishing an employee's full name invites recruiters and spam, and many businesses reasonably use first names only. Answering the phone live takes staff, and a fast written reply can do the same work when it is honest about hours; the guide to what should happen after a website inquiry sets out the options. Stock photography and a generic about page cost less and are easier to maintain.

The 2021 post that spread the theory asked its readers to believe that the people they met online were gone. The measurements since then say the people are still there, sharing the traffic with machines and often unsure which of the two made what they are reading. A business that lets a stranger check its phone number, its address and the people who work there gives those visitors a way to settle that question for themselves.

Frequently asked questions

Can an AI detector tell me whether a review or post was written by AI?

Not reliably for a single piece of text. In Graphite's May 2026 study, each of the three detectors it used wrongly flagged roughly 1.4 to 1.8 percent of articles published before ChatGPT's release, and the study did not test AI drafts that a person then edited. Figures from Originality.ai come from a company that sells the detector it used. A detector score is a lead to check against other evidence, such as whether the reviewer was a real customer.

Can bot traffic raise my website hosting bill?

On some platforms, yes. Vercel's shared responsibility model puts costs and resource consumption related to malicious traffic on the customer and lists setting a spending cap among the customer's duties. Vercel's bot protection ruleset is available on every plan but ships switched off, and until someone turns it on, matching traffic reaches the site. Cloudflare's free plan allows one rate limiting rule, which someone has to write.

Should I block bots from my website?

Only the harmful ones. Many bots do useful work. Googlebot, which made 4.5 percent of web page requests in Cloudflare's 2025 count, builds the index a site needs to appear in Google, and Google lists an unblocked Googlebot as a minimum requirement for indexing. Uptime monitors are bots too. Attacks such as a bot guessing passwords on a login page are what firewall rules and rate limits are meant to stop.

Latest from the Ledger

Continue reading

Return to the Ledger →