How Much of the Internet Is AI? What the Flood Means for B2B

I keep seeing the same argument about the influx of AI websites. Some people call it AI slop, others point to the dead internet theory, and everyone has a number: 10%, 24%, 35%, 50%, 74%.

So I went and checked. My job is AEO and SEO, and I don't like repeating a stat I haven't traced back to where it came from. All five of those numbers come from real research. They just answer different questions!!

Here is what I did, what I found, and what I would do if I ran a small B2B company right now.

What I checked

• The studies. I went through the main research on how much of the web is AI-written, plus what it means for rankings and AI citations.
• Our Search Console. We relaunched deligatr.ai on October 24, so I pulled the week after it.
• Our Cloudflare logs. I wanted to see which AI crawlers actually visit a site our size.
• Our own audits. I went back through our 12 industry AI visibility audits.

How much of the internet is actually AI?

My honest answer: no single figure is reliable. My best read is that roughly a quarter of web pages now contain AI-generated or AI-edited text (guilty, ours too!!). That comes from the largest study here, and the only one that corrects for detector error. Here is how it compares with the numbers you will see quoted.

Three things explain the spread, and once you see them the panic numbers calm down fast.

Old pages lower the average. Pew's 10% covers pages that existed long before ChatGPT. Among pages with a publication date after ChatGPT's release, Pew found signs of AI authorship on over one-third.

"Article" is narrower than "page." Graphite samples articles and listicles only. Product pages, directories and the rest of the web are outside its count.

"Contains AI" differs from "written by AI." This is where the 74% figure gets misused. Ahrefs found 74.2% of new pages contained some AI text, but only 2.5% were pure AI. Most were a blend. It also dates from April 2025, so any post calling it a current figure is out of date. If you see "74% of the web is AI," the source says something much narrower. I see this one quoted constantly, and it bugs me.

Detectors deserve some caution too. Pew warns they can misclassify individual pages, and Ahrefs says outright that no detector is perfect. The Visokay team calls a detector score a proxy, and their overall estimate moves a lot under the most extreme modeling choices. I treat every figure here as a range, and you should too.

Why the rise is speeding up

Visokay and colleagues estimate that the share of pages with AI text grew by 4.6 percentage points a year from early 2023 to January 2025. Since then it has grown by 8.1 points a year. Graphite's narrower count of primarily AI-generated articles, by contrast, has held at roughly 50% since early 2025.

Building a site has become cheap as well as writing one. Lovable, one of the AI app builders, claimed more than 100,000 new projects a day on its platform in December 2025. That is a company figure, and a project is not the same as a public website, but it shows how low the cost of shipping something has fallen. Registered domains reached 401.6 million at the end of the second quarter of 2026, up 8.1% from a year earlier, though Verisign gives no cause and we would not pin that on AI alone.

The flood is concentrated in the long tail

This is the finding that stopped me, because it matters most for smaller companies.

Visokay and colleagues split the web by how connected each domain is, using PageRank. The top tier is a tiny group, about 0.001% of all domains. The least central tiers include many small businesses, local organizations and personal sites.

The two lowest-connected tiers accounted for 78.8% of the net increase since December 2021. The authors are careful to say this is not a causal finding, and the paper has no breakdown by industry.

My reading is that an English-language .com for a small B2B company sits in the most crowded part of the web. I can't prove that from this data, but the direction is hard to miss.

What the research found, and what it didn't

The fear is that AI text makes the web less accurate. Dolezal and colleagues at Imperial College London, Stanford and the Internet Archive tested that directly and did not find it. They saw no significant link between AI likelihood and factual errors. They also found no significant drop in outbound links.

They did find two other effects. AI-flagged sites were about 33% more semantically similar to one another than other sites, and their average positive sentiment score was about 107% higher.

As a marketer, I read that as a very recognizable page: agreeable, confident and close to everyone else's. Accuracy is still your job, because Google's own guidance notes that generative models predict likely words instead of retrieving facts. But the measured penalty here is that you start to sound like your competitors.

Does AI content rank, and does Google penalize it?

AI content can rank, but it is under-represented at the top. Graphite studied 31,493 keywords in June 2025 and found AI-generated articles were 14% of what ranked in Google's first two pages and 7% of number-one results. In citations from ChatGPT and from Perplexity, they were 18%. Graphite did not measure AI drafts heavily edited by humans, and it says these are patterns, not proof of cause.

AI engines clearly don't filter this content out. A Northwestern University audit of 712 questions across four AI engines found about 16% of scrapable cited pages were AI-generated: 27.8% for Copilot, 14.7% for Gemini, 9.4% for Perplexity and 7.3% for ChatGPT. The authors treat those as lower bounds, and the questions covered politics, health and the environment, not B2B vendors.

Google's published guidance does not ban AI content or say it is penalized for being AI-written. That guidance, last updated on October 1, says generative AI can help with research and structure. The warning is about generating many pages "without adding value for users," which may breach its scaled content abuse policy.

So what counts is the value of the page, whichever tool helped make it, and that gap shows up most clearly at the very top of the rankings.

What I found in our own data

Studies are useful, but I wanted to see this on a real site. So I opened our Search Console.

We relaunched deligatr.ai on October 24 and added 48 URLs to the sitemap. Before that, we were getting around 200 impressions a day. Five days later we hit 684!!

From October 24 to 30 we appeared for 365 different searches. 249 of them had never shown up in our data before, not once since September 1, and they made up 44% of our impressions.

Before anyone gets excited: clicks barely moved lol. We got 11 clicks that week, and our average position is still in the 30s, which is page three or four. Impressions are the first sign that Google is testing you. They are not customers.

The part I found most interesting was the shape of the new searches. 49 of them were seven words or longer, and they read exactly like something you would type into ChatGPT. We sat at position 3 for "who offers ai visibility audits for service-based businesses?" and position 4 for "which agencies offer a free ai visibility audit for b2b saas?"

Those are tiny numbers, a handful of impressions each. But my guess is that we show up because we published specific audit data on that exact topic, and very few sites have that. That is one site and one week, so I would not call it proof. It does match what the research above suggests.

Then I checked Cloudflare. In the week to October 24, it counted about 5,550 requests from AI crawlers, including OpenAI, Google, Perplexity, Anthropic and Apple, on a site most people have never heard of. About 2,910 got a real page, and around 2,040 hit a 404.

That 404 number worried me until I read the paths. The top ones were things like /.env and /admin/.env: scanners hunting for leaked config files while calling themselves AI bots!! They got nothing, but it is a good reminder not to take any "AI traffic" number at face value. Go and look at what is actually in there.

Buyers use AI to shortlist, then check your site

Here is why I care about any of this: buying has changed.

G2 surveyed 1,076 B2B software buyers in March 2026. It found 51% now start research with an AI chatbot more often than with Google, up from 29% a year earlier. And 69% said they chose a different vendor than planned because of chatbot guidance. G2 sells review data, so weigh that with its interest in mind, but its survey also reports that 64% often or very often run into inaccurate AI recommendations. Buyers notice.

So they check. Semrush surveyed US B2B professionals who use AI at work and reported what they do after an AI recommendation. The figures add up to more than 100%, so many take several of these steps.

Now set that beside what we found across our 12 industry audits, from law firms to manufacturers. This is the number I keep coming back to.

So your website and your outside reputation do different jobs. AI answers about a category are built mostly from third parties, so your own site rarely gets you named. Once you are named, your site is where most buyers go to check you out. A generic page can lose them right there, in the short window after the AI put you on the list.

What I would do if I ran a small B2B company

This is my opinion, and Deligatr sells AEO services, so read it with that in mind.

1. Get named where AI reads. Across our audits, that means Wikipedia, directories, review sites and trade media more than your own pages. In G2's survey, 45% called review-site citations the most confidence-inspiring signal in an AI answer.
2. Build the page a buyer checks. Assume they arrive from a chatbot with one question: is this claim real? Show your pricing if you can, your process, your named people and specific proof.
3. Publish what a model can't average. Original data, client numbers you are allowed to share, and first-hand mistakes are hard to generate. Graphite's number-one results were overwhelmingly human-written. Our guess is that those pages had something new to say, though the study can't prove it.
4. Run the swap test. Put a competitor's name in your homepage headline. If the sentence still works, it isn't saying anything about you. Then state who you are not for. Sentiment data suggests AI drafts lean positive, and honest limits are the fastest way to sound like a real company.
5. Keep AI, and keep a person accountable. We use AI in our own work, and a person approves what goes out. Let it research and draft. Then have someone who knows the business check every claim, because a buyer's next step may be asking an AI to verify it.

What this data can't tell you

Three of the sources here need extra care. The Visokay, Dolezal and Allaham papers are preprints, and we found no sign that any of them has been peer reviewed. Pangram, whose detector appears in several of these studies, gave Dolezal's team a grant and API access, though the authors say it had no role in the design or the decision to publish.

Graphite, G2 and Semrush all sell into this market, and so do I. None of the prevalence studies breaks results out for B2B, and the citation audit used general-interest topics. Our own audits show who AI cites, not why it chose them.

My own numbers are one site, one week, right after a relaunch, so read them as an early signal and nothing more.

The direction across these sources is consistent, even where the exact numbers aren't. There is more generic content than ever, and buyers have new ways to test what you claim. Specific, checkable proof gets more valuable as a result.

FAQ

Related reading

• Where ChatGPT, Perplexity and AI Overviews get their citations
• What is answer engine optimization (AEO)? The complete guide
• We audited 30 B2B SaaS companies for AI visibility
• AI visibility research: every audit in one place
• B2B content marketing: the case for answer-first writing

Sources

• Google Search Console and Cloudflare AI Crawl Control data for deligatr.ai, pulled October 24 to 30, 2026
• Pew Research Center, How Much of the Internet Is Written With AI?, August 20, 2026, and Methodology
• Visokay et al., The uneven rise of AI-generated web content, arXiv preprint, October 27, 2026
• Dolezal, Alam, Graham & Bohacek, The Impact of AI-Generated Text on the Internet, arXiv preprint, April 2026
• Graphite, AI Now Writes as Many Online Articles as Humans, 2026
• Graphite, How Does AI-Generated Content Perform in Search and Answer Engines?, October 2025
• Ahrefs, What Percentage of New Content Is AI-Generated?, May 2025
• Allaham & Diakopoulos, Synthetic Sources? Auditing Generative Search Engine Citations for Evidence of AI-Generated Sources, arXiv preprint, May 2026
• Google Search Central, Using generative AI content on your website, last updated October 1, 2026
• G2, Half of B2B Software Buyers Now Start Their Research With AI Chatbots, April 15, 2026
• Semrush, How AI tools shape the B2B buying process, July 8, 2026
• TechCrunch, Vibe-coding startup Lovable raises $330M at a $6.6B valuation, December 18, 2025
• Verisign, Q2 2026 Domain Name Industry Brief
• Deligatr, AI visibility research across 12 industries