AI Is Killing Ad Revenue. That's Also AI's Problem.

The numbers coming out of the media industry right now are genuinely ugly.

Digital Trends went from 8.5 million monthly visitors to under 265,000. That's a 97% collapse. Stereogum lost 70% of its ad revenue. Business Insider cut 21% of its staff after organic search traffic fell 55% in three years. HuffPost lost roughly half of its search referrals. Ten major tech publications that collectively pulled 112 million US visits per month at their peak were down to under 50 million by January 2026.

These aren't edge cases or bad luck. This is a structural shift happening across the entire ad-supported web, and AI is the primary driver. Zero-click searches now account for 60% of all Google queries. When AI answers the question on the results page, nobody clicks through. No clicks, no page views. No page views, no ad revenue.

And here's the part that should make the AI companies nervous: if this keeps going, the web that trained them starts to rot. This isn't just a media industry problem. It's everyone's problem.

Jump to Section

• How Ad-Supported Media Actually Works
• What AI Overviews and Zero-Click Search Are Doing
• The Self-Destruct Loop Nobody Talks About
• Can Anything Fix This?
• What Types of Sites Actually Survive
• FAQ

---

How Ad-Supported Media Actually Works {#how-ad-supported-media-works}

The model is simple and has worked for 25 years: publish content, attract visitors, show ads, collect money.

A publisher earning $10 CPM (cost per thousand page views) on a million monthly visitors makes $10,000 a month from display advertising. Scale that up and it's a real business. Scale it down to a niche blog and it's a side income that covers hosting, maybe a writer or two, maybe a decent salary if traffic is strong enough. Either way, the entire thing collapses without one ingredient: people actually landing on the page.

Subscriptions work for a handful of outlets with massive brand loyalty, the New York Times being the obvious example. But most publishers can't get enough paying subscribers to offset lost ad revenue. The long tail of the web, niche sites, technical tutorials, independent journalism, developer resources, all of it runs on advertising or it doesn't run at all.

This is the foundation AI is currently eating.

---

What AI Overviews and Zero-Click Search Are Doing {#zero-click-search-ai-overviews}

Zero-click search isn't new, but AI has turbo-charged it. Rand Fishkin was documenting zero-click behavior back in 2019 when roughly 50% of searches ended without a click. By 2026, that number is 60%, and it got there fast. Google's AI Overviews, rolled out through 2024 and expanded aggressively in 2025, pushed zero-click from 55% to 60% in roughly 18 months, the largest single-year jump in the history of that metric.

When an AI Overview answers a question completely, there is no remaining reason for the user to click anywhere. They got what they came for. Click-through rate drops from 15% to 8% when an AI Overview is present, according to Pew Research. And only 1% of searches lead to a user clicking a link inside an AI Overview. The citation is there. Nobody clicks it.

Google isn't the only one doing this. ChatGPT handles an estimated 12% of Google's query volume but sends 190 times less traffic to external websites than Google does. Perplexity, Claude, Gemini, all of them are optimized to keep you in the conversation, not to send you somewhere else.

The IAB Tech Lab estimates AI-powered search summaries have reduced publisher traffic by 20% to 60% on average, with niche publications seeing losses approaching 90%. Across the 500 most-visited publishers globally, traffic dropped 27% year over year, representing roughly 64 million fewer monthly visits. The financial translation of that is approximately $2 billion in annual advertising revenue, already gone from the publishing sector.

And Google just announced at I/O 2026 that it's overhauling Search completely around AI mode with conversational follow-ups and autonomous agents. Every feature they announced further reduces the need to click through to a source. One SEO analyst called it a "devastating impact on the internet" and honestly that doesn't sound like an exaggeration.

Here's the part that stings though: Google's own total revenue hit $109.9 billion in Q1 2026, up 22% year over year. The company is doing great. It's just doing great by keeping users on Google instead of sending them to publishers. Google Network advertising revenue, the part that flows to external publisher sites, fell 4% in Q1 2026. Google wins, publishers lose.

---

The Self-Destruct Loop Nobody Talks About {#self-destruct-loop}

Here's where this gets legitimately weird.

AI models need training data. High-quality, human-written, well-sourced training data. The kind that has been produced for decades by funded journalists, expert writers, independent bloggers, and technical communities. Around 80% of the tokens used to train GPT-3 came from Common Crawl, a massive dataset scraped from the public web. More than 60% of all LLMs published between 2019 and 2023 used Common Crawl as a primary training source. The open web is the foundational ingredient.

When media sites lose ad revenue and shut down, or retreat behind paywalls that crawlers can't access, that ingredient supply shrinks. What fills the gap? AI-generated content. Cheap to produce, everywhere, and a serious problem for future model quality.

There's a documented failure mode called model collapse. When AI systems are trained increasingly on AI-generated content instead of human-written material, output quality degrades. The models produce text that's fluent and grammatically correct but becomes more generic, more repetitive, and more confidently wrong on nuanced topics over time. Researchers studying this have put it directly: "without enough fresh real data in each generation, future generative models are doomed to have their quality or diversity progressively decrease."

The feedback loop is straightforward. AI kills publisher traffic. Publishers lose ad revenue and shut down or paywall their content. The open web fills with lower-quality synthetic text. Future models train on that recycled garbage. Model quality declines. Users trust AI outputs less. The whole ecosystem degrades.

The Brookings Institution connected these dots explicitly: the large language models that scraped their training data from the open web are now using that data to eliminate the need to visit many of those same websites. AI is eating the hand that fed it.

The Nieman Lab put the licensing situation bluntly: as long as Google's search indexing and AI training operate as a unified system, publishers have no meaningful leverage, and the effective market price for their content is zero.

---

Can Anything Fix This? {#can-anything-fix-this}

There's no clean solution here, but several things are being tried.

Content licensing deals. OpenAI has signed agreements with AP, the Atlantic, the Financial Times, and Reddit. The idea is structured like music royalties: AI companies pay for the right to use content. The problem is these deals cover a thin slice of the web, the terms are often unfavorable, and Nieman Lab's analysis of where this is heading in 2026 is basically: don't count on meaningful revenue at scale.

Bot-level paywalls. Services like TollBit let publishers charge AI crawlers separately from how human visitors experience the site. Publishers can set a price for machine access and create different content versions for AI agents. Arc XP integrated TollBit in April 2026 specifically to help mid-size publishers who can't negotiate directly with the big AI companies. The Philadelphia Inquirer is piloting it. The concept is sound but it only works if it reaches critical mass.

AI-native advertising. Perplexity already runs sponsored answers inside chat responses. This is essentially the search ad model rebuilt for conversational AI. If publishers eventually get a cut when their content is the source behind an AI answer, that's a real revenue model. The infrastructure for that doesn't fully exist yet, and the economics aren't clear, but it's the most interesting path forward.

Lawsuits. The New York Times vs. OpenAI is the flagship case but it's not alone. These will take years, outcomes are genuinely uncertain, and even favorable settlements probably won't make publishers whole at scale.

Going full subscription. Some publishers are paywalling everything and accepting that AI crawlers won't be able to access it. This works for brands with genuine loyal audiences. It also accelerates the hollowing out of the free open web.

---

What Types of Sites Actually Survive {#what-types-of-sites-survive}

The sites in the most trouble are the ones that were essentially aggregating, summarizing, or repackaging information that already existed elsewhere. AI does that better, faster, and for free. If your value proposition is "we explain things," you have a serious problem.

What AI genuinely cannot replace is harder to define but you know it when you hit it: original reporting that required someone to show up somewhere, firsthand expertise from actually doing a thing rather than reading about it, community built around real relationships, and tools that people use rather than just read.

That last category matters a lot in developer contexts. An article explaining how JSON formatting works can be absorbed and summarized by AI in two seconds. A JSON formatter you actually run in your browser cannot. A guide to HTTP response headers can be paraphrased. An HTTP headers inspector that shows you live headers on any URL in real time is a different kind of thing entirely. The diagram editor here exists to be used. You can't summarize away a tool that does something.

You can check API structures and HTTP responses with tools that give you actual data rather than an AI's best guess about what headers a server is returning. That's the distinction that matters: can AI fully replace this, or does it require something interactive, real-time, or genuinely original?

If the honest answer is yes, AI can replace it completely, you're already on borrowed time and the traffic numbers probably show it.

---

FAQ {#faq}

Is AI actually causing media sites to lose traffic?
Yes, measurably. Chartbeat data from early 2026 shows small publishers down 60% in search referral traffic over two years, medium publishers down 47%, and large publishers down 22%. The Reuters Institute tracked a 33% decline in organic Google search traffic globally between November 2024 and November 2025.

What is zero-click search and why does it matter for ad revenue?
Zero-click search happens when a user gets their answer directly on the results page without visiting any website. About 60% of Google searches now end this way. Since ad-supported sites earn money from page views, zero-click searches generate zero revenue for publishers even if their content was used to formulate the answer.

How much ad revenue have publishers actually lost because of AI?
IAB Tech Lab estimates roughly $2 billion in annual advertising revenue losses across the publishing sector from AI search summaries reducing publisher traffic by 20% to 60% on average.

What is model collapse and why should AI companies care?
Model collapse is a failure mode where AI systems trained on AI-generated content instead of human-written material degrade in quality over time. As the open web fills with synthetic content from dying or paywalled publishers, future models have worse training data to learn from and produce lower quality outputs. AI companies have a direct financial stake in the health of the web they're currently helping to kill.

Are AI companies paying publishers for their content?
Some are. OpenAI has licensing deals with several major publishers. But according to Nieman Lab, as long as Google's search indexing and AI training function as a unified system, publishers have no real leverage, and meaningful licensing revenue at scale is unlikely in the near term.

What kinds of sites are most at risk from AI traffic loss?
Sites built around aggregating, summarizing, or repackaging information that already exists elsewhere face the highest risk. AI does that job better and free. Sites that produce original reporting, offer real functional tools, have loyal communities, or provide expertise from genuine hands-on experience are more defensible.

What is ChatGPT's impact on publisher referral traffic?
ChatGPT handles roughly 12% of Google's query volume but sends 190 times less traffic to external websites than Google does. Even as AI referral numbers grow, they're nowhere near offsetting what publishers have lost from traditional search.

Can paywalls protect publishers from AI traffic loss?
Partially. Paywalled content is harder for AI crawlers to legally scrape, which protects it to some extent. But it also removes content from the open web and cuts off new reader discovery through search. It works better for established brands with existing loyal audiences than for smaller or newer publishers.

What is TollBit and how does it help?
TollBit lets publishers charge AI crawlers for content access separately from human visitors, and create different content versions specifically for AI agents. Arc XP integrated TollBit in April 2026 to help mid-size publishers access bot-level monetization without needing to negotiate directly with AI companies.

What content actually survives the AI transition?
Original data and reporting, genuine expert knowledge from people who have actually done the thing, community and real relationships, and interactive tools that users run rather than just read. If AI can fully replace it with a paragraph of text, the content is already at risk. If using it requires actually doing something, that's defensible ground.