In my August post I went through the research on what has been happening to the open web and concluded that the internet is becoming an app; the web of pages still exists, but it is increasingly a back end rather than a destination. Three months on, the evidence points somewhere starker. We are entering the age of the new internet, where AI writes the answer and the page that supplied it often goes unvisited.
An app is still somewhere you go. In the new internet, you increasingly don’t go anywhere at all. The web is not being bypassed so much as repackaged, it’s being read by machines, summarised, and served back as an answer, usually without the reader ever arriving at the page it came from. And it is happening in the most ordinary place imaginable – the Google results page, which for most people is still the front door to the internet.
Last time I deliberately left out journalism and industry reports and stuck to academic work. I am going to break that rule here, because the most useful piece of work I have read on this since August is an industry analysis. RevScope, a company that sells web analytics, published a review of search, SEO and GEO on 9 November, and it rests on unusually solid foundations – a randomised controlled trial, Pew Research Center’s logs of real browsing, and DataForSEO’s monthly panel of 20 million keywords. The authors are also candid about their own position. They sell analytics, and they point out that almost every large dataset in this field comes from a company selling SEO data or software. That kind of honesty about where the numbers come from is rarer than it should be, in industry and in academia.
The DataForSEO data at the heart of RevScope’s analysis shows how quickly the answer box has spread. In September 2024, 5.93 percent of the Google results it tracks showed an AI Overview. By August 2026 the figure was 61.43 percent. Over the same two years the featured snippet, the quoted extract from a web page that used to sit at the top of the results, all but vanished, falling from 8.96 percent to 0.44 percent. Pew’s independent logging of real searches lines up with DataForSEO’s numbers for the months where the two overlap, which gives me some confidence this is not an artefact of one company’s method.
The detail that stopped me was the breakdown by intent. In August I wrote about a study from London Business School and UCLA which found that people who started using chatbots kept using search for navigation, typing a site’s name to get there, even as their question-style searches fell away. Search survived as an address book. AI Overviews now appear on 34.03 percent of navigational searches, almost exactly double the 17.04 percent of late 2025, and on 70.44 percent of informational ones. Even the address book now comes with an AI answer on top.
Nor is the long tail any safer, and the long tail is the generative part of the web in Zittrain’s sense, which is the part that welcomes change from anyone, anywhere. In the third quarter of 2026, keywords with between 201 and 300 monthly searches showed an Overview 58.59 percent of the time, against 48.62 percent for those searched more than 600 times a month. Health and finance both sit close to 70 percent, and those are exactly the areas where I would most want people to read the original source.
Prevalence is one thing; the real question is what an Overview does to the clicks. Here RevScope leans on the strongest evidence available, a preregistered field experiment by Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon University. They gave just over a thousand US participants a Chrome extension that randomly hid AI Overviews from some of them, then followed two weeks of their real searches. Where an Overview appeared, clicks through to outside websites fell by 39.8 percent, and the share of searches that ended without anyone leaving Google rose from 54 percent to 73 percent. Clicks on Google’s own ads did not change, which RevScope reads as a sign that Google is careful not to cannibalise its own revenue. Averaged across all searches, with or without an Overview, the authors put the loss at roughly 18.5 percent of the organic clicks a site gets from Google, and they expect it to grow as Overviews spread.
What stays with me is that the people who could see the Overviews were no happier for it. They rated their satisfaction, the quality of the information and the ease of finding it almost identically to the people who couldn’t. A second preregistered experiment, by Stephanie Wang, Jeffrey Gleason and colleagues at the University of Pennsylvania and Northeastern, points the same way – removing AI Overviews and AI Mode increased clicks to publishers, while routing every query through AI Mode reduced them and eroded people’s trust in what they found. Neither paper has been through peer review yet, and the first covers desktop only, but on a question like this randomised evidence is as good as it gets.
It is a different study, with a different method and a different question, and yet that 39.8 percent sits remarkably close to the 41.8 percent fall in traffic to the bottom quarter of websites that the London Business School and UCLA team found among chatbot adopters, which I wrote about in August. I would not make too much of the coincidence. What matters is that the direction has now been confirmed by a randomised experiment.
The observational data agrees. Pew found that people clicked an ordinary result on 8 percent of visits to a results page with an AI summary, against 15 percent without one, and that they clicked a link inside the summary itself on just 1 percent of visits. In DataForSEO’s panel, the click-through rate of the top organic result on desktop fell from 17.83 percent in September 2025 to 7.43 percent in August 2026. RevScope is careful to point out that the underlying report does not establish what caused that drop, but the practical meaning is blunt. The top spot on desktop now brings in well under half the clicks it did twelve months earlier.
Google disputes the whole picture. In August 2025 its head of Search, Liz Reid, wrote that total organic clicks had held roughly steady over the year and that Google was sending slightly more “quality clicks”, by which she meant visits where people do not come straight back to the results. She published no figures, and Google is the one organisation that could settle the question with its own data. The Agarwal and Sen experiment is the nearest thing we have to an independent test of that claim. People who clicked through after seeing an Overview spent about as long on the sites they visited, and bounced about as often, as people who never saw one. Fewer clicks, in other words, and no better.
If people read the summary and stop there, the question becomes who the summary cites, and the answer is telling. The five most cited domains in Google’s AI Overviews are YouTube, Reddit, Facebook, Wikipedia and, in fifth place, google.com itself. ChatGPT leans more on Wikipedia, Reddit, Fandom, Britannica and the US National Institutes of Health, and in 27.3 percent of its answers it cites nothing at all. Where there is a citation, it has become a stand-in for the visit.
Wikipedia shows what that means in practice. In August I noted in passing that the Wikimedia Foundation had reported a decline in human page views, without going into the reasons. The Foundation’s own explanation is worth knowing. The fall was roughly 8 percent against the same months a year earlier, and it put the decline down to generative AI and social media, and especially to search engines answering questions directly, often from Wikipedia’s own content. In DataForSEO’s sample, Wikipedia is now the most cited source in ChatGPT’s answers. As the Foundation put it, people are still reading what its volunteers write, all over the internet; they just aren’t visiting Wikipedia to do it. Fewer visits, it warns, may mean fewer volunteers to grow the content and fewer individual donors to pay for it.
The AI assistants, meanwhile, send very little back. Ahrefs data cited by RevScope found that assistants accounted for 0.28 percent of traffic across 74,752 websites in March 2026 – about 3.5 million visits, against 345 million from Google. That share is growing, and it is probably undercounted because some visits from apps are logged as direct traffic, but at roughly a hundred to one it is nowhere near enough to make up for what is being lost on the results page.
News publishers are already living with the consequences. Chartbeat data gathered for the Reuters Institute’s annual trends report showed organic search traffic from Google to more than 2,500 news sites falling by a third worldwide between November 2024 and November 2025, and by 38 percent in the United States. The 280 media leaders surveyed expect search traffic to fall by more than 40 percent over the next three years. The report is fair about the uncertainty and does not pin all of the decline on AI Overviews, but publishers are not waiting to find out. Many say they will produce less of the service and evergreen content that they expect chatbots to answer instead, even when the answer was originally theirs.
The industry’s response has a name – GEO, generative engine optimisation, the craft of getting your content cited inside AI-written answers. RevScope’s verdict is that, beyond ordinary SEO, only two things reliably seem to matter – showing up on the third-party sites that AI answers lean on, and writing pages that are easy to quote. They found little support for anything else, and the much-discussed llms.txt file, a proposed way of telling AI systems what a site contains, has shown no measurable effect. I find that oddly reassuring. The things that do seem to help are original data, named sources, and pages that are kept up to date and actually answer the question. That is what made a good web page in the first place. The full piece is worth reading for the detail, including a careful SISTRIX study that compares cited and uncited pages from the same sites.
But good craft is not a business model. A citation is not a visit, and it is visits that pay for journalism, bring new volunteers to Wikipedia, and give a small new site the chance to find its first thousand readers. If the answer engine takes the content, writes the answer, keeps the reader and loses none of its own ad clicks along the way, the open web is reduced to an unpaid supplier. That is the bargain at the heart of the new internet. In August I ended by saying we should go into this next phase with our eyes open about what is being wrapped and who holds the wrapper. The answer is becoming clearer. What is being wrapped is the web itself, and the wrapper is a single composed answer, chosen and phrased by a model, in place of a list of sources we could weigh for ourselves.
Zittrain’s fear in 2008 was that an internet anyone could change would be locked down into appliances that only their vendors could modify. What is happening is stranger. The web is still open, still being written and still being read; it is just increasingly read by machines, on our behalf. There is a technical irony in that. In 2024 Ilia Shumailov and colleagues showed in Nature that generative models trained indiscriminately on the output of other models degrade over successive generations, and that the first thing to go is the tails of the original distribution: the rare and the unusual. Genuine human-produced data, they concluded, becomes more valuable as generated content spreads, not less. The long tail of the web (the expert forum post, the hobbyist’s meticulous guide, the obscure personal blog) is precisely the part that is losing its readers. If the people who write those pages stop because nobody arrives, the answer engines will lose the very material that makes them worth asking.
So what follows for me as an educator? I direct computer science and data science degrees and teach HCI and UX, and two things stand out.
The first is about the students who will build the next version of the web. In August I argued that the unit of design is shifting from the page to the action. The evidence since then adds a second reader to every page we design. Our graduates will be writing for people and for the models that summarise their work for people, and they need to understand that being cited and being visited are now different outcomes, with very different economics. RevScope’s closing advice, to judge the work by revenue rather than by rankings, comes, naturally, from a company that sells revenue analytics. Underneath the pitch, though, is a point about honest measurement, and that is something we can teach.
The second is the shift from recall to judgement that I keep coming back to on this blog. Judgement depends on being able to check. When 73 percent of searches with an Overview end without a click, and people click a link inside a summary on 1 percent of visits, the habit of checking a source is withering at exactly the moment it matters most. I want students to treat a summary as a claim to be tested rather than an answer to be accepted, and that means actually opening the source. The Wikimedia Foundation’s advice to everyone is disarmingly simple, thy tell you that when you search, look for the citations and click through to the original. It is the cheapest support for the open web there is, and the data says almost nobody does it.
I am aware of the irony of writing all this on a small personal blog, which is about as long-tail as the web gets. Some of the people who might once have found this post through a search will now meet it, if at all, as a sentence inside somebody else’s answer. As I said in August, I am not only nostalgic, and the answers are often genuinely useful. But the machines need a living web to summarise, and at the moment the people keeping it alive are paying for it in lost readers. The question for the next few years is not whether AI will answer our questions, it already does, instead it’s if anyone will still be writing the pages it answers from.
Sources for this post, for anyone who wants to check my reading:
- Agarwal, S. and Sen, A. (2026). The Impact of Google AI Overviews on Publisher Traffic and User Experience: Evidence from a Field Experiment. Working paper, SSRN.
- Chapekis, A. and Lieb, A. (2025). Google users are less likely to click on links when an AI summary appears in the results. Pew Research Center.
- Miller, M. (2025). New User Trends on Wikipedia. Diff, Wikimedia Foundation.
- Newman, N. (2026). Journalism, Media, and Technology Trends and Predictions 2026. Reuters Institute for the Study of Journalism.
- Padilla, N., Lam, H. T., Lambrecht, A. and Hollenbeck, B. (2025). The Impact of LLM Adoption on Online User Behavior. Working paper, SSRN / AEA 2026.
- Reid, L. (2025). AI in Search is driving more queries and higher quality clicks. Google.
- RevScope (2026). Google’s AI Overviews cut clicks to websites by 40%. RevScope blog, 9 November 2026.
- Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R. and Gal, Y. (2024). AI models collapse when trained on recursively generated data. Nature, 631, 755-759.
- Wang, S. T., Gleason, J., Bart, Y., Wilson, C. and Metaxa, D. (2026). AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence. arXiv:2608.18352.
- Zittrain, J. (2008). The Future of the Internet and How to Stop It. Yale University Press.

Leave a Reply