It’s the End of the Web as We Know It (And I Don’t Feel Fine)

On August 4th, Katie Baker published an article in the Ringer declaring the demise of Google search (and search generally), and with it, the end of the internet as we know it. In this article, Baker describes the effects of Google’s pivot away from search in favour of “AI Overviews,” where AI summaries are delivered to the top of search pages, and “AI Mode,” which replaces search results altogether with a chatbox. With these changes, Google has displaced its core offering with services that, in Baker’s words, seek “to aggregate the entirety of, oh, all historical human knowledge into pithy (and not always accurate) bullet points.”

These changes may seem innocuous, in that users can choose to scroll past the AI summary to get to the search results or can engage in elaborate workarounds to remove AI from their search pages, but users are shifting en masse (if not by choice) from search results to AI summaries. Google reported in May 2026 that its AI Overview service had over 2.5 billion monthly users, and that its AI Mode had more than 1 billion, with usage growing rapidly. Baker argues that Google’s pivot to AI search and its related decision to tie its indexing for search to its indexing for AI summarization, has put content owners and hosts in a bind: they must allow Google to summarize their content (and potentially substitute for it) or have it withdrawn from Google’s search results altogether. Especially if the use of AI search is soon the primary way that people will seek information from the web, choosing to deny Google’s summaries might make content invisible. More recently, 404 Media reported on OpenAI and Microsoft communications unsealed in the plaintiff’s motion for summary judgment in the ongoing New York Times litigation, which featured a Microsoft executive describing this substitutive dynamic as a “doom loop” for the open web.

Even if the current experience of the web is, as Baker writes, a “depressing experience, filled with 404 errors and stubborn walled-garden paywalls and rotted links and shuttered publications and busted GIFs,” Google’s shift to AI search will only hasten the demise of an information ecosystem that once seemed boundless, and on which many organisations, institutions, and communities still rely. Technically speaking, Google’s AI search is very much built on its predecessor – the Googlebot index is repurposed for its Gemini LLM through what Google calls “Query Fan-Out,” whereby Gemini generates numerous sub-searches across its index to feed information back to itself in order to best “ground” its output to the content of its index. This “grounding” approach is like those of other LLM interface providers like OpenAI and Anthropic. The result? Bot traffic now outnumbers human traffic on the web, with the latest ratio in Cloudflare Radar showing a 64% to 36% advantage for automated requests over human ones as of October 6, 2026.

Google’s monopolistic dominance of search has in this respect given it a head start over other LLM interface providers, as it needs no separate crawled index to ground its AI outputs (OpenAI’s complex relationship with Microsoft allowed it access to Bing’s index of the web). Cloudflare’s “crawl-to-refer” ratio reports track the degree to which automatic crawls of the web actually produce a human click, and they show that Google is (unsurprisingly) ahead of other AI firms in terms of actually inducing a human referral – Google’s ratio is 4.8:1 compared with Anthropic’s 471:1. Users likely still come to Google expecting to click something, but in July 2025, the Pew Research Center published one of the first looks at how AI search in Google affects user behaviour and traffic to websites, and found that users presented with AI overviews clicked links about half as often as those presented with standard search result pages, and were more likely to end a search without clicking anything at all, or “zero-clicking.” Two field experiments in 2026 have backed up Pew’s findings, with one finding that use of AI Overview reduced outbound clicks by almost 40%, with another finding that the use of AI Mode reduced clicks even further, with neither finding users more satisfied. Academic publishers have found widespread declines in human referrals to their content, and are collectively working to integrate automated and agentic retrieval into the metrics which measure its value.

So what does this all have to do with copyright law, the nominal theme of this column series? Well, the answer lies in the always fuzzy claim that Google holds an “implied license” to index the entirety of the web to make it discoverable. In Field v. Google (2006) a court found that an author who could have blocked Google’s caching of his sites, but didn’t, had granted Google an “implied license,” and that caching was in any case a transformative fair use. Canadian law has offered slightly different routes to reduce liability for search, including statutory protections for internet intermediaries, as well as by offering protection from damages to providers of an “information location tool” via section 41.27 of the Copyright Act, which one court has read narrowly to protect tools that send users to content rather than tools that host it themselves (see Trader v CarGurus, 2017 ONSC 1841).

However, the introduction of AI into search has unsettled what seemed a settled question: the lawful scope of Google’s copying of the web. In the latest of several antitrust suits, Chegg and Penske Media argued that this implied bargain amounted to an agreement to exchange the right to reproduce content for resultant traffic, which Google broke by introducing AI search features that substitute its own generative answers for clicks. Google replied that there was no such deal and no promise of traffic, and that it could shape its search business however it liked. On September 30, Judge Mehta dismissed the publishers’ case, suggesting these were not questions for antitrust law but for policymakers and regulators (and, presumably, the public). Judge Mehta had earlier ruled that Google must make parts of its web index available to some competitors, potentially including other LLM providers (a ruling currently on appeal). The UK’s competition regulator took a different approach earlier this year, ordering Google to let UK publishers opt out of AI Overviews and AI Mode without also being removed from search.

While neither of these decisions directly resolve the copyright questions in indexing for generative summaries, both point towards possible legal or regulatory mechanisms to account for the effects of AI on the market for original works. In his decision in Kadrey v. Meta, Judge Chhabria’s “market dilution” theory suggested that AI training might fail fair use’s fourth factor test once the effects of the flood of LLM-generated works could be measured and submitted as evidence for a court to assess. The Authors Alliance argued, very persuasively in my view, that this theory expands copyright protection beyond expression to something like competition itself, since any new medium or work could be said to dilute the market for existing ones. For AI search, substitution may be even easier to measure, as it produces a summary of a specific work in place of a visit to it. But substitution itself isn’t infringement. In Canada, liability requires producing a substantial part of a work, which would mean a qualitative assessment under Cinar v. Robinson (2013), and it is very likely that summaries that convey facts and ideas in new words would often fall short of this standard, especially when AI firms have instructed their systems to quote sparingly and paraphrase, much as Google limited its snippets in prevailing in Authors Guild v. Google.

However, it is still unclear whether the scope of the questions that LLMs pose to existing copyright frameworks mean that we should expect continuity or possibly, some doctrinal changes down the road. For businesses, organizations, and hosts of works like publishers, libraries, and preservation repositories, especially those without the resources or inclination to litigate this rapid shift in the shape and nature of web traffic, the change to the web has posed all sorts of practical problems. These include managing bot traffic’s demands on servers, deciding whether and how to discriminate between human and bot traffic, and assessing the value of access to a given work or collection when traditional metrics used to do so are shrinking and novel standards for measuring LLM and agentic consumption haven’t yet been established. Blocking or discriminating access doesn’t come naturally to librarians, but increasingly, library users like authors, readers, organizations and communities are questioning the power imbalance resulting from the breaking up of the internet’s indexing bargain. If copyright can’t offer any quick or clear solutions to this imbalance for the time being, strategic approaches to, ways of thinking about, or outright rejections of the new shape will likely continue to emerge.

I’ve divided these emerging approaches into a quick typology: Close, Block, Poison, and Hide. Closing might mean the further re-embrace of paywalls for publishing, or the rejection of open licensing by scholars, or advocating for narrower user’s rights or new exclusive rights such as the “right to summarize.” But closing would also mean letting incumbents reset the terms of use of the web inequitably and encourage the creation of costly new licensing markets which may price out participation in the creation of new works for many, and would make the work of preserving the web more difficult. Blocking is already well underway, including via Cloudflare’s now default blocking and metering services, or via open-source tools like Anubis, which lets site owners wave through some bots while blocking or imposing tests on others. The UK regulator decision above is an example of Block being put into practice as a regulatory approach, as are components of the EU AI Act. Another fascinating and ironic example is the decision of Organisation for Transformative Works to gate access to its Archive of our Own platform. Poisoning – using tools like Nightshade and Glaze, to “scramble” what computers see in a visual work while maintaining its integrity to the human eye, is an active area of research within and outside of AI companies. Poisoning text, however, may be both more difficult and problematic, if even small tweaks in training data can lead model outputs to produce medical misinformation or propaganda, or because the lawfulness of “prompt injection” is still unclear. And finally, there is Hiding, which might mean more human cultural activity retreating to spaces outside of the reach of the crawl, such as to private Discord servers, decentralised platforms, in-person events, or the development of smaller-scale or alternate models for the dissemination of works. All of these approaches involve trade-offs, not least of which are the accessibility of these works to humans now and in the future, but in living with the end of the open web, many people and organisations may choose one or multiple of them while policymakers and the courts slowly deal with the social and cultural consequences of these new technologies.