Large Language Models Are Pushing the Web Toward Zero Clicks

A picture of how chatbot adoption leads to fewer web searches — and how that affects online publishers

Fewer than 1 in 3 Google searches in the U.S. still ends with a click on a blue link. According to SparkToro’s analysis of Similarweb data, in the first four months of 2026 as much as 68% of U.S. Google searches ended without the user clicking through to a publicly available website. 

You’ve probably read news accounts of the fallout. Business Insider’s organic search traffic fell 55% between April 2022 and April 2025, leading the news organization to cut 21% of staff. Vox Media began selling parts of its business in May 2026, due to the “decimation of search traffic.” And The Planet D, a travel blog, lost half its traffic in 2024 shortly after Google launched AI overviews, the AI-generated text summaries that appear at the top of some Google search pages.

“For publishers, Google Zero is already here,” Nilay Patel, editor-in-chief of The Verge, told The New York Times. Google Zero refers to a future when the search engine is no longer the gateway to the web, sending no traffic outward.

The mechanism of the web traffic collapse is clear. When people receive a summary for their query, they tend to stop searching. Pew Research Center tracked almost 70,000 Google searches by 900 adults in the U.S. in March 2025. Whenever an AI summary appeared, they found that users clicked a traditional result (a blue link) only 8% of the time. When no summary appeared, users clicked 15% of the time. They abandoned the session entirely on 26% of searches that contained a summary versus 16% without.

Watching as Everything Changes

Google used to be a directory that found the address where an answer lived and directed you how to get there. That trip was the business model of the open web. It spawned a gigantic industry (search engineer optimization) essentially to get websites onto the first page of Google results and forced businesses big and small to play the SEO game. But AI overviews have thrown a wrench into the works. 

Just as this sea change began, four marketing researchers studied how it was happening. A working paper by London Business School’s Nicolas Padilla, UCLA Anderson’s H. Tai Lam and Brett Hollenbeck, and London Business School’s Anja Lambrecht investigated what happened as people began to adopt large language models. 

The researchers tracked the desktop browsing of over 1 million U.S. users across 2022 and 2023, using the Comscore Web-Behavior Panel. In that panel, they found 2,041 users who began using ChatGPT, Bing Chat (now Copilot) or Bard (now Gemini) in the tools’ first year, and followed what became of the rest of their internet use.

This was a window into the early days of chatbots, given that ChatGPT was only introduced at the end of November 2022. The data documents use of first-generation tools, all of which invented facts freely. Nonetheless, people still began building a new habit around them.

Nothing — for Five Months

Padilla, Lam, Hollenbeck and Lambrecht defined LLM adoption as making at least one chatbot visit per week over three consecutive weeks. Because people began using the LLM-powered chatbots at different points across the year, those who had not yet started served as the comparison group for those who had.

The researchers’ first finding was noteworthy. Nothing happened.

For about five months, users’ search behavior was indistinguishable from before LLM adoption on a statistical basis. People had a new tool, used it every week, and went on googling pretty much as they always had.

The decline in their search traffic began around week 20. And it didn’t stop. Searches in traditional engines fell 21.7% relative to their pre-adoption average. This differs from the traffic reports for websites above since it measured how much less a single user searched, but a modest change per person is what becomes a severe decline per website once most of the audience has made it.

The authors interpret the 20-week delay as a learning period. It may represent time spent by users to discover which questions a chatbot handled well, and only after that experimentation did they stop double-checking answers with a web search. The data showed searches built on question words (how, what, why, where, when, which, who) fell sharply. The navigational searches where someone typed a website name to find it, held steady. This suggests users tended to keep using web search as an address book but stopped asking it questions.

Total browsing, which the researchers measured as the calls a browser makes to load pages, barely changed, but its distribution shifted enormously. If the dataset was sorted by most website traffic and cut into four equal slices, the top quarter of all web traffic included just seven websites, while the bottom quarter was spread across some 4.5 million. Traffic to the top 500 websites showed no significant decline, while calls to the generally small websites in the bottom quarter fell 41.8%.

Traffic to display-ad servers — these are the visual banners and images placed on third-party websites — fell significantly after LLM adoption. The drop was concentrated across users who shopped online the most, and this is the very audience advertisers pay premiums to reach. Search ad exposures were a different story. These ads continued riding the navigational queries that never went away, so they held steady.

Keeping the Questions

The study’s LLM adopters stopped entering their questions into Google’s search box and just entered them into the chatbot. Google noticed and, to get them back, it made its search box work like a chatbot, answering questions.

Clickstream data from Datos, a Semrush company, shows query length moving away from short keyword entries. The fastest growing query length was in the six- to-nine-word range and suggests people phrasing searches as though prompting a chatbot. With AI overviews appearing on 57.9% of question queries, they basically were prompting a chatbot. Ahrefs found that approximately 46% of queries containing seven words or more triggered an AI overview, compared with about 9.5% of single-word queries.

From this data, it’s plausible to make the interpretation that Google has been absorbing some chatbot-style behavior by answering longer, question-shaped searches directly. Since AI overviews require no special tool adoption, they may also shorten or eliminate some of the learning period observed in the study. Meaning people adopt LLM use faster. Also, the LLMs have gotten more accurate.

These changes could be why the 2023 web traffic decline looked survivable and 2024 did not. In 2023 the visible casualties were sites whose whole function was substitutable. These included Stack Overflow and the homework-help company Chegg, whose stock collapsed that May. Those websites served users who overlapped almost exactly with the tech savvy and the students who were early adopters of LLMs. For other websites, the adopters were too few and scattered to register at first.

Padilla, Lam, Hollenbeck and Lambrecht’s study noted that Wikipedia appeared to be unaffected in 2023. The authors suggested that its survivorship could be credited to being an authoritative source that a model that invents facts cannot replace. In 2025, though, Wikimedia Foundation reported human page views to be down roughly 8% year-over-year, something they blamed on generative AI and social media. 

What Google Says

Google’s defense is that AI overviews  don’t appear on every search. True, as far it goes. AI overviews appear on about a fifth of results pages. Where they don’t appear is mostly where a summary wouldn’t be helpful. For example, type in a navigational query like “reddit” and you get Reddit’s website instead of a summary. Of the keywords that do trigger AI overviews, 99.9% are informational and 0.1% navigational. Where AI overviews do appear is for the very searches the open web was built to answer.

Google’s answer-in-place may be moving into query types it once left alone. Semrush, tracking more than 10 million keywords through 2025, found the informational share of AI overview triggers falling from 91.3% in January to 57.1% by October, while navigational triggers rose from 0.84% to 10.33%. 

Organic search’s share of traffic at The New York Times slid to 36.5% in April 2025 from almost 44% three years earlier; Dotdash Meredith went from about 60% of its traffic arriving from Google in 2021 to roughly a third. 

AI overviews are comparatively rare on news queries, which is Google’s standing rebuttal, but Seer Interactive found that even queries without an AI overview lost 41% of organic click-through year-over-year. The cause of that change is unclear.

Corresponding to what the researchers’ study found regarding display ads falling while search ads did fine, Alphabet’s third-quarter 2025 report showed revenue declining year-over-year in Google Network, which places ads on other companies’ sites, while search advertising grew. 

For small creators, the losses are becoming terminal rather than just painful. Bloomberg reported that at least three businesses invited to a Google creator summit in October 2024 have since shut down. Raptive, which represents 5,700 creators, estimates that sites will lose a quarter of their traffic to AI overviews.

Google prefers to talk about total search volume. Liz Reid, its head of search, wrote in August 2025 that click volumes remain “relatively stable” in the billions and that reports of decline rest on “flawed methodologies.” 

On June 3, 2026, Britain’s Competition and Markets Authority ordered Google to attribute publisher content with clear links inside AI answers and to let publishers opt their content out of being used in Google Search’s AI features altogether. However, Google has up to nine months to implement it. The order followed the CMA’s designation of Google as having strategic market status. This is a finding that the company has substantial and entrenched market power in a strategically significant UK digital activity, not that it broke the law.

Regulators aren’t the only ones putting pressure on Google. Bloomberg columnist Parmy Olson reports that Cloudflare, the web infrastructure giant, intends to block Google’s mixed-use crawlers by default beginning in September 2026 for ad-supported customers.

Websites have been in an unfair situation for years. If they blocked Google’s bots to prevent their content from being used for AI training or AI answers, then they would vanish from search results. The mounting regulatory and private company pressure has pushed Google to offer a concession. They are now testing an opt-out feature in the UK that enables websites to block AI scraping without sacrificing their search rankings. The feature will eventually be rolled out globally.

The effect of the order and the opt-outs are yet to be seen. For now, the researchers’ work appears to have captured a glimpse of how LLM-powered chatbots substitute for the web, rather than feed it. What they discovered based on data ending in 2023 holds even stronger three years later.