Study: Wider ChatGPT Access Cuts Traditional Search Queries by Up to 17 Percent

New analysis of internet traffic in the United States shows that ChatGPT is not only taking over some queries from Google, but also directs people to websites in a radically different way than a search engine does. Researchers at Bocconi University calculated that expanding access to search in ChatGPT reduces the number of traditional queries by up to 17 percent after twenty weeks.

The paper, published as a preprint on arXiv, is based on Comscore clickstream data covering desktop browsing in the US. The researchers treated each ChatGPT conversation as a single search session and checked whether it ended with a click on a link leading outside the chat. They compared the result with the analogous behavior of classic Google users.

Fewer Outbound Clicks

The key figure is that 5.2 percent of ChatGPT sessions end with a move to another site, compared to 31.1 percent for Google. ChatGPT's monthly referral rate grew during the study period from about 2.5 to nearly 6.5 percent, but even that increase remains far behind the traditional search engine.

Where those few clicks land also matters. ChatGPT favors reference and knowledge sites, tools, SaaS platforms, and academic and developer sites. Ad-supported sites fare noticeably worse in this breakdown than in traffic from Google, losing 27.6 percentage points of share.

The Effect of Wider Access

The researchers took advantage of the fact that OpenAI gradually expanded access to the search feature in ChatGPT, from paying users, through free accounts, to access without logging in. Successive waves of rollout let them measure how the behavior of the same users changed as the use of AI for information search increased.

In this group, the weekly number of traditional queries fell by an average of 9.4 percent immediately after the access expansion, and after twenty weeks the difference grew to 17 percent. Among users who had already used ChatGPT Search before, the decline was smaller, from 4.9 to 8.2 percent over the same period, which suggests that the strongest effect applies to people who are only just discovering this new way of searching for information.

What It Means for Publishers

The paper's authors emphasize that they measure only the change in the distribution of observed traffic, not publishers' revenue or consumer welfare directly. They state this explicitly in the text so as not to suggest conclusions that go further than the data allow.

Our claim is deliberately narrower than a welfare claim: we measure the change in the observable distribution of traffic, not consumer surplus, publisher revenue, or long-term content production - Qiaoni Shi, Kai Zhu, Kai Gu, study authors

Despite this caveat, the conclusions are troubling for a business model based on advertising displayed alongside content used by AI systems. If a model answers a question without directing the reader to the source, the publisher loses an ad impression even though its content was used to generate the answer.

A Perspective for Polish Firms

For Polish publishers, editorial teams and e-commerce companies that build organic traffic from search, the study's results are another signal that it's already worth reckoning with the erosion of classic Google traffic as search in ChatGPT and similar tools grows in popularity. The biggest losses affect informational and reference content, exactly the kind that many industry and how-to portals rely on.

The marketing industry increasingly talks about optimizing for AI-generated answers instead of classic SEO, but the Bocconi study puts a number on the scale of the problem: even effectively reaching the AI model does not guarantee a click, because in most cases the model does not send the user onward at all.

Sources: Search Engine Journal (searchenginejournal.com), arXiv (arxiv.org)