For most of the past year, the discussion around AI search and publisher traffic has relied on observational data. Publishers see fewer clicks, AI answers become more prominent, and the obvious conclusion is that the two things are connected.

A new study gives that argument something more useful: a controlled experiment.

Researchers ran a preregistered field experiment involving 1,100 active participants using Google Search. Instead of comparing traffic before and after a product rollout, they changed the search experience itself. One group used regular Google, another had AI Overviews and AI Mode removed, while another was pushed into an AI Mode-only experience.

The direction of the result is difficult to ignore. Removing Google's AI features increased click-through to external publishers. Moving users into AI Mode reduced it. The paper also reports that the AI Mode experience did not compensate for those lost clicks with a better perceived experience; trust and several measures of user experience declined.

There are limitations. This is a preprint, the experiment covered Google and a relatively short period, and one experimental condition was affected when Google changed its HTML during the study. It should not be read as a universal measurement of every AI search experience.

But the broader question matters.

AI visibility and website traffic are increasingly becoming two different things. A publisher can be retrieved, understood and even cited while receiving fewer visits because the answer itself satisfies more of the user's need.

We recently looked at another side of the same transition in ChatGPT Is Changing the Types of Pages It Cites. Citation patterns are moving, but being selected as a source doesn't automatically reproduce the traffic model that supported traditional search publishing.

That may become one of the central problems of AI search: being visible without being visited.

Source: Wang et al., AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence, arXiv:2608.18352.
Read the original research on arXiv