AI Overviews correlate with fewer search clicks — but can researchers prove the AI itself, not something else, causes the drop?

This explores whether the drop in clicks on search results can be pinned on AI Overviews specifically, rather than on everything else changing alongside them. The short answer from this collection: not with one study. Researchers get closer by combining two kinds of evidence, each of which covers what the other can't. The corpus has only a few pieces on this, so treat what follows as a starting map, not a settled verdict.

The headline number comes from large-scale field data. Ahrefs found that the top-ranked result gets about 58% fewer clicks on searches that show an AI Overview than on searches that don't How much are AI Overviews actually reducing organic search clicks?. Read the wording closely, though: this is a correlation. The comparison is between different groups of searches, and Google doesn't show AI Overviews at random. They tend to appear on informational questions, the 'what is' and 'how does' kinds, where people often got their answer from the snippet and never clicked anyway. So part of that 58% may be the type of query rather than the Overview. A field comparison like this can show the size of the gap, but on its own it can't show the cause.

The second kind of evidence fills that gap. The RMIT and Microsoft eye-tracking study works the other way round: it controls the page in a lab and watches where people's eyes go Where do searchers look when AI Overviews appear?. With an AI Overview present, the share of attention going to the first-ranked result fell from 31% to 9%. That points to a mechanism, attention being physically pulled away, which a raw CTR chart can't show. The study also found that people trusted the Overview about as much as the ranked links. So they weren't looking at it out of suspicion. They were looking because it answered them. Neither study settles the question alone: the lab can't match real-world scale, and the field data can't rule out confounds. Together, the lab supplies a plausible cause and the field data shows that a gap of the matching size exists outside the lab.

There's a less obvious point in this question. Some 'confounds' may not be confounds at all. In a study of Indian and American writers using AI suggestions, the authors argue that the Indian writers' higher reliance on AI shouldn't be controlled away, because it is part of the phenomenon being studied Is higher AI use by Indian writers a confound to control?. The same logic applies here. If AI Overviews are deliberately placed on exactly the queries where an answer can replace a click, then 'query type' is part of how the Overviews do their damage, not noise to strip out. Trying to 'isolate' the AI Overview effect can end up hiding the very thing you're trying to measure.

One more caution, from an unexpected corner of the collection: small measurement choices can matter more than the variable you care about. In one study of LLM strategy advice, just changing the order the options were listed in shifted results more than changing the industry context did Do LLMs consistently favor the same strategic choices regardless of context?. For CTR research, the lesson is to check how keywords are grouped, how positions are counted, and which time window is used before trusting any single percentage. The collection doesn't yet hold a cleanly randomized field study, say one where Google turned Overviews on and off for comparable queries. Until something like that exists, the strongest case comes from lab mechanism plus field size, not from one decisive number.