Google AI Overviews cut clicks 40% - and the lost clicks weren't lower quality

Whenever AI Overviews are blamed for lost traffic, someone answers: 'but those were low-value clicks anyway.' A new controlled study takes that comfort away. AI Overviews cut organic clicks by about 40%, and the researchers found the clicks you lose are just as valuable as the ones you keep.

Most estimates of AI Overview traffic loss are correlations: sites with AIOs got fewer clicks. Correlations invite the "those clicks were low-value" rebuttal. This study is different, because it is a controlled experiment - and it tested the quality question head on.

What the researchers did

Saharsh Agarwal and Ananya Sen ran a randomised field experiment - the gold standard for isolating cause. When a Google AI Overview appeared, outbound organic clicks fell by about 39.8%, and zero-click searches rose by roughly 34.5%. Because the appearance of the AIO was the variable being tested, this is causal evidence, not a pattern that could be explained away.

Curious how AI engines describe your brand right now? Get a free visibility audit and see where you stand across ChatGPT, Gemini and Perplexity.

The finding that kills the usual defense

The reassuring story about AIOs is that they only skim off low-intent, low-value clicks - people who would have bounced anyway. The researchers checked. They found no meaningful difference in quality between the clicks kept and the clicks lost: bounce rates, quick returns to search, time on site, and survey-measured satisfaction were statistically indistinguishable.

In other words, AI Overviews are not conveniently trimming junk. They remove valuable visits at the same rate as everything else. The 40% is not 40% of the clicks you did not want; it is 40% across the board.

"AI Overviews are not conveniently trimming junk - they remove valuable visits at the same rate as everything else."

What it means for you

1. Take the loss seriously. You cannot dismiss AIO click loss as low-value traffic - a controlled study says the lost clicks are as good as the kept ones. Plan around a real reduction in valuable visits on affected queries.
2. Move weight to being in the answer. If the click is leaving, the exposure that remains is being cited inside the AI Overview. Optimise to be quoted there, with clear, liftable, factual content.
3. Defend high-intent queries. The loss concentrates on informational searches. Make sure the queries near a buying decision still give people a reason to click through, and that you are named when they do not.
4. Measure presence, not just clicks. Referral clicks now understate your visibility. Track whether you appear in the answer itself, because that is increasingly where the value is captured.

The takeaway

The value of a controlled experiment is that it settles an argument. The argument here was whether AI Overviews cost you traffic that mattered. The answer is yes - about 40% of it, and no less valuable than the rest. The response is not to mourn the click but to compete for the citation, because that is the exposure the AI answer still hands out.

What the study does not say

It is worth being precise about the finding, because the wrong reading leads to the wrong response. The study does not claim that AI Overviews destroy all your traffic, or that every query is affected equally. It measures the drop on the searches where an AI Overview actually appears, and those tend to cluster around informational, "explain this to me" intent. Transactional and navigational queries, where the searcher already wants to reach a specific place or buy a specific thing, are far less exposed.

Nor does the study say the click loss is permanent or beyond your influence. It measures what happens on average today, given how the answer is currently assembled and who it currently cites. If you are the source the overview leans on, some of that "lost" exposure is still yours - just in a different shape. The experiment settles the quality question; it does not settle your standing inside the answer, which is the part you can still change.

The mistake to avoid is fatalism. "AI Overviews cut clicks 40%" is not a reason to stop publishing or to write only for the ten percent of queries that still convert. It is a reason to be deliberate about which content earns a citation and which content earns a click, because those are now two different jobs.

A concrete way to think about it

Picture a company that sells project-management software and ranks on page one for "how to run a sprint retrospective". Historically that page pulled a steady stream of visitors, some of whom eventually became trial signups. An AI Overview now answers the question directly at the top of the results. Under the old assumption, the business shrugs: those were researchers, students, curious readers, not buyers, so losing them costs nothing.

The study takes that shrug away. The visitors who no longer arrive were, on the measured dimensions, no less engaged than the ones who still do. Some fraction of them would have explored the product, subscribed to the newsletter, or remembered the brand the next time a purchase was on the table. That downstream value does not show up in a bounce-rate column, which is exactly why the "junk traffic" defence felt safe and was wrong.

So what does the company actually do? It stops treating the retrospective article as a click-catcher and starts treating it as a citation-earner. That means writing the definitive, quotable answer to the question - clear definitions, a clean step list, a specific figure or two - so that when the overview assembles its response, this brand is the name it reaches for. A reader who never clicks still sees who the authority is.

"Earning a citation and earning a click are now two different jobs, and the same page rarely does both well by accident."

How to measure your exposure in the answer

If referral clicks now understate your visibility, the obvious next question is how to see the part they miss. Traditional analytics were built for a world where exposure and clicks were nearly the same event. They are not any more, so a few extra habits are worth adopting.

• Track citation share, not just click share. For the queries that matter to you, note how often an AI Overview appears and whether your domain is named or linked inside it. That percentage is a visibility metric your rank tracker will not give you.
• Segment queries by intent before you panic. Separate informational queries, where loss is expected, from high-intent ones near a decision. A drop on the former is the study playing out as predicted; a drop on the latter is a genuine problem to chase.
• Watch branded and downstream signals. If people meet your name inside answers without clicking, the effect surfaces later as branded searches, direct visits, and assisted conversions rather than as first-touch organic traffic. Judge the channel on those, not on referrals alone.
• Audit whether your content is liftable. Answers are assembled from content that is easy to quote. Well-structured claims, plain definitions, and concrete numbers get pulled in; buried, hedged, or padded prose does not.

None of this requires abandoning the metrics you already trust. It requires accepting that they now describe a shrinking slice of where you are seen, and adding the measurement that covers the rest. The brands that adapt first will be the ones with a clear, honest picture of how often they are named - long before their competitors notice the click numbers were only ever half the story.