The Clicks Are Leaving Search

A search result has long offered a path from a question to a company's website. An AI-written summary at the top of many Google search results can answer the question before the searcher visits a site. That changes what a business can learn from its traffic report.

I have built for the web since 1994. A traffic report can show visits that arrive from search, but it cannot show what a searcher read before deciding whether to visit.

What did the study find about clicks from Google search?

Researchers Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon University's Heinz College ran a randomized field experiment with 1,065 desktop Chrome users. The paper was posted to SSRN on April 3, 2026 and revised June 17, 2026. It observed 68,089 unique searches over two weeks. PPC Land reported on the findings.

Organic outbound clicks dropped by 39.8% when an AI Overview appeared, while searches with no click increased by 34.5%. These findings apply to the study's users and searches. They can't predict exactly how much traffic a particular company will lose.

An outbound organic click is a visit from an unpaid search result. A search that ends with no click can still have included an answer a person read. That means a fall in visits does not, by itself, show whether a business was visible or useful to searchers.

Traffic reports still help a business track visits that arrive from search. They cannot show what a searcher read in the summary or whether the business appeared there. A company that watches only visits may miss visibility in an answer that sends no one to its site.

What should a business measure instead of clicks?

Measure whether an AI answer names the business and whether its pages hold the facts the answer quotes. A business can appear in the summary even when the searcher does not visit its site. This is a different record from a count of site visits.

I use a locked library of a category's real buying questions. I ask ChatGPT, Claude, Perplexity, Gemini and Grok on a fixed schedule. Keeping the questions unchanged makes it possible to compare answers over time without confusing a prompt change with a change in visibility.

I check official domains so a same-named company does not count as a win. I also ask the model why it chose what it chose. I keep records in one format, so week 11 compares cleanly with week 1. That record lets me review whether the answer named the right company and which page facts it quoted.

What can change an AI recommendation rate?

In one engagement, Soapbox's AI recommendation win rate rose from 6.2% to 36.2% in 11 weeks. Trinzik delivered that work on the platform I built. The evidence is directional, not from a controlled study. In a separate program, I published into the gaps, and the win rate didn't change. I retest because a rate change by itself doesn't show what caused it.

For the review to be useful, I compare the same questions and check the same type of evidence each time. The answer records help me see what changed in a recommendation and which page facts were cited.

Can AI tools quote facts from your website?

A page needs to make its facts readable to systems that produce answers. Facts locked in PDFs and images can be hard for those systems to quote. I published my rubric for scoring a site's readability by agents. Any firm can use it to score itself and get the same answer.

Common questions