AI summaries often cite their sources — but do those citations actually send any traffic back?

This explores whether citations in AI summaries can make up for the traffic those summaries take from the sites they draw on. The short answer is that the corpus documents the loss clearly and has no evidence of a recovery. Nothing here measures whether citations bring clicks back, and the closest evidence suggests citations mostly do a different job.

The loss itself is well measured. Pew's browsing-panel data shows that people clicked a search result about 8% of the time when Google showed an AI summary and 15% of the time when it didn't. Searches with a summary were also more likely to end with no click at all Do AI summaries on Google reduce clicks to actual websites?. A natural experiment on Wikipedia, which is the most-cited source on the web, compared language editions that got AI Overviews with ones that didn't and found that English Wikipedia lost roughly 5% of its search referrals Does AI search summaries divert traffic away from Wikipedia?. Wikipedia is credited constantly and still loses traffic. The mechanism is that the summary answers the question before the reader feels any need to click, so a visible source link does little to change that.

This is where the citation evidence gets surprising. An analysis of 24,000 AI search interactions found that users preferred answers with more citations, and irrelevant citations raised preference almost as much as relevant ones Do users trust citations more when there are simply more of them?. That points to citations working mainly as a signal that makes the answer feel trustworthy. They aren't functioning as doorways people actually walk through. If readers aren't checking whether a citation fits, they probably aren't clicking it either. The citations also aren't always real: in one study, invented sources and evidence caused 39% of deep research agent failures, as the agents tried to look scholarly Why do deep research agents fabricate scholarly content?.

The lost clicks also cost readers something, not just publishers. Across seven experiments with more than 10,000 people, those who learned a topic from ChatGPT came away with shallower knowledge, felt less ownership of what they learned, and gave sparser advice than those who learned through web search Does learning from AI summaries produce shallower knowledge than web search?. The work of clicking through and comparing sources seems to be part of how people learn, so a summary that adds credit but removes that work doesn't fully replace it.

What the corpus doesn't have is any study of how often people click on cited links, or of design changes meant to send readers back to sources. One adjacent idea comes from a RAG system that refuses to answer unless it can point to grounded evidence Can RAG systems refuse to answer without reliable evidence?. That shows how an AI answer could be tied more closely to its sources, though nobody has tested whether doing so brings traffic back.