Do AI summaries send people to their sources?
Pew Research found Google users clicked a result in 8% of visits with an AI summary and 15% without. What that means for answers you need to check.
Overview
Mostly not. In Pew Research Center’s study of 900 US adults’ browsing in March 2025, people clicked a traditional search result in 8% of visits to a Google results page that showed an AI summary, against 15% when it did not, and clicked a link inside the summary in 1% of visits. If people rarely click through, the evidence has to travel with the answer itself.
Key points
- In Pew’s March 2025 browsing data, Google users clicked a traditional result in 8% of visits that showed an AI summary and in 15% of visits that did not.
- They clicked a link inside the AI summary in 1% of visits to pages with one, and ended their browsing session on 26% of those pages, against 16% without a summary.
- About 18% of the Google searches in the study produced an AI summary.
- When few people click through, the source record has to travel with the answer: the record, its date and what it does not say, not only a link.
AI summaries now sit above the links on many search pages, and assistants answer questions without sending anyone anywhere. That is convenient. It also changes who checks the answer. The best data on what people actually do comes from a Pew Research Center study published in July 2025, and its numbers are worth knowing if you build or rely on AI answers.
What did Pew find?
Pew analysed the browsing of 900 US adults in March 2025, covering 68,879 unique Google searches. About 18% of those searches produced an AI summary at the top of the results.
When a results page showed an AI summary, people clicked a traditional search result in 8% of visits. When it did not, they clicked one in 15% of visits, nearly twice as often. Links inside the summary itself were clicked in 1% of visits to pages that had one.
People were also more likely to stop. Pew found that 26% of results pages with an AI summary were the last page of a browsing session, against 16% of pages without one.
Results pages after which people ended their browsing session
Show the data
| Item | Value |
|---|---|
| Without an AI summary | 16% |
| With an AI summary | 26% |
Why fewer clicks matter for checking answers
A citation only helps if someone follows it. When summaries answer the question on the page, fewer people open the source, and the source’s job shifts from “where to read more” to “what this answer rests on”.
That would be fine if citations were reliable. They are not always. In a March 2025 test by the Tow Center at Columbia Journalism Review, eight AI search tools asked to identify the source of quoted news excerpts got more than 60% of 1,600 queries wrong. That measured one task, not overall accuracy, but it shows that a link next to an answer is a claim in its own right.
What should travel with the answer instead
If people will not go to the record, the record has to come to them. For each value an answer depends on, that means:
- The source and the record: not just a site, but the record that states the value, with its identifier.
- The record’s own date: when the source last attested the fact, which is often not when it was fetched.
- What the record does not say: the questions it leaves open, stated as unknowns rather than filled in.
This does not have to crowd the answer. The value comes first; the record sits one step away, behind a “view sources” control, for anyone who wants to look.
What this means for agents and answer engines
For people building agents, the lesson is practical. Assume most readers will never open your citations, and design so that the answer can be checked anyway: keep each value tied to its record, keep the record’s date, and say plainly what was not established. That is what we are building IKAG to return for the specialist data it reaches. IKAG is opening to early users and data partners first.
What the study does not tell us
The numbers come from one month, one search engine and US adults who agreed to share their browsing. They describe clicks, not satisfaction or trust: a person who does not click may have got what they needed, or may have given up. Pew’s figures are a strong signal, not the last word, and AI search has changed since they were collected.
Questions
What did the Pew study measure?
The online activity of 900 US adults who agreed to share their browsing data in March 2025, including 68,879 unique Google searches. Pew compared visits to results pages that showed an AI summary with those that did not.
Does this mean people trust AI summaries?
The study measured clicks, not trust. Fewer clicks can mean the summary answered the question, that people trusted it, or that they gave up. It does mean fewer people saw the original source.
How often did AI summaries appear?
About 18% of the Google searches in the study produced an AI summary, according to Pew.
What should an answer carry if people will not click?
The record behind each value: which source stated it, the record itself, when the source last attested it, and what it did not establish. A link alone asks the reader to do the checking.
Sources
Share this article