• AI visibility

Guide · AI search

Do people actually check the sources AI assistants cite?

Rarely. In a Pew Research Center panel that tracked the real browsing of 900 US adults, only 1% of visits to Google pages with an AI Overview led to a click on one of its cited sources. Smaller studies point the same way: people glance at one or two citations at most, and the presence of a citation, not its quality, seems to satisfy them.

By Underneath Team Guide Updated

The short version

  1. Only 1% of visits to Google pages with an AI Overview led to a click on a cited source, in a month of tracked browsing by 900 US adults (Chapekis and colleagues (opens in a new tab)).
  2. Expert users in a 2024 study looked at about 2 sources in an AI answer, against 12 on a regular Google page (Narayanan Venkit and colleagues (opens in a new tab)).
  3. When the AI answer agreed with their own view, those experts clicked a source 0.48 times per question on average (Narayanan Venkit and colleagues).
  4. A week of Google’s AI Mode cut click-through to other websites by 18.8 percentage points, in a field test with 1,100 US users (Wang and colleagues (opens in a new tab)).

How often do people click the sources in AI answers?

Almost never, in the best real-world data available. Pew researchers Chapekis and colleagues (opens in a new tab) tracked the browsing of 900 US adults from a representative panel throughout March 2025, covering 68,879 distinct Google searches. About 18% of those searches produced an AI Overview, the AI summary at the top of Google’s results.

Just 1% of visits to pages with an AI Overview led to a click on a cited source. Clicks on any result also fell: people clicked a search result on 8% of visits to pages with an Overview, against 15% on pages without one. They ended their browsing session on 26% of pages with an Overview, against 16% without.

This is observational data, so it shows association, not cause. It also counted only the first three cited sources, the most a desktop Overview showed without expanding.

Do people at least look at the sources?

Briefly, and far less than on a normal results page. In a 2024 study by Narayanan Venkit and colleagues (opens in a new tab), 21 participants with PhD-level expertise searched the same questions on an AI answer engine and on Google, thinking aloud as they went.

Behavior in the 2024 user studyAI answer enginesGoogle
Sources hovered over, on average212
Sources clicked, on averageusually 14

Agreement made checking even rarer. When the question matched the participant’s own opinion, they clicked a source just 0.48 times on average and hovered over 1.08. Questions that challenged their views prompted noticeably more checking.

Participants described the pull of a confident answer. One said: “It writes so confidently, I feel convinced without even looking at the source.” How that confidence compares with trust in Google results is covered in our guide to trust in AI search.

Does the checking that does happen catch bad citations?

Not reliably. In the same 2024 study, the authors found that misattributed citations were noticed only by the few participants who scrutinized the sources. Perplexity, for example, displayed 5.00 sources per answer on average but cited only 2.58 of them in the text.

A large experiment found the same blind spot. Li and Aral (opens in a new tab) gave some of their 4,927 US participants AI answers containing broken or irrelevant links. People who spent extra time and effort on a question showed no sign of noticing the bad links, and people who trusted AI more clicked more but spent less time evaluating what they found. The effect of those links on trust is covered in our guide on whether citations build trust.

Do people choose answers with better sources?

Not that researchers can detect. Yang (opens in a new tab) studied 1,534 head-to-head votes on an online arena where users compare two AI search answers. The quality and political lean of the news sources each answer cited made no detectable difference to which one users preferred.

What did matter was length: longer answers tended to win. Arena volunteers are not typical consumers, so this is supporting evidence, not proof.

Does AI search keep people from visiting websites at all?

It seems to, especially in conversational modes. In the field test by Wang and colleagues (opens in a new tab), switching 1,100 US users to Google’s AI Mode for a week cut click-through to other websites by 18.8 percentage points. The share of users clicking through to news sites fell 12.5 points, to Reddit 21.2 points and to Wikipedia 9.9 points.

The reverse also held. Hiding AI Overviews raised click-through by 8.8 percentage points, though only about half of Overviews were successfully hidden.

Standalone assistants show a similar pattern in panel data from Iannelli and Ai (opens in a new tab) of Scrunch AI, a company that sells AI visibility tools, in a paper not yet peer reviewed. User-weighted, 34.1% of sessions involving an AI assistant showed no visit to an outside website, against 19.5% of search sessions by the same people. More of the evidence is gathered in our guide to clicks lost to AI answers.

Why does this matter for brands?

Because most people read only the answer, what the answer says about you matters more than whether they click. The checking is done by the machine instead. In our study of hidden searches, ChatGPT ran a mean of 3.7 searches of its own per buyer question before answering, reading pages the user never sees.

Being cited is also a separate route from ranking. Xu and colleagues (opens in a new tab) found that 29.8% of the domains cited by Google’s AI Overviews did not appear anywhere in the first page of regular results for the same search.

What should you do about it?

Measure and manage what AI answers say, not just the traffic they send.

  1. Track whether and how you appear in the answer text for your buyers’ key questions. Referral clicks badly understate how many people read about you.
  2. Make the facts you most need buyers to know easy to quote, in plain sentences on pages AI assistants can read. Few readers will open the page to find the nuance.
  3. Fix inaccurate claims at their source, since readers are unlikely to click through and spot them.
  4. Treat AI referral traffic as a floor, not a measure of influence.

For help measuring how AI answers present your brand, see our generative engine optimization service.

What does the research not tell us yet?

The strongest data covers Google’s AI Overviews; much less is known about standalone assistants.

  • The Pew panel covers one month in 2025 and counts clicks on only the first three sources in each Overview.
  • The 2024 user study involved 21 experts in a think-aloud session, which may encourage more checking than everyday use.
  • The only large panel on ChatGPT-style assistants comes from a vendor and has not yet been peer reviewed.
  • Not clicking is not the same as not verifying. People may check elsewhere, later or by asking again, and no study measures that.
  • No study we found looks specifically at whether buyers check sources for product or brand questions.

Frequently asked questions

What percentage of people click on AI Overview sources?

About 1%. In Pew’s tracked panel of 900 US adults, just 1% of visits to Google pages with an AI Overview led to a click on a cited source.

Do people verify what ChatGPT tells them?

The evidence is thin. In one vendor panel, 34.1% of sessions involving an AI assistant showed no visit to any outside website, but that does not prove the answer went unchecked.

Does Google’s AI Mode send traffic to websites?

Less than regular search. In a 2026 field test, a week of AI Mode cut click-through to other websites by 18.8 percentage points.

If people do not click, does being cited still matter?

Yes. Citations raise trust in an answer even when nobody follows them: in one experiment, adding reference links lifted trust by 0.091 points on a 7-point scale.

Sources

Free strategy call

Some questions are easier to answer about your own business.

Bring the one that matters most. On a free 30-minute call we’ll take a first look at it and send you a short written read afterward.

Ask about my businessHow we work