• AI visibility

Guide · AI search

Do citations make people trust AI answers more, even when the links are wrong?

Yes. In a large US experiment, adding reference links to AI search answers raised people’s trust in them, and the boost was the same whether the links were valid or swapped for broken and irrelevant ones. That makes a citation a credibility signal in its own right, and a wrong attribution a quiet reputational risk.

By Underneath Team Guide Updated

The short version

  1. Among 4,927 US adults, adding clickable reference links to Google AI answers raised trust by 0.091 points and willingness to share by 0.121 points on a 7-point scale (Li and Aral (opens in a new tab)).
  2. Replacing a link with a broken or irrelevant one changed trust by just 0.021 points, a difference too small to distinguish from zero (Li and Aral).
  3. Citations themselves are often wrong: in a 2024 audit, only 49.0% to 68.3% of citations pointed to a source that supported the statement (Narayanan Venkit and colleagues (opens in a new tab)).
  4. In 1,534 head-to-head votes on real AI search answers, the quality and political lean of cited news sources had no detectable effect on which answer users preferred (Yang (opens in a new tab)).

Do citations make people trust AI answers more?

Yes, in the largest experiment on the question. Li and Aral (opens in a new tab) at MIT showed 4,927 US adults Google AI answers to nine public-affairs questions in spring 2024. One group of 785 people saw the answers with clickable reference links, while others saw the same text without them.

Links raised trust by 0.091 points on a 7-point scale and willingness to share by 0.121 points. Those are modest shifts in absolute terms. They are, however, about twice the size of the trust gap between AI answers and regular Google results in the same study. We look at that gap in whether people trust AI search less.

People shown links also clicked more and spent more time on each answer. That sounds like checking, but the next result suggests it did not work as checking.

Does it matter if the citations are broken or wrong?

Not measurably. Within the reference group, the researchers swapped one link for a bad one in 7 of the 9 tasks, chosen at random for each person. Most swaps were broken or invalid links; in two tasks the swap was a working link to an irrelevant page.

Trust moved by just 0.021 points when a link was bad, and by 0.009 points for an irrelevant link, neither distinguishable from zero. The authors also looked at people who spent extra time on their first question, and found no sign that they noticed invalid links either. They call this “the veneer of rigor”: citations create trust even when the references are not rigorous.

The original Google answers were not flawless either. In one, a booster every 10 years was described as “recommended for age 19-26” while the cited article recommended a much wider age range.

Who is most swayed by citations?

People with less formal education and people outside the tech industry. Li and Aral found reference links raised trust and sharing significantly more for participants without a college degree and for those not working in technology-related industries.

How often people used AI tools made no significant difference to the citation effect. In a later essay, Aral and colleagues (opens in a new tab) draw the conclusion directly: less educated and less tech-savvy consumers are more vulnerable to misrepresentations and errors in AI search.

The effect also varied by topic. Links raised trust significantly for questions on jobs and AI, police oversight and inflation, but not for gun control, immigration, taxes, healthcare, vaccines or climate change.

How often are AI citations actually wrong?

Often enough that a citation should not be read as proof. In the 2024 audit by Narayanan Venkit and colleagues, You.com’s citations pointed to a supporting source 68.3% of the time, BingChat’s 65.8% and Perplexity’s 49.0%. Even when a supporting source was in the list, engines often cited a different one.

Google’s AI Overviews do better, but not perfectly. Xu and colleagues (opens in a new tab) found 2,609 claims, in a 2026 sample of US searches, that were directly contradicted by a source the Overview itself cited. The authors note that a reader who sees only the summary has no way to detect this.

Aral and colleagues also cite a Columbia University test in which ChatGPT Search was “confidently wrong” in 146 of 200 attempts to credit quotes to their sources. And citations are not always shown at all: in their own global search data, Google’s AI answers showed references 92% of the time for general questions but only 50% of the time for Covid questions.

Do people reward better sources?

Apparently not, at least in what they choose. Yang (opens in a new tab) analyzed 24,069 conversations from an online arena where users compare two AI search answers and pick the better one. In 1,534 decided comparisons where both answers cited news sources, neither the quality nor the political lean of those sources predicted which answer won; answer length mattered most.

Yang concludes that users may not closely examine the sources cited, “provided that citations are present.” Arena volunteers are not typical consumers, and a vote is not a trust rating, so this supports the experiment rather than proving the same point.

Expert users describe the same habit. One participant in the 2024 user study put it plainly: “you see a citation, you assume it’s a valid source…I’ll just see that there is a source. That’s it. I don’t verify it.” Click data points the same way; see how often people check AI sources.

What does this mean for brands?

A citation borrows credibility from whoever is cited, whether or not the cited page says what the answer claims. In our study of “Is this brand legit?” answers, the engines attached a citation to 71.0% of the claims they made about brands, from 93.3% for Perplexity to 56.7% for Google AI Mode.

When we checked a sample of those cited claims against pages we could read, only 46.6% were fully supported, and others were partly supported or not supported at all. Combined with the experiment above, the implication is that readers likely give many of those claims the same credit, supported or not.

What should you do about it?

Assume readers take every citation of your brand at face value, and check what it is attached to.

  1. Read what AI answers say next to your links. A claim beside your URL gains credibility from your name even if your page does not support it. See how often AI misattributes claims to pages.
  2. Fix or redirect broken and outdated pages. Readers will not notice a dead link, but the assistant may still pair your name with an old claim.
  3. Publish the facts you want attributed to you in clear, quotable sentences, so a correct citation is easy for the engine to make.
  4. Watch the third-party pages most often cited about you, since their claims arrive with the same borrowed authority.

For help auditing how AI answers cite and describe you, see our generative engine optimization service.

What does the research not tell us yet?

The core finding rests on one well-designed experiment, with clear limits on how far it reaches.

  • Li and Aral tested nine public-affairs questions using 2024-era Google AI answers. The authors note that product and shopping searches may behave differently.
  • Each person saw each answer once. No study shows whether people learn to distrust citations after spotting bad ones.
  • Nobody has tested whether being cited makes readers trust the cited brand itself, as opposed to the answer.
  • The arena data measures preference among volunteers, not trust among buyers.

Frequently asked questions

Do people click on the sources in AI answers?

Rarely. In a Pew Research Center panel of 900 US adults, only 1% of visits to Google pages with an AI Overview led to a click on a cited source.

Are citations in AI answers reliable?

Not reliably. In a 2024 audit of three AI search engines, between 49.0% and 68.3% of citations pointed to a source that actually supported the statement.

Would people notice if an AI answer cited a broken link?

Mostly not, in the one experiment that tested it. Swapping in broken or irrelevant links changed trust by 0.021 points on a 7-point scale, and longer reading time did not help.

Does being cited by AI make my company look more credible?

Probably, but it is untested. The evidence shows citations raise trust in the answer; no study has yet measured the effect on trust in the cited company.

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