Direct answer
Citations can change how trustworthy an AI answer appears, and controlled studies have found this effect even when references were random, invalid, or hallucinated. That is evidence about perceived trust under specific experiments, not proof that citations improve factual accuracy or cause purchases. A citation can also invite checking, expose weak support, or have no effect for a given topic. Measure at least four outcomes separately: perceived credibility, source checking, belief or decision change, and downstream behavior. Good citation design makes the claim-source relationship easy to inspect and does not use decorative links as a trust signal.
A source link has two jobs that can conflict
A citation can provide a route to evidence. It can also act as a visual cue that tells the reader, “this answer has support.” The first function depends on the source and the claim. The second can operate before anyone opens the link.
This creates a difficult design problem. A well-placed citation can help a careful reader verify a statement. The same marker can increase confidence among readers who never inspect it. If the source is weak, irrelevant, or fabricated, the interface has added the appearance of rigor while making the answer no more reliable.
GEO teams often focus on acquiring citations as visibility events. Readers experience those citations as part of an answer interface. Understanding both sides requires evidence about human behavior, not assumptions about what a blue link must mean.
What the 303-participant experiment found
The study Citations and Trust in LLM Generated Answers ran a randomized controlled experiment with 303 Prolific participants. Participants asked ten open-ended questions and saw short ChatGPT-4 answers with zero, one, or five citations. Citation conditions used relevant links or random links drawn from previous queries.
The researchers reported that the presence of citations increased self-reported trust. Five citations did not significantly outperform one citation. In the citation conditions, the dataset contained 1,976 answers with at least one reference, and 193 citation checks were recorded, a 9.77% answer-level checking rate. Eighty-three of 197 participants in citation groups checked at least one citation during the experiment.
The design supports a conclusion about the interface treatment and measured trust in that setting. It does not show that citations made the answers more accurate or that participants made better decisions. The authors also note limitations: Prolific participants may differ from the general population, questions were user-selected, demographic power was limited, and trust used a one-item rating.
The practical lesson is uncomfortable. Citation presence can affect trust before citation quality is established.
What the 4,927-participant experiment added
The preprint Human Trust in AI Search: A Large-Scale Experiment used a preregistered randomized experiment with 4,927 participants selected to represent the adult US population. Participants evaluated information across nine consequential search topics. The design held the information constant while changing whether it was presented as generative or traditional search, and it randomized interface features including references.
In that experiment, reference links increased trust and willingness to share generative-search results. The increase did not significantly depend on whether the links were valid or invalid. Trust also varied by topic and participant characteristics. Participants reporting greater trust clicked more and spent less time evaluating generative-search content, while the reference condition itself encouraged more clicking and more time than the baseline generative condition.
The study does not establish a purchase effect. Its topics centered on public affairs and major social issues, and the authors explicitly state that product search or household tasks may behave differently. It was conducted in the United States in a lab-style environment, and generative-search designs continue to change.
Together, the two experiments show that citation interfaces can cause perceived-trust changes under defined conditions. They also show why a citation marker should not be used as a proxy for informed agreement.
Trust, accuracy, checking, belief, and action are different outcomes
| Outcome | Example measure | What it tells you | What it does not tell you |
|---|---|---|---|
| Perceived credibility | Multi-item trust rating | How believable or reliable the answer seems | Whether the answer is correct |
| Citation checking | Link open, hover, dwell, return | Whether a user interacted with a source | Whether the source was understood |
| Support recognition | User judges claim-source match | Whether the citation appears to support the claim | Whether the source itself is true |
| Belief change | Pre/post belief or confidence | Whether the response changed stated belief | Whether the belief is durable |
| Decision change | Choice before and after evidence | Whether the answer altered a decision | Whether the decision leads to action |
| Behavior | Click, share, trial, purchase, cancellation | What the user did in the observed path | Why the user did it |
Do not build a causal chain by assumption. Higher trust can coexist with low verification. More clicks can reflect curiosity, doubt, or a desire to inspect the evidence. A purchase can be driven by price or availability rather than the citation.
The support quality still matters
The trust experiments concern user response to interface cues. A separate question is whether citations support the answer. Evaluating Verifiability in Generative Search Engines formalized citation recall and precision and reported substantial support gaps in the specific systems and period it studied. Those historical percentages are not current platform benchmarks. The methods show why a source marker and claim support need separate evaluation.
For each verification-worthy claim, identify the associated source, locate the supporting passage, and judge full, partial, absent, or unclear support. Then evaluate source applicability: publisher, date, method, population, market, and whether the page is primary or derivative. A correct citation to an inapplicable source can still mislead.
The ALCE paper treats citation quality as a combination of correctness and completeness in long-form generation. The broader editorial lesson is that a good answer needs citations attached to the right claims and enough coverage for the important claims. Adding one impressive source to the end of a paragraph does not support everything in it.
Design citations for inspection
- Attach the citation to one clear proposition. Avoid placing a single marker after a paragraph containing several unrelated claims.
- Use descriptive source labels where the interface permits. Publisher and document title help readers decide whether to inspect a link.
- Link to the relevant location. A section anchor, page number, table, or evidence excerpt reduces verification cost.
- Preserve qualifiers in the answer. Population, date, market, method, and uncertainty should remain adjacent to the finding.
- Distinguish primary and derivative sources. A news report about a study and the study itself serve different roles.
- Show disagreement. When credible sources conflict, do not use citations to create a false appearance of consensus.
- Check the destination. Monitor redirects, withdrawals, paywalls, changed content, and broken anchors.
- Do not decorate. Remove links that do not support a claim or add useful context.
OpenAI's web-search documentation requires citations displayed to end users to be visible and clickable in applications using that tool. Visibility and clickability are necessary interface properties. Editorial teams still need to evaluate support and applicability.
Test the effect without manufacturing confidence
A publisher can study citation design ethically by comparing interfaces that all use valid, relevant sources. There is no need to expose readers to fabricated links.
- Define the target behavior: checking, comprehension, calibrated trust, decision quality, or another outcome.
- Create claim-equivalent versions of the same answer, such as inline descriptive links versus a source list.
- Validate every source and claim association before the test.
- Randomly assign eligible participants where feasible and preserve the assignment process.
- Measure baseline knowledge or confidence before exposure.
- Record source opens, time on source, return to answer, and completion without treating any one event as understanding.
- Ask support-recognition questions that require reading the source.
- Measure belief or decision immediately and, where useful, after a delay.
- Segment results only with a predeclared reason and sufficient sample.
- Report null results, exclusions, attrition, uncertainty, and external-validity limits.
The purpose should be calibrated understanding: justified trust rises when evidence is strong, and weak support becomes easier to detect.
A product-research example
Suppose a buyer asks whether a fictional air purifier is suitable for a 40-square-meter room. Version A of the answer says, “The manufacturer rates it for rooms up to 45 square meters.” Version B adds a link to the product home page. Version C links directly to the dated specification section and states the test conditions used for the rating.
A useful test would measure whether readers can identify the rated area, find the test conditions, distinguish manufacturer rating from independent performance testing, and choose an appropriate next verification step. Asking only “How much do you trust this answer?” would miss those abilities.
If Version B raises trust as much as Version C while readers cannot find the evidence, the home-page link is functioning mainly as a cue. If Version C improves correct support recognition, the more precise source path is doing evidentiary work.
The example is hypothetical. No purifier or interface was tested.
Commercial decisions require another layer of evidence
A recommendation or purchase involves fit, alternatives, price, risk, and availability. Citation trust research does not establish that cited AI answers increase conversion, and it certainly does not establish that any specific brand benefits.
To study commerce, define the decision population, product category, realistic alternatives, disclosure rules, and outcome. Distinguish a stated preference from a checkout, a checkout from a retained customer, and a retained customer from an appropriate decision. Protect participants from high-stakes or deceptive experiments.
For live-site analysis, treat the referral path as incomplete. Users can see an AI answer, search the brand later, visit directly, consult a distributor, or purchase offline. Use surveys, journey interviews, controlled campaigns, CRM fields, and analytics as complementary evidence rather than forcing every path into one referrer field.
Avoid trust theater
Trust theater uses citation density, logos, badges, or technical language to create confidence without making verification easier. It can produce a page that looks researched while relying on circular, outdated, or irrelevant sources.
Warning signs include a source list with no claim mapping, repeated links to one underlying report, citations to search results instead of original documents, and broad conclusions attached to narrow studies. Another sign is hiding the method while promoting a precise statistic.
A high-quality source note should state access dates, research scope, missing data, and which parts of the article are evidence versus editorial framework. This makes trust earned through inspectability rather than styling.
Citation-effect checklist
- Citation presence is not used as a proxy for answer accuracy.
- Trust, checking, belief, decision, and behavior have separate measures.
- Every important claim maps to a supporting source location.
- Source applicability is checked for date, population, market, and method.
- Interface tests use valid sources and a declared outcome.
- Link opens are not treated as proof of comprehension.
- Purchase claims require commerce-specific evidence.
- Historical study results retain their original system and sample scope.
- Broken, withdrawn, and changed sources trigger review.
- Null and adverse results remain in the report.
Frequently asked questions
Do more citations create more trust?
One 303-participant study found no significant trust difference between one and five citations, while citation presence increased trust relative to none. Other contexts may differ. Citation count should not replace source quality.
Do readers check citations?
Some do. In the same study, 193 checks occurred among 1,976 answers with references, and 83 of 197 participants in citation groups checked at least one. Interface, topic, stakes, and participant population can change checking behavior.
Does an invalid citation reduce trust?
The two experiments reviewed here found that citation presence could raise trust even when links were random or invalid, although details differed. This is a reason to validate sources carefully, not a tactic to exploit.
Can citation clicks measure GEO success?
They measure one interaction when the platform exposes it. They do not show comprehension, recommendation, purchase, or causal contribution by themselves.
What is the safest citation design?
Use a visible, specific link adjacent to the supported claim; preserve scope and uncertainty; label the source; and make the relevant passage easy to find. Test whether readers can verify the claim, not merely whether the answer looks credible.
Source and method note
Sources were retrieved on September 23, 2026. The two trust studies are preprints with defined experimental samples, interfaces, topics, and limitations. The verifiability and ALCE papers address citation support and generation evaluation rather than purchase behavior. OpenAI documentation applies to its web-search tool interface. The measurement model, hypothetical purifier example, test procedure, and trust-theater criteria are editorial proposals. No commercial conversion effect or named human review is claimed.