Online searches to evaluate misinformation can increase its perceived veracity

The study addresses how searching online to evaluate news (SOTEN) affects belief in misinformation—a central but underexamined question given search engines’ gatekeeping role in the digital information ecosystem. While social media’s role in spreading misinformation has been widely studied, less is known about search engines, despite widespread recommendations from platforms, civil society, and government to use search to verify information. The authors propose and preregister the hypothesis that encouraging online search when evaluating false or misleading articles may increase belief in those articles, contrary to common media literacy guidance. They also hypothesize that SOTEN may increase belief in true articles. The work situates this question within concerns about data voids, search engine optimization by low-quality publishers, and the timing of fact-check availability, especially shortly after publication of misinformation. The purpose is to provide causal evidence on SOTEN’s effects and to explore mechanisms—particularly exposure to low-quality search results—that could explain increased belief in misinformation.

The paper builds on literature about misinformation belief and mitigation, including ideological congruence, psychological factors, and digital media literacy interventions. Prior work has examined search engine result personalization and partisanship, and the role of search engines as modern gatekeepers influencing political knowledge and public opinion. The concept of data voids suggests that certain queries—often seeded by low-quality publishers through distinct phrasing and SEO—yield scarce credible coverage and abundant low-quality corroboration, especially shortly after publication when professional fact-checks lag. Research also highlights how coordinated media ecosystems and messaging strategies can create a propaganda feedback loop that saturates search results with ideologically aligned misinformation. Related studies indicate location-based personalization is common, and political queries can be personalized, potentially amplifying ideological congruence effects. Digital literacy research suggests more skilled users employ better search strategies, potentially avoiding low-quality result spaces. Source reputation and site features also influence perceived credibility and belief in online news.

The authors conduct five experiments to estimate the causal effect of SOTEN on belief in false and true news. Designs include between-respondent randomized controlled trials (studies 1 and 5) and within-respondent designs (studies 2–4) where the same respondent evaluates before and after an encouragement to search online. Articles (popular true and false/misleading from mainstream and low-quality sources) were selected via a preregistered pipeline and distributed to respondents within 48–72 hours of publication to capture the early information environment. Professional fact-checkers (six from national outlets) concurrently rated each article as true, false/misleading, or could not determine; modal ratings were used as ground truth.

• Study 1 (between-respondent): 3,006 US respondents via Qualtrics over 10 days; participants evaluated three articles within 48 hours of publication; treatment group encouraged to search online using standardized instructions; control not prompted. Outcomes: dichotomous (true vs false/misleading/cannot determine) and 7-point veracity scale. Analysis: OLS with article fixed effects; SEs clustered at respondent and article; controls for age, education, income, ideology congruence, gender.
• Study 2 (within-respondent): 4,252 respondents over 33 days; 1,010 evaluated one false/misleading article first without search, then after encouragement to search; same measures and models.
• Study 3 (within-respondent, delayed): 4,042 respondents; 982 evaluated the same set of articles 3–6 months after publication to test time-robustness.
• Study 4 (within-respondent, COVID-19 salience): 1,130 respondents over 8 days in June 2020; 386 evaluated one false/misleading COVID-19 article within 72 hours of publication.
• Study 5 (between-respondent with digital trace): 1,677 US respondents via Amazon Mechanical Turk over 12 days; presented three highly popular articles within 72 hours of publication; treatment required to search via Google before assessing veracity; custom browser extension collected URLs visited and top ten Google results; NewsGuard scores used to classify site reliability (score <60 unreliable; >60 reliable; >85 used to define very reliable in a subset analysis). Compliance verified for full compensation; differential extension compliance documented. Primary analysis used OLS with article fixed effects and clustered SEs for binary and ordinal outcomes. The authors controlled for demographics and ideological congruence unless otherwise noted. Mechanism analyses in study 5 related treatment effects to the quality composition of search results (unreliable share thresholds; quartiles of mean NewsGuard across top ten links) and examined the role of search terms (e.g., using article headline/URL) and individual covariates (digital literacy, ideology). Bayesian t-tests were used to corroborate some null findings. Preregistrations and detailed protocols are referenced in Supplementary Information.

Effect of SOTEN on belief in misinformation (false/misleading articles):

• Study 1: Encouragement to search increased probability of rating a false/misleading article as true by 0.057 (P=0.037; Cohen’s d=0.12; n=2,275), about a 19% increase. On the 7-point scale: +0.16 (P=0.154; d=0.09; n=2,275).
• Study 2 (within-respondent): +0.071 (P<0.0001; d=0.15; n=2,020) and +0.24 on 7-point scale (P=0.0004; d=0.13). Among those initially correct (false/misleading), 17.6% switched to true after searching; among those initially incorrect (true), 5.8% switched to false/misleading after searching. Those initially uncertain more often moved to true than to false/misleading after searching.
• Study 3 (3–6 months later): +0.066 (P=0.0018; d=0.14; n=1,964) and +0.23 on 7-point scale (P=0.0038; d=0.13).
• Study 4 (COVID-19): +0.067 (P=0.0452; d=0.14; n=772) and +0.26 on 7-point scale (P=0.0054; d=0.14).
• Study 5 (with digital trace): +0.107 (P=0.0143; d=0.21; n=1,485) on binary outcome; +0.16 on 4-point scale (P=0.0434; d=0.16); no significant effect on 7-point scale (P=0.201; d=0.10). Mechanism: exposure to low-quality search results (study 5):
• Queries about false/misleading articles more often returned unreliable links than queries about true articles: 38% vs 15% exposed to at least one unreliable link (22.5 pp difference; F=105.8; P<0.0001).
• Among treatment subgroup exposed to ≥10% unreliable links, belief in misinformation was higher than control (n=1,027; P=0.0035; d=0.29). Among those exposed only to very reliable links (top ten links all NewsGuard >85), no difference from control (n=940; P=0.926–0.927; d≈0.01).
• Quartiles of mean news quality (top ten links): lowest quartile showed increased belief vs control (n=1,006; P=0.0241) and second-lowest quartile (n=1,005; P=0.0116), while second-highest (n=1,006; P=0.801) and highest (n=1,008; P=0.420) showed no increase. Top 50% quality overall showed no increase vs control (n=1,113; P=0.429). Analyses acknowledged post-treatment nature of search results. Who gets unreliable results and why (study 5):
• Lower digital literacy correlated with exposure to unreliable results after adjusting for demographics; a 1 s.d. increase in ideological congruence associated with +0.037 probability of exposure (P=0.0827; d=0.08; n=501).
• Search terms mattered: using the article headline/lede or the unique URL increased exposure. 77% of such searches returned ≥1 unreliable link vs 21% for other queries (55.8 pp; F=157.8; P<0.0001). Excluding the original article: 57% vs 18% (39.7 pp; F=85.5; P<0.0001). A 1 s.d. increase in digital literacy decreased probability of using headline/URL by 0.034 (P=0.016; d=0.11; n=930).
• Case example: adding “engineered” to “famine” changed results; 0% of “famine” queries returned unreliable results vs 63% with “engineered famine”; 83% of queries that returned unreliable results contained “engineered famine”. Effect on belief in true news:
• Study 1: Overall +0.072 (P=0.0001; d=0.146; n=6,269). By source: mainstream +0.045 (P=0.0168; d=0.10; n=3,487); low-quality +0.105 (P=0.001; d=0.21; n=2,782).
• Study 2: Overall +0.0212 (P=0.083; d=0.044; n=6,046). Low-quality +0.081 (P<0.0001; d=0.16; n=2,596); mainstream −0.024 (P=0.069; d=0.05; n=3,450).
• Study 3: Overall +0.047 (P=0.0001; d=0.097; n=5,908). Low-quality +0.115 (P<0.0001; d=0.25; n=2,490); mainstream null (P=0.84; d=0.01; n=3,418).
• Study 4: Overall 0.03 (P=0.165; d=0.062; n=1,420). Low-quality +0.085 (P=0.044; d=0.17; n=516); mainstream null (P=0.92; d=0.01; n=904).
• Study 5: Overall +0.15 (P<0.0001; d=0.357; n=3,141). Low-quality +0.23 (P<0.0001; d=0.50; n=1,350); mainstream +0.091 (P=0.0008; d=0.24; n=1,791). Pattern: Between-respondent designs show similar search effects on true mainstream and false/misleading news, with larger effects for true low-quality news; within-respondent designs show little to no effect on true mainstream news but positive effects for true low-quality news and false/misleading news. A possible ceiling effect is noted for mainstream true news (65–80% correct in controls).

The findings show that encouraging or requiring online search when evaluating news can increase belief in misinformation, countering common media literacy recommendations. This effect persists across experimental designs, time since publication, and topics, including COVID-19. Mechanism evidence suggests that data voids and exposure to low-quality search results drive the effect: when searches return unreliable links, belief in misinformation increases; when results are very reliable, the effect disappears. The likelihood of falling into data voids is higher for queries about misinformation than true news and is exacerbated by low digital literacy and “lazy searching” strategies (e.g., copying headlines/URLs or using distinct misinformation phrases). SOTEN also increases belief in true news—especially from low-quality sources—while effects on true mainstream news are smaller or absent in within-respondent contexts, potentially due to ceiling effects. These results suggest that media literacy interventions that simply advise users to “search it” may unintentionally increase susceptibility to misinformation unless paired with strategies that improve search practices (e.g., lateral reading) and platform-side mitigations that address data voids and the quality of early search results.

Across five experiments, the paper demonstrates that searching online to evaluate news can increase belief in misinformation, primarily when search results include low-quality sources. Digital trace analyses implicate data voids and specific search behaviors (e.g., using headlines/URLs or unique misinformation phrasing) as pathways to low-quality exposure. The work also shows that SOTEN increases belief in true news, particularly from low-quality sources, with inconsistent effects for mainstream true news. Contributions include causal evidence on SOTEN’s unintended consequences, mechanism-consistent evidence linking search result quality to belief, and nuanced heterogeneity by source quality and experimental design. The authors recommend that media literacy programs ground guidance in empirically validated practices (e.g., lateral reading) and that search engines invest in solutions such as improving early-stage result quality and warnings where credible information is scarce. Future research should examine real-world dissemination of search-based guidance, allow user self-selection of content, and evaluate platform and educational interventions at scale in observational and field settings.

Key limitations include: (1) External validity and timing: effects are measured at the point in time investigated (often within 48–72 hours of publication) and may vary as the information environment evolves. (2) Article selection and self-selection: participants did not self-select content as in naturalistic settings; results focus on highly popular articles and may not capture individual content preferences. (3) Mechanism analyses condition on post-treatment variables (search result quality), limiting causal inference about mechanisms. (4) Study 5 experienced differential extension compliance between treatment and control due to technical differences, though balance checks indicated no substantive differences; still, potential selection cannot be fully ruled out. (5) Generalizability: samples were US-based (Qualtrics panels; MTurk in study 5) and used Google; findings may differ across countries, search engines, and populations. (6) Potential ceiling effects for mainstream true news may mask search impacts in some designs. (7) Fact-check availability and lag could influence early search environments; some false narratives are never fact-checked. (8) Measured digital literacy and ideological congruence are proxies and may not capture all drivers of search behavior and exposure.