Guess, A. M., Nyhan, B., & Reifler, J. (2020). Exposure to untrustworthy websites in the 2016 US election. Nature Human Behaviour, 4, 472–480. https://doi.org/10.1038/s41562-020-0833-x
Summary
This paper offers one of the first rigorous individual-level behavioral analyses of exposure to “fake news” from untrustworthy websites during the final weeks of the 2016 U.S. presidential campaign. By linking survey responses to passively collected, URL-level web traffic from a national YouGov Pulse panel, the authors show that consumption of factually dubious content followed a pattern of selective exposure — Trump supporters and those with strongly conservative information diets disproportionately visited pro-attitudinal untrustworthy sites. Crucially, however, such sites made up only a small fraction (~6%) of the average American’s news diet, and exposure was heavily concentrated among a narrow subgroup. The paper thus argues that alarm about the prevalence of fake news has been overstated, even as effects on turnout and vote choice remain too imprecise to estimate.
Key Contributions
- First comprehensive individual-level, behavioral (non-recall, non-forced-exposure) study linking real-world web traffic, candidate preference, and validated turnout for untrustworthy-site consumption in 2016.
- Quantifies the actual prevalence and concentration of fake news exposure, correcting overstated claims based on self-reports and aggregate data.
- Provides strong independent behavioral evidence of Facebook’s role as a distribution vector.
- Documents the “fact-checking mismatch”: corrective content almost never reaches those exposed to the misinformation it targets.
- Tests and finds no support for the claim that fake news crowded out hard news, and rules out only very large effects on the election outcome.
Methods
The authors combined an online survey (n=3,251; ~2,525 with desktop/laptop traffic data) with passively collected, time-stamped URL-level traffic from YouGov’s Pulse panel, October 7–November 14, 2016. Untrustworthy sites were defined via the Grinberg et al. classification of “black,” “orange,” and “red” domains, with robustness checks using Allcott & Gentzkow domains and fact-checked-false articles. Exposure was measured both as binary visits and as share of a news diet (hard-news plus untrustworthy sites). An information-diet slant measure (from Bakshy et al. sharing-based scores) sorted respondents into deciles. Facebook’s referral role was estimated via usage terciles and by inspecting the last three URLs preceding untrustworthy visits. Analyses used weighted OLS (with probit checks), difference-in-differences against 2015 traffic for crowd-out, TargetSmart voter-file validation for turnout, and CRT scores.
Findings
- 44.3% of Americans visited an untrustworthy website; such sites were only 5.9% of the hard-news diet (4.6% conservative, 0.6% liberal).
- 56.7% of Trump supporters visited an untrustworthy conservative site vs. 27.7% of Clinton supporters; these sites were 11.0% of Trump supporters’ diet vs. 0.8% for Clinton supporters.
- In the most conservative decile, 84.0% visited untrustworthy conservative sites (~20.9% of their diet, averaging 91 articles); 62% of all untrustworthy traffic came from the top two conservative deciles — echo chambers “deep but narrow.”
- Users spent more time per untrustworthy article (64.2s) than on hard news (42.1s); high-CRT Trump supporters consumed less, suggesting genuine engagement and belief.
- Facebook preceded 15.1% of untrustworthy visits vs. 5.9% for hard news, unlike Google or Twitter; heavy Facebook users had far higher exposure.
- Only 18.6% of untrustworthy-site visitors saw any fact-check, and just 3 of 111 who read a matchable false article read its corresponding fact-check.
- Hard news and untrustworthy consumption were positively associated (no crowd-out); effects on Trump support and turnout were imprecise, ruling out only effects around 9–10 points.
Connections
This study is a foundational behavioral account of misinformation exposure that later work builds on and refines; it draws directly on the source classification and diet-measurement approaches of Grinberg2019-ua and Guess2019-ym, and complements aggregate and platform-level findings in Gonzalez-Bailon2024-rq and Budak2024-ef. Its “deep but narrow” concentration thesis and skepticism about large persuasion effects connect to Allcott2017-yz, Guess2021-ym, and the broader misinformation-prevalence debate in Allen2020-nj, while its fact-checking-mismatch finding relates to corrective-information work such as Pennycook2021-jq and Nyhan2023-gb.