In ranking, recommending, and moderating content, platforms make governance decisions that affect billions of people —for instance, by controlling the reach of problematic content like misinformation. We present the first behavioral field study analyzing how different governance decisions made by a major online platform in the ordinary course of operation affect the reach of misinformation. Platforms have typically imposed restrictions on data access and research partnerships that limit causal research in this area to laboratory methods and survey measures, which can have limited external validity , especially for behavioral outcomes . We conducted a causal experiment with high external validity by partnering with a web search engine to collect nearly half a million in situ observations of user behavior on search results pages that contained links to misinformation websites. Our preregistered analysis shows that algorithmic deamplification reduced engagement with misinformation websites by over half, while informative interventions had statistically insignificant effects on engagement. These results suggest that research and platform priorities should shift away from informative interventions—which have been the overwhelming focus for the past eight years [4, 5]—and toward algorithmic interventions, which are comparatively understudied. We further conclude that co-designed studies with platforms are vital to improving scientific understanding of how platform algorithms and design affect society.
Platforms like search engines and social media play an essential role in the spread of speech today by ranking, recommending, and moderating content for billions of people . In this role, platforms must make difficult decisions like how to handle misinformation and other problematic content. The stakes for these decisions are high because when misinformation and other problematic content spreads widely on platforms, it can undermine democracies , empower authoritarians , and lead to violence and genocide [8, 9].
One reason that addressing the harmful consequences of platform-mediated speech is challenging is that platform companies resist public scrutiny. They suppress internal research investigating their products’ harms [10–12] and restrict independent research by limiting researchers’ access to platform data [13–15]. Data that platforms do share with researchers is often incomplete or inaccurate [16, 17]. This information asymmetry between platforms and the public impedes policymaking around acute global challenges like the spread of misinformation.
U.S. and E.U. regulators have pressed platforms for years to reduce engagement with misinformation [18, 19]. In response, platforms primarily highlight two types of interventions that they deploy: removal interventions, like content takedowns and account bans , and informational interventions, like fact checks, content labels, and information panels . Only a small fraction of misinformation on platforms qualifies for removal, so informational interventions are primarily what platforms point to to demonstrate their commitment to countering misinformation.
Platform companies have not demonstrated, however, that informational interventions effectively reduce engagement with misinformation. Platforms could conduct and release internal research on this issue, but they have not done so. Further, platforms have elided the role of another type of misinformation intervention they routinely deploy: algorithmic deamplification, which reduces the reach of content through changes to recommendation and ranking algorithms . The constant stream of announcements about informational interventions, contrasted with the scarce discussion of deamplification, has led some observers to conclude that informational interventions—rather than behind-the-scenes algorithmic tweaks—must be working . But platforms have not provided evidence to support this conclusion.
Lacking a complete picture from platforms, independent scholars have developed a large body of research to explore the effects of misinformation interventions (see SI6 for our review of 272 papers in this area). Researchers have extensively studied informational interventions—especially factual corrections, which comprise 202 (75 percent) of the papers in our literature review. These studies typically find that informational interventions cause modest improvements in the accuracy of users’ beliefs and users’ ability to discern misinformation [23–25].
But there are significant limitations to this prior work. Without access to platform data and systems, researchers have been largely confined to laboratory methods, survey experiments, and self-reported measures of behavioral intentions rather than direct observation of user behavior. These methods can have limited external validity, particularly when studying behavioral interventions , because participants may modify their behavior in laboratory settings  or inaccurately report their behavior [27, 28]. Further, there are no prior causal studies of algorithmic interventions like deamplification because it is exceedingly difficult to study the effects of changes to algorithmic systems—which are complex, heterogeneous, and display emergent behavior—without direct access to those systems .
In this work, we present one path forward for studying how platform interventions affect user behavior and the spread of problematic content. We co-designed an externally valid field experiment in partnership with a major platform, conducting the first scholarly study of how real interventions deployed in the normal operation of a platform affect user engagement with misinformation.Our partner in the study was DuckDuckGo, a privacy technology company whose search engine serves nearly 100 million daily queries to users around the world. In a preregistered experiment spanning six languages and ten countries, we tested three types of interventions that have been widely deployed by major platforms: information panels, a related news module that amplified trustworthy news, and algorithmic deamplification of misinformation. We randomly deployed these interventions to over 463,000 search results pages (SERPs) where links to misinformation websites appeared in search results, and using privacy-preserving methods, measured user engagement with news, misinformation, and non-news links.
The key finding from our preregistered analysis is that an algorithmic deamplification intervention reduced engagement with misinformation by more than 50 percent, while the information panels and related news module had statistically insignificant effects on engagement with misinformation. The related news module, which displayed links to news stories at the top of the SERP, did successfully amplify news, however, leading to a 40 percent increase in engagement with news links.
Prior field data shows that content ranking has powerful effects on users’ content choices [30, 31], so it is not surprising that deamplifying misinformation can decrease misinformation engagement and promoting news can increase news engagement. But placed in context with prior field studies of misinformation interventions, which have examined informational interventions and found small or insignificant behavioral effects [32, 33], our results are a strong signal that platforms, researchers, and regulators should direct more attention towards interventions that amplify desired content and deamplify problematic content.
While our collaboration with DuckDuckGo provides a model for future experiments co-designed by researchers and platforms, it is not a scalable approach because it relies on researchers’ personal connections to industry and the goodwill of platform companies. As regulators increasingly consider implementing researcher data access mandates for platforms, we urge based on our experience that these mandates include opportunities for co-designed research. Researchers (and other oversight groups like journalists and civil society organizations) should have input on what data is collected, how it is collected, and what interventions are experimentally evaluated.
This approach is not without ethical risks, however. Field experiments con- ducted by technology companies carry a risk of harm to participants, especially when the experiments test interventions that violate users’ normative expectations about their interactions with the platform. Researchers can take steps to minimize these risks, as we did, by testing benign interventions that carry little risk to users, proactively informing users about changes to platform practices, and prioritizing user privacy.
2. Behavioral Effects of Misinformation Interventions
We studied how three types of common misinformation interventions deployed on search results pages (SERPs) affected user behavior. Our study participants were search engine users who queried for information about the Russian invasion of Ukraine—a subject about which Russian state-affiliated media sources have been widely documented publishing disinformation . We measured whether or not users selected results, the types of results users selected, and whether users engaged with informational interventions. Full methodological details are included in SI.
In accordance with DuckDuckGo’s privacy guarantees, data collection was fully anonymized. We did not collect any user identifiers, search query terms, or any parts of the URLs contained in search results. To measure the types of results users were shown and selected, we collected a classification label for each result indicating whether the result domain was a news, misinformation, or other type of website. For the purpose of the study, we defined misinformation websites as Russian state-affiliated media outlets with a well-documented record of publishing false information, censoring facts, and ceding editorial control to the Russian government. We describe our classification and data collection schema in full detail in SI.
For each intervention, we hypothesized that we would observe decreased selection of misinformation results, increased selection of news and other results, and an increase in the proportion of SERPs where users navigated away without clicking any result. Following our preregistration plan, we used one- sided, two-sample z-tests at α = 0.05 to test these hypotheses: left-tailed for misinformation results, and right-tailed for the other outcomes. We present descriptive statistics and the results of these hypothesis tests in Table 1.
Table 1 Participants were randomly assigned to one of nine treatments or a control condition. We report observation counts, the rates at which users selected each type of result, and the rates at which users left SERPs without selecting a result. We also report the rates at which users clicked links in interventions. As a measure of effect size, we report the relative risk comparing each treatment to the control (with 95 percent confidence intervals). Effect sizes printed in bold are statistically significant at α = 0.05.
Table 1: Participants were randomly assigned to one of nine treatments or a control condition. We report observation counts, the rates at which users selected each type of result, and the rates at which users left SERPs without selecting a result. We also report the rates at which users clicked links in interventions. As a measure of effect size, we report the relative risk comparing each treatment to the control (with 95 percent confidence intervals). Effect sizes printed in bold are statistically significant at α = 0.05.
On SERPs assigned to a deamplification condition, the rank of each result classified as a misinformation website was reduced by a factor drawn from a distribution. In the strong deamplification condition, we drew this factor from a normal distribution centered on 8 with a standard deviation of 3. In the weak deamplification condition, the distribution was centered on 6 with a standard deviation of 2.
Strong deamplification more than halved the rate at which users selected misinformation results, from 5.84 percent in the control condition to 2.90 percent, a risk ratio of 0.497 ± 0.036. This effect was statistically significant (z = −20.265, p < 0.001). Weak deamplification decreased selection of misinformation results by nearly 30 percent (risk ratio of 0.711 ± 0.042), an effect that was also statistically significant (z = −11.589, p < 0.001). Both deamplification treatments also had significant effects on selection of news and other results, though the relative changes in behavior are small because the base rates are much higher. Deamplifying misinformation results did not significantly increase the likelihood that users left SERPs without selecting a result.
Figure 1: Participants had a strong tendency towards selecting highly ranked results. This tendency was more pronounced for misinformation results. We calculated odds ratios using a logistic regression model (p < 0.001 for all results).
Ranking and click behavior
Across all result types, participants had a strong tendency to choose top ranked results, selecting one of the top three results 62 percent of the time (Figure 1). On average, decreasing the rank of a result by one decreased the odds of it being clicked by 22 percent (logistic regression model, p < 0.001). The distribution of result clicks by rank aligns closely with prior search engine field data [30, 31].
Participants were more likely to click misinformation when it appeared at high ranks than they were to click news or other types of results at high ranks (Figure 1). We theorize that searchers may have found the misinformation results highly germane to their queries, because the study only included queries where the user was searching for information about Russia and/or Ukraine (see SI1). For these searches, it makes sense that Russian state media websites would have highly relevant content.
Decreasing the rank of a misinformation result by one led to a 26 percent decrease in click odds (logistic regression model, p < 0.001)—a larger decrease than we observed for news (21 percent) or other (22 percent) types of results. This trend held—and was actually more pronounced—when we excluded SERPs where a deamplification intervention changed the ranks of results (see SI5).
2.2. Information panels
The first type of informational intervention we tested was an information panel: a page-level module displaying a message and a redirection link to an authoritative source of information. Search engines and social media platforms have widely deployed information panels as misinformation interventions [35– 37]. The messages in these interventions typically aim to preemptively refute, or “prebunk” misinformation by providing factual information that counters misinformation and/or literacy tips explaining how to resist misinformation . Researchers using survey experiments to test information panels similar to those deployed by platforms have found positive, but very small effects on participants’ ability to discern misinformation [39, 40].
Figure 2: We show the mainline version of the humanitarian information panel intervention as it appears on desktop.
Each information panel intervention we developed displayed a message quoted from the Wikipedia article “2022 Russian Invasion of Ukraine” and a link to that article (see SI1). We chose Wikipedia as the message source and redirection destination because it is a reliable, authoritative source of information on a wide variety of topics , and major platforms routinely deploy information integrity features based on Wikipedia content [42–44]. We created three versions of the information panel, each with different messages (see SI3). For each message, we created a mainline version, which appears above organic results (Figure 2), and a sidebar version, which appears to the right of organic results (Figure S1, in SI1), for a total of six information panel conditions.
Two of the messages we selected aimed to preemptively refute Russian dis- information narratives. Russian leaders and media outlets have consistently reinforced the false claims that Russia did not invade or attack Ukraine, and that Russia’s goal in Ukraine is to protect the people living there . Both of our prebunking panels state that Russia did invade Ukraine; the condemnation panel cites the United Nations and International Court of Justice as calling for the withdrawal of Russian forces, and the humanitarian panel describes the invasion’s heavy toll on the civilian population of Ukraine.
We did not measure significant decreases in selection of misinformation results caused by either prebunking panel, nor did we observe significant in- creases in selection of news results. We also found that users rarely clicked the links in prebunking panels (between 0.06 percent and 0.27 percent of the time).
Figure 3: We show the related news module intervention as it appears on desktop.
The third information panel displayed a warning message stating that re- porting by Russian media on Ukraine is heavily censored. We selected this message to create an intervention that simulates a general misinformation warning, which is a type of informative intervention that alerts the user to a potential risk of misinformation  (rather than refuting or undermining specific misinformation, as a prebunking message does). In prior work, content- level misinformation warnings have shown insignificant or small effects on user beliefs and behavioral intentions [47, 48], as have general warnings . Interstitial content-level warnings have shown large, significant effects, however , as have contextual content-level warnings that also include a descriptive norm message .
We found that the warning panels significantly increased the likelihood that users selected news results (z = 3.296, p = 0.0005 for the sidebar panel; z = 3.0102, p = 0.0013 for the mainline panel). The relative change was under half a percent for both versions of the warning panel, however. As with the prebunking panels, users rarely clicked the redirection links in the warning panels (between 0.06 percent and 0.23 percent of the time).
Related News Module
Our final intervention was a related news module, which displayed a series of headlines and links to news stories relevant to the user’s query. Facebook has deployed a Related Articles module as an intervention against misinformation , and similarly Google includes of news carousels on SERPs as misinformation interventions . A laboratory study found that in the con- text of social media, a related news module significantly improved the accuracy of user beliefs , but no prior research has examined the behavioral effects of related news modules deployed in the field.
For our study, the related news module displayed a carousel of headlines and links to news stories relevant to the user’s search query (Figure 3). The related news module decreased selection of misinformation results compared to the control, but the decrease was not statistically significant after correcting for multiple comparisons.
Participants heavily engaged with the related news module, leading to an 40 percent increase in clicks on news links on SERPs. Due to an implementation error, we did not collect click data for the news module for the first four weeks of the study, so we imputed the missing data and ran a series of robustness checks to validate our imputation method (see SI5).
Informational interventions have shown promise in prior work, but we show evidence that they may not be effective in real deployments. Where laboratory studies of informational interventions have reported significant decreases in participants’ intention to share misinformation, we found no decrease in participants’ real engagement with misinformation. This discrepancy underscores the need for more ecologically valid field research on the behavioral effects of informational interventions.
Algorithmic deamplification interventions are significantly understudied, but based on our evidence, they are likely very effective at reducing engagement with misinformation in real-world deployments. To interrogate this claim, researchers need more rigorous information about how deamplifying different categories of content affects user engagement.
Platforms should publish detailed standards specifying what categories of content they will deamplify and the magnitude of deamplification they will apply. Platforms already publish standards explaining what content they remove; we propose they extend this approach to deamplification. Published standards for content deamplification would allow researchers to conduct large-scale observations of content reach and engagement. This research is currently hindered by the opacity of platform algorithms because researchers cannot distinguish the different factors that contribute to the reach of content.
Large-scale observation may not be sufficient to justify individual causal claims, however. Thus, we also recommend that platforms co-design studies with researchers so that researchers can study specific causal effects of interest. In the E.U., the Digital Services Act will soon require platforms to allow independent researchers access to data in order to conduct studies, but if researchers cannot participate in designing studies and data collection procedures, it will still be difficult to collect causal evidence. Platforms have previously cited privacy concerns to explain why they do not grant researchers exceptional access to systems or user data, but our study shows that careful research design can mitigate those concerns, as we were able to collect causal evidence while fully preserving user privacy and anonymity, as we discuss below.
3.1. Ethics and privacy
Internet research ethics is an evolving field, with best practices still emerging . Relevant to our study, there is a history of ethical concerns stemming from field experiments conducted by technology companies without affirmative consent from participants [55, 56]. The primary concern that has been raised is that such studies pose an increased risk of harm to participants . There are two reasons for this concern. First, studies conducted by private companies are not obligated to be independently reviewed for risks to human subjects (e.g., by an Institutional Review Board), so it can be left up to the researchers to decide whether the benefits of the study outweigh the risks . Second, users’ normative expectations about their interactions with an online service may be violated if the experiment involves changes to the normal practices of the service and users are not notified .
We took efforts to address both of these concerns in designing our study. To start, we deployed interventions with a low risk of harm to users. Our interventions are designed to improve information quality and are modeled on interventions commonly deployed by online platforms, so they are both benign and routine. We received approval to carry out this human subjects research from the Princeton IRB. Further, prior to deploying the experiment, DuckDuckGo proactively disclosed the new intervention created for the study and their policy for deploying the intervention. This disclosure was published in plain language in DuckDuckGo’s help pages .
We also designed the study with strong privacy protections, in accordance with DuckDuckGo’s promise to their users that “when you use our services, we have no way to create a history of your search queries or the sites you browse” . We did not collect any identifiers, even pseudonymous identifiers like IP addresses. We did not collect query terms, which could reveal information about a user. In recognition of the uniqueness of web browsing behavior , we did not collect any parts of the URLs contained in displayed search results, nor did we collect any parts of the URLs of the search results that users selected.
Due to our study’s privacy protections, we did not persist treatment assignment per-participant or measure for repeat observations of the same participant. Further, we did not collect information allowing us to detect if a user selected multiple results on the same SERP (e.g., by opening each result in a new tab) or clicked links in the same intervention multiple times.
Supplementary information. Files referenced below are available for download at https://github.com/citp/misinformation-intervention-study-SI.
SI1. Study Population
We collected anonymized event data on 463,024 SERPs served to DuckDuckGo users on desktop and mobile browsers for the twelve week span between January 18, 2023 and April 14, 2023. A SERP had to meet four criteria to be included in the study:
- The search query terms contained at least one keyword related to the Russia-Ukraine conflict (see below).
- The search language was either English, Russian, German, Spanish, Portuguese, or Italian.
- DuckDuckGo identified the query as non-navigational, meaning that the query did not include keywords indicating that the user’s purpose was to reach a specific website .
- The search results contained at least one result linking to a misinformation website (see below).
We designed our study with a focus on Russian propaganda about Ukraine. Russia has conducted extensive information operations as part of its military actions against Ukraine, and state-affiliated media outlets have been a key distribution channel for state propaganda . In creating a list of Russian state-affiliated media websites to label as misinformation for this study, we adhered to DuckDuckGo’s News Rankings policy , which describes how DuckDuckGo uses independent, non-political assessments of journalistic standards to inform ranking for news-related search results. The policy states that DuckDuckGo must see “at least three... non-governmental and non-political organizations that specialize in objectively assessing journalistic standards... independently assess a site as having extremely low journalistic standards and also see that none of these organizations have assessed the same site as having even somewhat robust journalistic standards.” We identified eleven websites that met this criteria and were also independently assessed as being under the editorial control of the Russian government. The websites are: rt.com, sputniknews.com, news-front.info, strategic-culture.org, usareally.com, journal-neo.org, katehon.com, pravda.ru, pravdareport.com, snanews.de, and southfront.org.
Query inclusion criteria
We developed a set of query terms to identify searches relevant to the Russian invasion of Ukraine. We started with an initial term list drawn from Chen and Ferrara . For each English-language term in their dataset, we sampled 300 searches (or as many searches as were available) from between April 2022 and August 2022 and manually labeled each search as either relevant or irrelevant to at least one of the following topics:
- The invasion of Ukraine
- Russian or Ukrainian politics, economics, history, military info, current events, notable people, landmarks, basic facts, or geography
- Global politics and events related to the invasion of Ukraine
- Any news coverage of Russia or Ukraine, or by Russian or Ukrainian outlets
We discarded any query terms with less than 85 percent query relevance, leaving 15 terms. We then conducted a round of snowball sampling to identify additional query terms that were frequently used in conjunction with the initial terms. We evaluated these additional query terms using the same method as for the initial list, resulting in 6 additional terms. Finally, we translated the terms into each of the languages our study supported. The query terms are:
- ukraine, ucrania, ucraˆnia, украина, украины, ukraina
- russia, russie, rusia, russland, ru ́ssia, россия, rossiya
- putin, poutine, путин
- soviet, sovi ́etique, sovi ́etico, sowjetisch, советский, sovetsky
- kremlin, kreml, кремль
- minsk, минск
- ukrainian, ukrainien, ucranio, ukrainisch, ucraniano, украинец, ukrainets
- NATO, OTAN, НАТО
- luhansk, lugansk, lougansk, луганск
- donetsk, donezk, донецк
- donbas, donbass, donb ́as, dombas, донбасс
- kyiv, kiev, kiew, kiiv, киев
- moscow, moscou, moscu ́, moskau, москва, moskva
- zelensky, zelenskiy, zelenski, зеленский
- KGB, КГБ
- crimea, crim ́ee, krim, crimeia, крым, krym
- kharkov, kharkiv, jarkov, charkow, carco ́via, харьков
- belarus, bi ́elorussie, bielorrusia, weißrussland, bielorru ́ssia, беларусь
- nova kakhovka, nouvelle-kakhovka, nouvelle kakhovka, nova kajovka, nova
- kachowka, новая каховка, novaya kakhovka
- kherson, kerson, cherson, херсон,
- duma, douma, дума
SI2. Data Collection
Our study triggered each time a DuckDuckGo user loaded a SERP meeting our inclusion criteria (see SI1). On each SERP load, we collected data describing which condition the SERP was randomly assigned to, whether it was a requeryor bounceback, the country or U.S. state to which the user’s IP address geolocated, the user’s search language, and the date. For each search result displayed, and for each result loaded as part of the back-end request but not displayed, we classified it clientside as either misinformation, news, or other based on the domain of the website to which the result linked. We also collected the displayed rank and the original rank (before any downranking interventions were applied) for each result.
Each time a user clicked a result on a SERP that triggered the experiment, we recorded information about the result selected (its type, displayed rank, and original rank). Similarly, each time a user clicked the redirection link in an infobox intervention or one of the news article links in the news module, we recorded a flag indicating that the user engaged with the intervention.
SI3. Information Panels
The English-language messages for each information panel were:
- Humanitarian: “On 24 February 2022, Russia invaded Ukraine... and caused Europe’s largest refugee crisis since World War II, with an estimated 8 million people being displaced within the country by late May as well as 7.7 million Ukrainians fleeing the country.”
- Condemnation: “On 24 February 2022, Russia invaded Ukraine... The United Nations General Assembly passed a resolution condemning the invasion and demanding a full withdrawal of Russian forces. The International Court of Justice ordered Russia to suspend military operations.”
- Warning: “On 24 February 2022, Russia invaded Ukraine... Protests occurred around the world; those in Russia were met with mass arrests and increased media censorship, including a ban on the words ‘war’ and ‘invasion.’”
Research assistants who were proficient in each of the languages included in the study assisted us in finding comparable quotes from foreign language-editions of Wikipedia.
Figure 4: Warning information panel, sidebar
SI4. News Websites
A key goal of our study was to measure whether interventions made users less likely to select Russian state-affiliated media (RSAM) outlets and more likely to select other, credible news results. We created lists of RSAM and credible news websites to classify search results client-side. The list of RSAM websites is included in SI1. For the dataset of news websites, we chose three sources:
- ABYZNewsLinks.com, a directory of links to online news sources from around the world. We collected 36,241 news links using an automated scraper.
- Wikipedia’s list of newspapers by country, a crowdsourced dataset of news outlets from around the world that includes websites for many entries. We collected 7,234 news links by manually traversing Wikipedia entries.
- DMOZ, also called the Mozilla Directory, which is a directory of web links maintained by a community of volunteers editors. DMOZ has not been updated since 2017 but the final directory is still accessible. We collected 35,302 domains by downloading a data dump and filtering for domains labeled as News, News & Media, or Newspaper.
We identified several other datasets and chose not to use them. Curlie (the successor to DMOZ) does not provide a data dump or permit automated data collection. World-Newspapers.com, AllMediaLink.com, and AllYouCanRead.com had significantly fewer links per country than ABYZNewsLinks.com, which we chose to use instead. Worldpress.org was not reliably available during the period of data collection. Listings of news websites on BBC.com were not indexed in a way that they could be easily traversed for data collection.
From 78,777 links, we performed filtering and ranking as follows:
- Liveness and redirection: we issued a GET request to each link. If we didn’t receive a response, or received an error response code, we discarded the link. If we did receive a response, we saved the original link and the response URL.
- Normalization: we normalized each URL to a domain name by stripping the path and subdomain, unless the domain was a publishing platform like wordpress.com or substack.com, in which case we preserved the subdomain. This left us with 58,254 domains.
- Ranking: we ranked the popularity of each domain using Tranco, a ranking dataset that is used widely in academic research and contains over 7.5 million domains.
- Classification and validation: to evaluate the quality of the dataset, we ran the top 10,000 ranked domains through WebShrinker, a domain classification service. WebShrinker classified 5,731 (57.3 percent) of the domains as news. We randomly sampled 500 of these domains to manually confirm classification and found only 7 false positives, which we left in the dataset. We manually checked all 4,269 non-news classifications, finding 1,437 false negatives. We added these false negatives to the classified news domains. This constitutes our high confidence, top ranked news dataset of 7,168 domains.
- Quality filtering: we removed 28 domains that qualify as extremely low- quality news sources based on DuckDuckGo’s News Rankings policy, leaving us with 7,140 domains.
- Partitioning: we selected the top 1,000 domains as our Top 1k dataset and the remaining 6,140 became our Top 1-7k dataset. Of the remaining 38,254 domains in our original dataset, 17,261 appear on the Tranco list, so this is our Long Tail dataset.
We provide full results from our hypothesis tests in an attached file (full_results.csv). Two effect size measures are commonly used in related literature: standardized mean differences (also called Cohen’s d or Cohen’s h) and risk ratios. We chose to present risk ratios as our main effect size measure. Risk ratios capture the relative change in participant behavior for low base rate events like clicking misinformation results as well as high base rate events like clicking news results. To allow comparison with prior work, we also include Cohen’s h for each hypothesis test.
We did not set any geographic filters when deploying the study, although we did set language filters. The study only triggered on SERPs where the user’s language was set to English, Russian, German, Spanish, Portuguese, or Italian, and this influenced the distribution of countries where we conducted observations.
In attached files, we present results broken down for each of the top ten countries by observation count: the United States, United Kingdom, Canada, Germany, Australia, Russia, Switzerland, The Netherlands, France, and Ukraine. For seven of the ten top countries, our top line result holds: strong downranking showed a significant downward effect on selection of misinformation results. For three countries—Australia, France, and Ukraine—the effect was not significant.
We also present results broken down by US state, for the top ten states by observation count: California, Texas, Florida, New York, Illinois, Virginia, Washington, Pennsylvania, Ohio, and Georgia. For eight of the ten top states, our top line result holds: strong downranking showed a significant downward effect on selection of misinformation results. For two states—Washington and Pennsylvania—the effect was not significant.
Figure 5: On SERPs where no deamplification was applied, participants still had a strong tendency towards selecting highly ranked results. This tendency was more pronounced for misinformation results. We calculated odds ratios using a logistic regression model (p < 0.001 for all results).
We collected data for twelve weeks: January 18, 2023 through April 12, 2023. In attached files, we present results subdivided by week. There are thirteen files; the first (weekly_results_2023_01-18.csv) contains data for five days and the last (weekly_results_2023_04-10.csv) contains data for three days; all other files contain seven days of data.
Our top-level result holds for every week of data: strong downranking showed a significant downward effect on selection of misinformation results.
Imputed news module engagement data
Due to an implementation error, we did not capture clicks on links in the related news module intervention for roughly the first month of the study (from the beginning of data collection on January 18, 2023 through February 20, 2023). This missing data is visible in the country-level, state-level, and week- level breakdowns provided above. For the key results presented in the body of the paper, and the full results provided above, we imputed the missing data using the following method. For each day that we did collect click data for the related news module intervention, we computed the click rate as the number of clicks divided by the number of SERPs where the news module was shown. For each day where we did not collect click data for the related news module intervention, we randomly sampled a click rate from the distribution produced above and used that rate to impute the number of clicks. This imputation does not affect any of our hypothesis tests, because we did not test hypotheses about engagement with interventions.
Ranking and click behavior
Figure 1 shows the distribution of clicks by rank for each type of result. Figure S5 repeats that analysis excluding SERPs where a deamplification intervention changed the rank of results. This excluded 26,074 SERPs (9.3 percent of our observations). The general trends hold: participants had a strong tendency towards selecting highly ranked results and were more likely to click misinformation when it appeared at high ranks than they were to click news or other types of results at high ranks. For misinformation results, the decrease in click odds corresponding to a one-rank decrease was even larger when excluding SERPs where deamplification was applied: 32 percent here compared to 26 percent when SERPs affected by deamplification were included.
SI6. Landscape of Misinformation Interventions
Below we overview the types of misinformation interventions platforms have deployed and relevant scholarly research. In describing misinformation interventions deployed by platforms, we draw from three databases of platform announcements containing a total of 417 intervention announcements [5, 65, 66] as well as our own knowledge of platform practices. To cover scholarly research, we draw on a survey by Courchesne et al. of 223 papers presenting causal evidence on the effects of misinformation interventions . We extended this corpus by collecting an additional 46 papers. We searched Google Scholar, SSRN, arXiv.org, OSF Preprints, and PubMed for papers containing the keywords “misinformation”, “disinformation”, “conspiracy”, “propaganda”, “fake news”, or “information operations” along with either the keyword ”intervention” or “countermeasure.” For each database and each pair of keywords, we read the titles and abstracts of the top 100 results and selected papers meeting the above criteria. For survey papers and meta-analysis papers, we also mined the bibliographies for additional papers for our corpus. We coded the full corpus of 272 papers according to the type of treatment delivered and the outcomes (or dependent variables) measured, including re-coding the original corpus. The coded dataset of papers is provided in the file Lit-Review-final.xlsx. We conducted this literature review in February 2023.
Below we highlight real-world usage and key research findings related to informative and reduction interventions.
Informative interventions provide information to users, such as a fact check on a news story, a label indicating that a social media account belongs to a state-affiliated media organization, or a notice with tips on spotting misinformation. Content-level informative interventions present information in context with a specific piece of content, account, page, or group, while page-level interventions provide a general message, typically at the top of a feed or search results page.
Platforms have widely deployed fact checks as content-level informative misinformation interventions [67–69]. 75 percent (207/272) of all papers in our literature review study fact check treatments. Importantly, only 15 percent of these papers (11 percent of the total corpus) study fact check labels, which we define as brief fact check messages presented directly alongside the content being fact checked.
The remainder of the papers study factual corrections presented in other ways (e.g., as a reading participants are asked to complete , or as a user comment on a simulated social media post ). Results are mixed, but the majority of papers find that fact checks significantly reduce belief in misinformation, with a meta-analysis of 30 studies finding an average effect of d = 0.29 , which is considered a small-to-medium effect. Key studies have also found significant effects on behavioral intentions: one study of fact check labels on simulated Facebook posts found that the labels decreased user intention to share false posts by nearly half (46 percent, an absolute decrease of 13.7 percentage points) .
Source information and rating labels
Platforms have deployed a variety of labels that disclose information about information sources, such as state-affiliated media labels [73, 74], publisher information labels [75, 76], and source credibility indicators  When re- searchers have studied real source information labels deployed by technology companies—including in a field study —they have not found that the labels reduce user misperceptions [77, 78] or cause users to consume more news from credible sources . Labels designed by researchers, however, have demonstrated significant effects on user behavior  and beliefs [78, 79] in laboratory studies.
Related articles modules
Related articles modules present links to article from credible sources, typically related to a piece of content or a user’s search query. Facebook has previously attached related articles modules to posts containing misinformation , while Google deploys page-level news carousels as part of their efforts to counter misinformation . Two papers in our corpus studied related articles modules, finding mixed results: one study found no effect on user beliefs , while another study found that user misperceptions could be corrected by as much as half a point on a seven-point scale—a significant but small effect .
Platforms have routinely deployed information hubs, which display a message and links to authoritative information. During the COVID-19 pandemic, for example, most major social media and search platforms added information hubs to their feeds and search result pages with public health information and links to the CDC, WHO, or other health organizations [36, 65, 81–85]. Platforms have also deployed information hubs as standalone resources, which they then link to from page-level or content-level interventions [67, 86]. To our knowledge, no prior research has studied the causal effects of information hub interventions.
Literacy tips and warnings
Platforms have also deployed literacy tips as page-level interventions . Lab- oratory studies of page-level literacy interventions have had mixed results, including finding insignificant effects on user belief in misinformation  and significant effects which range from very small (B = −0.08, n = 2, 994)  to more substantial (B = −0.196, n = 4, 907) . A field study found that YouTube users shown literacy interventions, then shown a misinformation headline, were slightly better at identifying which misinformation tactic the headline used, but the effect size was very small (h = 0.09, n = 22, 632) . This research has not found significant effects on users’ behavioral intentions [39, 46]. Several of these studies cite lack of user attention as an explanation for the small effect sizes [39, 40]. We note that by this same logic, the small effect sizes demonstrated in laboratories may be inflated compared to real-world effects, because users likely pay more attention to the interventions in laboratory settings where they have been instructed to read the interventions, unlike on real platforms where there are no such instructions.
Platforms use reduction interventions to limit the circulation of problematic content without fully removing it. Reduction often takes the form of preventing or controlling how content appears in algorithmic recommendations, promotional features, or search results. Platforms have also sometimes prevented users from engaging with or “resharing” problematic content, which also reduces the content’s reach .
Platforms have shared few details about the frequency and effectiveness of their reduction interventions . Sporadic disclosures and leaked documents from platforms indicate that reduction has powerful effects on user engagement [90–92], but there are no comprehensive, rigorous investigations of these effects. Our literature review did not find any experimental research evaluating the effects of reduction interventions.
© 2023, Benjamin Kaiser and Jonathan Mayer.
Cite as: Benjamin Kaiser and Jonathan Mayer, It’s the Algorithm: A Large-Scale Comparative Field Study of Misinformation Interventions, 23-10 Knight First Amend. Inst. Oct. 23, 2023, https://knightcolumbia.org/content/its-the-algorithm-a-large-scale-comparative-field-study-of-misinformation-interventions [https://perma.cc/G3LP-HST].
 Klonick, K.: The New Governors: The People, Rules, and Processes Governing Online Speech. Harv. L. Rev. 131, 1598 (2017). Available at SSRN: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2937985
 Amazeen, M.A., Benevenuto, F., Brashier, N.M., Bond, R.M., Bozarth, L.C., Budak, C., Ecker, U.K., Fazio, L.K., Ferrara, E., Flanagin, A.J., Flammini, A., Freelon, D., Grinberg, N., Hertwig, R., Jamieson, K.H., Joseph, K., Jones, J.J., Garret, R.K., Kreiss, D., McGregor, S., Mc-Nealy, J., Margolin, D., Marwick, A., Menczer, F., Metzger, M.J., Nah, S., Lewandowsky, S., Lorenz-Spreen, P., Ortellado, P., Pasquetto, I., Pen-nycook, G., Porter, E., Rand, D.G., Robertson, R., Swire-Thompson, B., Tripodi, F., Vosoughi, S., Vargo, C., Varol, O., Weeks, B.E., Wihbey, J., Wood, T.J., Yang, K.-C.: Tackling Misinformation: What Researchers Could Do With Social Media Data. Harvard Kennedy School Misinformation Review 1(8) (2020). https://doi.org/10.37016/mr-2020-49
 DellaVigna, S., Linos, E.: RCTs to Scale: Comprehensive Evidence From Two Nudge Units. Econometrica 90(1), 81–116 (2022). https://doi.org/10.3982/ECTA18709
 Courchesne, L., Ilhardt, J., Shapiro, J.N.: Review of Social Science Re-search on the Impact of Countermeasures Against Influence Operations. Harvard Kennedy School Misinformation Review (2021). https://doi.org/10.37016/mr-2020-79
 Yadav, K.: Platform Interventions: How Social Media Counters Influence Operations. Technical report at https://carnegieendowment.org/2021/01/25/platform-interventions-how-social-media-counters-influence-opera tions-pub-83698 (2021)
 Barrett, P.M.: Spreading The Big Lie: How Social Media Sites Have Amplified False Claims of U.S. Election Fraud. Technical report at https://bhr.stern.nyu.edu/tech-big-lie (2022)
 Paul, K.: Flood of Russian Misinformation Puts Tech Companies in The Hot Seat. The Guardian (2022). https://www.theguardian.com/media/2022/feb/28/facebook-twitter-ukraine-russia-misinformation
 Goel, V., Raj, S., Ravichandran, P.: How WhatsApp Leads Mobs to Mur-der in India. The New York Times (2018). https://www.nytimes.com/interactive/2018/07/18/technology/whatsapp-india-killings.html
 Mozur, P.: A Genocide Incited on Facebook, With Posts From Myanmar’s Military. The New York Times (2018). https://www.nytimes.com/2018/10/15/technology/myanmar-facebook-genocide.html
 Oremus, W.: Facebook Keeps Researching Its Own Harms—And Burying the Findings. The Washington Post (2021). https://www.washingtonpost.com/technology/2021/09/16/facebook-files-internal-research-harms/
 Mac, R., Frenkel, S.: Internal Alarm, Public Shrugs: Facebook’s Employees Dissect Its Election Role. The New York Times (2021). https://www.nytimes.com/2021/10/22/technology/facebook-election-misinfo rmation.html
 Dave, P., Dastin, J.: Google Told Its Scientists to ’Strike a Positive Tone’ in AI Research - Documents. Reuters (2020). https://www.reuters.com/article/us-alphabet-google-research-focus/google-told-its-scientists-to-strike-a-positive-tone-in-ai-research-documents-idUSKBN28X1CB
 Roose, K.: Inside Facebook’s Data Wars. The New York Times (2021). https://www.nytimes.com/2021/07/14/technology/facebook-data.html
 Ledford, H.: Researchers Scramble as Twitter Plans to End Free Data Access. Nature, 602–603 (2023). https://doi.org/10.1038/d41586-023-00460-z
 Kayser-Bril, N.: AlgorithmWatch Forced to Shut Down Instagram Mon-itoring Project After Threats From Facebook. AlgorithmWatch (2021). https://algorithmwatch.org/en/instagram-research-shut-down-by-facebook/
 Rosenberg, M.: Ad Tool Facebook Built to Fight Disinformation Doesn’t Work as Advertised. The New York Times (2019). https://www.nytimes.com/2019/07/25/technology/facebook-ad-library.html
 Timburg, C.: Facebook Made Big Mistake in Data It Provided to Re-searchers, Undermining Academic Work. The Washington Post (2021). https://www.washingtonpost.com/technology/2021/09/10/facebook-error-data-social-scientists/
McCabe, D., Kang, C.: Lawmakers Grill Tech C.E.O.s on Capitol Riot, Getting Few Direct Answers. The New York Times (2022). https://www.nytimes.com/2021/03/25/technology/facebook-twitter-google-capitol-riots-hearing.html
 Sandberg, S.: Testimony of Sheryl Sandberg. United States Senate Select Committee On Intelligence (2018). https://www.intelligence.senate.gov/sites/default/files/documents/os-ssandberg-090518.pdf
 Facebook: Threat Report: The State of Influence Operations 2017-2020. Meta Newsroom. Technical report at https://about.fb.com/news/2021/05/influence-operations-threat-report/ (2021)
 Gillespie, T.: Do Not Recommend? Reduction as a Form of Content Moderation. Social Media + Society 8(3) (2022). https://doi.org/10.1177/205.63051E+14
 Hsu, T.: Misinformation Defense Worked in 2020, Up to a Point, Study Finds. The New York Times (2023). https://www.nytimes.com/2023/04/13/business/media/misinformation-2020-election-study.html
 Porter, E., Wood, T.J.: The Global Effectiveness of Fact-Checking: Evidence from Simultaneous Experiments in Argentina, Nigeria, South Africa, and the United Kingdom. Proceedings of the National Academy of Sciences 118(37), 2104235118 (2021). https://doi.org/10.1073/pnas.2 104235118
 Walter, N., Cohen, J., Holbert, R.L., Morag, Y.: Fact-Checking: A Meta-Analysis of What Works and for Whom. Political Communication 37(3), 350–375 (2020). https://doi.org/10.1080/10584609.2019.1668894
 Pennycook, G., Rand, D.G.: Accuracy Prompts are a Replicable and Generalizable Approach for Reducing the Spread of Misinformation. Nature communications 13(1), 2333 (2022). https://doi.org/10.1038/s41467-022-30073-5
 McCambridge, J., Witton, J., Elbourne, D.R.: Systematic Review of the Hawthorne Effect: New Concepts are Needed to Study Research Participation Effects. Journal of Clinical Epidemiology 67(3), 267–277 (2014). https://doi.org/10.1016/j.jclinepi.2013.08.015
 Scharkow, M.: The Accuracy of Self-Reported Internet Use—A Validation Study Using Client Log Data. Communication Methods and Measures 10(1), 13–27 (2016). https://doi.org/10.1080/19312458.2015.1118446
 Araujo, T., Wonneberger, A., Neijens, P., de Vreese, C.: How Much Time Do You Spend Online? Understanding and Improving the Accuracy of Self-Reported Measures of Internet Use. Communication Methods and Measures 11(3), 173–190 (2017). https://doi.org/10.1080/19312458.2017.1317337
 Kitchin, R.: Thinking Critically About and Researching Algorithms. In-formation, Communication & Society 20(1), 14–29 (2017). https://doi.org/10.1080/1369118X.2016.1154087
 The Value of Google Result Positioning. Chitika Insights (2013). https://research.chitika.com/wp-content/uploads/2022/02/chitikainsights-valueofgoogleresultspositioning.pdf Accessed 2023-04-12
 Glick, M., Richards, G., Sapozhnikov, M., Seabright, P.: How Does Ranking Affect User Choice in Online Search? Review of Industrial Organization 45, 99–119 (2014). https://doi.org/10.1007/s11151-014-943
 Bowles, J., Larreguy, H., Liu, S.: Countering misinformation via whatsapp: Preliminary evidence from the covid-19 pandemic in zimbabwe. PLOS ONE 15(10), 1–11 (2020). https://doi.org/10.1371/journal.pone.0240005
 Nassetta, J., Gross, K.: State Media Warning Labels Can Counter-act the Effects of Foreign Misinformation. Harvard Kennedy School Misinformation Review (2020). https://doi.org/10.37016/mr-2020-45
 Bor, A., Osmundsen, M., Rasmussen, S.H.R., Bechmann, A., Petersen, M.B.: “Fact-Checking”’ Videos Reduce Belief in but Not the Sharing of “Fake News”’ on Twitter (2020). https://doi.org/10.31234/osf.io/a7huq
 Mosleh, M., Martel, C., Eckles, D., Rand, D.: Perverse Downstream Consequences of Debunking: Being Corrected by Another User for Posting False Political News Increases Subsequent Sharing of Low Quality, Parti-san, and Toxic Content in a Twitter Field Experiment. In: Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. CHI ’21 (2021). https://doi.org/10.1145/3411764.3445642
 Osadchuk, R.: Undermining Ukraine: How the Kremlin Employs Informa-tion Operations to Erode Global Confidence in Ukraine (2023). Technical report at https://www.atlanticcouncil.org/wp-content/uploads/2023/02/Undermining-Ukraine-Final.pdf
 Saltz, E., Leibowicz, C.: Fact-Checks, Info Hubs, and Shadow-Bans: A Landscape Review of Misinformation Interventions. Partnership on AI. Technical report at https://partnershiponai.org/intervention-inventory/(2021)
 Browne, R.: Google Launches Knowledge Panels in Search Results to Tackle Misinformation about COVID Vaccines. CNBC (2020). https: //www.cnbc.com/2020/12/10/google-search-panels-tackle-misinformat ion-about-covid-vaccines.html
 Ingram, D.: Twitter Launches ’Pre-Bunks’ to Get Ahead of Voting Mis-information. NBC News (2020). https://www.nbcnews.com/tech/tech-n ews/twitter-launches-pre-bunks-get-ahead-voting-misinformation-n1244777
 Lewandowsky, S., Van Der Linden, S.: Countering Misinformation and Fake News through Inoculation and Prebunking. European Review of Social Psychology 32(2), 348–384 (2021). https://doi.org/10.1080/1046 3283.2021.1876983
 Guess, A.M., Lerner, M., Lyons, B., Montgomery, J.M., Nyhan, B., Reifler, J., Sircar, N.: A Digital Media Literacy Intervention Increases Discernment between Mainstream and False News in the United States and India. Proceedings of the National Academy of Sciences 117(27), 15536–15545 (2020). https://doi.org/10.1073/pnas.1920498117
 Vraga, E., Tully, M., Bode, L.: Assessing the Relative Merits of News Literacy and Corrections in Responding to Misinformation on Twitter. New Media & Society 24(10), 2354–2371 (2022). https://doi.org/10.117 7/1461444821998691
 Mesgari, M., Okoli, C., Mehdi, M., Nielsen, F.A., Lanam¨aki, A.: “The Sum of All Human Knowledge”: A Systematic Review of Scholarly Research on the Content of Wikipedia. Journal of the Association for Information Science and Technology 66(2), 219–245 (2015). https://doi.org/10.1002/asi.23172
 Hughes, T., Smith, J., Leavitt, A.: Helping People Better Assess the Stories They See in News Feed with the Context Button. Facebook Newsroom (2018). https://about.fb.com/news/2018/04/news-feed-fyi-more-context
 Harrison, S.: Twitter Wants to Use Wikipedia to Help Determine Who Gets a Blue Checkmark. Slate (2020). https://slate.com/technology/202 0/12/twitter-checkmark-verification-wikipedia-notability.html
 Solon, O.: YouTube Will Use Wikipedia to Help Solve Its Conspiracy Theory Problem. The Guardian (2018). https://www.theguardian.com/technology/2018/mar/13/youtube-wikipedia-flag-conspiracy-theory-videos
 Ankel, S.: Russia Repeats Claim It Didn’t Attack Ukraine, Saying It Was Forced to Defend Pro-Kremlin Regions. Insider (2022). https://www.businessinsider.com/russia-repeats-claim-did-not-attack-ukraine-2022-3
 Clayton, K., Blair, S., Busam, J.A., Forstner, S., Glance, J., Green, G., Kawata, A., Kovvuri, A., Martin, J., Morgan, E., et al.: Real Solutions for Fake News? Measuring the Effectiveness of General Warnings and Fact-Check Tags in Reducing Belief in False Stories on Social Media. Political behavior 42, 1073–1095 (2020). https://doi.org/10.1007/s11109-019-09533-0
 Capraro, V., Celadin, T.: “I Think This News Is Accurate”: Endorsing Accuracy Decreases the Sharing of Fake News and Increases the Sharing of Real News. Personality and Social Psychology Bulletin (2022). https://doi.org/10.1177/01461672221117691
 Ternovski, J., Kalla, J., Aronow, P.M.: Deepfake Warnings for Political Videos Increase Disbelief but Do Not Improve Discernment: Evidence From Two Experiments (2021). Preprint at https://osf.io/dta97/
 Kaiser, B., Wei, J., Lucherini, E., Lee, K., Matias, J.N., Mayer, J.R.: Adapting Security Warnings to Counter Online Disinformation. In: 30th USENIX Security Symposium (USENIX Security 21), pp. 1163–1180 (2021). https://www.usenix.org/conference/usenixsecurity21/presentation/kaiser
 Andı, S., Akesson, J.: Nudging Away False News: Evidence from a Social Norms Experiment. Digital Journalism 9(1), 106–125 (2020). https://do i.org/10.1080/21670811.2020.1847674
 Smith, J., Jackson, G., Raj, S.: Designing Against Misinformation. Face-book Design (2017). https://medium.com/facebook-design/designing-a gainst-misinformation-e5846b3aa1e2
 Snir, I., Hebbar, N.: Five New Ways to Verify Info with Google Search (2023). https://blog.google/products/search/google-search-new-fact-checking-misinformation/
 Bode, L., Vraga, E.K.: In Related News, That Was Wrong: The Correction of Misinformation Through Related Stories Functionality in Social Media. Journal of Communication 65(4), 619–638 (2015). https://doi.org/10.1 111/jcom.12166
 Zimmer, M., Kinder-Kurlanda, K.: Internet Research Ethics for the Social Age: New Challenges, Cases, and Contexts. Peter Lang International Academic Publishers, New York (2017). https://doi.org/10.3726/b11077
 Kramer, A.D., Guillory, J.E., Hancock, J.T.: Experimental Evidence of Massive-Scale Emotional Contagion through Social Networks. Proceed-ings of the National Academy of Sciences 111(24), 8788–8790 (2014). https://doi.org/10.1073/pnas.1320040111
 Rajkumar, K., Saint-Jacques, G., Bojinov, I., Brynjolfsson, E., Aral, S.: A Causal Test of the Strength of Weak Ties. Science 377(6612), 1304–1310 (2022). https://doi.org/10.1126/science.abl4476
 Selinger, E., Hartzog, W.: Facebook’s Emotional Contagion Study and the Ethical Problem of Co-opted Identity in Mediated Environments Where Users Lack Control. Research Ethics 12(1), 35–43 (2016). https://doi.org/10.1177/1747016115579531
 Meyer, M.N.: Everything You Need to Know About Facebook’s Controversial Emotion Experiment. Wired (2014). https://www.wired.com/2014/06/everything-you-need-to-know-about-facebooks-manipulative-experiment/
 Flick, C.: Informed Consent and the Facebook Emotional Manipulation Study. Research Ethics 12(1), 14–28 (2016). https://doi.org/10.1177/1747016115599568
 DuckDuckGo: News Rankings. DuckDuckGo Help Pages. https://help.duckduckgo.com/duckduckgo-help-pages/results/news-rankings/ Accessed 2023-04-12
 Olejnik, L., Castelluccia, C., Janc, A.: Why Johnny Can’t Browse in Peace: On the Uniqueness of Web Browsing History Patterns. In: 5th Workshop on Hot Topics in Privacy Enhancing Technologies (HotPETs 2012) (2012). https://hal.inria.fr/hal-00747841
 Broder, A.: A Taxonomy of Web Search. In: ACM SIGIR Forum, vol. 36 , pp. 3–10 (2002). https://doi.org/10.1145/792550.792552. ACM New York, NY, USA
 Chen, E., Ferrara, E.: Tweets in Time of Conflict: A Public Dataset Tracking the Twitter Discourse on the War Between Ukraine and Russia. Preprint at https://arxiv.org/abs/2203.0748 (2022)
 Saltz, E., Leibowicz, C.: Shadow Bans, Fact-Checks, Info Hubs: The Big Guide to How Platforms Are Handling Misinformation in 2021. Technical report at https://www.niemanlab.org/2021/06/shadow-bans-fact-s-info-hubs-the-big-guide-to-how-platforms-are-handling-misinformation-in-2021/ (2021)
 Harman, G., Tarrant, R., Tolbert, A., Ungerleider, N., Wolf, C.: Disinfodex. Available at https://web.archive.org/web/20230213222216mp_/ https://disinfodex.org/ (2018)
 Roth, Y., Pickles, N.: Updating Our Approach to Misleading Information (2020). https://blog.twitter.com/en_us/topics/product/2020/updating-our-approach-to-misleading-information
 Constine, J.: Instagram Hides False Content Behind Warnings, Except for Politicians. TechCrunch (2019). https://techcrunch.com/2019/12/16/instagram-fact-checking/
 Gibbs, S.: Google to Display Fact-checking Labels to Show If News Is True or False. The Guardian (2017). https://www.theguardian.com/techhttps://www.theguardian.com/technology/2017/apr/07/google-to-display-fact-checking-labels-to-show-if-news-is-true-or-false
 Nyhan, B., Reifler, J.: Displacing Misinformation about Events: An Ex-perimental Test of Causal Corrections. Journal of experimental political science 2(1), 81–93 (2015). https://doi.org/10.1017/XPS.2014.22
 Vraga, E.K., Bode, L.: I Do Not Believe You: How Providing a Source Corrects Health Misperceptions Across Social Media Platforms. Information, Communication & Society 21(10), 1337–1353 (2018). https://doi.org/10.1080/1369118X.2017.1313883
 Pennycook, G., Bear, A., Collins, E.T., Rand, D.G.: The Implied Truth Effect: Attaching Warnings to a Subset of Fake News Headlines Increases Perceived Accuracy of Headlines Without Warnings. Management science 66(11), 4944–4957 (2020). https://doi.org/10.1287/mnsc.2019.3478
 Gleicher, N.: Labeling State-Controlled Media On Facebook (2020). http s://about.fb.com/news/2020/06/labeling-state-controlled-media/
 Kofman, A.: YouTube Promised to Label State-Sponsored Videos But Doesn’t Always Do So. ProPublica (2019). https://www.propublica.org/article/youtube-promised-to-label-state-sponsored-videos-but-doesnt-always-do-so
 Anker, A., Su, S., Smith, J.: New Test to Provide Context About Articles. Facebook Newsroom (2017). https://about.fb.com/news/2017/10/news -feed-fyi-new-test-to-provide-context-about-articles/
 Hebbar, N.: Check the Facts with These Google Features (2022). https://blog.google/products/news/fact-checking-misinformation-google-features/
 Aslett, K., Guess, A.M., Bonneau, R., Nagler, J., Tucker, J.A.: News Credibility Labels Have Limited Average Effects on News Diet Quality and Fail to Reduce Misperceptions. Science Advances 8(18) (2022). https://doi.org/10.1126/sciadv.abl3844
 Gao, M., Xiao, Z., Karahalios, K., Fu, W.-T.: To Label or Not to Label: The Effect of Stance and Credibility Labels on Readers’ Selection and Perception of News Articles. Proc. ACM Hum.-Comput. Interact. 2(CSCW) (2018). https://doi.org/10.1145/3274324
 Kirchner, J., Reuter, C.: Countering Fake News: A Comparison of Possible Solutions Regarding User Acceptance and Effectiveness. Proceedings of the ACM on Human-computer Interaction 4(CSCW2 ), 1–27 (2020). https://doi.org/10.1145/3415211
 Hatmaker, T.: Facebook Will Put a New Coronavirus Info Center on Top of The News Feed. TechCrunch (2020). https://techcrunch.com/2020/03/18/facebook-coronavirus-information-center-zuckerberg/
 Singh, M.: Whatsapp Unveils $1m Grant, Info Hub to Fight Coronavirus Rumors. TechCrunch (2020). https://techcrunch.com/2020/03/18/what sapp-unveils-1m-grant-and-info-hub-to-fight-coronavirus-rumors/
 Li, Y., Guan, M., Hammond, P., Berrey, L.E.: Communicating COVID-19 Information on TikTok: A Content Analysis of TikTok Videos From Official Accounts Featured in the COVID-19 Information Hub. Health Education Research 36(3), 261–271 (2021). https://doi.org/10.1093/her/cyab010
 Ozoma, I.: Bringing Authoritative Vaccine Results to Pinterest Search. Pinterest Newsroom (2019). https://newsroom.pinterest.com/en/post/bringing-authoritative-vaccine-results-to-pinterest-search
 Shu, C.: Twitter Launches New Search Features to Stop the Spread of Misinformation about Vaccines. TechCrunch (2019). https://techcrunch.com/2019/05/14/twitter-launches-new-search-features-to-stop-the-spread-of-misinformation-about-vaccines/
 Kraus, R.: Facebook Labeled 180 Million Posts as ’False’ since March. Election Misinformation Spread Anyway. Mashable (2020). https://mashable.com/article/facebook-labels-180-million-posts-false
 Mosseri, A.: A New Educational Tool Against Misinformation. Meta Newsroom (2017). https://about.fb.com/news/2017/04/a-new-educational-tool-against-misinformation/
 Roozenbeek, J., van der Linden, S., Goldberg, B., Rathje, S., Lewandowsky, S.: Psychological Inoculation Improves Resilience Against Misinformation on Social Media. Science Advances 8(34), 6254 (2022). https://doi.org/10.1126/sciadv.abo6254
 Morrison, H.: Twitter Disables Ability To Retweet or “Like” ’ President Donald Trump’s Tweet “Due to the Risk of Violence”’ After US Capitol Stormed. MassLive (2021). https://www.masslive.com/politics/2021/01/twitter-disables-ability-to-retweet-or-like-president-donald-trumps-tweet-due-to-the-risk-of-violence-after-us-capitol-stormed.html
 Clegg, N.: You and the Algorithm: It Takes Two to Tango. Medium (2021). https://nickclegg.medium.com/you-and-the-algorithm-it-takes-two-to-t ango-7722b19aa1c2
 Cameron, D., Wodinsky, S., DeGeurin, M., Germain, T.: Facebook Papers Directory. Gizmodo. See, e.g., https://www.documentcloud.org/documents/21600352-tier0_rank_ir_0120 (2023). https://gizmodo.com/facebook-papers-how-to-read-1848702919
Our study focuses on a specific type of misinformation: Russian state propaganda. As such, we define misinformation websites as Russian state-affiliated media outlets with a well-documented record of publishing false information, censoring facts, and ceding editorial control to the Russian government. See SI1 for details.
A requery occurs when a user initiates a new search from SERP.
A bounceback occurs when a user navigates back to a SERP after visiting one of the search results.
Benjamin Kaiser is a Ph.D. candidate in computer science at Princeton University in the Center for Information Technology Policy.
Jonathan Mayer is an assistant professor of computer science and public affairs at Princeton University.