Deceptive Claims Using Fake News Advertising: The Impact on Consumers

Abstract

Fake news advertising—advertising that mimics legitimate news articles—can be harmful if it misleads consumers to take actions they otherwise would not have taken (e.g., purchase an inferior product). However, little is known about whether fake news ads bring in new customers or are merely viewed by people already in the market for the advertised products. The author exploits a Federal Trade Commission–enabled shutdown of fake news advertisements for various products such as acai cleanses and teeth whiteners (but where the product sites continued to remain operational) to identify the extent of consumer interest in the presence and absence of fake news advertising. The findings indicate that interest wanes after the shutdown of fake news advertising, with the probability of a product site not receiving any new visits increasing by 22%. The overall decline in visits caused by the absence of fake news ads occurs despite some substitution by consumers to regular advertisements.

Keywords

natural experiments public policy,legal advertising history,events,and cases

The use of fake news advertising has been prevalent for a long time. Advertising that mimics the format of its surrounding news content in both digital and print media, promotional segments aired on news programs without disclosure to the public,¹ and fake news sites (i.e., advertisement sites posing as legitimate news channels) are all examples of firms using fake news stories to deceive consumers into buying their products. Such deceptive practices are so prevalent that the Federal Trade Commission’s (FTC's) guidelines explicitly state that ads are deceptive if they convey that they are independent, impartial, or from a source other than the sponsoring advertiser.² However, little is known regarding whether fake news advertising affects consumers. Do such fake news stories change consumers' purchase propensity, or are they viewed by a select set of consumers already in the market for such products? The former, if true, is more harmful because it implies a direct treatment effect of fake news advertising on consumers.

In this article, using a set of fake news sites rendered inoperational by the FTC, I measure the impact of such campaigns on consumers' visits to product domains that used fake news sites to market their products. Moreover, because these product domains were using both regular advertising and fake news advertising, I can disentangle the impact of fake news advertising from that of regular advertising. In April 2011, the FTC brought to a halt the operations of ten fake news website operating companies. These ten companies operated over 150 fake news websites with names such as onlinenews6.com and consumerdigestweekly.com. The product domains (also referred to as merchants or advertising companies) to which these fake news sites referred consumers typically sold purported weight-loss products and colon-cleansing products. There are 174 such merchants in the data. To propagate their products, these companies used both direct advertising through sites such as Google and fake news “ads.” The fake news sites falsely reported, in a journalistic manner, the positive impact of using the products. The FTC's actions were directed toward the operators of the fake news sites, and not the advertising companies, which were under no restraint and could continue to operate. The difference in product-domain visits before and after the FTC action identifies the causal impact of the visits due to the fake news ads.

If fake news advertising were the only means of reaching these product sites, the answer to “What happens to product site visits after the fake news ads shutdown?” is trivial (i.e., products would receive no visits after the FTC shutdown). However, consumers can reach these sites through organic links (e.g., Google search) as well as regular, nonfraudulent ads (e.g., legitimate sponsored ads on Google). If, after the shutdown, consumers continue finding these sites at the same rate, then fake news advertising did not play an important role in product discovery. If this rate decreases substantially, it implies that the fake news ads were an important driver of product discovery, without which consumers are no longer able to find these products. Therefore, apart from quantifying the effect of the FTC shutdown, this research aims to identify the extent of substitution to other pathways (if any).

Using detailed browsing data from comScore, which are well-suited for this purpose because referral domains are also tracked, I can identify whether consumers visited the product's website directly, were referred from a fake news ad site, or were referred via regular ads (e.g., sponsored sites on Google). Such data help identify whether consumers, after the shutdown of the fake news sites, were still able to reach the product website—and, if so, how. If consumers no longer visit the product websites, the fake news sites played an important role, without which consumers are no longer able to find the sites. Note that apart from visits via fake news ad sites, which will mechanistically decline after the FTC shutdown, visits can also decline via other channels if there are positive spillovers between fake news ads and these channels. If consumers still visit the product sites, I can identify whether they reach them directly or are now being referred there by regular ads. To the extent that consumers reach these sites directly (e.g., through organic search results), advertisements are not influential in this setting; to the extent that consumers use regular ads to reach these sites, fake news ads and regular ads are likely to be substitutes. In either case, the implication is that the fake news sites are not crucial to product discovery and that consumers are still able to reach these product sites. Quantifying the degree to which fake news sites had an effect is important in understanding the extent of harm such campaigns cause and, therefore, the extent of regulatory oversight required.

The findings indicate that the probability of merchants not receiving any visit in a given month after the FTC shutdown increases by 22%–41%. To understand whether the intensity of the decline is different across various pathways, I break down the total visits into organic visits, fake news ad referrals, and regular ad referrals. I find that referrals via fake news ads drop (mechanistically), and direct visits drop. The drop in direct visits suggests positive spillovers between organic searches and fake news ads. However, referrals via regular ads exhibit heterogeneity across merchants, with some merchants seeing a reduction in visits and others seeing an increase in visits. For those merchants that see an increase in visits, the average increase is 25 visits per month and is statistically significant, suggesting that fake news ads and regular ads are substitutes, at least in these product domains.

Quantifying the effect of the drop in visits coming from each pathway, I find that the drop directly attributable to fake news ad referrals is three visits per domain per month. For reference, the average monthly visits per domain in the year before the shutdown (2010) was 34 visits. The drop, accounting for fake news ad referrals and organic demand, is 15 visits, suggesting positive spillovers between these two channels. Finally, adding in regular ad referrals, the drop is smaller, at 12.5 visits. Therefore, regular ads are able to bring in at least 2.5 new visitors, an increase of 17%. However, this increase is statistically insignificant and much smaller in magnitude than the decline in demand occurring through fake news ad referrals and organic demand. Taken together, these findings imply that the regulatory action had a significant effect on reducing the number of visitors who arrived at these merchant sites.

I supplement this analysis with data from consumer-complaint boards, which are indicative of actual purchases and amount spent on these product websites. I find that the probability of a merchant receiving a complaint declines by 8% following the FTC shutdown, further confirming the impact of the regulatory action. Using two other proxies for purchase from the browsing data—duration spent at the merchant site and visits to potential order management sites—I find a significant decline in such metrics. Taken together with visits and the complaints data, these provide evidence on the impact of the shutdown on product purchases.

I also analyze supply-side responses in terms of changes in ad spend and the introduction of (potentially) new fake news ad campaigns after the FTC-enabled shutdown. I do not find evidence supporting a change in advertising metrics such as spend, duration, and impressions, nor do I find evidence that merchants introduced new fake news ad campaigns. Moreover, the demand-side results are robust to controlling for advertising intensity.

As with any event study, the interpretation of the effect requires the assumption that the only change after the FTC shutdown is the absence of the fake news ads. Confounds such as publicity effects of the FTC press release could contribute to the decline in demand. I verify that negative publicity (arising from the FTC's press release or news public relations [PR] effects) alone does not explain the decline in visits. I do so by exploring heterogeneity among consumers as defined by their level of news consumption, finding that the largest impact is on those who consume the least news. If news effects were driving the decline, the group consuming more news would be expected to have the largest decline. This finding is consistent with the hypothesis that those with low news consumption are likely to be more susceptible to false claims. Finally, the effect could include changes that search engines such as Google might have made to their algorithms in response to the shutdown (e.g., Google no longer ranks these merchants high on its search results page because of the merchants' questionable practices). Therefore, the measured decline is likely to be an upper bound on the true impact of the removal of the fake news advertising campaigns.

The specific setting of the research involves products that were scamming consumers (charging their credit card for larger amounts than authorized) and that were using fake news advertising as one means of marketing. However, only later on did consumers and regulators³ discover that the products were scams. Therefore, these findings would likely apply to legitimate products using such a fake news advertising strategy. Future work in other settings would help establish the generalizability of these findings.

The rest of the article is organized as follows. The next section briefly describes the related literature in this domain. Subsequently, I present a framework illustrating the various pathways a consumer can use to reach the merchant. I then describe the data and the institutional setting and present empirical analysis on browsing data. Using data from complaint boards, I analyze purchase behavior after the shutdown. I then analyze possible firm-side responses. The final section concludes.

Literature Review

Advertisements designed to look like editorial content have been prevalent since newspapers began national circulation circa 1879 (Petty 2013). However, this form of deceptive advertising has received little empirical attention until recently. Chiou and Tucker (2018) study the impact of Facebook's ban on fake news ads on sharing of fake news articles. This article contributes to this limited literature by studying the effect of a regulatory ban on fake news ads on consumers' site visit behavior and proxies for purchase. Moreover, by observing the path a user takes to reach a product domain, I can analyze whether consumers nevertheless find these product domains without fake news ads either directly or through regular online ads.

This research is also closely related to the literature on deceptive practices that firms undertake to win consumers' wallet share. This literature has received attention from both the behavioral literature and, more recently, the empirical literature. Early work in marketing (Olson and Dover 1978; Shimp and Preston 1981) shows that evaluative (nonfactual) claims in advertising are more likely to deceive consumers by influencing their beliefs. Recent empirical work on deceptive practices includes review fraud (Anderson and Simester 2014; Luca and Zervas 2016; Mayzlin, Dover, and Chevalier 2014), false claims (Chiou and Tucker 2021; Rao and Wang 2017; Zinman and Zitzewitz 2016), and use of deceptive ad formats (Aribarg and Schwartz 2020; Edelman and Gilchrist 2012; Sahni and Nair 2020). The current research contributes to this growing stream of literature by studying ads designed to appear like news, which are deceptive in both format and content.

The impact of fake news has received a lot of attention in the political literature (e.g., Guess, Nagler, and Tucker 2019; Guess, Nyhan, and Reifler 2018). Fake news has been studied by researchers aiming to understand the factors that influence the perceived accuracy of such news (Pennycook, Cannon, and Rand 2018, 2020), whether such beliefs can be corrected (Porter, Wood, and Kirby 2018), and the reasons for the spread of such news (Vosoughi, Roy, and Aral 2018). Fake news sites have recently garnered a lot of attention because of their potential impact on the 2016 elections. Sites such as Facebook and Google are under fire for their role as distributors of fake news. However, the empirical impact of fake news on actual outcomes is unknown. In the context of the 2016 U.S. presidential election, Allcott and Gentzkow (2017) point out that fake news may have merely strengthened voters' predetermined beliefs and did not change their voting behavior. For example, people likely to vote for Donald Trump are the ones who believe pro-Trump fake news. Measuring the true impact of the fake news is empirically challenging because we do not observe voting behavior prior to the exposure to fake news. This research extends this literature by studying the impact of fake news ads in the domain of consumer goods, with the added benefit this particular setting provides, which allows me to observe outcomes before and after the fake news ad campaigns.

Work in the political setting has also shown that being persuaded in general is hard (Berelson, Lazarsfeld, and McPhee 1954; Kalla and Broockman 2018) and that being persuaded by misinformation might be harder (e.g., Little 2018) either because receivers discount such information or because such misinformation forms a small part of all other messages a receiver sees. This article aims to understand the extent of such persuasion by misinformation in the context of marketing. Specifically, do fake news ads influence visits to product sites, or are such fake ads viewed by consumers already in the market for such products?

Several settings have shown that consumers might choose to consume content geared toward their preferences: Gentzkow and Shapiro (2010) and Simonov and Rao (2021) in the context of news consumption, and Blake, Nosko, and Tadelis (2015) and Chiou and Tucker (2021) in the context of paid search ads. This article aims to understand such selection in the context of deceptive fake news ads in a market where consumers are likely to be more susceptible, where understanding whether such campaigns cause harm and thus need increased regulatory surveillance is important.

This research studies advertisements geared to appear like news, bringing together the literature studying news consumption and the literature studying firms' deceptive practices and their impact on consumers' purchase decisions. Because the timing of such fake news ad campaigns is unlikely to be exogenous to demand, and because one cannot create such exogenous variation using field experiments in this setting (federal law prohibits false or misleading advertising), this article uses an event-study approach using the timing of the FTC shutdown as an exogenous event.

Framework

Consumers can reach a product domain in three ways: (1) directly, (2) via referrals through the fake news sites, or (3) via referrals through other regular advertisements. Direct visits occur when the consumer either directly types the product domain's address into the search bar or reaches the product site through an organic search. Referrals through fake news sites occur when the consumer clicks on a domain such as Channel6HealthBeat.com and then clicks on a link to the product domain present on that fake news site. Referrals through regular advertisements occur when a consumer clicks on an ad such as LeanSpaAcai.com and then reaches the product domain. Examples of regular ads are provided in Web Appendix W1. This article asks what happens when the fake news ad path is shut down: Do consumers no longer find the product domains, or do they merely substitute to finding these sites through other channels such as regular advertisements and organic links? The former, if true, implies a treatment effect of the fake news sites; that is, consumers find the product domains because of the fake news sites. The latter, if true, implies a selection effect; that is, consumers interested in these weight-loss sites will reach them through other means, and the fake news sites were just one way of doing so.

To understand the two effects of interest (i.e., the treatment and selection effects), I present a framework in this section. Let $Q_{m}$ represent the demand (e.g., clicks) for merchant m's site, consisting of $Q_{org}$ , $Q_{fn}$ , $Q_{reg}$ —the demand through organic search/direct visits, fake news ad referrals, and regular ad referrals, respectively. That is,

Q_{m} = Q_{org} + Q_{fn} + Q_{reg} .

(1)

The firm can influence demand through fake news advertising,

A d_{fn}

, and regular advertising,

A d_{reg}

. Let the demand from each pathway be expressed as follows:

Q_{org} = α_{org} + γ_{1}^{fn} A d_{fn} + γ_{1}^{reg} A d_{reg},

(2)

Q_{fn} = α_{fn} A d_{fn} + γ_{2}^{reg} A d_{reg} I (A d_{fn} > 0),

(3)

Q_{reg} = α_{reg} A d_{reg} + γ_{3}^{fn} A d_{fn} I (A d_{reg} > 0) .

(4)

In this specification, the coefficients corresponding to

α

capture the direct effects:

α_{org}

in Equation 2 is the organic demand that would occur in the absence of any advertising,

α_{fn}

in Equation 3 is the direct effect of fake news ads on visits to the merchant coming from fake news ad referrals and

α_{reg}

in Equation 4 is the direct effect of regular ads on regular ad referrals. The coefficients corresponding to

γ

capture the spillover effects:

γ_{1}^{fn}

and

γ_{1}^{reg}

in Equation 2 capture spillovers from fake news ads to organic demand and from regular ads to organic demand, respectively;

γ_{2}^{reg}

in Equation 3 captures the spillovers from regular ads to fake news ad referrals; and

γ_{3}^{fn}

in Equation 4 captures the spillovers from fake news ads to regular ad referrals. These spillovers can be positive or negative. An example of a positive spillover from fake news ads to organic visits (i.e.,

γ_{1}^{fn} > 0

) is when exposure to fake news ads leads to increased word of mouth causing others to directly visit the merchants. The indicator function

I (A d_{fn} > 0)

in Equation 3 depicts that spillovers from regular ads to fake news ad referrals can occur only if fake news ads are present (i.e., if there are no fake news ads, consumers cannot click on them). Similarly,

I (A d_{reg} > 0)

in Equation 4 is an indicator function.

Treatment Effect

Using this illustrative framework, fake news advertising has a direct treatment effect if it generates demand via any channel. There are three ways this effect can manifest itself: a direct effect through $α_{fn}$ , a positive spillover to organic searches through $γ_{1}^{fn}$ , and a positive spillover to regular ad referrals $γ_{3}^{fn}$ . Therefore, an absence of fake news ads not only will cause fake news ad referrals to mechanistically drop to zero but also may result in a decline in demand at other channels if there are positive spillovers between fake news ads and these other channels. Such a spillover from ads across channels has been shown in Rutz and Bucklin (2011), where generic ads increase branded ad clicks and impressions, and Thomas (2020), where national advertising has spillovers to local markets.

Selection Effect

In this scenario, consumers visit the merchant sites in both the presence and absence of fake news ads. In the absence of fake news ads, users continue to reach the merchants because the regular ad/organic link is still persuasive (albeit not as persuasive as fake news ads). This effect would manifest as an increase in demand via other channels. Such a substitution effect has been documented in Goldfarb and Tucker (2011), who show that banning offline ads for alcohol caused consumers to substitute to the online ad channel. Both effects can exist in the market, and the current research aims to understand whether this specific setting has more consumers for whom regular ads/organic links are a reasonable substitute or whether regular ads/organic links do not suffice. Understanding this split has implications for how harmful fake news ads are in this domain: For example, if all consumers just substitute to regular ads and continue reaching the site, fake news ads might have been more persuasive but did not alter consumer behavior drastically. In contrast, if consumers completely stop visiting the merchant site, it implies that fake news ads were drastically altering consumer behavior. The exogenous variation (i.e., the shutdown of the fake news ads) creates shifts in $A d_{fn}$ , enabling identification of the direct effects $α_{fn}$ and the spillover effects $γ_{1}^{fn}$ and $γ_{3}^{fn}$ . Because variation in regular ads is not exogenous, the empirical part will not focus on $γ_{1}^{reg}$ , $γ_{2}^{reg}$ , and $α^{reg}$ . Because firms could potentially alter their regular advertising levels after the FTC shutdown, I investigate this possibility in the “Supply-Side Responses” section.

Data and Descriptives

In April 2011, the FTC identified ten companies as operating fake news websites (FTC 2011). Table 1 lists a subset of the fake news websites operated by the ten companies, as recovered from the FTC complaint files accessed using the Bloomberg Law database. Figure 1 shows an excerpt of such a fake news advertisement posing as a news article, extracted from one of the FTC court dockets (FTC v. Coulomb Media Inc, Declaration of Loretta Kraus). The figure highlights four features common to these campaigns: the presentation of facts by a journalist and the presence of legitimate news-channel names such as CNN and ABC. However, neither the channel name (“Health News,” in this example) nor the website (consumerhealthwarning.com) are legitimate news sites. In short, in these fake articles, the “journalists” voice their skepticism of the products, claim to try it for themselves, and report their “fake” findings of weight loss. Web Appendix W2 shows two other examples of such articles.

Figure 1.

Example of a fake news advertisement from consumerhealthwarning.com.

Table 1.

A Subset of Fake News Domains and Their Visit Count.

Company	Fake News Domain Names	2010	2011	2012
Ambervine	usahealthnews.org	8	0	0
Ambervine	usahealthnewsreports.com	8	0	0
CircaDirect	channel8health.com	313	0	1
	clkrd.com	428	0	0
	online6reports.com	961	1	0
	onlinenews6.com	826	4	0
	online6health.com	191	437	0
	online8report.com	252	148	0
Beony	channel6reports.com	423	15	0
	healthnews10.com	43	0	0
	consumertipsdaily6.com	90	37	0
	rachelrayblogs.com	50	0	0
	consumersdigestweekly.com	580	0	0
	consumerstipweekly.com	85	0	0
	internetjobnews.net	60	0	0
	kljsdf.com	47	11	0
Coulomb Media	news6reports.com	982	0	0
	fdx8news.com	90	0	0
	new6reports.com	86	127	0
DLXM	health8news.com	51	43	0
	dlxm.net	84	36	0
	channel9newstoday.com	41	1	0
	channel5healthnews.com	70	66	1
Charles Dunlevy	acai-berry-trial-offers.com	206	0	0
Charles Dunlevy	colon-cleanse-trial-offers.com	44	0	0
Garrett Vaughn	channel9newstoday.com	41	1	0
	health8news.com	51	43	0
	channel6reports.com	423	15	0
	news6report.com	259	1	0
IMM Interactive	channel9healthbeat.com	131	3	0
	nbsnewsat6.com	51	0	0
	consumerproductsdaily.com	137	0	0
	fdxnewsat8.com	217	9	0
	consumeracaicleanse.com	47	47	0
Ricardo Jose Labra	consumerstricksweekly.com	169	0	0
Ricardo Jose Labra	consumerworkjournal.com	49	1	0

I combine this information on the identity of the fake news site with consumer browsing data in 2010 and 2011 from comScore. comScore tracks detailed browsing and buying behavior for 50,000 internet users across the United States. The panel is based on a random sample from a cross-section of more than 2 million global internet users who have given comScore explicit permission to confidentially capture their web-wide activity.

Table 1 also lists the number of yearly site visits to these fake news domains, using the comScore data. Following the FTC order in April 2011, these sites were inoperational. To show that after the FTC order, the fake news websites indeed discontinued operation, Figure 2 records the site visits across all fake news websites at the monthly level. This figure provides evidence that the fake news sites were inoperational after April 2011.

Figure 2.

Fake news domains' visit counts.

Consumers can either reach the fake news site organically or be referred there by another website/advertisement. In the data, more than 50% of traffic came in organically, whereas the rest were referred by advertising platforms such as AdShuffle, Facebook, and Rubicon Project. Because some of the organic searches might have come from users who discovered the site previously, through a referral, I keep only the first visits per user domain and find the percentages to be fairly identical; that is, over 50% of visits are organic. Figure 3 shows examples of each of these paths. Users do not spend much time at these fake news websites: the median number of visits per user is one website, and the median duration spent is one minute. The 90th percentile fake news site received 256 visits in 2010 and served as a referral to a product merchant 64 times, while the median fake news site received 25 visits in 2010 and served as a referral 6 times. These numbers suggest that a few fake news sites appear to contribute the most.

Figure 3.

Organic links and sponsored ads linking to fake news sites.

Product Domains/Merchants/Advertising Companies

Of direct relevance are the domains to which these fake news sites referred users. The comScore data record not only the domain visited but also the referrer domain that led the user to a given domain. I can therefore observe the set of product domains that were using the fake news sites as referrers. Extracting all domains to which the fake news sites referred users results in a set of 191 product domains, including sites such as fibradetox.com, getslimpackage.com, and tryacaiberrypure.com. Some of these referred-to sites are normal domains (e.g., accuweather.com, live.com, newsvine.com) and a few legitimate ad networks (e.g., crwdcntrl.net). I conduct the analysis excluding these normal domains, resulting in 174 product domains that likely used affiliate marketers to advertise their products via fake news sites. I use the term “product domains,” “merchants,” or “advertising companies” to refer to the websites to which these fake news sites referred consumers.

I assume these product domains were operational after the FTC order in April 2011. This assumption is valid because the FTC order only required the affiliate marketers to shut down their news website operations. The merchants themselves were not ordered to shut down (two of the merchants, LeanSpa and NutriSlim, were subsequently investigated and were required to shut down in December 2011⁴). However, to be precise, I also restrict analysis to the set of merchants mentioned in the FTC documents, resulting in a smaller set of 48 product domains. To ensure that the product domains were operational after the FTC order, I further restrict attention to those that received at least one consumer visit in the data after April 2011, resulting in a much smaller subset of 17 merchants. Note that this measure is extremely conservative, because comScore keeps track of only a sample of consumers, and all 48 sites are extremely likely to have had visits that might not be visible in the sample.

The most frequently sold category relates to weight-loss products (including products related to acai berry and colon cleansing), followed by merchants that engage in earn-money schemes, skin care, and teeth-whitening products. Finally, there are sites that appear to be ad trackers, which could be either jump links to the main site or a tracking site, as well as what appear to be order-placing portals. The remaining are either categorized into “Other” or “Not classified.” “Other” consists of smaller categories such as quit smoking, cure cancer, and health sites. “Not classified” sites have no information about them on the web or on the Wayback Machine internet archive. Table 2 provides these statistics across the three samples studied in the article. The nature of the products/services was inferred from historical data available through the Wayback Machine and/or the name of the domain. Web Appendix W3 provides examples of sites in each category from screenshots taken from Wayback Machine. Web Appendix W4 provides the classification and number of visits for each domain in the sample “Cited by FTC.”⁵ Given the difficulty in finding parent-company-related information for the merchant domains, it appears that many of these are short-term businesses aiming to make profits in the short run by luring in susceptible consumers with their claims.

Table 2.

Classification of Merchants by Type of Service.

	(1) All	(2) Cited by FTC	(3) Active	Examples
Weight loss	31	17	7	getleanspa.com
Earning money online	35	7	4	workfromhomeheaven.com
Skin care	20	7	3	dermitage.com
Teeth whitening	13	3	0	mysuperstarsmile.com
Offers/savings/coupons	10	3	0	dropdowndeals.com
Ad malware/services/jump links	22	2	1	dmcpmtrack.com
Order management services	10	2	1	securesiteorders.net
Other	11	1	1	ecigsbrand.com
Not classified	22	6	0	akjhf.com
Total	174	48	17

Pathways to the Merchants

Using data at the individual user level, Table 3 shows the percentage of visits coming from each of the three different paths for all domains (All), as well as restricting attention to those sites cited by the FTC in the court documents (Cited by FTC) and those that are definitely active after the FTC shutdown order (Active). This table shows that fake news sites form a small portion, about .8%–4.2%, of the way consumers reach these product domains. Because attribution is difficult—subsequent visits could have been informed by previous visits—I keep only the first visit per individual per domain. The last three columns of Table 3 show that fake news sites still represent a small portion of how consumers reach the domains. Regular advertisements, followed by organic visits, seem to be the most common pathway by which consumers reach these domains. The magnitude of this statistic is in line with the extant literature; for example, Allcott and Gentzkow (2017) estimate that the average U.S. adult saw about one fake news story in the months before the election; Guess, Nyhan, and Reifler (2018) estimate that about one in four Americans visited a fake news website from October to November 2016 and that fake news consumption is concentrated in a small subset of people; Guess, Nagler, and Tucker (2019) show that 91.5% Facebook users do not share even one fake news link, contrasting this to regular article shares, where a large majority (61.3%) share 100 to 1,000 links.

Table 3.

How Do Consumers Reach Product Domains.

	All Visits			First Visits
	All	Cited by FTC	Active	All	Cited by FTC	Active
Number of unique websites	174	48	17	174	48	17
Total visits per individual-domain-year-month	2.39	1.41	1.47	1	1	1
Direct/organic	38.4%	34.2%	26.6%	44.9%	39.6%	31.9%
Referred by fake news	.8%	3.0%	1.9%	2.4%	4.2%	2.7%
Referred by other	60.8%	63.2%	71.6%	52.7%	56.2%	65.5%
N observations	149,103	29,333	23,228	110,134	26,243	20,471
N individuals	50,277	19,035	16,784	50,277	19,035	16,784

Notes: “All” refers to all domains to which the fake news sites referred people, “Cited by FTC” restricts attention to those sites cited in the FTC court documents, and “Active” refers to those that are definitely active after the FTC shutdown order in the comScore data.

Aggregating the individual-level data, using each individual's first visit per domain, to the domain-month level, Table 4 provides the statistics for the various pathways across the main data set (all 174 domains). Statistics for the other two subsets are similar and not reported for brevity.

Table 4.

Summary Statistics for the Dependent Variables Used in Analyses.

Dependent Variable	Mean	SD	Min	Max	N Obs.	N Domains
Browsing Behavior ^a
Number of visits, total	30.8	170.4	0	5,843	3,306	174
Number of visits, direct	14.2	68.3	0	1,748	3,306	174
Number of visits, fake news ad referral	.8	5.8	0	175	3,306	174
Number of visits, regular ad referral	15.9	129.8	0	5,746	3,306	174
Purchase Data ^b
Complaint (yes/no)	.1	.3	0	1	966	46
Number of complaints	.1	.9	0	17	966	46
Total spent ($)	16.1	135.4	0	2,976	966	46
Proxies for Purchase ^c
Duration spent, total	62.0	392.2	0	16,667	3,306	174
Number of visits, tracking	87.6	267.3	0	2,505	380	20
Number of visits, tracking/ordering	67.9	220.2	0	2,505	608	32

Browsing behavior presents, for all 174 domains, the average monthly visits via each pathway per domain. Only individuals' first visit per domain used.

Purchase data obtained from complaint boards presents for all domains where there was a complaint issued, whether there was a complaint in a given month, the total number of complaints and the dollar amount cited in that complaint.

Proxies for purchase include duration spent and number of visits at sites with names containing “trk” and “trac,” and names containing “secure,” “clk,” “click,” “order,” “trk,” and “trac.”

Notes: The table presents summary statistics for all dependent variables used in the article. Each observation is at the domain-year-month level.

Purchase Data

The comScore data, which also track transactions, show that no one in the sample actually purchased the purported weight-loss products. However, per the FTC documents, consumers were subject to considerable harm and injury, likely in the order of $4.68 million, which was the judgment issued across all ten defendants investigated. This judgment implies that consumers were deceived into purchasing the products. To quantify the effect of the fake news advertisements on actual purchases, a source of purchase data is needed.

Because the comScore data cannot speak to purchases, I turn to a secondary source of data: consumer-complaint boards. Using two main sources, Ripoff Report and Complaints Board, I collect, for each of the 174 product websites, the date and content of each complaint. The following is an excerpt of a complaint for theadvancedcleanse.com, written in April 2010:

I ordered the Acai Berry Pure and Advanced Cleanse trial products based off the internet report from Newschannel 6 that states the offer is legitimate. The site states you will receive the product in a few days and if not interested you must cancel within 14 days…. Needless to say I never received the Acai Berry but I did receive a deduction of $149.95 from my bank account.

Of the 174 websites, 46 had at least one complaint that was filed against them in these two complaint boards between the years 2010 and 2012; the others had no complaints filed against them in this period.

Figure 4 plots the total number of complaints on these sites from 2010–2012, as well as the amount consumers reported having been charged wrongfully. The dark solid line indicates the date of the FTC order. A decline in purchases appears to exist, but this occurs nearly two months after the FTC order. Such a lag in decline is due to consumers often not immediately reporting a scam; consumers usually take a few weeks to realize they have been scammed and subsequently post a complaint on a board (e.g., the credit card is charged one month after a trial subscription). To empirically verify this, I extract the product purchase date from the text of the complaints data and compare it with the date of the complaint. Web Appendix W5 illustrates this process using two examples. Across all instances where such data are available, the average difference between complaint and purchase date is 2.18 months, and the median is 1.50 months, suggesting that such a lag is indeed a feature of the data. The lighter solid line in Figure 4 is therefore representative of the effective FTC shutdown date for these data.

Figure 4.

Number of complaints and total stated amount spent on product websites.

I also use two additional proxies for purchases. First, I use duration spent at each site as a proxy for purchases. Second, I use visits to what could be order management sites indicating that the user moved to an ordering stage: (1) visits to websites that have “trac” or “trk” in their name and (2) visits to sites that have “click” or “clk,” “trac” or “trk,” “order” or “secure” in their name. These sites include those classified as “order management services” (e.g., securesslcenter.com, secure81.com, securesiteorders.net, securesiteoffers.com, orderwave.com, cpatrac.com, trkcpa.com, xyztrk.com, securedorderweb.com, imatrack.com).

Table 4 provides the statistics for the purchase data from complaint boards as well as the statistics for duration spent and visits to order management sites.

Empirical Analysis

I first present results using a simple before–after analysis, relative to a placebo (previous) year. I then investigate, in the next subsection, the pathways through which site visits change.

Dependent Variable

In all regressions, I use an individual's first visit to a domain to construct the dependent variable. I do so for two reasons. First, the number of first visits to a domain is a direct measure of the number of new users a site is able to attract. Knowing whether new users are able to reach these domains even in the absence of the fake news ads is a measure of the effectiveness of fake news marketing. If users are unable to reach the merchant sites, it implies that the fake news ads had a large treatment effect, without which users are unable to find the merchants. However, if users are still able to reach the merchants, it implies a lesser degree of harm. Therefore, knowing whether the first visits drop after the shutdown of fake news ads is crucial from a regulator's perspective. Second, the first visit is a measure of a user's exposure to the product domain, after which she might have stored it in her browsing history, bookmarked it, or reached it from recall. For example, consider an individual who was referred to a merchant by a fake news website and subsequently visited the site organically. The second visit should not be counted if the fake news site was the reason she discovered this site in the first place. Using only the first visit helps correct for any wrongful attribution.

Difference-in-Differences Analysis

Using the FTC order date as an event, I run a before–after regression on all product domains to which the fake news sites referred people. In this regression, I also compare this estimate with a placebo date.⁶ To understand the impact at the domain level, I conduct the analysis on data aggregated to the domain-year-month level. To account for the excess zeros in the data, I conduct the analysis using a zero-inflated Poisson model (Lambert 1992).

I first illustrate the basic differences-in-differences approach using a linear specification and then use this setup to specify the zero-inflated Poisson model. The basic differences-in-differences setup is specified by Equation 5 as

y_{jt} = β \times {Post}_{t} + β^{FTC} \times {Post}_{t} \times {FTC}_{Y} + α_{m} + α_{m 1} \times {FTC}_{Y} + α_{m 2} \times t + ε_{jt} .

(5)

Here,

y_{jt}

represents the number of visits received by domain j in month t.

{Post}_{t} = 1

if t is between May and July (the three months after the FTC shutdown month in either the treated or the untreated years) and

FT C_{Y}

is an indicator for the actual year when the shutdown occurred (2011).

β

captures the average visits attributable to the months May–July relative to the rest of the year, in any year.

β^{FTC}

is the treatment effect and captures the additional change attributable to the FTC shutdown. Domain j is classified into merchant type m as specified in the “Data and Descriptives” section.

α_{m}

is the merchant type fixed effect that allows some merchant types to have more visits than others,

α_{m 1}

is the merchant type year fixed effect specific to the year 2011, and

α_{m 2}

is the merchant-type-specific monthly time trend.

Figure 5 illustrates the idea behind this specification. The three (red) filled dots after the FTC shutdown are compared with not only the (black) filled dots before the shutdown (with appropriate year and month controls) but also the three (red) filled dots after the placebo shutdown. Although this figure shows a decline in new visits to merchant sites, the estimation equation helps control for merchant type fixed effects and merchant-type-specific time trends. Because the object of interest is the short-term treatment effect, the unfilled dots after the treatment window are not included in the analysis.⁷

Figure 5.

First visits: measuring the change after the FTC fake news ads shutdown.

Because the data consist of a large number of zero-visits (nearly 60% of observations at the domain-month level), I estimate a zero-inflated Poisson model specified by Equation 6.

y_{jt} \sim {\begin{matrix} 0 & with probability p_{jt} \\ Poisson (λ_{jt}) & with probability 1 - p_{jt} \end{matrix} .

(6)

Here,

y_{jt}

represents the number of visits received by domain j in month t, and

p_{jt}

is the probability that domain j receives a zero-visit in month t. Conditional on a domain getting a visit, the distribution is assumed to follow a Poisson process with mean

λ_{jt}

The probability of a zero visit, $p_{jt}$ , is modeled using a logit model, such that $p_{jt} = \frac{\exp (γ X)}{1 + \exp (γ X)}$ where $γ X$ follows the same specification in the right-hand side (RHS) of Equation 5. The mean, $λ_{jt}$ , of the Poisson distribution (for nonzero visits) is specified such that $\ln (λ_{jt}) = β X$ also follows the same specification as the RHS of Equation 5. Standard errors are clustered at the domain level.

I run this regression separately for the three subsets of the data described previously: (1) “all,” excluding the normal domains, comprising 174 unique product domains; (2) “cited by the FTC,” those websites that are mentioned in the FTC court documents, comprising 48 unique domains; and (3) “active,” those websites that are definitely operational post-April 2011, comprising 17 domains.

An observation in this regression is at the domain-year-month level, created by aggregating data at the individual browsing level. The comScore data set records only sessions of active browsing. To allow for no visits, I expand the data set so that every domain has an observation for every month of the data, accounting for zero visit counts when no visits to domains occurred. This expansion allows search behavior to change; that is, domains might experience a decline in visits after the shutdown but they might also experience a complete stop to their visits, resulting in zero visits. Not accounting for zero visits would result in an incorrect estimate.

Table 5 presents the results of these regressions. Because of the nonlinear nature of the specification, I also present the marginal effects of each component (zero visits and number of events) along with the total marginal effect. All marginal effects are computed at $t = 5$ (i.e., one month after the shutdown [May]). First, consider the probability of a zero visit. Across all specifications, the interaction term Post × FTC is positive and significant, indicating that the probability of a zero visit increases after the FTC shutdown. The marginal effect for the increase in probability of a zero visit occurring after the FTC shutdown ranges from 22%–41% across the three subsets of the data. Next, conditional on a visit, domains do not experience a change in the total number of visits. The marginal effect is not statistically different from zero. The total marginal effect, which combines both the zero-visit probability and the number of events conditional on a visit, exhibits a decline (albeit a statistically insignificant one). The biggest effect of the shutdown therefore is on the probability that a domain receives a visit (and not on the number of visits). I next examine the visits via the various pathways, as outlined in the “Framework” section.

Table 5.

Difference in Domains' Visits After Fake News Sites Shutdown.

	(1)		(2)		(3)
	All		Cited by FTC		Active
N Unique Treated Websites	174		48		17
	Coef.	t-Stat.	Coef.	t-Stat.	Coef.	t-Stat.
Pr(zero visits): Logit
Post	.151	1.56	.266	1.16	−.015	−.06
Post × FTC	1.000	5.87	1.722	3.95	1.934	2.60
Marginal effect	22%	6.04	33%	4.28	41%	2.7
Number of events: Poisson
Post	−.191	−1.44	−.153	−.43	−.250	−.57
Post × FTC	.136	.54	.253	.41	.360	.37
Marginal effect	11.19	.54	17.99	.39	37.61	.35

Total marginal effect	−12.46	−1.47	−17.78	−1.16	−31.37	−.51
N observations	3,306		912		323
N	174		48		17
Controls	Merchant type			Year
	Merchant type × Year			Month
	Merchant type × Month
Cluster	Domain

Notes: The table presents, using first visits to each domain across individuals, results of a before–after zero-inflated Poisson regression, relative to a placebo year. The dependent variable is the count of visits per month. The three-month window after the FTC shutdown is the treatment period. The coefficients corresponding to Post × FTC are the relevant treatment effects (in boldface). “All” refers to all domains to which the fake news sites referred people, excluding normal domains, “Cited by FTC” restricts attention to those sites cited in the FTC court documents, and “Active” refers to those that are definitely active after the FTC shutdown order in the comScore data. Due to insufficient observations, merchant-type-specific effects cannot be estimated in the subset “Active.” Data are aggregated to the domain-year-month level. Standard errors are clustered at the domain level.

Pathways to the Merchants

The results show that new visits to product-domain sites dropped after the FTC's efforts to halt these fake news operations. Although visits to the domains from the fake news sites have to go to zero (by definition, because the fake news sites do not exist anymore), whether direct visits to these sites increase or decrease after the shutdown is not clear. Similarly, it is unknown whether referrals coming in through other ads/sources such as Facebook increase or decrease. As illustrated in the “Framework” section, these alternative pathways could have positive spillovers from fake news ads or function as substitutes. Therefore, understanding the pathways by which these site visits change is useful. Knowing the total impact is crucial, because if consumers find their way to these product domains through other means, the impact of the policy (i.e., the shutdown) is unclear.

To this end, I classify visits as “Referred via Fake News Ads,” “Referred via Regular Ads,” and “Direct.” The total visits to a product domain consist of one of these three forms of visits. Table 6 presents the result of the zero-inflated Poisson regression with this classification for the main data set⁸ (all 174 domains).

Table 6.

Difference in Domains' Visits Across Various Pathways.

All	Total		Direct		Referred by Fake News Ad		Referred by Regular Ad
All	Coef.	t-Stat.	Coef.	t-Stat.	Coef.	t-Stat.	Coef.	t-Stat.
Pr(zero visits): Logit
Post	.151	1.56	.156	1.47	.096	.41	.100	.98
Post × FTC	1.000	5.87	.837	4.68	3.036	4.75	.734	4.23
Marginal effect	22%	6.04	17%	4.77	32%	5.14	15%	4.26
Number of events: Poisson
Post	−.191	−1.44	.004	.03	−.434	−1.16	−.232	−1.63
Post × FTC	.136	.54	−.339	−1.11	−1.240	−1.71	.501	2.38
Marginal effect	11.19	.54	−17.24	−1.05	−8.10	−1.81	25.07	1.98

Total marginal effect	−12.46	−1.47	−12.66	−2.17	−3.03	−4.35	1.05	.31
N observations	3,306
N	174
Controls	Merchant type
	Merchant type × Year
	Merchant type × Month
Cluster	Domain

Notes: The table presents, for only the first visit to each domain across individuals, for each possible path to a domain, results of a before–after zero-inflated Poisson regression, relative to a placebo year. Results under the “Total Visits” column reproduce results from Table 5. Subsequent columns break these total visits into direct, referred by fake news ads, and referred by regular ads. The dependent variable is the number of site visits per month. The coefficients corresponding to Post × FTC are the relevant treatment effects (in boldface). “All” refers to all domains to which the fake news sites referred people, excluding normal domains. Data are aggregated to the domain-year-month level. Standard errors are clustered at the domain level.

First, consider the probability of a zero visit. Across all pathways, the coefficient for the term Post × FTC is positive and significant, indicating that the probability of a zero visit increases after the FTC shutdown. This increase in zero visits suggests positive spillovers between fake news ads and other channels. Next, conditional on a visit, the number of visits increases only for “referred by regular ads” after the FTC shutdown: a statistically significant increase of 25.07 visits. This pattern for regular ad referrals suggests an underlying heterogeneity across merchants: some merchants see an increase in zero visits after the shutdown and other merchants experience an increase in visits. This increase in visits via regular ad referrals for some merchants implies that consumers are reaching these merchants through this new path now that the fake news sites are shut down. In other words, the absence of the fake news advertisements does not entirely prevent consumers from finding these merchants; regular ads serve as a good substitute.

I next quantify the impact attributable to each pathway to understand whether consumers substitute to any particular pathway after the shutdown. Note that because of the nonlinearity of the zero-inflated Poisson model, one cannot simply add up the total marginal effect from each additional pathway.

Quantifying the effect

To quantify the effect from each pathway, I evaluate the cumulative effect of each additional individual pathway. Starting with visits coming from fake news ad referrals, which mechanistically have to drop, I evaluate the drop in direct visits and fake news ad referrals, and finally the drop in direct visits as well as fake news ad and regular ad referrals, taken together. To do so, I construct cumulative visits to the merchant starting with fake news ad referrals and subsequently adding direct visits and, finally, regular ad referrals. In other words, I run the same zero-inflated Poisson regression on three dependent variables constructed as

\sum_{p} Visit s_{p} where p \in {\begin{matrix} Fake News \\ Fake News, Direct \\ Fake News, Direct, Regular \end{matrix} .

Table 7 presents the results of this analysis. The first column, “Fake News Ads Only,” presents the change in visits occurring via only fake news ad referrals and is identical to the column “Referred by Fake News Ad” in Table 6. The drop in fake news ad referrals is 3 visits. The second column, “Fake News Ad + Direct Visits,” presents the change in visits occurring via both these paths. The total marginal effect of the FTC shutdown shows that total visits drop by 15.02 visits when considering fake news referrals and direct visits, suggesting positive spillovers between fake news advertising and organic demand. However, when regular ad referrals are included (column 3 of Table 7), this drop is smaller, at 12.46 visits, implying that the regular ad referrals are able to bring in 2.55 new visitors per month. Note that the drop in direct visits could be cushioned by regular ads as well. For example, if the mere exposure to regular ads leads users to visit the merchant directly (positive spillovers between regular ads and organic visits), then the absence of regular ads would imply a drop larger than 15 visits. Therefore, regular ads are able to bring in at least 2.55 new visitors, an increase of 17%. This increase is statistically insignificant and much smaller than the decline occurring via direct visits, suggesting that the selection effect is small (i.e., substitution to the regular ad channel is outweighed by the drop in organic demand and fake news ad referrals).

Table 7.

Cumulative Effect of Each Additional Pathway.

All	(1)		(2)		(3)
	Fake News Ad Only		Fake News Ad + Direct Visits		Fake News Ad + Direct Visits + Regular Ad
	Coef.	t-Stat.	Coef.	t-Stat.	Coef.	t-Stat.
Pr(zero visits): Logit
Post	.096	.41	.170	1.5	.151	1.56
Post × FTC	3.036	4.75	1.051	5.6	1.000	5.87
Marginal effect	32%	5.14	22%	5.83	22%	6.04
Number of events: Poisson
Post	−.434	−1.16	−.017	−.12	−.191	−1.44
Post × FTC	−1.240	−1.71	−.348	−1.08	.136	.54
Marginal effect	−8.10	−1.81	−16.54	−1.03	11.19	.54

Total marginal effect	−3.03	−4.35	−15.02	−2.36	−12.46	−1.47
N observations	3,306
N	174
Controls	Merchant type
	Merchant type × Year
	Merchant type × Month
Cluster	Domain

Taken together, these findings suggest that fake news ads had a large impact on consumers' site visitation behavior, both directly and through spillovers to organic searches. Although regular ads for some merchants see an increase, this increase is relatively small, suggesting that the extent of substitution between fake news ads and regular ads is low. From a policy perspective, this finding indicates that fake news ads were crucial in product discovery and the FTC shutdown was impactful. Note that these interpretations require that the only change at the time of the FTC shutdown was the absence of the fake news ads. To the extent that other factors cause a decline in demand, the observed decline in organic demand is an upper bound of the treatment effect and the observed increase in regular ad referrals at some merchants is a lower bound (i.e., factors causing a decline should have further reduced demand via regular ads). I discuss such possible factors next.

Discussion

Factors other than the shutdown could have contributed to the estimated decline. In Appendix B, I provide evidence that news-driven PR effects are unlikely to contribute to the decline. Specifically, I consider consumers with high versus low levels of news consumption and find the effect to be stronger for those with low news consumption. If PR effects were driving the decline, one would expect those with higher news consumption to see the largest decline. I also show that the declining popularity of acai during this time period does not contribute to the decline following the shutdown. Moreover, any other macro trends facing the industry are controlled for using the monthly time trend. I also point out that merchants were not likely to have increased fake news advertising but could have altered their levels of regular ad spending. I control for such changes to advertising spend and find that the effect continues to hold. I show that an information effect, where the mere presence of the fake news ad might be sufficient for consumers to visit the merchant site (clicking on the ad but not using it as a referral, not clicking on the ad but absorbing the information in the ad) might contribute to the decline. I also show that the shutdown of the fake news advertising can have spillovers to other merchants that an individual would have otherwise visited. These two effects are implicitly effects of the shutdown. Finally, changes made by search engines such as Google could have had an impact on the ranking of these merchants and, thus, their visits.

Click-throughs are a relevant metric used by the industry as well as in various academic studies (e.g., Aribarg and Schwartz 2020; Blake, Nosko, and Tadelis 2015; Chatterjee, Hoffman, and Novak 2003). Visits to the site have also been used as proxies (e.g., Ilfeld and Winer 2002; Sherman and Deighton 2001). However, because actual conversion to purchases might be a small fraction of click-throughs (see, e.g., Manchanda et al. 2006), I next examine the impact of the shutdown on stricter proxies for purchase, such as complaints arising from those who purchased, duration spent at the merchant sites, and visits to potential ordering sites.

Purchases

Using data collected from complaint boards, I run a logit⁹ regression to examine the change in propensity of receiving a complaint after the FTC shutdown. The probability of merchant j receiving a complaint in month t is specified by Equation 7 as

\Pr ({complaint}_{jt}) = \frac{\exp (u_{jt})}{1 + \exp (u_{jt})},

(7)

where

u_{jt}

follows the same specification in the RHS of Equation 5. The only difference here is that I take into account the two-month lag in complaints (because complaints occur two months after the actual purchase) described in the “Data and Descriptives” section (i.e., the treatment period under consideration is two months after the FTC shutdown; July–September). Table 8 presents the results of the analysis, showing that the probability of receiving a complaint dropped significantly after the FTC shutdown. The coefficient for the interaction term Post × FTC is negative and statistically significant, and the marginal effect indicates a significant decline of 8%. In Appendix C, I present results from two additional proxies for purchases, duration spent and visits to possible order management sites, which further confirm the decline in purchases.

Table 8.

Change in Probability of a Complaint After Fake News Sites Shut Down.

Complaint (Yes/No)	Coefficient	t-Statistic
Complaint (Yes)
Post	−.656	−1.32
Post × FTC	−1.605	−2.04
Marginal effect	−8%	−1.97
N observations	957
N	46
Controls	Merchant type
	Merchant type × Year
	Merchant type × Month
Cluster	Domain

Notes: The table presents results of a logit (0: No complaint, 1: Complaint) regression on the period after the FTC shutdown relative to a placebo year (2010). No complaint is treated as the reference. Data aggregated to the domain-year-month level and is expanded to include months of no complaints. The coefficient corresponding to Post × FTC is the treatment effect (in boldface).

Taken together with visits (which have been used as a proxy for purchases in the extant literature), these provide further evidence on the impact of the shutdown on product purchases.

Supply-Side Responses

The impacted merchants might have increased or decreased their ad expenditure on regular advertisements after the FTC shutdown. To measure a change in ad spend, I use the Nielsen Ad Spend data, and extract any ads relevant to the merchants. To do so, I search for various combinations of keywords pertaining to the merchant's website name in the Nielsen Brands database. For example, the merchant site fibradetox.com corresponds to the brand Fibra Detox Nutritional Supplement in the media data. I run a difference-in-differences analysis, specified in Equation 8, on total ad expenditure, frequency, and impression at the brand-month level.

{Ad}_{jt} = γ \times {Post}_{t} + γ^{FTC} \times {Post}_{t} \times {FTC}_{Y} + δ_{m} + δ_{m 1} \times {FTC}_{Y} + δ_{m 2} \times t + ε_{jt},

(8)

where

A d_{jt}

is the corresponding measure of advertising by merchant j in month t.

Pos t_{t} = 1

if t is between May and July—the three months after the FTC shutdown month in either the treated or the untreated years.

FT C_{Y}

is an indicator for the actual year when the shutdown occurred, 2011.

δ_{m}

is the merchant-type-specific fixed effect that allows some domains to have more ads than others.

δ_{m 1}

is the merchant type year fixed effect specific to 2011 and

δ_{m 2}

is the merchant-type-specific monthly time trend.

Table 9 presents the results of this analysis showing no significant change in advertising across all measures used. In a robustness check presented in Web Appendix W6, I control for advertising intensity and find the results are robust.

Table 9.

Changes in Ad Measures for the Treated Brands.

	Frequency		Spend ($)		Impressions
	Coef.	t-Stat.	Coef.	t-Stat.	Coef.	t-Stat.
Post	−5.63	−2.26	−1,803.18	−.4	−280,425	−.33
Post × FTC	4.49	1.04	1,614.19	.2	320,188.4	.22
N observations	589
N brands	31
Controls	Merchant type
	Merchant type × Year
	Merchant type × Month
Cluster	Domain

Launch of Other Fake News Sites

Firms may open and operate new fake news websites with different names after the FTC shutdown of the 150 websites. To identify if firms engage in such conduct, I use two means of identifying potential fake news websites. First, I extract all websites that contain the terms health, news, report, today, review, channel, daily, weekly, evening, or journal, resulting in over 163,000 sites. Most of the original fake news websites contained at least one of these terms.

In the first approach, I use the Wayback Machine to capture the start date of the websites.¹⁰ Figure 6 provides an example of how such start dates are captured: for each domain, Wayback Machine tracks not only screenshots but also the start date of the site. In this example, below “27 captures” is the start date (November 1, 2009) of the site weekly-bulletin.com. However, scraping the large number of sites is infeasible both because the Wayback Machine does not allow such large-scale scraping and because it would take a long time to collect all this information. A middle ground was to scrape a smaller sample of sites from the Wayback Machine: I chose a random 10% sample and, for this sample, obtained the start date of each domain.

Figure 6.

The Wayback Machine provides the start date of each website.

In the second approach, I consider all domains that have a number in their domain name (e.g., news6reporter.com). This approach relies on the fact that many FTC shutdown sites had such names (e.g., online8report.com, news6report.com).

I then run a difference-in-differences analysis on the resulting count of unique domain names at the monthly level. If firms indeed started new alternative fake news sites, then we should see an increase in sites classified as potential fake news sites. Table 10 shows no such increase under both approaches: (1) start dates and (2) contains number. As a reference, I present the regression for the fake news sites that were actually shut down, which shows a significant decline. However, if merchants come up with domain names very different from those used preshutdown, then these two approaches do not capture the launch of new fake news domains. To the extent such fake news domains exist, they are currently captured under regular ad referrals.

Table 10.

No Evidence for Increase in Number of New Potential Fake News Sites.

	Coefficient	t-Statistic
Start Dates ^a
Post	14.67	.15
Post × FTC	−197.00	−1.39
Contains Number ^b
Post	−46.67	−1.14
Post × FTC	−155.00	−2.68
Fake News Shut Down by FTC ^c
Post	7.33	5.11
Post × FTC	−27.33	−13.48
N months	12

Counts the number of domains that started operation in a given month and contain terms such as “health.”

Counts those domains that contain terms such as “health” and also a numeral in their name.

Sites the FTC actually shut down.

Notes: Table presents a difference-in-differences of the change in the total number of unique domains per month after the FTC shutdown, using the previous year as a placebo.

Conclusion

This article examines the role of fake news as advertisements and finds that fake news ads can cause increased interest in product domains. The FTC shutdown of the fake news ad campaigns appears to have a treatment effect, emphasizing the role of a regulator in a context where users are likely to be more susceptible. This research studies the path through which users arrive at the merchants that employ such fake news advertisements. The absence of fake news ads results in a drop in users arriving not only through fake news ad referrals but also through direct visits, suggesting that the two paths have positive spillovers. However, the shutdown of fake news ads seems to have diverted some traffic to regular advertisements, indicating that the two are substitutes to some extent.

Some of the decline in direct visits could be due to consumers' prior exposure to fake news advertisements and spillovers to other merchants they would have otherwise visited in the presence of fake news ad referrals. These two effects are implicitly effects of the shutdown. Other possible factors (which I rule out) are the declining popularity of acai and the effects of negative publicity. Finally, actions taken by Google and other search engines might have impacted organic rankings. I also show that merchants did not increase fake news advertising after the shutdown and were unlikely to do so in light of the FTC investigation. However, these sites might also have used fewer/more regular ads following the FTC shutdown. Although the ad-spend data suggest otherwise, and the results are robust controlling for such advertising changes, to the extent that there exist other potential supply-side responses (that are unobservable to the researcher), the effect measured includes such supply-side actions.

To understand the impact of the absence of the fake news ads on consumer interest, this research uses consumer visits to product domains as well as proxies for purchase such as consumer complaints. These outcome measures come from different data sources (visits from comScore and proxies for purchase from consumer complaint boards). Ideally, both outcome measures, visits and purchases, would have come from the same data source. However, this research and setting represent a step toward understanding the impact of fake news ads on consumers' purchase decisions.

Appendix A: Difference-in-Difference-in-Differences Analysis

In this section, I include control sites to control for any industry trends beyond seasonality that might influence the results. The specification follows the same zero-inflated Poisson distribution as specified in the main analysis, but the mean of the Poisson process, $λ_{jt}$ , is now specified such that

\begin{aligned} \ln (λ_{jt}) = & β \times {Post}_{t} + β^{FTC} \times {Post}_{t} \times {FTC}_{Y} \\ + β^{Fake} \times {Post}_{t} \times {Fake}_{j} + β^{FTC, Fake} \times {Post}_{t} \\ \times {FTC}_{Y} \times {Fake}_{j} + α_{m} + α_{m 1} \times {FTC}_{Y} + α_{m 2} \\ (A 1) & \times t + ε_{jt} . \end{aligned}

(A1)

Here,

Pos t_{t} = 1

if t is between May and July—the three months after the FTC shutdown month in either the treated or the untreated years;

FT C_{Y}

is an indicator for the actual year when the shutdown occurred (2011).

Fak e_{j}

is an indicator for those domains impacted by the FTC shutdown.

β

captures the average additional visits in the months May–July relative to the rest of the year, in any year.

β^{FTC}

is the treatment effect and captures the additional change across all domains during the year of the FTC shutdown (2011).

β^{Fake}

and

β^{FTC, Fake}

capture the additional change in these estimates specific to the domains impacted by the FTC shutdown.

α_{m}

is the merchant type fixed effect that allows for the fact that some merchant types have more visits than others,

α_{m 1}

is the merchant type fixed effect specific to 2011, and

α_{m 2}

is the merchant-type-specific monthly time trend. Control domains are included as a separate merchant type. The portion that accounts for the excess zeros,

p_{jt}

, is modeled using a logit model, such that

p_{jt} = \frac{\exp (γ X)}{1 + \exp (γ X)}

, where

γ X

is the same specification used in the RHS of Equation A1.

Ideal control sites would be domains/merchants that used fake-news-style advertising but did not face an FTC shutdown during the period of analysis. Although such merchants exist in the postshutdown period, they did not exist in the preshutdown period (January–March 2011), preventing a difference-in-difference-in-differences style analysis. These merchants were those whose affiliate networks faced shutdown orders in 2012 and 2014. In addition, because these merchants were in the initial stages in the treated period (May–June 2011), the results of a simple difference analysis with the treated and control sites might not be representative. Instead, to control for time trends in this industry, I use sites related to regular weight-loss products such as Weight Watchers, My True Weight Loss, and Weight Loss for All. In the comScore data set, I search for all domains that include the word “weight” in them and treat the resulting domains as control sites. In Appendix B, because acai products form a major portion of the sold products on the merchants' sites, I also use sites that contain the term “acai” as controls.

Table A1 present the results of this regression. The results are qualitatively identical to the difference-in-differences analysis presented in the “Empirical Analysis” section. To ensure parallel trends in the pretreatment period between the control sites and the treated sites, I compare data patterns of the control domains with the treated sites. A visual verification shows that the pretreatment trends look fairly similar across control and treated sites. I further add a pretreatment term in the estimation equations for the control and treated groups and find that the treated groups are not significantly different, validating the parallel-trends assumption.

Table A1.

Difference in Domain First Visits After Fake News Sites Shutdown: Difference-in-Difference-in-Differences.

	All
N Unique Treated Websites	174
	Coef.	t-Stat.
Pr(zero visits): Logit
Post	−.081	−2.44
Post × FTC	.031	.57
Post × Fake	.240	2.38
Post × FTC × Fake	.950	5.43
Marginal effect	10%	5.42
Number of events: Poisson
Post	−.243	−2.76
Post × FTC	−.023	−.12
Post × Fake	−.075	−.37
Post × FTC × Fake	−.435	−1.24
Marginal effect	−18.76	−1.16

Total marginal effect	−4.81	−2.18
N observations	72,390
N	3,810
Controls	Merchant type
	Merchant type × Year
	Merchant type × Month
Cluster	Domain

Notes: The table presents, for only the first visit to each domain by an individual, results of a before–after zero-inflated Poisson regression, relative to a placebo year and relative to control sites related to weight. The dependent variable is the number of site visits. The coefficients corresponding to Post × FTC × Fake are the treatment effects (in boldface). “All” refers to all domains that to which the fake news sites referred people, excluding normal domains.

Appendix B: Robustness Checks

News Consumption and Heterogeneous Effects

One alternative explanation for the decline in direct visits is the effect of PR following the FTC shutdown, where bad press might have reduced consumers' propensity to visit these domains. However, I show below that the biggest impact of the shutdown is on those who consume the least news. If news effects were driving the decline, one would expect the group consuming more news to see the largest decline. This finding is consistent with the hypothesis that those with low news consumption are likely to belong to the susceptible population.

To examine whether news consumption (FTC's efforts or news PR effects) explains the decline in visits, I explore heterogeneity among consumers as defined by their level of news consumption. I create an indicator variable, $LowNew s_{i} = 1$ , if individual i's news consumption is below or equal to the median consumption across users. I then aggregate the data to the group-domain-month level, where all individuals with $LowNew s_{i} = 1$ are grouped together into group $Low$ and the remainder are grouped into another group. I use three measures that are indicators of consumers' news consumption: (1) visits to the FTC's website (ftc.gov), (2) visits to news sites that covered the FTC's press release in April 2011, and (3) top news sites such as CNN. The median number of visits to (1) ftc.gov is 0, (2) the 17 news sites covering the release is 0, and (3) the top news sites is 4.

Equation B1 illustrates the basic differences-in-differences approach in a linear specification:

\begin{aligned} y_{gjt} = & β \times {Post}_{t} + β^{FTC} \times {Post}_{t} \times {FTC}_{Y} \\ + β^{Low} \times {Post}_{t} \times {Low}_{g} + β^{Low, FTC} \times {Post}_{t} \times {FTC}_{Y} \\ \times {Low}_{g} + α_{m} + α_{m 1} \times {FTC}_{Y} + α_{m 2} \times t \\ + α_{m}^{Low} \times {Low}_{g} + α_{m 1}^{Low} \times {FTC}_{Y} \times {Low}_{g} \\ (B 1) & + α_{m 2}^{Low} \times t \times {Low}_{g} + ε_{gjt} . \end{aligned}

(B1)

Here,

y_{gjt}

represents the number of visits received by domain j in month t by group g (i.e., low-news consumers or high-news consumers). The rest of the terms are as specified in the main regression analysis of the article. The only difference between Equation B1 and the main regression specification is that each term is now interacted with the group.

Because of the excess zeros in the data, I then estimate a zero-inflated Poisson regression, where

y_{gjt} \sim {\begin{matrix} 0 & with probability p_{gjt} \\ Poisson (λ_{gjt}) & with probability 1 - p_{gjt} \end{matrix} .

(\rm B2)

Here,

y_{gjt}

represents the number of visits received by domain j in month t by group

g

(i.e., low-news consumers or high-news consumers) and

p_{gjt}

is the probability that domain j receives a zero visit in month t by group g. Conditional on a domain getting a visit, the distribution is assumed to follow a Poisson process with mean

λ_{gjt}

The probability of a zero visit, $p_{gjt}$ , is modeled using a logit model, such that $p_{gjt} = \frac{\exp (γ X)}{1 + \exp (γ X)}$ , where $γ X$ follows the same specification in the RHS of Equation B1. The mean, $λ_{gjt}$ , of the Poisson distribution (for nonzero visits) is specified such that $\ln (λ_{gjt}) = β X$ also follows the same specification as the RHS of Equation B1.

The results reported in Table B1 show that those who visit fewer news sites—either ftc.gov or sites that cover the FTC announcements—drop their visits to the questionable product domains. Considering the probability of a zero visit, the coefficient for the interaction term Post × Low × FTC is positive and significant for sites that cover the FTC press release (column 2) and marginally significant for visits to ftc.gov (column 1), indicating that the probability of a zero visit increases for those who consume less news. (There is no difference between users who consume more vs. less popular news; column 3). This result provides evidence that this drop is unlikely to be driven by the press effect (e.g., consumers read news and therefore decide to visit the domains less often), because these users consume very little news by construction.

Table B1.

Analysis Interacted with News Consumption: Zero-Inflated Poisson Model with Aggregate Data.

	(1) ftc.gov		(2) News on Press Release		(3) Top News
	Coef.	t-Stat.	Coef.	t-Stat.	Coef.	t-Stat.
Pr(zero visits): Logit
Post	.062	.52	.149	1.19	.108	1.04
Post × FTC	.697	3.28	.724	3.56	.950	5.44
Post × Low	.090	.88	−.013	−.13	.047	.79
Post × Low × FTC	.314	1.67	.326	2.18	.112	1.04
Marginal effect	6%	1.67	7%	2.17	2%	1.04
Number of events: Poisson
Post	−.078	−.54	−.191	−1.19	−.176	−1.25
Post × FTC	.165	.71	.187	.77	.218	.88
Post × Low	−.123	−1.66	−.016	−.21	−.029	−.69
Post × Low × FTC	−.016	−.1	.004	.03	−.049	−.6
Marginal effect	−1.32	−.16	−.20	−.03	−2.23	−.57

Total marginal effect	−2.90	−1.14	−2.83	−1.2	−1.78	−1.36
N observations	6,612
N domains	174
Controls	Merchant type
	Merchant type × Year
	Merchant type × Month
Cluster	Domain

Notes: “Low” is defined as those who visit fewer than the median number of visits of (1) ftc.gov, (2) news sites that covered the FTC's press release, and (3) top news sites such as CNN. Data are at the domain-year-month level. Standard errors are clustered at the domain level.

This test helps rule out the press effect and also helps recover an important layer of heterogeneity: it suggests those who consume less news are likely more influenced by the fake news ads and are the group of users for whom the regulation is impactful.

Popularity of Acai

Another source is the popularity of acai (which forms a major portion of the sold products on the merchants' sites), which might be declining in the period corresponding to the FTC release. To test if acai-driven popularity might be contributing to this effect, I use Google search volume data on acai as a control, as well as regular “acai” sites as controls.

Acai trend

Using Google Trends, I capture the search volume per month for the search term “acai.” If popularity of “acai” is following a downward trend, this might explain the drop in visits around the FTC press release. Figure B1 plots this search volume for the 2010 and 2011, exhibiting a decline over time. I therefore use the acai search volume as an additional control in the analysis. The monthly time trend captures any other trends that might be occurring in the industry. Table B2 presents the results of this regression. The coefficients for the search volume are statistically insignificant in both the zero-probability and number-of-events distributions, suggesting that the merchant-specific monthly time trends capture most of the decline. Considering the probability of a zero-visit, the main coefficient of interest corresponding to Post × FTC is still positive and significant, indicating that the probability of a zero visit increases after the FTC shutdown.

Figure B1.

Google Trends: acai search volume.

Table B2.

Difference-in-Differences Analysis Controlling for Search Volume of Acai.

	Coefficient	t-Statistic
Pr(zero visits): Logit
Acai search volume	.011	.56
Post	.166	1.5
Post × FTC	.994	5.78
Marginal effect	22%	5.96
Number of events: Poisson
Acai search volume	−.025	−.93
Post	−.229	−1.44
Post × FTC	.143	.57
Marginal effect	12.39	.56

Total marginal effect	−12.65	−1.43
N observations	3,306
N	174
Controls	Merchant type
	Merchant type × Year
	Merchant type × Month
Cluster	Domain

Notes: Search volume corresponds to the monthly search volume for the term “acai” captured from Google Trends data.

Acai sites as controls

If acai products saw a downward trend during this period, all sites containing the term “acai” should experience such a decline and would serve as a good control because they represent the industry trend. Moreover, a PR effect (if any) will likely affect normal acai sites if consumers cannot easily distinguish between a merchant using a fake news site and a regular merchant, and there is a negative spillover to the normal acai sites. I therefore extract all sites that contain the term “acai” (but are not merchants associated with the fake news ads), including sites such as takeacainow.com, acaimaxcleanse.com and acaiberryforsale.net. Table B3 presents the results of the difference-in-difference-in-differences analysis using legitimate “acai” sites as controls. The results indicate that only the treated sites see a decline in this period, providing an additional robustness check to the main finding.

Table B3.

Difference-in-Difference-in-Differences Analysis Using Normal Acai Sites as Controls.

	All
N Unique Treated Websites	174
	Coef.	t-Stat.
Pr(zero visits): Logit
Post	.229	2.51
Post × FTC	−.149	−.68
Post × Treated	−.078	−.59
Post × FTC × Treated	1.149	4.15
Marginal effect	14%	4.19
Number of events: Poisson
Post	−.104	−.58
Post × FTC	.438	1.21
Post × Treated	−.086	−.39
Post × FTC × Treated	−.302	−.68
Marginal effect	−16.32	−1.08

Total marginal effect	−5.41	−2.07
N observations	18,734
N	986
Controls	Merchant type
	Merchant type × Year
	Merchant type × Month
Cluster	Domain

Notes: The table presents, for only the first visit to each domain by an individual, results of a before–after zero-inflated Poisson regression, relative to a placebo year and relative to control sites related to acai. The dependent variable is the number of site visits. The coefficients corresponding to Post × FTC × Treated are the treatment effects (in boldface). “All” refers to all domains to which the fake news sites referred people, excluding normal domains.

Past Exposure to Fake News Ads

An information effect, where the mere presence of the ad is sufficient for consumers to visit the site (clicking on the ad but not using it as a referral, not clicking on the ad but absorbing the information in the ad), might be responsible for the decline in site visits. One measure of consumers' past exposure are observations where they clicked on the fake news ad but did not use it as a referral to visit the merchant site. Such observations form .83% of the data (compared with 2.4% fake news ad referrals shown in Table 3), suggesting that consumers read these fake news advertisements and then visit the site subsequently (without a direct referral).

While this provides a direct measure of exposure, other means of exposure, such as viewing the ad without clicking and reading the fake news ad, are unmeasurable but likely to contribute to consumers' exposure as well.

Spillovers to Other Merchants

On average, using the pretreatment year 2010, an individual visits 2 merchants per year. Restricting attention to the subset of users who used fake news ad referrals once, the average number of visits to merchants by such individuals is much higher, at 3.25 visits. This statistic implies that users who came in through fake news ad referrals are visiting more merchants, suggesting that there can be spillovers of the shutdown to other domains an individual might visit. That is, an individual does not just stop visiting the merchant she became aware of through the fake news advertisement but also might stop visiting other such merchants.

Role of Search Engines

Another alternative explanation is that Google and other search engines might have stopped ranking these now-questionable merchants after the shutdown, causing a drop in direct visits because of a change in the search engine's algorithm and not because of the absence of the fake news ads. This possible explanation would also lower the effect of regular ad referrals, implying that the increase is a conservative estimate.

Supply-Side Factors

It is unlikely that firms increased any advertising of the fake news kind because the FTC would penalize such actions in the light of the investigation. However, firms could have increased or decreased other kinds of advertising. The analysis using the ad expenditure data in the “Supply-Side Responses” section provides evidence that there was no significant ad spend increase. I also show that the results are robust after controlling for changes in advertising intensity. However, to the extent that there exist other potential supply-side responses (that are unobservable to the researcher), the effect measured includes such supply-side actions.

Appendix C: Additional Proxies for Purchases

Here, I use two additional proxies for purchases, duration spent and visits to possible order management sites. Table C1 presents the results of difference-in-differences analyses on duration spent and visits to tracking sites (websites that have “trac” or “trk” in their name) and visits to possible ordering sites (sites that have “click” or “clk,” “trac” or “trk,” or “order” or “secure” in their name). The results indicate a significant decline in purchases using both these metrics. Specifically, the probability that a domain receives a zero visit after the shutdown increases by 15%–22% across these three proxies, and this increase is statistically significant.

Table C1.

Analyses Using Proxies for Purchase: Duration Spent, Visits to Possible Ordering Sites.

	Duration		Tracking		Ordering
	Coef.	t-Stat.	Coef.	t-Stat.	Coef.	t-Stat.
Pr(zero visits): Logit
Post	.159	1.62	.181	1.08	.150	1.03
Post × FTC	1.029	5.8	1.186	2.84	.712	2.18
Marginal effect	22%	5.92	27%	3.02	15%	2.2
Number of events: Poisson
Post	−.200	−1.06	−.003	−.02	.115	.65
Post × FTC	−.101	−.34	.295	1.4	.055	.23
Marginal effect	−15.94	−.35	48.10	1.39	6.73	.23

Total marginal effect	−41.20	−2.25	−19.05	−.89	−14.94	−.9
N observations	3,306		380		608
N	174		20		32
Controls	Merchant type
	Merchant type × Year
	Merchant type × Month
Cluster	Domain

Supplemental Material

sj-pdf-1-mrj-10.1177_00222437211039804 - Supplemental material for Deceptive Claims Using Fake News Advertising: The Impact on Consumers

Supplemental material, sj-pdf-1-mrj-10.1177_00222437211039804 for Deceptive Claims Using Fake News Advertising: The Impact on Consumers by Anita Rao in Journal of Marketing Research

Footnotes

Acknowledgments

The author thanks seminar participants at Boston University, Northwestern Law and Economics Seminar, University of Chicago Law and Economics Workshop, and University of Florida and participants at the 2019 Marketing Dynamics Conference, 2019 Marketing Science Conference, 2019 Workshop on Economics of Advertising and Marketing, 2019 Competition Policy Enforcement Conference, 2019 Midwest IO Fest, and the 2019 Frank M. Bass UT Dallas FORMS Conference for their valuable suggestions. The researcher’s own analyses were calculated (or derived) based in part on data from The Nielsen Company (US), LLC and marketing databases provided through the Nielsen Datasets at the Kilts Center for Marketing Data Center at The University of Chicago Booth School of Business. The conclusions drawn from the Nielsen data are those of the researcher and do not reflect the views of Nielsen. Nielsen is not responsible for, had no role in, and was not involved in analyzing and preparing the results reported herein.

Associate Editor

Dina Mayzlin

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship and/or publication of this article.

ORCID iD

Anita Rao

Online supplement:

Notes

References

FTC (2011), “FTC Seeks to Halt 10 Operators of Fake News Sites from Making Deceptive Claims About Acai Berry Weight Loss Products,” press release (April 19), https://www.ftc.gov/news-events/press-releases/2011/04/ftc-seeks-halt-10-operators-fake-news-sites-making-deceptive.

FTC (2013), “FTC Staff Revises Online Advertising Disclosure Guidelines,” press release (March 12), https://www.ftc.gov/news-events/press-releases/2013/03/ftc-staff-revises-online-advertising-disclosure-guidelines.

Allcott

Hunt

Gentzkow

Matthew

(2017), “Social Media and Fake News in the 2016 Election,” Journal of Economic Perspectives, 31 (2), 211–36.

Anderson

Eric T.

Simester

Duncan I.

(2014), “Reviews Without a Purchase: Low Ratings, Loyal Customers, and Deception,” Journal of Marketing Research, 51 (3), 249–69.

Aribarg

Anocha

Schwartz

Eric M.

(2020) , “Native Advertising in Online News: Trade-Offs Among Clicks, Brand Recognition, and Website Trustworthiness,” Journal of Marketing Research, 57 (1), 20–34.

Berelson

Bernard R.

Lazarsfeld

Paul F.

McPhee

William N.

(1954), Voting: A Study of Opinion Formation in a Presidential Campaign. Chicago: University of Chicago Press.

Blake

Thomas

Nosko

Chris

Tadelis

Steven

(2015), “Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment,” Econometrica, 83 (1), 155–74.

Chatterjee

Patrali

Hoffman

Donna L.

Novak

Thomas P.

(2003), “Modeling the Clickstream: Implications for Web-Based Advertising Efforts,” Marketing Science, 22 (4), 520–41.

Chiou

Lesley

Tucker

Catherine E.

(2018), “Fake News and Advertising on Social Media: A Study of the Anti-Vaccination Movement,” Working Paper 25223, National Bureau of Economic Research, https://www.nber.org/papers/w25223.

10.

Chiou

Lesley

Tucker

Catherine E.

(2021), “How Do Restrictions on Advertising Affect Consumer Search?” Management Science (published online September 1), https://doi.org/10.1287/mnsc.2021.3980.

11.

Edelman

Benjamin

Gilchrist

Duncan S.

(2012), “Advertising Disclosures: Measuring Labeling Alternatives in Internet Search Engines,” Information Economics and Policy, 24 (1), 75–89.

12.

Free Press (2011), “FCC Fines Local TV Stations for Airing Fake News,” (March 25), https://www.freepress.net/news/press-releases/fcc-fines-local-tv-stations-airing-fake-news.

13.

Gentzkow

Matthew

Shapiro

Jesse M.

(2010), “What Drives Media Slant? Evidence from US Daily Newspapers,” Econometrica, 78 (1), 35–71.

14.

Goldfarb

Avi

Tucker

Catherine

(2011), “Advertising Bans and the Substitutability of Online and Offline Advertising,” Journal of Marketing Research, 48 (2), 207–27.

15.

Guess

Andrew

Nagler

Jonathan

Tucker

Joshua

(2019), “Less Than You Think: Prevalence and Predictors of Fake News Dissemination on Facebook,” Science Advances, 5 (1), eaau4586.

16.

Guess

Andrew

Nyhan

Brendan

Reifler

Jason

(2018), “Selective Exposure to Misinformation: Evidence from the Consumption of Fake News During the 2016 U.S. Presidential Campaign,” working paper, Dartmouth College.

17.

Ilfeld

Johanna S.

Winer

Russell S.

(2002), “Generating Website Traffic,” Journal of Advertising Research, 42 (5), 49–61.

18.

Kalla

Joshua L.

Broockman

David E.

(2018), “The Minimal Persuasive Effects of Campaign Contact in General Elections: Evidence from 49 Field Experiments,” American Political Science Review, 112 (1), 148–66.

19.

Lambert

Diane

(1992), “Zero-Inflated Poisson Regression, with an Application to Defects in Manufacturing,” Technometrics, 34 (1), 1–14.

20.

Little

Andrew T.

(2018), “Fake News, Propaganda, and Lies Can Be Pervasive Even if They Aren’t Persuasive,” Comparative Politics Newsletter, 28 (2), 49–55.

21.

Luca

Michael

Zervas

Giorgos

(2016), “Fake It Till You Make It: Reputation, Competition, and Yelp Review Fraud,” Management Science, 62 (12), 3412–27.

22.

Manchanda

Puneet

Dubé

Jean-Pierre

Goh

Khim Y.

Chintagunta

Pradeep K.

(2006), “The Effect of Banner Advertising on Internet Purchasing,” Journal of Marketing Research, 43 (1), 98–108.

23.

Mayzlin

Dina

Dover

Yaniv

Chevalier

Judith

(2014), “Promotional Reviews: An Empirical Investigation of Online Review Manipulation,” American Economic Review, 104 (8), 2421–55.

24.

Olson

Jerry C.

Dover

Philip A.

(1978), “Cognitive Effects of Deceptive Advertising,” Journal of Marketing Research, 15 (1), 29–38.

25.

Pennycook

Gordon

Bear

Adam

Collins

Evan T.

Rand

David G.

(2020), “The Implied Truth Effect: Attaching Warnings to a Subset of Fake News Headlines Increases Perceived Accuracy of Headlines Without Warnings,” Management Science, 66 (11), 4944–57.

26.

Pennycook

Gordon

Cannon

Tyrone D.

Rand

David G.

(2018), “Prior Exposure Increases Perceived Accuracy of Fake News,” Journal of Experimental Psychology: General, 147 (12), 1865–80.

27.

Petty

Ross D.

(2013), “From Puffery to Penalties: A Historical Analysis of US Masked Marketing Public Policy Concerns,” Journal of Historical Research in Marketing, 5 (1), 10–26.

28.

Porter

Ethan

Wood

Thomas J.

Kirby

David

(2018), “Sex Trafficking, Russian Infiltration, Birth Certificates, and Pedophilia: A Survey Experiment Correcting Fake News,” Journal of Experimental Political Science, 5, 159–64.

29.

Rao

Anita

Wang

Emily

(2017), “Demand for ‘Healthy’ Products: False Claims and FTC Regulation,” Journal of Marketing Research, 54 (6), 968–89.

30.

Rutz

Oliver J.

Bucklin

Randolph E.

(2011), “From Generic to Branded: A Model of Spillover in Paid Search Advertising,” Journal of Marketing Research, 48 (1), 87–102.

31.

Sahni

Navdeep

Nair

Harikesh

(2020), “Sponsorship Disclosure and Consumer Deception: Experimental Evidence from Native Advertising in Mobile Search,” Marketing Science, 39 (1), 5–32.

32.

Sherman

Lee

Deighton

John

(2001), “Banner Advertising: Measuring Effectiveness and Optimizing Placement,” Journal of Interactive Marketing, 15 (2), 60–64.

33.

Shimp

Terence A.

Preston

Ivan L.

(1981), “Deceptive and Nondeceptive Consequences of Evaluative Advertising,” Journal of Marketing, 45 (1), 22–32.

34.

Simonov

Andrey

Rao

Justin

(2021), “Demand for Online News Under Government Control: Evidence from Russia,” working paper, Columbia Business School.

35.

Thomas

Michael

(2020), “Spillovers from Mass Advertising: An Identification Strategy,” Marketing Science, 39 (4), 807–26.

36.

Vosoughi

Soroush

Roy

Deb

Aral

Sinan

(2018), “The Spread of True and False News Online,” Science, 359 (6380), 1146–51.

37.

Zinman

Jonathan

Zitzewitz

Eric

(2016), “Wintertime for Deceptive Advertising,” American Economic Journal, 8 (1), 177–92.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

10.70 MB