Learning from Incidental Exposure to Political Information in Online Environments

This article aims to explain learning outcomes of incidental exposure (IE) to political information in online environments. Drawing on the Political Incidental News Exposure Model, we predict learning outcomes by distinguishing between first-level (i.e., scanning of incidentally encountered information) and second-level IE (i.e., effortful processing of incidentally encountered information appraised as relevant). Furthermore, we conceptualize intention-based IE (i.e., while looking for non-political content) and topic-based IE (i.e., while looking for different political content). In a 2x2 experiment ( N ¼ 290), we manipulated respondents’ initial processing goal (i.e., political or non-political information) and low (i.e., first-level IE) versus high (i.e., second-level IE) relevance of the incidentally encountered information. Results show stronger learning effects for second level than for first-level IE. Learning effects do not differ between topic-based and intention-based IE, but second-level IE decreases learning related to the initial processing goal. Theoretical implications are discussed.

Against this background, political communication scholars pointed at the internet's capacities to expose citizens to politics even when they did not intend to see such information. Such so-called incidental exposure (IE) refers to situations in which individuals encounter political content although they did not actively search for it in the first place (Bode, 2016;Lee & Kim, 2017). For instance, numerous online media present political next to non-political information. Especially social network platforms can be considered as the "spaces in which flows combine and intertwine" (Thorson & Wells, 2016, p. 320). As another example, political news is displayed on most of today's login pages of email providers or internet portals.
However, current research on IE sparked some criticism (Kaiser, Keller, & Kleinen-von Königslöw, 2018;Matthes, Nanz, Stubenvoll, & Heiss, 2020;Thorson, 2020;Vraga, Bode, Smithson, & Troller-Renfree, 2019;Weeks & Lane, 2020). First, research often does not distinguish between the mere scanning and subsequent skipping of incidentally encountered content and the intensive processing of relevant information that was encountered incidentally. This distinction is important because the scanning of information and the intensive processing of information may lead to totally different learning outcomes. Second, by primarily focusing on IE while people were explicitly not looking for political information, previous research neglects the possibility that individuals stumble upon political information about a certain topic while they are looking for political information on another topic. This notion is important, because in high-choice media environments, individuals are facing increased possibilities to select those political topics they are most interested in, while other topics can be circumvented or ignored. Yet, IE can explain learning outcomes with respect to topics recipients were not looking for or even avoiding.
In response to these criticisms, we take a more fine-grained approach. Following the Political Incidental News Exposure (PINE) model , we define IE as exposure to information that people encounter without actively seeking for it. Importantly, we distinguish two levels of IE: First-level IE, which is the scanning of incidentally encountered information, and second-level IE, defined as the effortful processing of incidentally encountered information. As we will argue in the following, this distinction is crucial to understand the effects of IE. Furthermore, the PINE model recognizes intention-based IE (i.e., IE while looking for non-political content) and topic-based IE (i.e., IE while looking for different political content). The present study is the first to test the PINE model's core assumptions.

Incidental Exposure to Political Information
Numerous studies shed some light on the antecedents and consequences of IE (e.g., Heiss, Knoll, & Matthes, 2019;Lee & Kim, 2017;. Despite these efforts to understand IE's effects on political outcomes, three major criticisms have been expressed in the literature (Kaiser et al., 2018;Matthes et al., 2020;Vraga et al., 2019). First, IE research lacks a clear-cut conceptualization. IE is often used as some kind of umbrella term for a set of diverse situations that involve exposure to information that respondents were not looking for. On the one hand, briefly glimpsing at an incidentally encountered headline or scrolling past information someone stumbled upon is considered as IE. On the other hand, scholars also denote situations in which individuals read a full article after clicking on an incidentally encountered link as the same phenomenon (e.g., Fletcher & Nielsen, 2018;Tewksbury, Weaver, & Maddex, 2001). Kaiser and colleagues (2018, p. 3) therefore speak of "a lack of differentiation." We argue that effects of IE will differ for situations in which incidentally encountered news are either briefly scanned or processed with full attention. However, this notion has never been put to test.
In line with this, scholars are not clear in their operationalization of what it means to "encounter" or "come across" information incidentally. As Matthes and colleagues (2020) noticed, survey research often leaves the definition of these terms to the respondents. As expressed by Vraga and colleagues (2019, p. 237), "it is unclear to what extent people pay attention to and remember content they do not deliberately choose, rather than ignore or skip over it." Experimental research on IE has similar shortcomings. Most experiments apply forced exposure designs in which they expose participants to mock webpages with political information next to nonpolitical content. Then, after exposure, participants respond to the dependent measures, as for instance recognition. However, hardly any studies assessed what participants are actually doing during exposure (for exceptions see Lee & Kim, 2017;Vraga et al., 2019). That is, participants may read everything carefully, or they may not even notice the IE item. It is also possible that they dedicate full attention only to the IE items while ignoring the information they were exposed to in the first place. Thus, a respondent's behavior during news exposure needs to be taken into account. Exposure alone has limited explanatory power .
Second, and closely related, the theoretical foundation of IE research often appears to be underdeveloped. Various studies on IE build their argument on passive learning (e.g., Bode, 2016;Lee & Kim, 2017), which refers to situations with a lack of motivation to learn (Krugman & Hartley, 1970). While this explanation fits when individuals only briefly glance at content, it may not explain other situations in which individuals focus their attention on incidentally encountered content. For example, individuals seeking relaxation may stumble upon a story on the president's latest comment. They may be intrigued to learn what the president said and decide to click on the news story in order to read it. Obviously, this kind of knowledge acquisition cannot be explained by passive learning theory.
Third, previous research focused exclusively on what scholars have called intention-based IE. According to Yadamsuren and Erdelez (2016), such intentionbased IE occurs when individuals see political information while they are using media for non-political purposes. In fact, most surveys asked participants whether they encountered political information while they were "on-line for a purpose other than to get the news" (Tewksbury et al., 2001, p. 548). Similarly, most experiments on learning from IE present the political information next to non-political social information (e.g., Bode, 2016;Lee & Kim, 2017).
However, individuals may get incidentally exposed to political information on a certain topic when they look for information on another political topic. This has been called topic-based IE (Yadamsuren & Erdelez, 2016). In a survey by Pew Research Center (2017), online news consumers reported that they happen to encounter news on their main news topic while they get news on other topics on average 24% of the time. Such topic-based IE is important because it can explain learning outcomes regarding information that recipients were not actively looking for or even avoiding. Furthermore, the notion of topic-based IE may explain why individuals may get distracted from political information on a topic they are looking for in the first place. In what follows, we present a theoretical model that addresses these three criticisms.

The Political Incidental News Exposure (PINE) Model
There are four key assumptions of the PINE model . First, individuals possess a processing goal at all times during media reception. With the term processing goal, we refer to the purpose of an individual to cognitively engage with content. However, the term processing goal cannot be equated with generic content selection only. In contrast to uses and gratifications approach's motivations (see Rubin, 2009), processing goals refer to the engagement with the content individuals want to see and not the underlying gratification sought. This distinction is important because a common uses and gratifications motivation like pass time might be fulfilled for some individuals by consuming funny videos while others may pursue it by watching a political debate. Also, motivations as theorized by the uses and gratifications approach can be understood as generic categories that explain content selection, but not information processing. Processing goals, by contrast, refer to the directed engagement with content during media use.
Processing goals need to be understood as dynamic, that is, they may constantly change during reception. Also, individuals may pursue multiple goals (e.g., multiple informational needs) at the same time. In such a situation, the dominant goal (i.e., the strongest goal) is considered to be the processing goal. The PINE model distinguishes between specific political processing goals (e.g., seeking information on the president's latest remarks) and non-political processing goals (e.g., looking for entertainment content). For instance, people may have a non-political processing goal at a given time during reception, and may be confronted with political news incidentally.
While processing goals are understood as dynamic, they may also become chronically accessible. A decent share of media use is influenced by habits. Repetition of goal directed behavior is a key determinant of habit formation (Wood & Rünger, 2016). That is, if individuals regularly engage in a non-political processing goal during reception, this may lead to the formation of a habit. In other words, processing goals can become chronic. Importantly, even if media consumption is habitualized, there is still an underlying processing goal. Also, the pursuit of a chronic goal must not be conscious (Dijksterhuis & Aarts, 2010). Although more general content preferences and media habits may constantly shape individuals' processing goals and subsequently the selection and processing of information, there is evidence that short-term goals influence media use substantially. Karnowski et al.'s results (2017), for instance, show that reading intentions for a news story depend on topical interest and prior knowledge about the topic, while general frequency of news use is not a significant predictor of reading intentions.
Second, the PINE model theorizes that, during reception, individuals constantly engage in a so called relevance appraisal. That is, when individuals encounter information online, they permanently check whether the information at hand is relevant. Relevance appraisals do not require substantive amounts of cognitive resources. Individuals may start reading the first few words of the content to decide whether it is worth the effort (see Bode, Vraga, & Troller-Renfree, 2017). Individuals have to engage in this process for every bit of information they encounter in order to determine whether the content fits to the current processing goal or not. It is important to note that fit should be understood as a continuum, rather than a yes-no distinction. That is, a given content may fit a goal to varying degrees. Individuals engage in relevance appraisals because they consider the process of a relevance appraisal as a mean of assessing the fit between processing goal and encountered content. 1 There are three possible outcomes: As a first outcome, the content is in line with the processing goal. One may call this intentional exposure. As a second outcome, the encountered content is not in line with the processing goal and is not appraised as more relevant than the current processing goal. For these two outcomes, there will be no change in processing goals. As the third outcome, the content is not in line with the processing goal but, during the process of checking the relevance of the content, the individual appraises the content as more important than the original processing goal. This would lead to a switch of the processing goal.
Rephrased, in case an individual encounters information that does not fit the current processing goal, it is theorized that the relevance appraisal may lead to a switch of processing goals if incidentally encountered information is regarded as more relevant than the information in line with the initial processing goal. For example, at a certain moment during reception, an individual may be inclined to consume entertaining content but stumble upon an article about the president's recent comments. Importantly, the relevance appraisal is not only driven by the topic. Individuals may consider this article as more relevant for various reasons. For instance, genuine interest in the subject, seeing the name of one's home state in one of the president's remarks, a partisan source cue, or a friend's comment below the article could drive perceived relevance of incidentally encountered information. If this article is considered to be more relevant than the entertainment content, the individual will dedicate time and cognitive effort to process the article. In other words, there is a switch from a non-political processing goal to a specific political processing goal. This switching process can constantly occur during reception, depending on the current processing goal and the outcome of the relevance appraisal.
Third and in line with the notion of a relevance appraisal, the PINE model distinguishes between first-level IE, which describes the mere scanning of information regarded as irrelevant, and second-level IE, which incorporates the more effortful processing of incidentally encountered information regarded as relevant. Clearly, first-level IE happens under the conditions that (a) individuals encounter content that does not align with their processing goal and (b) that they appraise this content as irrelevant. Thus, because they lack a motivational driver they opt for scanning content with low engagement. The PINE model assumes that this mere exposure to incidentally encountered information should only lead to passive learning (Krugman & Hartley, 1970), and therefore, marginal knowledge gains. In contrast, second-level IE occurs only when individuals regard incidentally encountered information as relevant. In that case, individuals shift their cognitive resources toward incidentally encountered information.
Fourth, the PINE model accounts for intention-based as well as topic-based IE. For topic-based IE, individuals may be exposed to political information incidentally while they are looking for other political information. In line with Reinemann, Stanyer, Scherr, and Legnante (2012), we consider information to be political if any of the following four aspects is mentioned: (a) societal actors, (b) decision-making authorities, (c) activities of planning, deciding or realizing programs related to issues important to society, and (d) information about groups concerned with political decisions. As argued by Matthes et al. (2020Matthes et al. ( , p. 1037, "[t]opic-based IE may have the same effects as intention-based IE. The reason is that the incidentally encountered information is processed in very similar ways." However, this notion has never been put to test.

Testing the Political Incidental News Exposure Model
Negative Relevance Appraisal: First-Level IE In online news reception, individuals are typically exposed to a headline and a short teaser consisting of two or three sentences, conveying some information about the content. In line with the PINE model's notion of continuous and fast relevance appraisals during reception, individuals briefly scan these chunks of information to determine whether they want to have a closer look. Often individuals may keep moving on to the next piece of content after encountering political content incidentally. The PINE model posits that such first-level IE may lead to learning effects. That is, in case individuals see information incidentally but appraise this information as irrelevant, the information is still processed with minimal amounts of attention. Yet such processing may leave memory traces. The theoretical mechanism behind this effect is passive learning (e.g., Bode, 2016;Tewksbury et al., 2001).
Passive learning refers to a process of information acquisition when recipients are not interested in gaining knowledge, so their attention is not directed toward a stimulus (Krugman & Hartley, 1970). For instance, in an experiment by Lee and Kim (2017), even respondents who did not click on an incidentally encountered news banner were able to recognize that they saw such a news story while recall was dependent on clicking on the news banner. Their results suggest that the mere scanning of incidentally encountered information (i.e., first-level IE) may leave memory traces. Based on this reasoning, we expect that individuals may process the information transmitted by a headline, even though they regard the headline and the story behind it as irrelevant.
Taken together, in cases when several headlines are presented, first-level IE would predict that an incidentally encountered headline would still yield some amount of recognition. However, there is a certain likelihood that respondents report recognition simply by chance. Should first-level IE occur, PINE predicts that the recognition of the incidentally encountered information should be higher than by chance. In our study, first-level IE is manipulated with low relevance of the incidentally encountered information. Thus: H1: The headline recognition rate of respondents in the low relevance condition should be higher than recognition rates by chance.

Positive Relevance Appraisal: Second-Level IE
The PINE model argues that incidentally encountered information can be appraised as relevant. While scanning headlines and teasers, individuals may react by dedicating increased cognitive effort to relevant content. In line with this notion, an eyetracking study by Bode and colleagues (2017) suggests that those with high political interest are less likely to skip political social media posts the earlier political words occur in the posts. In other words, while political words can serve as cues that lead to a negative relevance appraisal for those with low interest, attention of those with high political interest sticks to content with political words for a longer amount of time. Similarly, factors like partisan slant or topical interest may alter the likelihood of a positive relevance appraisal (Bakshy, Messing, & Adamic, 2015;Karnowski et al., 2017). In case the relevance appraisal is positive, individuals engage in what the PINE model calls second-level IE. That is, the incidentally encountered content receives more attention during the reception process. In the case of online information environments, second-level IE has consequences in terms of clicking behavior and learning. It can be assumed that second-level IE increases the likelihood that recipients click on a headline. They do so because they have appraised the incidentally encountered information as relevant. It follows: H2: Respondents in the high relevance condition are more likely to click on incidentally encountered content than those in the low relevance condition.
The PINE model postulates that second-level IE should also lead to learning. To reiterate, one of the PINE model's main arguments is that first-level IE and secondlevel IE differ regarding outcomes. While the brief scanning of irrelevant information may leave memory traces (e.g., via priming; see Knoll, Matthes, & Heiss, 2020), incidentally encountered information appraised as relevant should increase learning. In other words, knowledge gains should primarily occur for second-level IE. The abundance of the information available online forces individuals not only to be selective in their exposure (i.e., clicking only on interesting stories) but also in their processing of information. Even for TV news, it has been shown that viewers are by far not able to recall all news stories they saw (Neuman, 1976). Clearly, because "not all perceived information can be processed (i.e., behaviorally utilized or stored for later retrieval and behavioral utilization), processing also entails selection" (Zillmann & Bryant, 1985, p. 1). Various scholars stressed that although exposure to political information is a precondition for learning, "it alone does not determine how much will be learned or how well" (Eveland, 2001, p. 588). Following this argument, we theorize that incidentally encountered information appraised as relevant is more likely to be processed thoroughly and subsequently leads to increased knowledge compared to non-relevant information.
More specifically, we look at three different learning outcomes: (a) headline recognition, (b) content recognition, and (c) story recall. IE research often stressed that exposure to headlines and teasers may lead to learning (Fletcher & Nielsen, 2018;Tewksbury et al., 2001). However, the full story is often hidden behind a link. Thus, we can assess learning outcomes with respect to the headline and the content of the full story, in case respondents clicked on that content. Furthermore, research on learning typically differentiates between recognition and recall. Lang (2000) highlights that recognition indicates that a bit of information was encoded (e.g., while scanning) while recall may serve "as an index of how thoroughly a specific bit of information was stored" (p. 56). Taken together, we theorize that incidentally encountered political information appraised as relevant should have a positive effect on headline recognition and recall.
H3: Respondents in the high relevance condition score higher on (a) headline recognition, and (b) story recall of incidentally encountered political information than those in the low relevance condition.
Although headlines and teasers often contain important information, much of the information is only given in the full text behind the link. To learn and subsequently recognize such information, individuals have to be exposed to this additional information by clicking on the link. This is the reason, why we do not expect a direct effect of a positive relevance appraisal on the third learning outcome, content recognition. However, we may expect a mediated effect of a positive relevance appraisal on learning outcomes through clicking on IE content for two reasons. First, reading the full story fosters processing of information that was given in the headline or teaser. The content behind the link often reiterates and contextualizes the headline. By reading the story, individuals process more information which is closely related to the headline. According to Lang (2000), storage of information is improved by additional links to related information in memory. Subsequently, retrieval (i.e., recognition and recall) should be easier the better the information was stored. Thus, we expect a partially mediated effect of a positive relevance appraisal on headline recognition and recall through clicking on IE content.
Second, clicking on IE content exposes individuals to new information. Headlines and teasers transport the most important part of a story while they omit details and context. However, the information hidden behind the link can be crucial for understanding. For example, a headline saying "Government spends 1 billion on employment measures" does not reveal anything about the measures itself. Typically, specifics which may also help individuals to form an opinion are discussed behind the link. Thus, exposure to this additional information allows individuals to learn information which they could not access without clicking on the IE content. Because exposure to information is a non-negligible precondition of knowledge acquisition, we expect a fully mediated effect of a positive relevance appraisal on content recognition through clicking on IE content. Taken together, effects on all three learning outcomes should be strengthened by clicking and being exposed to the actual content of the story.
H4: The effects of the relevance manipulation on (a) headline recognition, (b) story recall, and (c) content recognition are mediated through clicking on the IE content.

Learning Outcomes with Respect to the Original Processing Goal
So far, we discussed learning of incidentally encountered information. Yet the PINE model also postulates that IE can affect outcomes related to the initial processing goal. This may be negligible for political communication scholars, if IE is intentionbased. However, topic-based IE may also concern political learning outcomes. A positive relevance appraisal will lead the attention away from the original content to the incidentally encountered content. Individuals may skip the information that is unrelated to a new salient goal. For example, individuals may look for the latest news on presidential candidates but stumble upon information on foreign politics. In case they appraise news on foreign politics as relevant, they will shift their attention away from information about the candidates. Thus, learning outcomes regarding the candidates may be diminished. This effect can be explained by the Limited Capacity Model (Lang, 2000). Because cognitive resources are limited, an attention shift away from any original content will deteriorate the processing of the original content. Therefore, we assume that second-level IE should hinder individuals from achieving their initial processing goal. Regarding political outcomes, this is crucial because individuals may get distracted by IE while they were looking for the political information in which they were interested in the first place.
H5: Clicking on incidentally encountered information decreases the recognition of content related to the initial processing goal.
Intention-based versus Topic-based IE Intention-based IE refers to incidentally encountered political information while individuals had the goal to process non-political information. In contrast, topicbased IE means that people's original goal was to process a specific political topic. In both cases, the incidentally encountered content was not searched for in the first place. However, the original processing goal is different. We theorize that incidentally encountered political information leads to stronger knowledge outcomes if the original processing goal was also related to political content (i.e., topic-based IE). In that case, the incidentally encountered information is congruent to the original processing goal. One reason to expect such a congruency effect is that political information needs might often origin from similar higher-level goals (e.g., surveillance, watching for possible threats, acquiring political information for later use in discussions). According to goal systems theory, subgoals and higher-level goals are linked in a hierarchical network (Kruglanski et al., 2002). For example, the higher-level goal which drives individuals to inform themselves about an upcoming election might be more similar to the higher-level goal which motivates an individual to attend to information about foreign policy than the higher-level goal which drives exposure to a funny video. In other words, incidentally encountered information similar to an existing higher-level goal is more relevant than incidentally encountered information unrelated to a higher level goal. Due to the larger amount of links between two political processing goals compared to political processing goals and non-political processing goals, spread of activation and, subsequently, goal pursuit (e.g., cognitive processing, or reading) is more likely. Congruency to the original processing goal will thus lead to increased processing of the IE content, and therefore, enhance knowledge outcomes. It follows: H6: Topic-based IE leads to stronger effects on (a) headline recognition, (b) story recall, and (c) content recognition than intention-based IE.

Design and Sample
In a 2Â2 online experiment, we manipulated the processing goal (intention-based vs. topic-based), and the relevance appraisal (high relevance, i.e., second-level IE vs. low relevance, i.e., first-level IE). After the study, participants were thoroughly debriefed. Based on representative quotas for gender, age, and education, 341 respondents were recruited from a German online panel by Dynata. Out of the raw data, 51 respondents were excluded because they deactivated JavaScript, reported zip codes and states did not match, took less than 7 minutes, or more than 30 minutes for the study, leaving N ¼ 290 respondents. All analyses including the omitted cases yielded the same results. Participants were 51% female and M ¼ 47.74 years old (35.52% less than a high school, 12.41% high school, and 52.07% above high school).

Manipulation and Stimulus Material
Each participant saw two webpages in random order (see Figure 1, Online Supporting Materials document). Each displayed seven headlines. Webpages consisted of four articles about the processing goal topic, two articles regarding regional politics and one filler (about the British royals). The regional articles were fourth and sixth in one version and in the other version sixth and seventh from top. We informed respondents that clicking on a headline revealed the full article. Exposure to multiple articles at the same time was impossible.
Processing goal manipulation: Prior to exposure to the webpages, we told them to inform themselves about a political or non-political topic while hinting that they will take a quiz at the end of the study. Respondents in the intention-based group were advised to dedicate their attention to "news from Hollywood and cinema" and "news from sports." We chose these two topics because we consider them to be non-political. Respondents in the topic-based group were asked to inform themselves about "US trade war" and "rent, housing, and living conditions in Germany." We consider these two topics to be political. While articles for the topic-based group stressed the political dimension of the topic, none of the articles for the intentionbased group mentioned any political implications of the article's subject. Each webpage displayed four articles about the topic the respondents should inform themselves about. For example, a webpage for the intention-based group included four articles about rent, housing, and living conditions. The articles shown on the webpage matched the processing goal respondents were advised to pursue. Articles were based on real newspaper articles and of similar length (M ¼ 151.06 words, SD ¼ 8.47).
Relevance appraisal manipulation: We altered the regional articles (from here on we call them IE articles) based on the respondent's initially measured zip code. The high relevance group incidentally encountered three IE articles that mentioned cities and villages close to their place of living and one that mentioned the state they were living in. Respondents in the low relevance group received the very same articles including names of places far away from their place of living (see Knobloch-Westerwick et al., 2005, for a similar procedure). We believe that manipulating relevance via geographical proximity is a reasonable choice to test the framework for the first time. In contrast to other manipulations (e.g., issue salience), this manipulation is hardly confounded with political variables. Prior to the study, we matched each zip code in Germany (approximately 8,300) with its state and three villages or cities close to it. These places were shown to participants in the high relevance group. We decided to show respondents living in large cities like Berlin or Munich their own city. Participants in the low relevance group received a set of places that were located in a state that did not border the state they were living in. None of the low relevance cities was a nationally important city (like Berlin). The content of the article was identical, only the place varied.
IE articles: Four IE articles about regional politics were drafted. Each headline mentioned the city or state in which the story took place. In the "fraud story," it was reported that a former state legislator was convicted of fraud in a city. The "construction story" warned that a road construction will lead to massive traffic jam close to a city. The "tax story" reported that a state has to pay back its citizens. In the "water story," the tap water of a city was discolored and unhealthy. In all four stories, politicians and/or officials were cited and the political dimension of the topic was stressed. Articles were of similar length (M ¼ 150 words, SD ¼ 4.69).

IE article clicked:
We tracked whether respondents clicked on the IE articles by implementing JavaScript code into the web questionnaire. On average, respondents clicked on M ¼ 1.30 (SD ¼ 1.45) of the four IE articles.
Number of processing goal articles clicked: Similarly, the number of clicks on processing goal-related articles was tracked, i.e., how many of the articles about sports (U.S. trade war) and Hollywood and cinema (rent, housing, and living conditions in Germany) a respondent clicked on. On average, respondents clicked on M ¼ 4.59 (SD ¼ 2.93, a ¼ .89) of eight articles.
Story recall: Participants indicated whether they recognized any story. Those who recognized a story were asked to recall all details (per story, 1 ¼ story recall, 0 ¼ no story recall). An author and a student assistant coded all responses (N ¼ 758, Krippendorff's a ¼ 0.95).
Headline recognition: Respondents were asked to identify the headline they just saw from a list of four. Respondents were advised to take the "do not know"-option if they were not able to recognize the correct answer. We ensured that all response options were rather similar. We recoded responses to a dichotomous variable (1 ¼ correct headline recognition). Across all four articles, respondents correctly identified an average of M ¼ 38.44% (SD ¼ 0.31) headlines. Our operationalization actually probes for recognition. In studies by Bode (2016) and Lee and Kim (2017), respondents reported whether "they remember seeing" (Bode, 2016, p. 34) or "have seen" (Lee & Kim, 2017, p. 1010) IE content by ticking yes or no (Bode, 2016, also offered "don't know"). These measures are prone to generate false positives-for example, because of social desirability. In comparison, our measure was a multiple choice question that offered four similar headlines and a "don't know" option. Respondents had to choose one of the answers.
IE content recognition and processing goal content recognition: Respondents received a randomized list including two statements (one true, one false) for each of the four IE articles and each of the eight processing goal articles, totaling 24 statements. While the statements for the processing goal articles differed depending on the processing goal (topic-based or intention-based), recognition of the IE articles was assessed identically in all groups. Respondents indicated whether they thought the statement was true or false according to the information they saw before. Importantly, statements included information that was not conveyed by the headline itself. Thus, respondents must have read the articles to rate the statements correctly. Respondents were advised to tick "do not know" if they did not remember. We summed the scores (correct ¼ 1, wrong and DK ¼ 0) for each IE article and for all processing goal articles. With respect to IE content recognition, respondents were able to identify M ¼ 0.41 (SD ¼ 0.62) of two statement related to the IE article on average. Regarding the processing goal content recognition, the intentionbased group correctly identified M ¼ 5.65 (SD ¼ 3.81, a ¼ .82) while the topicbased group correctly recognized M ¼ 5.58 (SD ¼ 3.63 a ¼ .79) of 16 possible statements.
Manipulation check: Participants rated the distance of their place of living to each place mentioned in the stimuli on a 7-point scale from "far away" (1) to "very close" (7) (M ¼ 3.81, SD ¼ 2.73). Respondents indicating that they did not know the name of the city were recoded as 1. 2 We asked respondents about the topics they focused on "while they saw the two websites with news articles." For each of the three sets of articles (i.e., national politics, non-political, and regional), we asked two statements on a 7-point scale from "completely disagree" (1) to "completely agree" (7): "My attention was focused on articles about [national political topics/non-political topics/regional incidents]," "I spent the most time reading articles about [national political topics/non-political topics/regional incidents]." We created three mean scales: focus on national political topics (M ¼ 3.95, SD ¼ 1.93, a ¼ .92, r ¼ .85), non-political topics (M ¼ 3.96, SD ¼ 1.92, a ¼ .94, r ¼ .88), and regional incidents (M ¼ 3.75, SD ¼ 1.89, a ¼ .93, r ¼ .86).

Analysis
We had data for clicking behavior and knowledge questions for each IE articles separately. This allowed us to rearrange the data so that we had multiple observations per respondent. Therefore, we turned to multilevel logistic regression models for H2, H3a, H3b, H6a, and H6b and multilevel Poisson regression for H6c (both are applications of generalized linear mixed models, GLMM). 3 In our models, four observations for the IE articles were nested within each of the 290 respondents. We replicated hypothesis tests with simpler models as well (i.e., non-hierarchical Poisson regression model). 4 Direction of coefficients and their significance remained the same for all hypotheses. Additionally, we checked whether results differed for the four articles. Direction of coefficients remained the same across all tests. We used the "mediation" package for R (Tingley, Yamamoto, Hirose, Keele, & Imai, 2014) to estimate the mediation proposed in H4. For the analysis of H4c, IE content recognition was dichotomized. 5 Multilevel mediation can potentially lead to confounded estimates when within-group effects differ from between-group effects (Zhang, Zyphur, & Preacher, 2009). Therefore, we checked the robustness of our results with non-hierarchical mediation models. 6 Direction of coefficients and their significance remained the same for H4a, H4b, and H4c. The results were robust.

Results
First of all, we checked whether the pairing of places and zip codes worked. The high relevance condition (M ¼ 6.44, SD ¼ 0.8) reported to live closer to the stimuli cities than the low relevance group (M ¼ 1.13, SD ¼ 0.58), t(267) ¼ À64.59 (Welch-Satterthwaite), p < .001. 7 Hence, matching respondents' zip codes with cities close to their place of living worked.
Then we ran three ordinary least squares (OLS) regressions predicting focus on national political topics, focus on non-political topics, and focus on regional incidents with dummies for relevance manipulation, processing goal manipulation, and an interaction term as predictors. As intended, only the coefficient for the processing goal variable was significant when it came to predicting focus on national political topics (b ¼ À1.72, p < .001) and focus on non-political topics (b ¼ 2.20, p < .001). Thus, the processing goal manipulation worked. Similarly, only the relevance manipulation was a significant predictor for focus on regional incidents (b ¼ 0.68, p < .05), with no significant interaction terms. Hence, our relevance manipulation worked.
In H1, we hypothesized that respondents in the low relevance group (i.e., firstlevel IE) would be able to recognize the headlines correctly at a level better than chance. If respondents would choose one of the five answer options (three wrong headlines, one right headline and one "do not know"-option) at random, we would expect a mean of .2 correct answers (by chance, respondents would pick the right answer for 20% of the questions). Respondents had to select the correct headline for each of the four IE articles which leads us to expect .8 as mean for randomly responding respondents. A one-sample t-test revealed that respondents in the low relevance group (N ¼ 143; M ¼ 1.13, SD ¼ 1.19) recognized statistically significant more than .8 headlines on average, t(142) ¼ 3.35 (Welch-Satterthwaite), p < .01, supporting H1.
In H2, we expected that participants are more likely to click on incidentally encountered political information (i.e., regional IE news articles) if appraised as relevant. Model 1 in Table 1 shows a multilevel logistic regression with the dichotomous clicking variable as dependent. The coefficient for the relevance manipulation was positive and significant (b ¼ 0.98, p < .05). Holding all other predictors at fixed values, our model expected a 166% increase in the odds of clicking when the IE articles named cities close to respondents. Thus, H2 is supported.
We then looked at headline recognition (H3a). In Model 2 in Table 1, we regressed correct headline recognition on a set of predictors. We added a variable indicating whether participants clicked on the article (i.e., the dependent variable of  Model 1 in Table 1). The coefficient for clicking was positive and significant (b ¼ 1.27, p < .001), indicating that respondents who clicked on the article were 3.55 times more likely to correctly recognize the headline than respondents that did not click on the article. Importantly, the relevance manipulation was also a significant and a positive predictor of headline recognition (b ¼ 0.78, p < .01). Respondents in the high relevance group were more likely (118% increase) to recognize the headline correctly even after controlling for the effect of actual exposure to the full article. Hence, H3a is supported. H3b predicted that participants in the high relevance condition would have higher story recall than those in the low relevance condition. Results of a multilevel logistic regression are shown in Model 3 in Table 1. Similar to H3a, we decided to include a dummy indicating whether respondents clicked on the article. Again, actual exposure to the article (i.e., clicking) led to increased story recall (b ¼ 2.45, p < .001). Importantly, even when we control for clicking, the relevance manipulation remained a significant predictor of story recall (b ¼ 1.76, p < .001). All else equal, the likelihood of recalling an IE articles increased by 483% in the high compared to the low relevance condition, supporting H3b. We then tested whether the relevance manipulation's effect on headline recognition, story recall, and IE content recognition was mediated through actual exposure to the full article (i.e., clicking on the article). To test H4, we estimated three multilevel mediation models with quasi-Bayesian confidence intervals based on 5,000 simulations. Our analysis yielded that effects of the relevance manipulation were significantly mediated through clicking on IE content for (a) headline recognition (indirect effect ¼ 0.04, 95% CIs ¼ [0.01, 0.06], p < .01), (b) story recall (indirect  (Table 2). In H5, we expected that clicking on incidentally encountered articles should decrease learning related to the initial processing goal. In Model 1 of Table 3, our dependent variable indicates how many of the 16 statements about the processing goal articles respondents classified correctly. We included an interaction term for our manipulations into the non-hierarchical OLS regression model. The number of processing goal articles a person clicked on was a highly significant predictor (b ¼ 0.58, p < .001). In line with H5, we found a negative and significant effect of clicking on IE content on processing goal content recognition (b ¼ À0.73, p < .001). Finally, in H6 we expected that topic-based IE leads to more learning of incidentally encountered information than intention-based IE. Results are reported in Model 2, Model 3, and Model 4 of Table 1. We did not find any difference between the two processing goals when it came to learning of IE content. All three interaction coefficients-for (a) headline recognition (b ¼ 0.27, n.s.), for (b) story recall (b ¼ 0.22, n.s.), and for (c) IE content recognition (b ¼ 0.23, n.s.)-were not significant. We conclude that people learn from incidentally encountered political information regarded as relevant in a similar way regardless whether IE was topic-based or intention-based. H6 is rejected.

Discussion
We found that second-level IE leads to fundamentally different learning effects as compared to first-level IE. In particular, there are three main theoretical contributions of this study. First, we find unambiguous support for the notion that individuals will not only attend to, but also process incidentally encountered information more thoroughly if information is appraised as relevant. Second-level IE therefore leads to learning in online environments via two paths: First, memory for the presented information itself is increased. That is, when individuals read information snippets and headlines and appraise them as relevant, this information will stick to memory because it is processed more effortfully. Second, compared to firstlevel IE, second-level IE leads individuals to click on incidentally encountered content. This subsequently increases knowledge gains by exposing individuals to further information. Rephrased, there are two paths to learning from second-level IE: One is selective attention. Individuals will allocate attention to incidentally encountered information they appraise as relevant. So even without clicking on any additional information, learning processes are fostered by second-level IE. This insight is relevant because clicking on incidentally encountered information requires more resources (e.g., time) and cognitive capacities than processing short information snippets. Yet in today's media environments, individuals might often lack the capacities or resources to click on information they encounter incidentally, even if the information is deemed as important (Costera Meijer & Groot Kormelink, 2015). Nevertheless, our results suggest that substantial learning from incidentally encountered teasers, headlines, and snippets might still occur under the condition of second-level IE. In conceptual terms, these insights suggest that clicking on information cannot be treated as a measure of second-level IE because second-level IE can also occur when individuals do not click on the information that is appraised as relevant. As a second path to learning, second-level IE leads individuals to selectively expose themselves to additional content. This is arguably very important in terms of learning effects because such additional content typically contains new information that individuals most likely would never see under first-level IE conditions. As the second contribution, we looked at the differences between intentionbased and topic-based IE. Previous research only considered situations in which individuals looked for non-political content and stumbled upon something political. However, our findings suggest that whether individuals encounter political information incidentally while they are looking for non-political content or political content does not affect learning through IE at all. According to our findings, intentionbased and topic-based IE do not differ in terms of processes and effects. Nevertheless, we argue that neglecting topic-based IE is like turning a blind eye on a large share of the phenomenon (Pew Research Center, 2017). In fact, there is a myriad of research questions on topic-based IE worth to explore. For instance, according to the PINE model, recipients may constantly switch their processing goals during reception in online environments. For topic-based IE, they may switch between incidentally encountered topics and the original topic, back and forth. The more people switch, the more they get distracted from either topic, leading to a decrease in learning outcomes. As another example, the thematic, evaluative, and emotional congruence between the original topic and the incidentally encountered topics may matter for second-level IE. This could be tested in experimental research varying topical congruence.
Third, this study sets the stage for a new branch of research by highlighting distraction effects of IE. We found that second-level IE to political information harms learning of information related to the initial processing goal. Clearly, second-level IE requires cognitive resources. From a normative democratic perspective, this might hardly be troubling for situations in which citizens pursue non-political processing goals. However, when people actively search for a political topic, second-level IE to other political information can distract individuals and subsequently decrease knowledge gains related to the topic they were originally searching for. We can also ask the question whether IE to non-political information may lead to distraction from political processing goals . For example, individuals may search for information on an upcoming election but stumble upon entertainment content they regard as more relevant than the political processing goal. That is, in order to understand IE, not only the content and effects of IE need to be taken into account, but also the content and effects of the original information under the condition of IE.
Beyond these three contributions, our findings provide some evidence that firstlevel IE can lead to learning. That is, even when the presented information was not appraised as relevant, memory traces were higher than by chance. Although this is in line with previous research (Bode, 2016;Lee & Kim, 2017), better designs are needed to corroborate this claim, considering that we did not have a proper control group. Thus, this finding should be interpreted with caution. We find that memory traces from first-level IE cannot be compared to the more substantive learning outcomes generated by second-level IE. Nevertheless, especially in online environments and given a large share of audiences is uninterested in politics, first-level IE may explain why the politically uninvolved, who tune out of politics, are still connected to political news.
Prior to concluding remarks, it is important to stress the study's limitations. First, we did not explore the reasons for a positive relevance appraisal. By manipulating the geographical proximity of news events, we altered relevance based on the perceived utility of the information (Knobloch-Westerwick et al., 2005). Information on events far away from one's place of living are less important because it may less likely touch one's own life. However, in most online environments a range of competing cues can drive one's relevance appraisal (e.g., recommendations, content types ranging from text to videos, partisan cues). Future studies should replicate our findings and manipulate the relevance appraisal in more diverse ways (for further examples, see Kaiser et al., 2018) and should test the model by using inherently political manipulations like partisan cues. Second, we did not examine boundary conditions for entering second-level IE. Future studies should consider possible hindrances of engaging in deeper processing (i.e., second-level IE) of relevant information (e.g., time constraints, cognitive fatigue, Matthes et al., 2020; see also Weeks & Lane, 2020). Third, we used a processing goal manipulation that does not represent the variety of processing goals that may occur. Specifically, characteristics of the processing goal may influence the depth of first-level IE. For example, the amount of information an individual has to process to determine relevance may differ between processing goals. Similarly, processing goal characteristics may influence the likelihood of engaging in second-level IE. For instance, individuals pursuing a very strong processing goal may not engage in second-level IE at all. Future research should try to manipulate more characteristics of the processing goal. Additionally, we induced processing goals by telling respondents to focus on certain content. However, most individuals may also have more stable processing goals. In an additional post-hoc analysis, we thus probed for an interaction between the relevance manipulation and general political interest. We did not find a significant interaction effect on any of the learning outcomes. Future studies may consider more stable processing goals. Fourth, our design did not include a control group. Thus, in contrast to previous studies (Bode, 2016;Lee & Kim, 2017), we cannot show that respondents incidentally exposed to information learn more than respondents that did not see such information. Rather, we compared the low relevance group's recognition score with randomly responding individuals. A control group would allow a more robust test of the effects of first-level IE. Fifth, our experiment uses journalistic news articles as stimuli and we suggest that future research may use more diverse information environments. Sixth, the PINE model was explicitly designed for IE in social media environments. Our experiment employed a mock webpage (see also Lee & Kim, 2017). User behavior and reception situations may differ for social media sites. However, we are rather confident that the basic mechanisms should be the same regardless of the specific online information environment. Finally, the diachronic processes theorized in the PINE model were not fully taken into account in the present study. In fact, all current processing goals and relevance appraisals are shaped by preceding goals and appraisals. Future research should employ truly diachronic designs in order to trace the temporal dynamics theorized in PINE. Different methodological approaches, such as eye-tracking (King, Bol, Cummins, & John, 2019) or mobile experience sampling (Naab, Karnowski & Schlütz, 2019), may improve our understanding of IE phenomena (for a discussion of methodological implications, see Matthes et al., 2020).

Broader Implications for the Field
While this article primarily discusses IE to political information, our findings have significant implications for other areas of communication research. In online environments, the notion of IE is relevant to any kind of information, no matter if related to, for instance, health, risk, advertising, science, or the environment. Whatever the specific content is, we theorize that individuals have a processing goal and they constantly engage in relevance appraisals. That is, whenever individuals incidentally stumble upon information, they check this information for relevance. For example, multiple studies have investigated learning through IE to health information on the internet (e.g., Tian & Robinson, 2009). However, this strand of research typically does not consider first-and second-level IE. If individuals appraise incidental health information as relevant, the effects on all health-related outcomes such as learning, health-related behavior, or attitudes will be larger compared to a negative relevance appraisal. Likewise, in advertising research, click-through rates of banner ads can be conceptualized as second-level IE, and the PINE model can be used to predict clicking on banner ads as well as learning from banners appraised as irrelevant (see Yoo, 2009). Yet the PINE model would not conceptualize ad clicking as a static behavior (i.e., clicked or not) but conceptualize exposure to ads diachronically. The notion of constant relevance appraisals helps to better understand the dynamics of attention allocation to ads during reception. That is, individuals may click on a banner ad leading them to new content, yet this content will also be automatically appraised for its relevance, leading to effortful processing or the skipping of that content.
Beyond learning from incidentally encountered information, our research has also implications for the notion of distraction from content related to the initial processing goal. In most areas of the field, survey researchers ask respondents about the perception of media content (i.e., late-night comedy; climate change information, advertising, etc.), and such perceptions are typically correlated with outcome variables such as cognitions, affects, or behaviors. Yet exposure to incidentally encountered information unrelated to the initial processing goal is mostly ignored. We would argue that, especially in online information environments, there is always the possibility of distraction from the content related to the initial processing goal, and such effects need to be taken into account. For instance, when asking individuals how often they see political content on social media in order to explain political participation, scholars typically ignore (incidentally encountered) non-political content (e.g., . Such non-political content, however, can lead to the opposite effects as compared to political content (i.e., dampening participation). Again, this logic does not only apply to political content, but to any content that individuals are exposed to. Ultimately, incorporating distraction effects by IE may help research in other subfields to get a more fine-grained picture of media effects.
Finally, our research may also be relevant selective exposure research (see Knobloch-Westerwick et al., 2005). Selective exposure research primarily refers to the selection (or avoidance) of content categories, such as a particular channel or webpage. PINE, in contrast, takes a diachronic perspective and focuses on the reception process, arguing that people may continuously switch between processing goals during reception. Yet selective exposure scholars could adapt the logic of PINE by theoretically distinguishing between processing goals, and what one could call consistency appraisals (i.e., is the content consistent with my ideology). In case of a negative consistency appraisal, one would expect selective avoidance (see Bode et al., 2017).

Conclusion
We conclude that IE is more nuanced and leads to more complex learning effects than previously assumed. The mere scanning of incidentally encountered political information is by far less substantial compared to the processing of incidentally encountered content appraised as relevant. IE exposure to political information can deteriorate the learning outcomes of the information that was in line with the original processing goal. Understood as a dynamic concept, IE can thus explain various learning outcomes spanning attention and exposure processes, calling for a diachronic theoretical and methodological perspective.

Supporting Information
Additional Supporting Information may be found in the online version of this article.

Funding
This research was supported by the Austrian Science Fund (FWF), project P 31081-G29.

Conflict of interest
On behalf of all authors, the corresponding author states that there is no conflict of interest. 5. Respondents answering at least one of the two IE content recognition items correctly received the value 1. 6. We summed all first-level variables per respondent and ran a mediation analysis with participants instead of IE articles as unit of observation. 7. Even if we excluded respondents indicating that they "do not know" the place, the mean difference between groups remained highly significant.