A novel coronavirus initially named 2019-nCoV emerged in Wuhan, China, and was formally reported to the World Health Organization (WHO) on December 31, 2019 [1-3]. Further scientific evidence soon unveiled the semblance of the 2019-nCoV’s genome sequence to a previous epidemic, the severe acute respiratory syndrome (SARS), a disease epidemic caused by SARS-CoV, which broke out in Foshan, China, in 2002 [4,5]. Some initial studies also identified similar features that were related to the Middle East respiratory syndrome (MERS) epidemic caused by MERS-CoV as the causative agent [6].

The outbreak was originally named 2019-nCoV on January 13, 2020, the same day that the first imported case outside China occurred in the Philippines and other countries [7]. The spread of 2019-nCoV continued across many countries, causing the WHO to declare the outbreak a pandemic [6]. The 2019-nCoV was later renamed SARS-CoV-2 and identified as the causative agent of COVID-19 in February 2020 [8,9]. The highly contagious COVID-19 spread rapidly globally and caught the world unprepared. With no adequately planned health communication strategies, panic ensued, while confirmed cases of infections and deaths from COVID-19 increased rapidly worldwide [3]. The public rushed to internet platforms to learn about the outbreak through Google searches, online news outlets, and social media platforms [10-15].

In March 2020, the WHO launched a free online introductory training course in different languages, including English, French, Spanish, and Chinese, to make the public aware of the contagious COVID-19 [6]. However, it is unclear how many people knew about or used the free training lessons about COVID-19 that the WHO had made available via its website [6]. Instead, several studies suggest that the public flocked to the internet to learn about SARS-CoV-2 and COVID-19 through web searches, online news outlets, and social media [16,17]. Analyzing how people search and navigate the World Wide Web and other internet platforms for health-related information can provide valuable insights into the health-related behavior of populations [18-20]. The public’s preference for online health information closely matches the field of infodemiology, a term that is a portmanteau of information and epidemiology. According to Eysenbach [18], the term is defined as the science of distribution and determinants of information in an electronic medium—specifically the internet—or in a population, with the aim to inform public health and public policy. Considering the global spread of COVID-19, using the internet to learn or gain information about the pandemic in this digital age is not surprising, as internet use has become pervasive worldwide [21,22]. Several studies have examined social media’s influence on what people learned and the appropriate behaviors toward misinformation and conspiracy theories [14,15]. Similarly, Sulyok et al [12] examined the impact of web searches on the confirmed cases of COVID-19 in Europe, while Neely et al [23] investigated information-seeking behaviors on social media regarding the pandemic.

Miller [14,15] identified political leaders’ failure in sensitizing the public as a motivating factor that pushed people to the internet as an alternative information source to learn about the COVID-19 pandemic. Misinformation had started flooding the web right from the initial stage of the emergence of the novel coronavirus, mainly from user-created content on social media [24]. Thus, as people turned to the web to search for information, there was limited nontechnical information for the nonexpert public about the coronavirus. Rather, people either got exposed to learning incorrect information about SARS-CoV-2 and COVID-19 or embraced fake news, misinformation, and conspiracy theories, with grave consequences [16]. Some of the unfounded misinformation included misconstruing COVID-19 as a “common cold” or as a hoax, which made people have a false sense of immunity, while others ignored any public health safety measures [11]. Similarly, the conspiracy theories propagated online included that COVID-19 was a bioweapon, that China intentionally released the virus to reduce the world population, and that 5G technology contributed to the fast spread of the pandemic. These beliefs initially led to the hoarding of essential goods as well as racial attacks against Chinese and other Asian people [14,15,25,26].

Other studies examined the role of social media and internet news outlets in generating misinformation, disinformation, fake news, and conspiracy theories about COVID-19 [25,27-29]. These studies tend to leave out the aspect of web searching, such as the use of Google searches, which constitutes a major channel through which the public obtain health-related information [30,31].

This paper undertakes the first empirical investigation using a web search to learn about SARS-CoV-2 and COVID-19 and people’s attitudes toward public health guidelines as expressed in the following research objectives:

These objectives are developed into research hypotheses in the sections that follow.

This study used an infodemiology approach to evaluate the use of web searching to learn about SARS-CoV-2 and COVID-19. As an area of scientific research, infodemiology is a method or technique designed to measure and track health information “demand” automatically (eg, by analyzing search queries) as well as “supply” on the internet [18,20,47]. The goal is to inform public health policy and practice. This study uses data from Google Trends, a freely available online resource that provides information on what was and is trending based on actual users’ Google queries [11,48,49].

Google Trends offers various search options, such as “Trending Searches” (ie, trending queries for daily search trends and real-time searches in a selected region) or “Year in Search” (ie, what was trending in a specific area in a particular year). Another option is to “Explore,” which allows an investigation of an area of interest based on keywords over the selected periods and regions. This study uses the “Explore” option, which allowed data to be retrieved directly from the Google Trends “Explore” page in comma-separated values format. It is also important to note that Google Trends data points are normalized to have a maximum value of 100 and a minimum value of 0. We normalized the data set by dividing each data point by total searches of the geography and the time range it represents to compare relative popularity. Note that the value 0 does not necessarily indicate no searches but represents a significantly low search volume that does not warrant inclusion in the results [47].

In this study, we captured the worldwide Google Trends data covering the initial months of the SARS-CoV-2 outbreak from January 1 to June 30, 2020 (ie, 182 daily data points for each search term). Regarding the search terms, this paper employed 25 keywords and phrases used by the public to learn about the COVID-19 pandemic through web searches. We identified the search keywords (Table 1) through a literature survey of published documents indexed on the Web of Science. Six search terms were related directly to the ongoing pandemic: “nCoV,” “2019-nCoV,” “SARS-CoV-2,” “COVID-19,” “pandemic,” and “coronavirus.” Another six keywords addressed previous viral or coronavirus epidemics: “SARS-CoV,” “SARS,” “MERS-CoV,” “MERS,” “virus,” and “influenza.” The third category of search terms represented public health safety measures that experts recommended as guidelines to limit the spread of COVID-19: “social distancing,” “wear a facial mask,” and “wash hands.” The final category of keywords represented misinformation and conspiracy theories, such as “China virus,” “common cold,” and “bioweapon” (Table 1 [2,3,6,7,9,11,14,15,25-27,42,46,50-60]).

Data analysis employed statistical trends and graphical visualization, correlation and regression, principal component analysis (PCA), and predictive models [12,56,62-64]. The statistical trends and analyses involved evaluating relationships among the listed variables using the statistical trends, including graphical display, correlation, and PCA, which helped determine the predictiveness of the learning attributes and learners’ actions toward public health guidelines. We used the JMP 15 package from SAS software (SAS Institute Inc) [65] for statistical analysis and Microsoft Excel 2019 to create the charts and graphs. We also used SPSS software (version 27; IBM Corp) to compute the correlation matrix and the PCA as well as statistical packages in R (version 4.0.5; The R Foundation) for the linear modeling. The evaluation helped establish the correlation between the study attributes.

The data analyzed in this study came from the Google Trends worldwide index covering the period from the initial outbreak of COVID-19 on January 1, 2020, up to June 30, 2020, when the pandemic became widely known [2]. The outbreak had been reported to the WHO’s office in China on December 31, 2019 [6]. The reason for focusing on the first 6 months of the pandemic was to capture what people learned during the early days of the outbreak as well as the possible impacts of what people learned through web searching on individuals’ attitudes toward public health safety measures.

To better understand the characteristics of Google Trends data, we have presented the summary statistics of the daily search index for each of the 25 keywords or search terms and phrases (Table S1 in Multimedia Appendix 1). The average normalized scores for the terms varied from 2.65 (“ingesting bleach”) to 39.75 (“SARS-CoV-2”), as shown in Table S1 in Multimedia Appendix 1.

The keywords employed to conduct web searches indicate what people learned about the COVID-19 pandemic [66]. As presented in Table 1, some of the search keywords addressed the novel coronavirus directly, while others examined misinformation and conspiracy theories.

Figure 1 presents the first category of web search terms that people used to learn about the COVID-19 pandemic. In the early period, most people used keywords and phrases that explain previous coronavirus epidemics, including “influenza,” “MERS,” “MERS-CoV,” “SARS,” and “virus.” Although scientists ruled out the past epidemics, the WHO officials highlighted those terms as examples of past coronavirus outbreaks during press briefings [6]. The use of those keywords in the web searches nosedived after the WHO formally named the novel coronavirus and the disease (ie, “nCoV,” “2019-nCoV,” “SARS-CoV-2,” and “COVID-19”). The coefficients of determination (R²) of the keywords are 0.37 for “SARS-CoV-2,” 0.36 for “COVID-19,” 0.27 for “influenza,” 0.24 for “2019-nCoV,” 0.18 for “nCoV,” 0.12 for “SARS,” and 0.11 for “SARS-CoV,” indicating the proportion of the variation in the search index over the period for the listed keywords. Similarly, the search terms “MERS,” “MERS-CoV,” “pandemic,” “virus,” “coronavirus,” “nCoV,” and “2019-nCoV” had R² values of less than 0.1.

The second category of keywords involved misinformation and conspiracy theories (Figure 2). The variation in the use of the terms was measured using coefficients of determination (R²), which were 0.15 for “China virus,” 0.10 for “common cold,” and 0.09 for “5G” or “@5G.” Most searches in the initial months of the outbreak used the keywords “common cold” (“cold2020”), “biological weapon,” and “China virus,” thus encouraging the misconception about SARS-CoV-2 as a “common cold,” a “biological weapon,” or a “China virus” [50]. Some studies explain that the purpose of releasing the coronavirus was to reduce the world population [14,25]. However, web searches using these terms fell continuously over time to a near-zero search index, while new words (“5G” and “COVID-19 hoax”) surfaced and increased significantly (Figure 2).

The third segment of the trend analysis involved web searches to learn about public health measures (Figure 3). The results show that there was little or no interest in learning about wearing a facial mask (“wear mask”) and maintaining social distancing (“social distancing”) at the start of the pandemic. But the trends changed quite quickly, recording a dramatic increase from a search index of 0 at the beginning of the outbreak to achieving a maximum search index of 100 in March and April 2020, as the pandemic spread worldwide. The coefficients of determination were as follows: “wear mask” (R²=0.56), “social distancing” (R²=0.13), and “quarantine” (R²=0.09). The increases, especially regarding “wear mask” and “social distancing,” were sustained for a long time.

The keywords identified above approximate what people learned about SARS-CoV-2 and COVID-19 through web searches. However, some of the search keywords existed before the ongoing pandemic, while some terms referred to previous coronavirus epidemics (eg, “SARS,” “SARS-CoV,” “MERS,” “MERS-CoV,” “influenza,” “virus,” “pandemic,” and “coronavirus”). It is plausible to argue that the search index for the pre-existing keywords listed represent purposes other than learning about COVID-19. Based on this assumption, we conducted a dependent two-sample t test to examine the difference in the mean search index of the pre-existing keywords in the previous years before the COVID-19 outbreak and during the ongoing pandemic. Figure 4 compares the mean search index before and after the outbreak for each keyword.

The null hypothesis was that the search indexes of each learning term—“SARS,” “SARS-CoV,” “MERS,” “MERS-CoV,” “influenza,” “virus,” “pandemic,” and “coronavirus”—before and after the outbreak of COVID-19 would be equal, and the alternative hypothesis was that they would be unequal. The results, as seen in Figure 4, show that the differences in the mean search indexes before and during the ongoing pandemic were more than 60% in all cases. Also, the P values were close to zero for all the variables. We rejected the null hypothesis and concluded that the significant differences in the mean search indexes of the variables were due to the ongoing COVID-19 pandemic. Unexpectedly, the mean search index for some pre-existing keywords (eg, “flu,” “influenza,” “SARS,” “pandemic,” and “virus”) declined during the pandemic.

PCA was employed to evaluate the underlying latent variable of the search terms that contributed to learning about COVID-19. Based on the scree plot and the elbow rule, we can limit the factors extracted to the first two principal components (Figure 5): the keywords that address previous epidemics (pastCoVepidemics) and the keywords that explain the ongoing pandemic (presCoVpandemic). A scree plot is a graphical representation of the percentage variability explained by each principal component.

The first two underlying components explained a total of 88.78% of the variation in learning terms, with the first (pastCoVepidemics) and second (presCoVpandemic) components determining about 65.2% and 23.58%, respectively, of the information about COVID-19 from the 12 search keywords (Table 2).

Linear combinations of the two components (pastCoVepidemics and presCoVpandemic) are as follows:

Table 3 shows the weights (loadings) of the terms for the two components. Note that we record loadings greater than 0.6 to combine only search keywords that have a high correlation with the component in the linear combinations.

We identified eight search keywords from the literature that denote misinformation and conspiracy theories (Table 1). We also performed a PCA to evaluate the search terms that fueled misinformation and conspiracy theories. The results, as seen in Table 4 and Figure 6, identified three principal components and their variabilities based on the elbow rule: principal component 1 (48.17%), principal component 2 (22.65%), and principal component 3 (14.03%).

The three components explained 84.85% of the variation in the search keywords under misinformation and conspiracy theories. The first component represented misinformation. We quantified the daily number of misinformation terms searched using the linear combination as follows:

The second and third components addressed two conspiracy theories (ConspTheory1 and ConspTheory2), which speculated that 5G technology contributes to the spreading of COVID-19, that COVID-19 is a “China virus” that was intentionally created and released, and that ingesting or injecting bleach can cure COVID-19 infection or kill the virus. The results present these variables in the separate components (Table 5) as follows:

This section investigates the impacts of what people learned through web searches on behaviors toward public health safety measures against COVID-19. Based on the elbow rule, the PCA identified two public health measure components, which we labeled as PubHealthMes1 and PubHealthMes2. The two components accounted for the variability in the search index of keywords used to learn about the public health measures against the spread of COVID-19 (Figure 7). The first and second components explained more than 75.4% and 18.7% of the variability, respectively; that is, the first two components explained about 94% of the variability (Table 6).

The first component, PubHealthMes1, includes the keywords “social distancing,” “wash hands,” “isolation,” and “quarantine.” In the second component, PubHealthMes2, the keyword “wear mask” explained 84.7% of the variability (Table 7):

This section presents further analysis that tested the hypothesis raised in the earlier section using built predictive models. The variables identified the linear combination of search keywords significantly correlated (loading >0.6) to the principal components discussed in the Results subsections above. Also, Tables 2 to 7 present the underlying latent variables of the 25 search terms used to learn about COVID-19, the misinformation and conspiracy theories, and the public health measures. As stated, Figure 8 shows the underlying variables.

Here, we examine how the underlying variables and the search terms impacted learning and behavior toward the public health measures: learning about COVID-19 (pastCoVepidemics and presCoVpandemic), misinformation and unproven or misleading assertions (misinformation), conspiracy theories (ConspTheory1 and ConspTheory2), and public health safety measures (PubHealthMes1 and PubHealthMes2).

The results show three essential highlights from the daily search index. First, the most popular search terms used at the initial outbreak of the pandemic in early January 2020 were terms representing misinformation and past epidemics. The search keywords that represented conspiracy theories were not used until May 2020. Also, the use of learning terms that directly explained COVID-19 (ie, presCoVpandemic) corresponded with the WHO’s naming and renaming of the coronavirus and the disease (ie, “2019-nCoV,” “SARS-CoV-2,” and “COVID-19”) in January, February, and March 2020.

We employed correlation analysis among the variables, scatterplots, and their histograms to examine the relationship between what people learned and their attitudes toward public health measures (Figure 9).

PubHealthMes1 has a robust positive relationship with pastCoVepidemics (r=0.80) and a moderate positive relationship with presCoVpandemic (r=0.73), which implies the effectiveness of learning keywords associated with past coronavirus epidemics (eg, “SARS,” “SARS-CoV,” “MERS,” “MERS-CoV,” and more) and the ongoing pandemic (eg, “2019-nCoV,” “SARS-CoV-2,” and “COVID-19”). Similarly, PubHealthMes2 (ie, wearing a facial mask) has a moderate positive relationship with presCoVpandemic (r=0.71). There is a strong association between actions taken and the information learned. Figure 9 shows a correlation matrix.

Given such a strong linear relationship between the search terms and people’s behaviors and actions, a multiple linear regression model seems acceptable as a predictive model. But the data failed the assumption of normality, as shown by the quantile-quantile (Q-Q) plots in Figure 10, A and B. A normal Q-Q plot helps to compare two probability distributions, by plotting the residuals against theoretical quantiles. Plots A and B in Figure 10 show that most residuals are not lying on the diagonal line; hence, the data are not normally distributed.

From Figure 9, we can observe that the PubHealthMes1 histogram suggests a right-skewed distribution, while the PubHealthMes2 histogram suggests a bimodal distribution. To describe and explain the relationship between the public health measures and the terms that contributed to learning about COVID-19, misinformation, and conspiracy theories, we used the log-linear predictive model, since the data are based on a number of occurrences or frequency and not normally distributed. Also, the log-linear model does not need to satisfy any assumptions, which we represent as follows:

where x_i is the covariate and β_i is the parameter to be estimated.

The data obtained through the PCA have been fitted to a log-linear regression model using RStudio tools from R (version 4.0.5; The R Foundation). The regression model obtained becomes the following:

The R² of the model is 0.93 and the P value is <.001. All the coefficients in this model are significant at an α level of 0. From the model, we see that there is a 1.10 percentage (100 × [exp^0.011 – 1]%) increase in the public health measure for every 1% increase in presCoVpandemic (ie, “2019-nCoV,” “SARS-CoV-2,” and “COVID-19”). This may be explained by links contained in recent COVID-19 pandemic articles on public health measures. For every 1% increase in past epidemic searches, there is a 0.25% increase in the searches of public health measures. Similarly, the regression model for PubHealthMes2 obtained becomes the following:

The R² of the new model is 0.50 and the P value is <.001. The two models show that both PubHealthMes1 and PubHealthMes2 are significantly predictive using the presCoVpandemic keywords. It shows that learning about the present pandemic creates an incentive for people to learn about public health measures, which we can approximate to a desired intention to comply with the measures.

This section examines the relationship between conspiracy theories, misinformation, and public health measures using correlation and predictive analyses. As discussed earlier, we recategorized the search terms representing misinformation and conspiracy theories into two principal components. Hypothesis 3A, as defined in the Research Hypotheses section, focuses on misinformation and public health guidelines. At the same time, Hypothesis 3B addresses the impact of conspiracy theories on people’s behavior toward the same safety measures.

Using correlation analysis to evaluate the association between conspiracy theories and public health measures, we observed a moderate positive linear association between ConspTheory2 and PubHealthMes2 (r=0.66) and a moderate linear relationship between ConspTheory1 and PubHealthMes1 (r=0.53). The log-linear analysis showed that conspiracy theories are not significant predictors of PubHealthMes1 (P=.62) but are for PubHealthMes2 (P=.008). Thus, the null hypothesis that there is no association between conspiracy theories (ie, ConspTheory1) and people’s responses to public health (ie, wearing facial masks to limit the spread of COVID-19) is rejected. We can conclude that conspiracy theories are predictive of people’s behaviors in wearing facial masks.

We also analyzed the relationship between misinformation and behavior toward public health measures. Although the correlation and the predictive analyses showed a moderate positive linear relationship between misinformation and PubHealthMes1 (r=0.83), it is not predictive. The relationship between the two variables is just mathematical but not causal. Despite a negligible negative linear relationship between misinformation and wearing a facial mask (r=0.11), the log-linear model shows that misinformation is not a significant predictor for both PubHealthMes1 (P=.27) and PubHealthMes2 (P=.13). Notwithstanding the strong linear relationship between web searches to learn about misinformation and public health measures, there is no sufficient evidence to reject the null hypothesis. We can conclude that there is no association between learning about COVID-19 misinformation and people’s response to public health guidelines.

We found that people used search keywords related to past coronavirus epidemics (pastCoVepidemics) and the ongoing pandemic (presCoVpandemic) to learn about SARS-CoV-2 and COVID-19. However, the attention accorded to the pandemic led to less focus on terms relating to perennial illnesses (eg, “common cold,” “flu,” and more). These results corroborate studies reporting the unintended positive consequences of COVID-19 leading to declines in cases of influenza, flu, and similar infections (eg, Soo et al [46]). Other learning terms employed were keywords that addressed the pandemic directly. The average search indexes for those keywords were 19.01 for “nCoV” and “2019-nCoV,” 39.75 for “SARS-CoV-2,” and 32.40 for “COVID-19.”

Studies examining learning by web searching emphasized the significance of the search terms or phrases on what the users intended to learn. A trending word on the web indicates what information people are interested in learning [11,61,66]. This study identified the 25 most-used keywords to learn about SARS-CoV-2 and COVID-19 through web searches.

Regarding the impacts of what people learned on their behavior toward public health measures, the PCA identified three latent variables, classified as misinformation, ConspTheory1, and ConspTheory2. Only ConspTheory1 (“@5G”) directly and significantly influenced people’s behavior toward public health measures (ie, PubHealthMes2 [“wear mask”]). The conspiracy that 5G technology enhances the easy spread of COVID-19 [14] highlights danger, which can cause people to take precautions. A different study [15] identified erroneous beliefs in the 5G conspiracy theory as leading to the hoarding of essential goods during the initial period of the SAR-CoV-2 outbreak. Although there was a high correlation between misinformation (Tables 4 and 5) and behavior toward public health measures, this was not statistically significant based on the web search index. Also, as the pandemic lingers, thereby causing severe health and social crises, strains in family relations, and economic and business losses, many people are becoming increasingly aware of COVID-19 dangers [21,57,67-69]. Through direct impacts or by experience, this can cause changes in people’s behavior irrespective of whether they believed the misinformation or not.

Internet platforms continue to play a significant role in health communication during the ongoing COVID-19 pandemic. Some studies attribute the increase in misinformation and conspiracy theories about COVID-19 in different countries to web searches, social media use, and online news media platforms that are used to learn about SARS-CoV-2 and COVID-19 [10,11,29]. However, most studies were anecdotal with no empirical evidence. Using Google Trends data, this study provides the first empirical evidence to this discourse. In the era of big data, the analysis of Google queries can be envisioned as a valuable tool for researchers to explore and predict human behavior, especially as studies suggest that online data can correlate with actual health data [70,71].

Infodemiology studies have their limitations too. While Google search keywords are short and easy to classify automatically, interpreting the terms semantically can be challenging. It is not clear why people are searching for these keywords. Furthermore, when using Google Trends, the sample is unknown and may not be representative, and individuals using the internet are not representative of the entire population. They are more likely to be younger, more educated, earn higher incomes, and reside in urban areas [18]. Individuals who are more likely to be severely affected by COVID-19 are not usually represented by this population [72,73]. Despite the identified limitations, previous studies suggest that web-based data provide valuable and valid results in exploring and predicting behavior and highly correlate with actual data [70,71]. Further, there are reports of rapid penetration of internet access and usage in different parts of the world, except for in regions with low internet penetration or countries with low scores in freedom of speech [22,74,75].

The results of this empirical infodemiology study showed that a good portion of the global population learned about the outbreak of SARS-CoV-2 and COVID-19 through web searches, particularly in the early period of the pandemic. The period covers the initial days, weeks, and months from the emergence of the novel coronavirus in January 2020 up to June 30, 2020, when the public became more aware of the pandemic, especially after the first wave [1].

The PCA showed that people used the web to learn about the ongoing COVID-19 pandemic in two ways, namely, using pastCoVepidemics keywords and using presCoVpandemic keywords. The use of pastCoVepidemics keywords in web searches nosedived as the WHO formally named the novel coronavirus and the disease (ie, “nCoV,” “2019-nCoV,” “SARS-CoV-2,” and “COVID-19”) and, therefore, as these terms became available. The trends analysis showed that web searches used to learn about COVID-19 followed a similar trend as learning about public health measures, implying that the more that people focused their attention on learning about SARS-CoV-2 and COVID-19, the more they also learned about public health measures, and vice versa. Interestingly, learning about the conspiracy theory (ConspTheory1) that 5G technology contributes to the fast global spread of COVID-19 is a predictor of people’s behavior toward public health measures (PubHealthMes2). This erroneous belief makes people take precautionary measures, such as wearing a facial mask, although borne out of fear [14,15]. The same studies using the survey method also identified the same 5G-related conspiracy theory as making people respond out of fear to take precautions. This factor contributed to stockpiling of goods in the early days of the pandemic [15]. This study is the first to examine what people learned through web searches and how these influence people’s social behavior toward public health safety guidelines.