Capturing Clicks: How the Chinese Government Uses Clickbait to Compete for Visibility

ABSTRACT

The proliferation of social media and digital technologies has made it necessary for governments to expand their focus beyond propaganda content in order to disseminate propaganda effectively. We identify a strategy of using clickbait to increase the visibility of political propaganda. We show that such a strategy is used across China by combining ethnography with a computational analysis of a novel dataset of the titles of 197,303 propaganda posts made by 213 Chinese city-level governments on WeChat. We find that Chinese propagandists face intense pressures to demonstrate their effectiveness on social media because their work is heavily quantified–measured, analyzed, and ranked–with metrics such as views and likes. Propagandists use both clickbait and non-propaganda content (e.g., lifestyle tips) to capture clicks, but rely more heavily on clickbait because it does not decrease space available for political propaganda. Government propagandists use clickbait at a rate commensurate with commercial and celebrity social media accounts. The use of clickbait is associated with more views and likes, as well as greater reach of government propaganda outlets and messages. These results reveal how the advertising-based business model and affordances of social media influence political propaganda and how government strategies to control information are moving beyond censorship, propaganda, and disinformation.

KEYWORDS:

• Propaganda
• clickbait
• social media
• quantification
• China
• WeChat

Acknowledgments

Our thanks to Mengmeng Guo, Yueyi Li, Zeyu Li, Yuqian Lin, Yitian Liu, Sujing Lyu, Lu Tian, Chenyi Zhang, Jing Zhang, Bowen Zhao and others for superb research assistance; Erin Baggott, Angèle Christin, Charles Crabtree, King-wa Fu, Holger Kern, Jon Krosnick, Carlo Horz, Dan Mattingly, Tamar Mitts, Rich Nielsen, Molly Roberts, Wei Wei, and others for helpful comments and suggestions; and to the NSF (Award #1831481) and the Stanford Graduate Fellowship for research support. Replication materials can be found at https://doi.org/10.7910/DVN/TALJOT.

Disclosure Statement

No potential conflict of interest was reported by the authors.

Data Availability Statement

The data described in this article are openly available in the Open Science Framework at https://doi.org/10.7910/DVN/TALJOT.

Open Scholarship

This article has earned the Center for Open Science badge for Open Materials. The materials are openly accessible at https://doi.org/10.7910/DVN/TALJOT.

Notes

1. There are many different definitions of propaganda. Parry-Giles (2002) describes propaganda during the Cold War as “strategically devised messages that are disseminated to masses of people by an institution for the purpose of generating action benefiting its source” (xxvi). Pratkanis and Turner (1996) describe propaganda as “attempts to move a recipient to a predetermined point of view” (190). Carey (1997) describes corporate propaganda as “communication where the form and content is selected with the single-minded purposes of bringing some target audience to adopt attitudes and beliefs chosen in advance by the sponsors of the communication” (20). Jowett and O’Donnell (2018) define propaganda as “the deliberate, systematic attempt to shape perception, manipulate cognition, and direct behavior to achieve a response that furthers the desired intent of the propagandist” (6). The common theme among these definitions is that propaganda consists of messages aimed at shaping the perception, cognition, and behavior of receivers according to the desires of the propagandist. We focus on political propaganda, thus narrowing this definition to messages aimed at shaping political opinion, preferences, and behaviors.

2. The concept of affordance comes from ecological psychology (Gaver, 1991; Gibson, 1977; McGrenere & Ho, 2000), and we use it to refer to the possibilities suggested by the material characteristics of technology that influence human behavior and dynamics. A wide range of researchers have used affordance-based approaches to study social media (boyd, 2011; Chadwick, 2007, 2017; Ellison et al., 2011; Resnick, 2002; Treem & Leonardi, 2013; Wellman, 2001).

3. In the mid-1990s, 200 to 250 million people tuned in each night to the state television broadcast (Economist, 2016), and in 2019, only 55 to 60 million did so (from http://www.csm.com.cn/cpfw/(Accessed April 19, 2019). The number of Internet users in China increased fifty-fold between 2000 and 2019, reaching 854 million in 2019 (see https://bit.ly/31Bk0OP and https://bit.ly/2uoOKGJ (Accessed January, 2020)).

4. This system includes website blocking in the form of the Great Firewall, where people located within China cannot access websites such as Google, Facebook, and the New York Times; search filtering where certain results do not appear on search engines such as Baidu and Bing because they are deemed objectionable by the government; keyword blocking where content containing certain words or phrases cannot be publicly posted; and removal of posts that have already appeared online.

5. See https://zd.net/2SNaXEk (Accessed Sept. 26, 2019).

6. These downsides are less likely if authoritarian governments are spreading disinformation and manipulating metrics of engagement outside of their borders.

7. See https://nyti.ms/3cuH3PO and https://nyti.ms/2RKaFk4 (Accessed March 1, 2020).

8. Whether or not the third pathway is in effect depends on the specific technical features of a social media platform.

9. See http://www.xinhuanet.com//2018-08/26/c_1123331382.htm (Accessed Sept. 28. 2019).

10. See “Government Website and Government New Media Inspection Indicators” (政府网站与政务新媒体检查指标) and “Annual Assessment Indicators for Government Website and Government New Media Supervision” (政府网站与政务新媒体监管工作年度考核指标) https://bit.ly/2nwpKtt (Accessed Sept. 28, 2019).

11. In 2013, the State Council issued “Opinions of the General Office of the State Council on Further Strengthening Government Information Disclosure in Response to Social Concerns and Enhancing Government Credibility” (国务院办公厅关于进一步加强政府信息公开回应社会关切提升政府公信力的意见), which required all local governments to establish social media accounts (see http://www.gov.cn/zhengce/content/2013-10/18/content_1219.htm (Accessed Sept. 28, 2019)). In 2016, Xi Jinping stressed the importance of social media for “two-way interaction” see http://cpc.people.com.cn/n1/2016/1010/c64094-28763907.html (Accessed Jan, 2020).

12. See 2018 State Council “Opinions on Promoting the Healthy and Orderly Development of New Media in Government Affairs” (关于推进政务新媒体健康有序发展的意见).

13. When policy documents refer to social media, they are referring to WeChat, Weibo, and government-specific applications. This definition of social media is described in“Opinions On Comprehensively Promoting the Work of Open Government Affairs” (关于全面推进政务公开工作的意见) and “Key Points of Government Affairs Work in 2018” (2018年政务公开工作要点).

14. See https://www.chinainternetwatch.com/28566/weibo-fiscal-2018/(Accessed Sept. 28, 2019).

15. See https://support.weixin.qq.com/cgi-bin/mmsupport-bin/getopendays (Accessed Sept. 28, 2019) and http://www.cioall.com/uploads/f2019080910242398145.pdf (Accessed Sept. 28, 2019).

16. See https://bit.ly/2m0S46N (Accessed Sept. 28, 2019).

17. See https://bit.ly/2lWMHp0 (Accessed Sept. 28, 2019) and https://bit.ly/2nydGI2 (Accessed Sept. 28, 2019).

18. It is also possible to encounter government content through the Top Stories feature WeChat launched in December 2018 and March 2019; however, because of the relative newness of this feature, we do not illustrate it here.

19. Moments, or Friend’s Circle on WeChat, allows users to post and browse the updates from their friends (Chen et al., 2018).

20. Teams include those that were part of the local propaganda department, the local government information office, and local media outlets. The physical offices of these teams included those both within and outside of government buildings.

21. While it is possible that our presence could alter the day-to-day work of the teams we observed, we think this is unlikely because the teams had tasks they had to complete and would not have allowed us to be present had it interfered with their work.

22. All government Official Accounts contain the term fabu (发布), which literally means “to release” or “to promulgate,” in their account name. There are 333 city-level administrations in China; 27% of cities do not have a city-government Official Account and 9% of cities have accounts but are not certified as a government WeChat Official Account, or have not posted since April 2018 (or ever). The cities that do not have city-government Official Accounts typically have lower-level county-government WeChat accounts or local department-led accounts. By examining the account information of city accounts that could be collected, we know that 120 accounts (56%) are affiliated with the city propaganda department, 43 accounts (20%) are affiliated with the information office of the local government (政府新闻办公室), and the remaining accounts are associated with other government administrative offices or local media outlets.

23. See Appendix for the plot of account creation.

24. See https://weixin.sogou.com/. In scraping process, we followed the terms of services of weixin.sogou.com and adhered to requirements outlined in their robots.txt file. See Appendix for more details on this data collection process.

25. A constraint of our data collection method is that we are not able to collect, and hence analyze, the thumbnail image associated with the title, which is also shown to users before their click. The thumbnail image urls shown on Sogou Weixin redirects to links hosted on WeChat. Accessing these urls to download the image triggers WeChat’s anti-spider mechanisms. As such, our conclusions are limited to the effects of text on user behavior.

26. We used both the http://qinwenfeng.com/jiebaR JiebaR package (Qin & Wu, 2019) and the Chinese Word Segmenter (Tseng, Chang, et al., 2005). JiebaR produced better segmentation based on our qualitative evaluation of a sample of segmented titles. We do not remove infrequent words because trimming leads to the removal of names of government organizations, officials, and locations, which are important to the topics.

27. Content related to motivational message combines messages unrelated to politics with messages reflective of CCP’s emphasis on “positive energy” (正能量).

28. We excluded sampled titles that had no identifiable message.

29. Wh-pronouns in Chinese are who, whose, whom, which, what, where, and how.

30. We are worried about false positives because some of the general nouns in the dictionary could be used in other grammatical contexts. For example, “work” is a general noun related to the government (e.g., “government work was emphasized”) but “work” could also be used as the verb (e.g., “He is going to work on the project”).

31. We use the Stanford POS Tagger; see Toutanova et al. (2003); Tseng, Jurafsky, et al. (2005).

32. Examples of a false positive hyperbolic word would be a word that is part of a book title; an example of a false positive for slang would be the slang term being used in a literal sense.

33. See https://www.digitaling.com/articles/32729.html (Accessed Sept. 10, 2019).

34. If we exlude clickbait titles which are classified as such because they contain one puncuation mark, 56% of government WeChat titles contain clickbait. If we exclude all clickbait titles classified as such because they contain any punctuation mark, 49% of government WeChat titles contain clickbait.

35. We selected the three most popular accounts with different types of ownership and affiliation (influencer, company, and commercial media) based on WeChat Official Account rankings. One account run by a well-known investment expert, Hao Zhan. Another is a healthcare and wellness account managed by Dingxiangyuan Company, and the third account is that of the magazine Lifeweek. See Appendix for more details on account selection and characteristics.

36. The presence of this low-arousal content may relate to cultural differences in emotion (Lim, 2016). An example is a title describing a disabled man who made a living playing the piano with his feet: “Ren Jianwei, the boy who lost his arms from Jun county, gains a fortunate life with his feet.” (浚县小伙儿任建伟失去双臂,却用双脚奏出幸福人生).

37. We cannot assess the third algorithmic pathway with our current observational data because assessing this pathway would require gathering usage and content data from WeChat users.

38. This time lag ensures that the views and likes data of all of the titles are comparable. Most posts gain their views and likes within a week or two of posting. Our collection of views and likes data takes place 4 months after the last post in our dataset. We found views data for 58,411 observations of our sampled data, and we found likes data for 16,384 observations. We have much fewer observations for likes because between December 2018 and March 2019, WeChat replaced the previous like feature “Praise” (赞) with the new like feature “Wow” (在看) through two software updates. When the “Wow” feature was implemented, previous “Praise” data was no longer displayed on WeChat. As a result of change in WeChat’s technical features, our data on likes consists of “Wows” on titles posted between March 14, 2019 and May 25, 2019.

39. We use a negative binomial regression because we have count data and views are overdispersed (see Appendix). The AIC score of the negative binomial model is 231 times smaller than that of the Poisson model.

40. We use a zero-inflated negative binominal because the “Wow” data are overdispersed count data and many posts have no “Wow’s” because it was a new feature. There are no substantive changes if we use a negative binominal regression. And the AIC score of the zero-inflated negative binomial model is 5% lower than the normal negative binomial model.

41. WCI is published by Qingbo Big Data Corporation (清博大数据). WCI rankings are available on Gsdata.cn. WCI rankings are also sometimes posted in the media. For example, Foshan city’s WeChat rank was posted by The Paper at http://www.thepaper.cn/newsDetail_forward_4106112. We collected WCI scores on September 5, 2019. In total, we found WCI scores for 199 cities in our dataset.

42. For examples, see https://bit.ly/2o31Ohz and https://bit.ly/2msL96r.

43. For more details, see http://www.gsdata.cn/site/usage.

44. We normalized WCI data.

45. These city controls come from the 2018 China City Statistical Yearbook published by the National Bureau of Statistics of China, which provides complete data on 185 cities; see http://www.stats.gov.cn/tjsj/ndsj/2018/indexeh.htm.

Additional information

Funding

This work was supported by the National Science Foundation [1831481].

Notes on contributors

Yingdan Lu

Yingdan Lu is a Ph.D. Candidate in the Department of Communication at Stanford University.

Jennifer Pan

Jennifer Pan is an Assistant Professor in the Department of Communication at Stanford University.