What's in a tweet? Optimizing social media impressions

The Internet has facilitated instant dissemination of information. In particular, the use of social media has fundamentally changed how scientific literature is engaged with and consumed. In 2017, most urologists were using social media, with 53% of survey responders reporting a Twitter profile. These accounts are used for various reasons, including professional networking, socializing, attending virtual journal clubs and debating topics on the academic frontier. Journals frequently share newly published works on social media, and professional societies—such as the European Association of Urology (EAU)—have used Twitter to boost awareness vis-à-vis guideline recommendations. Visual abstracts (VAs) have been created to present information in an easily digestible, visually appealing manner for social media posts. A VA is a pictorial representation of the background, methodology and key findings of a study. VAs typically accompany social media posts highlighting newly published articles. Preliminary work in the British Journal of Urology International (BJUI) reported a difference in likes and retweets between posts containing a VA and those without it. Herein, we build on that preliminary evidence by evaluating engagement (total likes, retweets and replies) and examining the factors associated with successful tweets, including the presence of VAs. Our study aimed to determine the how the composition of a tweet affected engagement using the official BJUI Twitter account (@BJUIjournal). We analysed all tweets from November 2019 to October 2020, as this was the only era that contained VAs. In total, data on 421 tweets were extracted, 17 of which contained a VA. We compared engagement between tweets that contained a VA and those that did not. We found that the median engagement was 201 for those that contained a VA and 57 for that did not; the distributions in the two groups differed significantly (Mann–Whitney U = 777.0, p < 0.001). We then performed a linear regression to predict engagement using the following independent variables: presence of a VA, number of characters in the tweet, time of day (categorically defined as one of four 6-h periods during a day) and number of hashtags. Because engagement (dependent variable) was not normally distributed, it was transformed into its logarithmic equivalent: log10(engagement). The assumptions of linear regressions held true: absence of multicollinearity, normally distributed residuals, homoscedasticity and no autocorrelation (Durbin–Watson test: 2.03). We found that VAs were positively associated with engagement (β = 0.601; p < 0.001) and the remaining variables were nonsignificant (Table 1). The independent variable coefficients are interpreted as 1 unit increase results in a 10 fold increase in engagement. Thus, the presence of a VA increases the engagement fourfold (10). Overall, the model had an R of 0.09 and was significant (p < 0.001). Our study is unique as it investigates the composition and timing of a tweet to help optimize engagement. We find that VAs are the only significant factor associated with engagement. Though a modest 9% of the variation in engagement is explained by our model, we posit that this is substantial. Our model explains nearly 10% of a tweet’s engagement without consideration of the scientific content, especially considering there is only one statistically significant variable. The benefit we found in leveraging VAs aligns with previously reported data involving randomized trials and crossover trials investigating the role DOI: 10.1002/bco2.180