A Data-Driven Look at Two Decades of Super Bowl Ad Strategies

36315 Final Group Project

Author

Stephanie Kim, Christina Tang, Bradley Xu

Introduction

There are few marketing venues and corporations with large-scale budgets like the Super Bowl. A single 30-second advertisement in the 2024 Super Bowl cost millions of dollars, yet creative decisions revolving around what type of advertisement to display remains driven by intuition: Should we hire a celebrity? Tell a joke? Raise morale and boost American pride? To determine the answer to these questions, we will analyze which creative advertisement choices actually drive engagement with the Super Bowl, and how those choices have evolved over time.

Dataset Description

The dataset used in this analysis is the “Super Bowl Commercials” dataset from the #TidyTuesday project on GitHub. Each row of the dataset represents one commercial aired during the Super Bowl. There are 247 commercials total, covering the 10 brands with the most Super Bowl appearances between 2000 and 2020.

The key variables that span the columns of the dataset are:

Ad Trait Variables (categorical: TRUE/FALSE):

  • funny: Ad contains humor.

  • show_product_quickly: Product is shown early in the ad.

  • patriotic: Ad contains patriotic themes.

  • celebrity: Ad features a celebrity.

  • danger: Ad contains dangerous situations.

  • animals: Ad contains animals.

  • use_sex: Ad uses sexual appeal.

Engagement Metric Variables (quantitative):

  • like_count: Number of YouTube likes on the ad video.

  • dislike_count: Number of YouTube dislikes on the ad video.

  • comment_count: Number of YouTube comments on the ad video.

Other Variables:

  • brand: The brand being advertised.

  • year: The year the ad aired during the Super Bowl.

Research Questions

We have analyzed the following three research questions in our report:

  1. How does the presence of certain ad traits impact Super Bowl ads with the most engagement?
  2. What types of Super Bowl ads have changed the most over time?
  3. Do brand commercials with celebrities consistently outperform ones without celebrities?

Engagement is a key metric marketers use to justify the multi-million dollar cost of a Super Bowl advertisement spot. Identifying the ad traits that significantly boost comment engagement will help advertisers understand what elements and traits of an ad bring in the most viewers. Further, tracing which ad traits have risen in popularity and which have fallen over time can reveal how brand strategy adapts to shifting audience preferences, societal pressures, and social media dynamics. Finally, using celebrity features in ads command large appearance fees; so, understanding whether their presence yields measurably higher engagement can guide budget allocations and talent negotiations. Overall, analyzing these three research questions will provide insights that will inform creative planning and decision-making for future Super Bowl campaigns.

Question 1

How does the presence of certain ad traits impact Super Bowl ads with the most engagement?

We are interested in identifying which individual ad traits — funny, show_product_quickly, patriotic, celebrity, danger, animals, use_sex — tend to correspond to the highest engagement on YouTube, measured by comment_count. To observe this relationship, we first examine a violin plot that visualizes the distribution of comment counts for ads containing each ad trait among the top 10 most-commented ads. We then explore how these traits co-occur using a correlogram, and analyze any notable relationships between traits in the most engaging commercials.

We will first look at the violin plot that highlights how the presence of each ad trait relates to YouTube comment counts.

Below is the violin plot:

From the violin plot, we can observe that ads featuring show_product_quickly and funny traits both have tight, left-skewed distributions with low medians, indicating that these traits are associated with lower levels of engagement. The comment counts for these traits are densely concentrated in the lower ranges. Ads featuring celebrity and animals show a slightly wider spread, but these traits still lean left with moderate engagement as their medians remain relatively low.

In contrast, the danger trait exhibits a wide and even spread across a broad range of comment counts, with a median near the center. This suggests that danger-themed ads are frequent among top-commented commercials and produce a variety of engagement outcomes. The patriotic trait stands out with a right-leaning distribution and the highest median comment count of all traits. This indicates that patriotic ads consistently drive the strongest viewer engagement.

Overall, traits like show_product_quickly, funny, celebrity, and animals are associated with lower or moderate engagement, while danger and patriotic ads achieve higher and more impactful levels of viewer response.

While the violin plot reveals which individual traits are most associated with higher engagement, it does not show how these traits might overlap within the same advertisements. To explore relationships between traits, we examined a correlogram of ad traits among the top 10 most-commented commercials. Before calculating the correlation matrix, we removed any traits that showed no variation across these top-commented ads since they cannot meaningfully contribute to the analysis.

Below is the correlogram:

The correlogram reveals several notable relationships between the traits. A strong positive correlation is observed between funny and show_product_quickly (0.8), suggesting that funny ads often also feature their products early. A moderate positive correlation exists between show_product_quickly and animals (0.43), suggesting that ads that show products quickly sometimes include animals. Similarly, danger and animals are moderately correlated (0.36), implying that danger-themed ads often include animals as well. Patriotic and celebrity show signs of moderate correlation (0.36), suggesting that patriotic ads frequently feature celebrities.

Several strong negative correlations are also present. Patriotic and show_product_quickly exhibit a perfect negative correlation (-1), meaning that patriotic ads never focus on displaying the product early. Patriotic and funny show a strong negative correlation (-0.8), indicating that patriotic ads almost never use humor. A moderate negative correlation between celebrity and animals (-0.53) suggests that ads tend to feature either celebrities or animals, but rarely both in the same ad.

Other relationships include moderate negative correlations between celebrity and show_product_quickly (-0.36), and between patriotic and animals (-0.43). Very weak positive or negative correlations, such as between funny and animals (0.09) or between show_product_quickly and danger (0.09), suggest that those traits do not have strong relationships.

Overall, these patterns in the correlogram suggest that successful Super Bowl ads often combine certain traits strategically, such as humor and early product showcasing, while keeping other traits, such as patriotism and humor, deliberately separate.

All in all, the results of the violin plot and correlogram show that certain creative choices are much more effective than others in driving audience engagement. Traits like patriotism and danger are consistently associated with higher comment counts, whereas humor and early product reveals, although common, tend to generate more modest engagement. Furthermore, the way traits combine within an ad, such as through the strong pairing of humor and quick product showcasing or the deliberate separation of patriotism and humor, reflects strategic creative decisions. These insights suggest that advertisers aiming to maximize Super Bowl ad engagement should carefully select not only individual traits, but also how different creative elements are combined.

Question 2

What types of Super Bowl ads have changed the most over time?

Now we are interested in analyzing which individual ad features – funny, show_product_quickly, patriotic, celebrity, danger, animals, and use_sex – have exhibited the largest shifts in prevalence from 2000 to 2020. This points us toward the most significant changes and pivots in Super Bowl advertising. To observe this change in ad feature over time, we can first look at a time series line plot, analyze the significance of certain ad traits in the time series using linear regression analysis, then look at a PCA analysis, scree plot, and biplot, and test whether there indeed exists a shift in ad features in earlier years versus later years.

We will first look at a time series line plot that highlights the change in trend of two key ad features from 2000 to 2020:

From this time series plot, we can see that each line shows the change in the proportion of ads containing a specific trait in a given Super Bowl year. Two key ad traits were highlighted – celebrity and funny. Upon initial observation of the plot, these two ad traits appeared to have the most dynamic changes in their trends over time. Other ad traits like show_product_quickly, animals, or danger seemed to remain relatively stagnant through the given time period.

In 2000, nearly 100% of the sampled Super Bowl ads that year used humor or funny content. However, from 2000 to 2006, the use of humor in advertisements suddenly declined, dropping from 100% in 2005 to nearly 60% in 2006. After 2006, the decline continues, reaching a drastic low of around 20% in 2017-18 – i.e. by 2018, only 20% of the sampled Super Bowl ads in that year were using funny content to engage with viewers. By 2020, the proportion of ads using funny content managed to reach 55%, but the magnitude of the total decline in proportion of Super Bowl ads that use humor was so significant that it overshadows any other ad feature that has a negative trend.

Conversely, nearly 0% of the sampled Super Bowl ads in 2000 used celebrities or patriotic content. The use of celebrities in Super Bowl ads appears dynamic through the years, fluctuating between 20% to around 50% in earlier years (i.e. 2000 to 2010) then hitting another low of 0% in 2012. However, after 2012 the use of celebrities in Super Bowl ads began increasing, hitting a high of around 64% in 2016-17, declining again to around 27% in 2018-19, followed closely by a record-high of 100% in 2020 – 100% of Super Bowl ads in 2020 in the data set used celebrities. By 2020, there was a universal adoption of using celebrities in Super Bowl advertisements, yielding a magnitude of change in proportion of Super Bowl ads that use celebrities of 100%.

Ad features like show_product_quickly remain around the 75% line from 2000 to 2020, which suggests that tactics like showing the product faster are commonly used in ads throughout time. Other features like animals, danger, and use_sex fluctuate around low baselines without much obvious direction. The muted color palette for these ad features ensure that these trends are not nearly as significant as the trends for celebrity and funny.

Statistical Analysis

To test whether the changes in trends of funny and celebrity are significant over time, we ran a linear regression analysis of each trait on year (scaled in calendar units). The following shows the summary statistics of funny linearly regressed on year:

 
Call:
lm(formula = funny ~ year, data = youtube)

Residuals:
    Min      1Q  Median      3Q     Max 
-0.9235 -0.4959  0.1771  0.3281  0.5544 

Coefficients:
             Estimate Std. Error t value Pr(>|t|)    
(Intercept) 51.253241   9.605432   5.336 2.16e-07 ***
year        -0.025152   0.004778  -5.264 3.08e-07 ***

---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 0.4392 on 245 degrees of freedom
Multiple R-squared:  0.1016,    Adjusted R-squared:  0.09794 
F-statistic: 27.71 on 1 and 245 DF,  p-value: 3.083e-07
 

The following shows the summary statistics of funny linearly regressed on year:

 
Call:
lm(formula = celebrity ~ year, data = youtube)

Residuals:
    Min      1Q  Median      3Q     Max 
-0.4006 -0.3107 -0.2465  0.5930  0.8305 

Coefficients:
              Estimate Std. Error t value Pr(>|t|)   
(Intercept) -25.521289   9.799428  -2.604  0.00977 **
year          0.012839   0.004875   2.634  0.00898 **
---

Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 0.4481 on 245 degrees of freedom
Multiple R-squared:  0.02753,   Adjusted R-squared:  0.02356 
F-statistic: 6.936 on 1 and 245 DF,  p-value: 0.008983
 

Given the outputs of the regression analyses, we see that the estimated coefficient of funny on year is -0.025 and the estimated coefficient of celebrity on year is 0.013. This means that the proportion of Super Bowl ads using humor falls around 2.5% annually, whereas ads using celebrities increases around 1.3% annually. The p-values for each of these estimated coefficients are approximately zero – meaning that the negative trend in ads using funny content over time and the positive trend in ads using celebrities over time is statistically significant. These highly significant slopes formally confirm the visual trends seen in the time series line plot.

As mentioned above, ad traits like show_product_quickly, animal, and danger visually appeared to be relatively stagnant over time with few major fluctuations in trends. However, ad features like using patriotism or using sexual appeal were a bit ambiguous in that their trends appeared far less dynamic than celebrity or funny, but still appeared to have a positive or negative change in trend over time. Again, we use a linear regression to test whether the changes in trends of patriotic and use_sex is significant.

 
Call:
lm(formula = patriotic ~ year, data = youtube)

Residuals:
     Min       1Q   Median       3Q      Max 
-0.31482 -0.20862 -0.13276 -0.04173  0.97344 

Coefficients:
              Estimate Std. Error t value Pr(>|t|)    
(Intercept) -30.331216   7.933904  -3.823 0.000167 ***
year          0.015171   0.003947   3.844 0.000154 ***

---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 0.3628 on 245 degrees of freedom
Multiple R-squared:  0.05688,   Adjusted R-squared:  0.05303 
F-statistic: 14.78 on 1 and 245 DF,  p-value: 0.0001544
 
 
Call:
lm(formula = use_sex ~ year, data = youtube)

Residuals:
    Min      1Q  Median      3Q     Max 
-0.4524 -0.2888 -0.1798  0.5658  0.8929 

Coefficients:
             Estimate Std. Error t value Pr(>|t|)    
(Intercept) 36.797401   9.431568   3.902 0.000124 ***
year        -0.018173   0.004692  -3.873 0.000138 ***
---

Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Residual standard error: 0.4313 on 245 degrees of freedom
Multiple R-squared:  0.0577,    Adjusted R-squared:  0.05385 
F-statistic:    15 on 1 and 245 DF,  p-value: 0.0001379
 

From the output of the linear regression models, we can see that the estimated coefficient of patriotic on year is 0.015 and the estimated coefficient of use_sex on year is -0.018. This means that the proportion of Super Bowl ads using patriotic content increases by 1.5% annually, whereas the proportion of Super Bowl ads using sexual appeal falls by 1.8% annually. The p-values for these coefficients are 0.00015 and 0.00014, respectively, which indicate that the positive trend over time in ads using patriotic content is statistically significant and the negative trend over time in ads using sexual appeal is statistically significant. Thus, we can conclude here that Super Bowl ads using patriotic features and ads using sexual appeal have significant changes in their trends over time.

Our time series graph pinpoints how often each ad trait appears in every Super Bowl year, but does not revel exactly when those appearances are most concentrated, or whether the timing patterns of different ad traits overlap. To address this concern, we analyzed a density plot of Super Bowl year by ad trait.

The density plot shows the distribution of Super Bowl years for every ad feature. Ridges that are shifted to the right (i.e. have a left-skew) represent ad traits that became popular in more recent years, whereas curves shifted towards the left (i.e. have a right-skew) represent ad traits that were more popular in earlier years. From the plot, we can see that patriotic is almost exclusively a later-decade popular ad trait, with a single peak at around 2016-18. celebrity also seems like it is shifted slightly more towards the right, with two peaks – one at around 2005 (which aligns with the time series graph, with the highest peak in earlier years being in 2005) which indicates early popularity, but another peak at around 2018, where the true population surge began. We can also see that funny has a broad curve, but it is dominantly shifted towards the left. This suggests that humor dominated early years and lost popularity gradually. use_sex also appears to be shifted left, peaking at around 2002 and having a density of nearly zero after 2015 – this indicates that the popularity of ads using sexual appeal was dominant in earlier years, but receded steadily and virtually disappeared more recently.

Across two decades of Super Bowl advertising, the creative decisions that go into creating Super Bowl ads has swung decisively from light-hearted humor and the use of sexual appeal towards patriotism and celebrity influence. Humor and sexual appeal now appear in less than half of today’s Super Bowl commercials, falling at statistically significant rates of around 2-2.5% per year. In contrast, patriotic themes and celebrity appearances have climbed steadily – patriotism rising the most – making these essential ad features of the late-2010s. The density plot confirms that these ad traits rising in popularity remain unchanged – the use of patriotism and celebrities in Super Bowl ads has indeed gained more popularity in recent years. We can conclude that there is a clear strategic pivot – recent Super Bowl ads increasingly seek emotional resonance with viewers through national identity and recognizable celebrity faces, leaving funny and provocative approaches in the past.

Question 3

Do brand commercials with celebrities consistently outperform ones without celebrities?

It is important to properly analyze Super Bowl advertisement success, especially the ones with celebrity appearances since it is a characteristic that has a significant investment commitment. There are also many other effects that we can go into.

Financial Factors

A Super Bowl commercial slot is the most expensive advertising real estate available in television. Companies will invest millions of dollars for the short seconds of airtime and anything that can increase the positive effects of the commercial should not be taken lightly. Celebrities in particular are a large part of marketing throughout the TV era and seeing the effects and justifying their appearance is very important for maximizing our advertising efficiency. This way brands can decide whether it is necessary to commit to a celebrity appearance and resource allocations for these high stake commercials.

Cultural Significance and Consumer Psychology

There is also an element of cultural significance and consumer psychology that we can analyze from the celebrity influence. By looking into the way that these celebrity endorsements have done across the years we can track a consumer’s interest in fame and authority. This research helps us understand the ways that celebrities are viewed and we can analyze society as a whole to see if this is a trend that is taking place. This type of analysis can be conducted across the years and we can see if the culture has shifted in regards to the impact of celebrity advertising.

Overall the insights that we gather from this research will allow us to be able to make industry decisions regarding whether it is needed to commit to a celebrity appearance in order to maximize our return on investment.

Analysis metric:

We decided to create a metric as the like count to dislike count ratio in order to properly measure the effect we are trying to capture. This ratio properly captures the audience sentiment more than a simple statistic such as the total view count. A statistic like views would only be able to capture the reach of the commercial, but we would not be able to see how the consumers actually receive the commercial. Many commercials in the super bowl have ended up gaining significant reach but receiving negative reactions. By looking into the like to dislike ratio we can see how the audience is responding while normalizing our data. Some of the brands have significantly larger audiences for their ads on YouTube and ratio allows us to compare different size brands vs a simple like count.

For our first graphic we decided to use a side-by-side bar graph:

From this bar graph we can see that across the board, all brands that have had celebrities featured in their commercials have done significantly better than those that do not. The standout difference is with Budweiser where celebrity impact shows a ratio 15x higher for commercials with a celebrity. We can see that for brands like Hyundai and Kia there are also substantial improvements.

There is an important thing to note that E-Trade has never had a commercial with a celebrity. This is interesting for E-Trade and the ideology they have gone with for their Super Bowl commercials. They also do have a relatively high non-celebrity like-to-dislike ratio. For some of the larger brands, however, we can see that using a celebrity shows only a marginal improvement. This makes sense as the brand itself is already so large there may not be a need for celebrity appearances. Brands like the NFL and Bud Light are already so large that there is really not much need to add more fame to the brand.

We created this time series visualization to track like-to-dislike ratios for Super Bowl commercials from 2000 to 2020. We also developed a faceted version of this graph to facilitate brand-specific analysis, allowing us to examine performance patterns more precisely. The key insights from our analysis reveal varied celebrity impact across different brands:

Some of the key insights we can observe are the varied celebrity impacts across different brands. The faceted graphs allow us to examine each brand more closely.

For Kia, celebrity commercials consistently outperform non-celebrity ads, with celebrity ads reaching ratios of 80 in 2015-2016 compared to non-celebrity ads that is around 20.

Hyundai has an interesting pattern as the commercials with the celebrities were generally performing better than the ones without but there was a spike in the year of 2014 with around a ratio of 70 for the non-celebrity commercial.

Pepsi demonstrates the most volatility, with a dramatic spike for celebrity ads reaching approximately 80 in 2002, followed by inconsistent performance.

We can see that E-Trade stayed rather constant in their commercial performance as they never opted to get a celebrity appearance in their commercial.

Toyota showed the have the most preference towards non-celebrity commercial and had great results as well. They peaked around a ratio of 80 for their non celebrity commercials around 2012.

These patterns indicate that while celebrity presence can boost commercial performance for some brands, but there are definitely other factors coming into play. This could be something that is specific towards each individual brand and there is a need to tailor the use of celebrities towards each brand.

Conclusion

From our analysis of Super Bowl advertisements from 2000 to 2020, we were able to reveal many significant and important insights for brands and their marketing teams. Our findings demonstrated the key ad traits that would drive the viewer engagement among the comments of the YouTube videos. Humor and celebrities appeared to be the most powerful influential features. There is also an extreme evolution in advertisement strategy over the decades changing from humor and sexual content towards using celebrities in the later years. There are also patriotic themes that emerged from the more recent times. These changes can be categorized as a cultural change and the differences in viewer preference with the new age that has come. The development of social media and growth is truly so impactful towards these viewers. Finally our brand specific analysis confirms our assumptions on the celebrity culture influence on reception from the audience. Many brands saw a significant uptick in audience reception from their celebrity commercials suggesting the importance of contemporary Super Bowl advertisers investing in celebrity usage.

Limitations and Future Research

There were several limitations that we found throughout this study that could be points to tackle in future research. In regards to our dataset that we were provided, certain statistics were missing for some of the data points which made it difficult to capture the true results. Also the data that we were given was using the YouTube video of these commercials in order to analyze the effects of the Super Bowl commercials. This is vastly different from the live reaction to the commercial as people may have just watched the commercial live and not looked it up on YouTube. In order to capture a closer genuine sentiment of the audience, it would be beneficial to analyze data across social media platforms that have more content than YouTube may provide. A platform like X/twitter would be able to give the live reaction as these commercials are played and gathering the engagement from that would be like a live reaction from the consumers. Also there could have been more categories and covariates for the dataset when dealing with the celebrities, further categorizing them by level of fame or with humor and breaking it down by the humor style.

The clear future research that could be done would just be to simply take the more recent Super Bowl commercials into account. Some alternative directions would be to incorporate the data from Instagram, TikTok and twitter to give a larger and more complete view of the audience reaction to each ad. The categorization for celebrities would be deeply impactful to understand the difference that an A list celebrity would make vs someone less famous. There also could be a specific analysis of the brand sales figures to go along with the timeline of these ads to see if they are driving revenue directly.