Digital Advertising in the 2022 Midterms

1 Data and Methods

In this paper, we examine the spending and content of advertising on Meta (which includes Instagram and Facebook) and the spending on Google (which includes YouTube, search, display and banner) by general election candidates in federal races (including coordinated spending between candidates and parties), and we also compare candidate digital spending and content to candidate advertising on television. More specifically, we draw on ad data from Meta (https://www.facebook.com/ads/library/) and Google (https://transparencyreport.google.com/political-ads/) for digital, which are described in further detail below. Television data come from Kantar/CMAG (please see Franz, Ridout, and Fowler 2022 in this issue of The Forum for more information on the television data).

For Meta’s digital spending totals, we utilize each page name and disclaimer combination from the platform’s daily aggregate reports (Meta Ad Library Report). For Google’s spending totals, we use weekly reports from the platform’s Transparency Report, which report Sunday to Saturday totals every week. In both cases, we calculate cumulative spending by subtracting the reported totals on the day before our desired starting date from the reported cumulative spending of the end date (so for Meta, for example, general election spending is defined by subtracting the reported cumulative spending in the September 4 report – because we want spending starting on September 5, 2022 – from reported cumulative spending in the November 8 report). In this manuscript, we report on three different time periods, which are defined as follows: (1) cycle-to-date information is defined as January 6, 2021, through November 8, 2022, for Meta and January 10, 2021, through November 5, 2022, for Google, (2) election year only is defined as January 1 through November 8, 2022, for Meta and January 2 through November 5, 2022, for Google, and (3) the general election (post Labor Day) period, which we define as September 5 through November 8, 2022, for Meta and from September 4 through November 5, 2022, for Google. When we report on television data, we match the television data to the dates listed for Meta.

To identify activity by federal candidates, we rely on a 2022 candidate list from OpenSecrets and the Wesleyan Media Project, and we restrict the universe to ads sponsored by general election candidates for U.S. House and U.S. Senate. We search aggregate reports from both platforms (and the Meta Ad Library API) using keywords, including candidate names and generic terms such as “senate,” “senator,” “congress,” and “representative” to identify and match candidates to their respective advertising. We collapse any identified coordinated activity (where a candidate jointly sponsors advertising with their party) with the candidate, and include both spending and content from coordinated advertising in the candidate analysis below for digital (and in comparable television figures).

Beyond spending information, when we examine digital ad content, we focus our attention on Meta. The Facebook API provides ad metadata such as the range of ad spending and impressions, region distribution, and demographic distribution for both Instagram and Facebook advertising. We include all ads from pages and entities (page ID and funding disclaimer pairings) that have been identified through the process described above, and we filter this collection to include all ads as long as they were running between September 5, 2022, and November 8, 2022 (the general election period only).

To inform our classifiers, we collect text fields from the API, which include page name, disclaimer, ad creative body, ad creative link caption, ad creative link title, and ad creative link description. We also obtain the transcription of the ad audio through automatic speech recognition (ASR) via Google’s speech-to-text API (enhanced ‘video’ model). Using this textual information, we are able to detect candidate mentions from which we can calculate national politician mentions and traditional ad tone; we also use the text to classify the goal of the ad along with specific issues.

To detect which candidates and national politicians are mentioned, we apply named entity recognition to each text field of an ad available through the API, as well as its ASR transcript. Named entities are then compared to a curated knowledge base of candidates and other major figures. If a named entity matches more than one person in the knowledge base, the match is resolved by calculating the word embedding similarity of the text field to sentences containing a basic description of each candidate.

Ad tone is determined on the basis of who is mentioned in an ad (excluding the page name and funding entity disclaimer fields), following a longstanding method in the discipline (Goldstein and Freedman 2002). If an ad mentions only the sponsoring candidate, it is coded as ‘Promote.᾽ If it mentions only their opponent, it is coded as ‘Attack.’ If it mentions both, it is coded as ‘Contrast.’

We determine the goal of an ad based on a random forest classifier we trained on hand-labeled data (performance on a held-out test set is 99% F1-score). There is one binary classifier for each of a number of goals, but in this research, we focus on ‘Donate’ specifically. The classifier is applied to tf-idf features of all text fields plus ASR.

Similarly, issues are also determined by a random forest classifier. The main difference between the two is that for ad goal, we rely on a larger training dataset of 2020 data (because the way donate ads are framed does not change fundamentally between the years), but because abortion and inflation are particularly salient in 2022 and likely different in the way they were discussed in 2022 than other recent years, we require more up-to-date training data. Therefore we use labels created by Kantar/CMAG for 2022 TV ads. Performance on a held-out test set is 96% F1-score for abortion, and 94% for inflation.

As we will discuss, the content results for digital advertising should be considered preliminary as they are built on initial, text- and ASR-only classifiers, which may differ (sometimes greatly) from results that incorporate visual information inherent in digital advertising. They also only include Meta (Facebook and Instagram), and a different picture may emerge if Google advertising were included.

2 Initial Findings on Digital Ad Spending in 2022 by Federal Candidates

We start our analysis by looking at Meta and Google advertising totals for U.S. Senate and U.S. House candidates (including coordinated spending) for the 2021–2022 election cycle, the calendar year 2022, and the fall general election period. These totals are reported in Table 1. All told, we identify just under $150 million in candidate spending on digital advertising in federal races on these two platforms. In the last two months of the campaign, over $65 million was spent by federal candidates on digital ads. Regardless of the time period, about 60% of total spending was in Senate races.

Although these numbers seem large, they are underestimates, possibly severe, of total digital spending during the 2022 election cycle. As noted above, this is because they do not include spending on platforms other than those supported by Meta or Google. Additionally, the numbers in Table 1 exclude spending by parties or groups in federal races, nor do they include spending in races for governor or other races down ballot.

We look more closely at this spending in Figure 1 by comparing the distribution of candidate advertising spending in U.S. House and U.S. Senate races for Google and Meta and then comparing them to spending in 2020. Our universe only includes candidates who have spent money on at least one of the two platforms. In many cases candidates spent money on both Meta and Google, but in other cases candidates spent money on only one platform. The time frame in these tables is the fall election period (see the methods section for our definitions).

Figure 1:

Distribution of candidate spending on digital (Meta and Google) in U.S. House and U.S. Senate Races (general election only). Source: Meta Ad Library and Google Transparency Report with analysis by the Wesleyan Media Project.

Three broad patterns are evident. First, the distributions are very long-tailed, as most candidates spent very little, and in many cases, nothing at all. This stands in contrast to several extreme outliers, which drove most of the spending. Second, U.S. House candidates preferred Facebook and Instagram, while U.S. Senate candidates in 2022 gravitated more towards Google. Third, Google’s share of the market in 2022 has increased compared to 2020, potentially at the expense of Meta.

For example, in 2020, the median U.S. House candidate spent $7839 on Meta, and $0 on Google (again, this only includes candidates who have spent money on at least one of the two platforms). In 2022, these numbers were $7188 for Meta, and $1100 for Google. For U.S. Senate candidates, the shift towards Google was more extreme. In 2020, the median Senate candidate spent $23,864 on Meta, and $8450 on Google. In 2022, the median candidate’s spending on Meta decreased slightly, to $22,795, while median spending on Google increased more than five-fold, to $47,200. In other words, Meta spending has stagnated (if not slightly decreased) for both U.S. House and U.S. Senate candidates, while Google spending has increased. We discuss spending on these platforms (combined) relative to spending more widely on radio and television in the next section.

Table 2, which covers the full 2021–2022 election cycle, reveals that the top candidate spender on digital ads in races for U.S. Senate was Raphael Warnock of Georgia, who spent $15.7 million in the 2022 cycle. This does not include money spent on advertising for the December 2022 runoff election in Georgia. Number two on the list is Democratic Senate candidate Val Demings of Florida, who spent almost $11.4 million on digital ads. She was followed by Arizona Senator Mark Kelly, with $7.8 million in spending. One striking observation is that Democrats hold eight of the top 10 spots on this list. Not only were Democrats often advantaged in terms of the volume of television ads they aired (as discussed in Franz, Ridout, and Fowler 2022 in this issue of The Forum), but this advantage appears to extend to digital ad spending as well.

Table 3 shows the top digital spenders in competitive U.S. House races. The top spender, by far, on Meta and Google was Democratic representative Katie Porter, with $3.2 million in digital spending. Porter, from Southern California, not only faced a tight reelection campaign, but her district is in the Los Angeles media market, which is one of the most expensive (and inefficient) in the country. Thus, digital advertising, which can be much more narrowly targeted than broadcast television, is especially important in large media markets. Porter’s heavy digital spending may also have helped prepare for her run for the U.S. Senate, which she announced in January 2023. As with U.S. Senate candidates, the top digital spenders in the U.S. House were all Democrats, a pattern we will return to more systematically below.

While Table 3 shows digital spending in competitive U.S. House races, Table 4 shows spending on Meta and Google in all House races. The lists are largely similar but with some exceptions. One is Marcus Flowers, who was the Democrat running against controversial Rep. Marjorie Taylor Greene of Georgia; Flowers presumably was able to use his willingness to run against Greene to raise substantial sums nationally. New York Rep. Alexandria Ocasio-Cortez, who is known for being an outspoken progressive and for her digital-first tactics, also spent over $2 million on digital advertising. Republican House leader Kevin McCarthy also was among the top 10 digital spenders.

Figure 2 displays the relationship between Democratic and Republican opponents’ spending on digital advertising in U.S. Senate races during the fall election period. Those observations above the line represent races in which the Democratic candidate spent more than the Republican candidate, while those observations below the line represent the opposite. The story here is clear: in almost all competitive U.S. Senate races, the Democratic candidate spent more on digital than the Republican candidate. And in some instances, the differences were stark. In Georgia, for example, Warnock’s campaign spent well above $6 million while his Republican opponent, Herschel Walker, spent around $1 million. Similar imbalances were seen in Wisconsin, Pennsylvania, and even Ohio, where Republican J.D. Vance was favored throughout the entire campaign. (It is important to reiterate that these totals exclude supportive expenditures from outside groups, which may have been substantial in some cases. Nor do they include any independent expenditures from the party committees.)

Figure 2:

Spending on digital (Meta and Google) in U.S. Senate Races by Democratic and Republican candidates (general election only). Source: Meta Ad Library and Google Transparency Report with analysis by the Wesleyan Media Project. The Louisiana and Alaska races are excluded as they feature more than one candidate from the same party.

The same data are shown for competitive U.S. House races in Figure 3. A similar pattern emerges, with Democratic candidates spending well more than their Republican rivals on digital ads in many of the most competitive races, including California’s 47th district (the Katie Porter v. Scott Baugh race), Minnesota’s second district (Democrat Angie Craig v. Republican Tyler Kistner) and Virginia’s seventh district (Democrat Abigail Spanberger v. Republican Yesli Vega). One race in which the Republican had a digital spending advantage was Pennsylvania’s 7th district, where Democrat Susan Wild faced Republican Lisa Scheller.

Figure 3:

Spending on digital (Meta and Google) in competitive U.S. House Races by Democratic and Republican candidates (general election only). Source: Meta Ad Library and Google Transparency Report with analysis by the Wesleyan Media Project. Competitive contests are defined as those rated toss-up in the Cook Political Report from November 7, 2022. The Alaska at-large district is excluded as it features more than one candidate from the same party.

3 Digital Versus TV

Next, we look explicitly at how digital spending on Meta and Google by candidates compared to their other ad spending on television and radio in 2022. We totaled ad spending by candidates on local broadcast stations, local cable stations, satellite television, and AM/FM radio stations using data from the Wesleyan Media Project. We collapse that spending and compare it to the Meta and Google totals. We show this for U.S. Senate races in Figure 4 and for “toss-up” U.S. House races in Figure 5. The totals show ad spending for the period between September 5 and Election Day, and they include spending by the candidate directly and in coordination with a formal party organization.

Figure 4:

Media spending in U.S. Senate Races (general election only). Source: Meta Ad Library and Google Transparency Report with analysis by the Wesleyan Media Project.

Figure 5:

Media spending in competitive U.S. House Races (general election only). Source: Meta Ad Library and Google Transparency Report with analysis by the Wesleyan Media Project. Note: Competitive contests are defined as those rated toss-up in the Cook Political Report from November 7, 2022.

In general, traditional ad spending accounts for the lion’s share of total ad spending in both U.S. House and U.S. Senate races. It is clear that campaigns still put great value on reaching voters through traditional mass media.^[1] But comparisons across mediums are tricky. Television is generally more expensive than targeted Facebook ads (especially in media markets that include denser urban areas), and the reach of television compared to digital is hard to quantify. Moreover, the returns on a digital buy per impression might be higher and/or easier to measure than the impact of a television ad for each viewing. Indeed, research on the impact of digital ads remains small. Still, in the most expensive U.S. Senate campaigns—for example, those of Democrats Warnock, Kelly, and Fetterman—digital spending was significant. Warnock spent more than $40 million on political ads in the fall election period, with about $7 million of that on Facebook and Google.

How does the share of digital spending on Meta and Google in 2022 compare to past election cycles? We showed earlier that there were some important shifts in spending across Meta and Google between 2020 and 2022. Here we consider spending on both platforms combined in comparison to ad spending on television and radio.

Over-time comparisons are not straightforward to calculate, but we can get a pretty good sense of the trajectory of digital spending by comparing the amount spent on digital to the amount spent on television advertising. For the 2018, 2020 and 2022 election cycles, we calculated the digital percentage by dividing the amount spent by candidates on digital ads on Meta and Google in U.S. Senate races and in “toss-up” U.S. House races during the general election period (the numerator) by the amount spent on those two platforms and broadcast television during the general election period (the denominator).^[2] We calculated weighted percentages to account for the varied levels of spending across House and Senate campaigns.^[3]

In 2018, an average of 11.4% of a U.S. Senate campaign’s ad spending was on digital (Fowler, Franz, and Ridout 2022).^[4] By 2020, that percentage had risen to an average of 18.7%.^[5] In U.S. House races, an average of 8% was on digital in 2018, and that rose to 10.6% in 2020. In 2022, however, the share committed to Meta and Google declined in both House and Senate campaigns.^[6] In competitive U.S. House campaigns, the average for Meta/Google spending was just 5.9% of digital and broadcast television spending. In U.S. Senate races, the average share was 13.5%.^[7]

Having reviewed trends in digital spending, we next turn our attention to content with a specific focus on advertising across Meta technologies (Facebook and Instagram).

4 Ad Goals

Unlike political advertising on television, which is almost always focused on persuasion (Fowler, Franz, and Ridout 2022), campaigns pursue many goals with their digital advertisements (Ballard, Hillygus and Konitzer 2016; Ridout, Fowler, and Franz 2021), including fundraising, acquiring contact information from voters, and mobilization. Past research has suggested that appeals for donations are common in online political advertising—appearing in a quarter of presidential ads in 2016, for instance (Ballard, Hillygus, and Konitzer 2016)—but to what extent do fundraising appeals vary across campaigns? To investigate this, we calculated the percentage of ads that ask for a donation in a variety of U.S. Senate campaigns. This is shown in the red circles in Figure 6, which displays the top 20 highest spending candidates on Meta (across Instagram and Facebook) for the general election period. The column at the right of the figure shows the total spent on Meta ads during this time period.

Figure 6:

Proportion of donation and in-state ads for top 20 U.S. Senate candidate spenders on Meta (general election only). Source: Meta Ad Library with analysis by the Wesleyan Media Project.

Clearly, there is substantial variation in the percentage of total ads devoted to fundraising. Over 75% of ads contained fundraising appeals in the case of candidates like Michael Franken of Iowa, Val Demings and Marco Rubio of Florida, and John Fetterman of Pennsylvania. At the other end were long-time incumbents like Chuck Schumer of New York and Lisa Murkowski of Alaska whose Facebook ads did not ask for money.

Also plotted in Figure 6 (in the blue triangles) is the proportion of ads from each campaign that were seen by in-state audiences.^[8] One can see a strong relationship between the proportion of fundraising ads and the proportion of in-state ads. As in-state ads increased, donation ads declined. It is clear that a candidate like Murkowski was very much focused on persuading Alaska voters with her Facebook ads, while a candidate like Franken in Iowa was very much focused on raising funds from out-of-state donors. Other candidates, of course, deployed a more even mix of out-of-state fundraising ads and in-state non-fundraising (persuasion, acquisition and mobilization) ads.

Overall, nearly half of the spending on U.S. Senate advertising on Facebook and Instagram was committed to donation appeals (47%), which was just below the money spent on in-state advertising (51%). Not surprisingly, the proportion of donation advertising by candidates in U.S. House races was lower than in U.S. Senate contests, with nearly a third (31%) of spending on donation and fundraising appeals, and more U.S. House candidate advertising was concentrated in the candidate’s state (70%) compared to Senate candidates. (The ad-level data from Meta show distributions of impressions only by state, and so we are limited to in-state and out-state comparisons for House candidates.)

5 Ad Tone

Prior analyses of Facebook and television advertising have shown that digital advertising is much less negative than television (Fowler et al. 2021). As a reminder, ad tone here refers to mentions of the opponent and not to the valence or sentiment of the language employed by ad sponsors. To say that ads on Meta are more positive than ads on television means (in this context) that candidates are less likely to refer to their opponent.

Table 5 shows that the finding that Meta advertising is more positive than television advertising holds true in 2022. Interestingly, however, U.S. Senate candidate advertising on Facebook and Instagram does appear to be using a non-trivial number of pure attack ads (those that save for the page name and disclaimer solely mention the opposing candidate). Still, when combining attack and contrast ads on Meta compared to television, there are a lot fewer mentions of the opposing candidate on Meta.

This is not necessarily an expected finding. Because digital ads can be highly targeted, one might predict that candidates would stoke partisan animosities more in their digital ads given that the likely targets are sympathetic partisans—all the more true, perhaps, in a deeply polarized political environment. Indeed, we might expect candidates to think that targeting an opponent’s perceived weaknesses would be a good way to mobilize supporters, either to vote or to donate. Still, this is not borne out by the evidence in Table 5; candidates devote the majority of their spending on Meta to ads that ignore the opposing candidate.

6 Biden and Trump Mentions

Even if candidates are more likely to ignore their opponents in digital advertising than on television, it doesn’t necessarily mean that their ads are devoid of partisan content. In fact, an analysis of ads from 2018 showed that Facebook advertising is more partisan (Fowler et al. 2021). One way in which election advertising can heighten partisan content is through a focus on party leadership. Midterm elections are often seen as referenda on incumbent presidents and their party, which is also why the president is frequently invoked in election advertising—particularly by the party out of power.

Candidate advertising in 2022 was no exception. Republican candidates were particularly likely to mention President Biden in their advertising (see Table 6). Republican advertising for U.S. House seats was more likely to reference Biden than U.S. Senate advertising on both Meta and TV, although references in digital ads were lower than in TV ads for both races. Democratic ads were much less likely to invoke the president than Republican ads. Democratic U.S. House candidates mentioned President Biden more often in their digital ads (as a proportion of total spending) than in their television ads (as a proportion of airings).

Former President Trump did not feature prominently in federal advertising sponsored by Republicans on either Meta or television. But Trump did appear at much higher rates among Democratic senatorial advertising, especially on Facebook and Instagram where nearly 16% of all Democratic advertising mentioned the former president. Democrats running for U.S. Senate mentioned Trump less frequently on television (8% of airings), but that percentage was still much higher than for Republican U.S. Senate candidates who mentioned Trump in only 0.2% of their television airings.

7 Issue Mentions

Prior analyses comparing digital and television advertising found that ads on Facebook are much less likely to mention substantive policy issues (Fowler et al. 2021) with only two exceptions: abortion and the environment. Of course, mentions of abortion on TV in earlier cycles was very low. So how did issue discussion of two of the most prominent topics from the 2022 midterms on TV compare to advertising on Meta? As shown in Table 7, although mentions of abortion on Facebook and Instagram were less prominent than on television (where they featured in roughly a quarter (26%) of Democratic candidate advertising), online ads mentioning abortion by Democratic general election candidates made up nearly one fifth of their spending. In other words, abortion featured prominently in 2022, both on television and online. On the Republican side, mentions of inflation occurred in nearly half (45%) of candidate advertising on TV, and while attention on digital was lower, inflation still comprised nearly one fifth of ad spending by Republican candidates.

8 Limitations

This research is our first post-election look at digital political advertising for the 2022 midterm elections and therefore provides preliminary results with several limitations. First, our content analysis relies solely on text fields from the Facebook Ad Library API and audio transcription. This means that we are not capturing words that appear in ad images and videos, which would require the use of Optical Character Recognition (OCR). Thus, we may be under-reporting Biden and Trump references in digital advertisements because they may be only invoked in the text of an image. Our results for ad tone presented here may also differ with the incorporation of OCR as they are determined based on the mentions of sponsoring candidates and opponents. Second, visual cues, including facial recognition of images (including visuals in videos) may change our classifications of party leader references and tone because Biden, Trump and other politicans may only be pictured and not mentioned. Additional visuals like symbolic objects and featured scenes, if available, may change our machine learning classification results for issues. This is especially true if candidates strategically use visual cues to make reference to certain issues without explicitly discussing them. Finally, we conduct content analysis only on Meta ads. Candidates may employ different campaign strategies for ads on other digital platforms such as Google.

9 Summary and Conclusions

Our analysis uncovered declines in federal candidate spending on digital advertising (defined as the combined spending on Meta’s and Google’s platforms) in 2022 as a share of media spending compared to 2020. This finding may be driven more by decreases in spending on Meta in particular as there is some evidence that Google spending increased. Still, general election candidates in U.S. House and U.S. Senate races devoted nearly $150 million to digital ads on these platforms over the full 2021–2022 election cycle. Our analysis excludes spending by outside groups in congressional races, and it excludes digital ad spending on other platforms, including on streaming channels. Ads placed on streaming services (referred to as OTT or “over the top”) through Hulu, Roku or other services are delivered over the Internet but can be seen on television sets, tablets, smart phones or other mobile devices. If candidates and allied groups and parties have started to diversify their digital ad buying—which we suspect is likely—then the Meta and Google platforms may not account for the lion’s share of digital ad spending in cycles to come. This means that digital ad spending, inclusive of all possible platforms, is likely to grow in importance in coming cycles, even if the two biggest platforms in 2022 appeared to matter a little less. Meta and Google provide important, though imperfect, ad archives, which allow for deeper discussions of digital ad content, but most other platforms—most notably streaming channels—do not allow for systematic tracking of spending and analysis of ad content.

Similar to television advertising (see Franz, Ridout, and Fowler 2022 in this issue of The Forum), and consistent with prior assessments of partisan differences in digital investments (Sheingate, Scharf, and Delahanty 2022), we demonstrate that Democrats in most U.S. Senate and tossup U.S. House contests held large advantages in digital spending on Meta and Google compared to their Republican opponents’ spending. Importantly, our analysis incorporates coordinated spending, but excludes outside group investments. Still, the imbalance in spending between Republican and Democratic candidates is noteworthy.

Our analysis of the ad content on Meta uncovered a few important findings. For one, there is a lot of diversity across campaigns in the goal of paid advertising. Many campaigns, especially in the U.S. Senate, focused on citizens across the country, far from their home constituencies, with the goal of raising money. We know from prior research that there are limits to digital advertising when it comes to voter persuasion. Campaigns can track how digital ads result in sign-ups or donations, but there is not much evidence that digital ads actually have a large effect on voters’ choices at the ballot box (Coppock, Green, and Porter 2022), and recent work also suggests that there is limited effect on mobilization as well (Aggarwal et al. 2023). As such, many candidates chose to use the platform to raise money and build email lists for follow-up solicitations.

Additionally, we confirm prior research on the 2018 election that found Facebook ads to be less negative and less focused on issues compared to television ads. Using the traditional definitions of positive, negative, and contrast ads in American elections, digital ad campaigns on Meta in 2022 were more positive than ads on television. Moreover, while abortion featured heavily in Democratic ads on Meta, and inflation was a focus for many Republican ads, these issues were much less likely to be referenced in digital ads than in a 30-second television spot. In short, there is not much evidence that sponsored Facebook and Instagram ads aim to engage voters with more extensive issue appeals than what viewers see when they are watching television.

The research in this paper is in many ways preliminary, as processing of ads on Meta and Google (including processing of important visual cues and text) is on-going as of this writing. There also remain a host of important questions beyond what we have asked here. How much did outside groups spend in 2022 compared to 2020, for example? Did outside groups also use these platforms to raise money, or did they focus disproportionately more on persuasion than candidates? Moreover, there are limits to the questions we can ask without more transparency from other platforms about ad spending and ad content. Indeed, the landscape of political advertising continues to change, and this will likely inspire a new generation of scholarship on the distribution and impact of advertising beyond the spots we see on television.