Abstract
Using advertising campaign data from a large US-based mobile game developer, the authors study a global advertising shutoff in the context of mobile app install ads. Contrary to prior studies in search advertising—which reveal a major over-attribution problem where paid advertising takes credit for organic traffic that would have occurred otherwise—this study shows the opposite: paid ads generate positive spillovers to organic installs. Event study analysis shows that the shutoff decreased organic installs by 20–30%. Fixed-effects panel models estimated on longer-term data find that every $100 spent is associated with 32 paid installs and 2.2 organic installs, highly consistent with the event study estimates. Further analysis strongly suggests that this positive paid-to-organic spillover operates through a ranking mechanism: paid installs boost app store category rankings, thereby increasing organic visibility. Combining campaign and ranking data, the authors find that (1) ad spend has a statistically and economically significant relationship with store rankings; and (2) the relationship between organic installs and ad spend disappears once these rankings are factored in, indicating that they absorb the relationship. These findings demonstrate that mobile app install ads are more effective than paid install metrics alone indicate, implying that developers may systematically underinvest in marketing. Keywords: app advertising, field experiments, app store rankings, organic spillovers, mediation analysis 1These authors contributed equally to this work. 2Corresponding author. 4 .1 1 v N ] J u l Prior studies of high-profile ad shutoff experiments in paid search find that ads cannibalize organic search traffic for well-known brands, with organic recovering anywhere from 37% to 99.5% of paid effects (Blake et al., 2015; Coviello et al., 2017; Golden and Horton, 2020; Simonov and Hill, 2019). Fundamentally, these results point toward an underlying over attribution problem where paid ads are taking credit for organic traffic that would have occurred otherwise (Berman, 2018; Li and Kannan, 2014). Whether this issue extends to other digital environments is an open question. Consumers often use search navigationally (e.g., typing a brand name to reach its website), so paid and organic clicks are more likely to compete for the same underlying intent (Golden and Horton, 2020; Simonov et al., 2018). We study this question in the context of mobile app install ads. Mobile apps have become a major consumer medium, with U.S. adults spending about 4 hours per day on mobile devices (eMarketer, 2024) and global app-store consumer spending reaching $150 billion in 2024 (Sensor Tower, 2025). In this environment, developers compete primarily through paid user acquisition. 1 Millions of apps compete for visibility, with thousands released each day, while global app marketing spend reached $109 billion in 2025 (AppsFlyer, 2025). Measuring the returns to that spending therefore matters, especially for firms operating on thin margins in markets with highly skewed consumer spending. However, industry standard last-touch / click attribution credits only the final ad touchpoint before conversion (Gordon et al., 2021). As search is a major user acquisition channel, a similar over-attribution problem might also be present in this context. However, app store ranking lists and personalized recommendations offer additional discovery channels. Field experiments can address these biases (Lewis et al., 2015; Gordon et al., 2019, 2023, 2026), though they remain costly and have historically been available mainly to large advertisers. In this study, we leverage the nearly 2 years of advertising campaign data of a major US-based mobile game developer we will refer to as GameSpace. During this time period, 1Throughout, we use “developers” to refer to both app makers and app publishers. Our usage of “pub lishers” will refer to advertising publishers and platforms such as Google, Facebook, and Vungle, unless explicitly noted otherwise. GameSpace conducted a global advertising spend shutoff covering search, social, display, and in-app advertising. Using an event study around this global ad shutoff, we find that when the developer halted ad spending, contrary to findings in previous studies, organic installs dropped by 20–30%, depending on bandwidth. To more precisely estimate effect sizes by dollar amounts, we also analyzed several fixed-effects panel models. We found that every $100 of ad spend was positively associated with 32.3 paid and 2.2 organic installs. Lagged fixed-effects specifications suggest modest next-day paid carryover (approximately 2.8 installs per $100, p ≈ 0.05, see Web Appendix A), consistent with delayed attribution, while lagged effects on organic installs are not statistically significant. These contemporaneous associations are qualitatively consistent with our event study findings. By combining our advertising campaign data with app store ranking data, we find that ad spend is very strongly associated with app store rankings: during the shutoff, our event study analysis finds a 2.3–3.3-fold deterioration in raw chart position, corroborated by fixed effects panel analysis. More crucially, we also find that the relationship between ad spend and organic installs disappears once we control for app ranking, indicating that rankings are mediating the effect of ad spend. These results strongly suggest that the positive paid-to organic spillovers in our context are the result of an underlying ranking mechanism: paid advertising drives up installs, which then drive up app ranks, which leads to more downstream organic discovery and installs. While our work relies on app store ranking data, this ranking mechanism potentially extends to other forms of algorithmic ranking such as personalization and search as well. If the model training of such systems doesn’t distinguish between paid and organic signals, then a similar loop to that described above occurs: paid advertising drives up installs, models train on this data and subsequently score apps higher, which leads to more downstream organic discovery and installs. Our study contributes to both academic research and advertising practice. Our study provides empirical evidence of positive paid-to-organic spillovers in platform discovery envi ronments, contrasting with the cannibalization and substitution effects widely documented in paid search. By identifying app store ranking as the key mediating mechanism, we highlight a critical methodological challenge: because rankings are determined at the platform level, localized advertising changes can spill over globally, violating SUTVA. Standard incremen tality tests—which withhold ad spend from holdout users—may therefore suffer systematic downward bias by suppressing the counterfactual ranking lift that would otherwise drive organic installs. Taken together, our findings suggest that mobile app advertising is more effective than standard attributed install numbers imply, indicating that mobile advertisers may be systematically underinvesting in marketing. Before advertising shutoff experiments, research on search advertising examined the impact of organic and sponsored search on ad performance. For example, Ghose and Yang (2009) analyzed six months of panel data on search metrics across numerous keywords, revealing a direct correlation between conversion rates and search result rankings: the higher the rank, the better the conversion rates. Moreover, Yang and Ghose (2010), through an ad pulse experiment, substantiated that paid search advertisements significantly enhance click through rates and revenue. In their pioneering study using a shutoff in ad spend, Blake et al. (2015) (henceforth BNT) produced strong evidence that paid search advertising substitutes for organic traffic. Using a series of large-scale field experiments conducted at eBay, they showed that when branded search ads were suspended, most of the lost advertising traffic simply shifted to organic search traffic. One interesting heterogeneous effect was that non-branded search ads had a positive effect on new or infrequent consumers. On average, however, returns on advertising were ultimately negative since most ads were delivered to frequent consumers who, at least in the short run, were not affected by ads. Later, Coviello et al. (2017) were concerned with the generalizability of BNT’s findings given that eBay was particularly well-known at the time. They replicated an experiment at Edmunds.com similar to BNT’s at eBay. They found the same substitution effect but at a far smaller magnitude: less than half of paid search traffic was recovered through organic search. Such an effect suggests that paid search may provide a positive return, but it still points to substitution of paid advertising with organic traffic. Simonov and Hill (2019) and Golden and Horton (2020) further confirmed Coviello et al. (2017)’s results. Although both studies were primarily focused on understanding how a firm’s paid search advertising impacts their competitors’ outcomes, their direct results show that paid search is indeed effective but that it crowds out organic traffic to some degree. In the case of Golden and Horton (2020), they noted that paid advertising is approximately 63% efficient, implying that 37% of paid advertising is substituted by organic traffic. Similarly, Simonov and Hill (2019) found that paid search ads cannibalized about 37.8% of a brand’s organic search traffic. Other studies also examine the impact of search result competition on conversion performance, noting attenuation in performance with increased competition (Agarwal et al., 2016; Bhattacharya et al., 2021). Our work builds on this existing literature in several ways. First, our work is not limited to paid search advertising. While paid search is still a dominant channel in digital advertising, other channels continue to gain market share each year, reaching 61% of all digital ad spend in 2019 (Marin Software, 2019). Second, unlike prior ad shutoff experiments, our results uncover evidence of positive organic spillovers from paid advertising rather than cannibalization. While positive digital advertising spillovers have been documented in other contexts—such as competitive spillovers to non-advertised brands (Sahni, 2016), online-to-offline spillovers (Kalyanam et al., 2018), and consumer native ad spillovers (Sahni and Nair, 2020)—our results demonstrate a positive organic spillover back to the focal firm’s own digital channel under a global ad shutoff. Third, our work brings ad shutoff experiments to the mobile channel. As mobile use and mobile advertising are both growing at unprecedented rates, understanding the extent to which paid advertising generates organic spillovers on the fastest growing devices is essential. In a sense, our work brings the ad shutoff experiment literature from the desktop era to the mobile era. Despite growing interest in mobile advertising, surprisingly little work evaluates mobile app install advertising using field experiments. Existing research has focused mainly on machine learning for bidding and targeting (Ma et al., 2016; Bhamidipati et al., 2017; Sahu et al., 2018). Our work is among the first to assess install-ad effectiveness using a large-scale spending shutoff experiment. A related stream of work studies how published app store rankings shape consumer demand and firm strategy. Carare (2012) uses daily rank data from Apple’s App Store to show that today’s bestseller rank causally raises tomorrow’s willingness to pay by roughly $4.50 for a top-ranked app, with effects declining down the top-100 list. Deng et al. (2023) study freemium launches in the App Store and attribute part of the spillover from a free to a paid version to enhanced app discovery through top-chart visibility, using category ranking data as a key outcome. These studies establish that rankings matter for demand and discovery, but neither examines whether paid install advertising moves category rankings or uses an ad shutoff to identify the advertising-to-ranking link. A smaller literature asks whether firms can strategically manipulate chart position. Li et al. (2016b) model and estimate how developers “buy downloads” to climb top-app lists, finding that $100 of purchased downloads improves an app’s ranking by roughly 2.2%. Dover and Neslin (2015) study Facebook advertising for a paid smartphone app sold on the App Store and document a “rank amplification” mechanism where advertising boosts sales, which improves the published sales rank, thereby driving subsequent sales. While their study pro vides an elegant modeling and simulation-based analysis for a single paid app, our work differs in several critical dimensions. First, we examine user acquisition (organic and paid installs) in the context of free-to-play apps on category-level free charts rather than purchase decisions on paid bestseller charts. Second, while Dover and Neslin (2015) rely on observa tional data and simulation, we leverage a massive, clean, and exogenous multi-platform ad shutoff to provide direct, causal, and experimental evidence of this mechanism. Finally, we trace and quantify these links (from advertising to category rank, and from category rank to organic installs) using a combination of event studies, lagged panel regressions, and fixed effects mediation analysis. To our knowledge, our study is the first to provide comprehensive empirical evidence of this ranking-mediated mechanism using large-scale field experimental variation. This platform-mediated discovery process represents a structural departure from consumer-level spillover mechanisms documented in other digital environments (Sahni and Nair, 2020), where spillovers operate through consumer cognitive processing and subsequent recall. In contrast, the spillover we document operates through a system-level algorithmic feedback loop: paid advertising drives a surge in install velocity, which alters the public ranking state and subsequent organic discovery. The study of attribution models in digital advertising is key to understanding their impact on online marketing strategies and budget allocation (Li et al., 2016a; Danaher and van Heerde, 2018; Danaher et al., 2020; Lewis et al., 2025). In particular, Berman (2018) highlighted the drawbacks of last-touch attribution in multi-publisher online advertising, which can lead to excessive ad exposures and distorted market pricing. However, standard last-touch models often fail to capture indirect spillovers across channels (Gordon et al., 2021). While attribution models were not our primary focus, our work relates the literature on attribution models to the literature on ad effectiveness, in the context of mobile apps. First, ad effectiveness is estimated through attribution models. While previous studies show that attribution can overestimate ad effectiveness when search ads function as navigational substitutes, our work demonstrates how attribution can underestimate ad effectiveness in ranking-mediated mobile ecosystems where ads drive organic rank spillovers. Moreover, Li and Kannan (2014) show that last-touch attribution significantly underestimates the contribution of e-mails, display ads, and referrals to conversions. They demonstrated that some customers saw ads on one channel and searched on a search engine as a navigational tool, as also shown by Golden and Horton (2020) and Simonov et al. (2018). We report exploratory platform-decomposed regressions in Web Appendix C that are consistent with this broader attribution concern—last-touch models may understate non-search channels such as Facebook relative to search—though we cannot identify causal platform-level effects from our aggregate shutoff variation. Mobile app install ads are advertisements designed to drive installs of a mobile app. Although they can appear across the entire spectrum of digital channels (e.g., search, social media, display, in-app, video), they generally link to an app’s listing in an app store to allow consumers to install the app from the ad directly. Moreover, they are generally mobile-only to make the app installation process as frictionless as possible. Figure 1 illustrates four examples of such ads. We conducted our analysis on the historical advertising campaign data of a major game developer, which we refer to as GameSpace. These data were provided to us through our collaboration with a US-based startup, which we call AdTech, that manages and optimizes digital advertising spend on behalf of GameSpace and other clients. The data track the daily spend, impressions, clicks, and installs for all the digital advertising campaigns set up by GameSpace across 85 different ad publishers for 500 days in 2018 and 2019 (calendar dates are withheld due to privacy concerns). Paid installs are installs that occur as a result of a user clicking on a paid advertisement. We used last-touch attribution, crediting the Figure 1: Examples of Mobile App Install Ads last ad impression with the install rather than any previous impressions. Organic installs are installs that occur without a user clicking on any paid advertisements. Overall, organic installs account for just over 40% of all installs. To provide a sense of the scope of this data, over 100 billion impressions were served worldwide during our time period. To investigate the underlying mechanism of organic lift, we also collect daily category ranking history from Appfigures, a leading third-party mobile market intelligence platform, for each of our six games across both iOS and Android platforms in the United States storefront. The ranking data cover the same 500-day panel window as the advertising data, allowing for a complete merge with our primary advertising panel. We track positions on each app’s primary category-level Free chart: iOS Puzzle Free (category ID 7012) and Android Games Free (category ID 43). These represent the main storefront lists where the games actively competed for organic visibility, and are the primary categories under which the developer’s games are listed and where they rank highest on average (matching the specific puzzle subcategory on iOS and the broader games category on Android). Appfigures does not report a rank on days when an app falls off the tracked category chart. Since standard app store chart rankings are numerically inverted (where 1 represents the top chart position and larger numbers represent worse rankings), we log-transform the raw chart position, denoting the result log Rankijt for app i on operating system j on date t. Higher values of log Rankijt therefore indicate worse chart positions. Category ranks are also heavy-tailed: moving from rank 5 to rank 10 conveys much more visibility than moving from rank 500 to rank 505. A one-unit increase in log Rankijt therefore corresponds to an e-fold deterioration in raw chart position (e.g., rank 20 to rank ≈ 55). Web Appendix B replicates the rank-mechanism results using − log(Rankijt), − Rankijt, and 1/Rankijt, with qualitatively unchanged conclusions. For our analysis, we aggregated this historical campaign data by summing installs and spending for all campaigns for each app and operating system (OS) combination at a daily level (so each observation of the data is uniquely identified by an app, OS, and date combi nation). We restricted our analysis to 6 particular mobile apps out of GameSpace’s larger portfolio, leading to a dataset with 6000 observations (6 apps × 2 OS × 500 dates). We fo cused specifically on these 6 since they had the most complete and active advertising activity throughout our data. Furthermore, all these apps were relatively mature, with the youngest being released over 6 months before the beginning of our observation period, allowing us to avoid any inflation in organic install numbers due to media mentions. Despite all apps being relatively mature and having active advertising activity, there was still some heterogeneity with the most popular app garnering around 4 times the spend and installs of the least popular app. The estimation sample excludes dates with app-store featuring spikes, which would oth erwise confound organic-install dynamics, leaving 5829 app-OS-date observations. For rank mechanism analyses, our main specifications use observed Appfigures chart positions wher ever reported (approximately 93% of these observations); off-chart days are omitted because Appfigures does not report them. Web Appendix B shows that the rank-mechanism results are robust to alternatively coding off-chart days as rank 401, just below Appfigures’ default top-400 cutoff. To visualize the data, we plotted the scaled values for spend, paid installs, and organic installs for the 12 different app-OS combinations in Figure 2. Scaled values were computed by dividing each series by the standard deviation of that metric for that particular app. Thus, the scaling factor is preserved between device OS (e.g., App1 Android spend and App1 iOS spend are divided by the same number) but will differ across metric and app (e.g., App1 iOS spend, App1 iOS paid installs, and App2 iOS spend are all divided by different numbers). The key to our identification strategy is a global ad shutoff experiment similar to the branded search shutoff experiments conducted by BNT at eBay. GameSpace implemented a coordinated shutoff across all apps and platforms between days 75 and 98 of our data collec tion (highlighted in dark gray shading in Figure 2). Figure 3 examines this period in more detail to give a better sense of the underlying dynamics. Specifically, GameSpace stopped 0 2 4 6 8 A p p 1 Android iOS 0 2 4 6 8 A p p 2 0 2 4 6 8 A p p 3 0 2 4 6 8 A p p 4 0 2 4 6 8 A p p 5 100 200 300 400 500 Date 0 2 4 6 8 A p p 6 100 200 300 400 500 Date V a lu e Spend Paid Installs Organic Installs Figure 2: Scaled Time Series. The scaled time series for spend, paid installs, and organic installs by app and OS. Each row denotes an app and each column denotes an OS. Scaled values are produced by dividing the app-OS-date summed numbers by the standard deviation of that metric across both OSes. This means that each metric is scaled separately within app–OS pairs. Note that there are several instances of unusually high organic installs. These spikes occurred when an app was somehow featured in the respective app store. adding additional spending to their accounts several days earlier. As many campaigns still had positive balances, it took several days for the remaining budget to be exhausted. Looking closely at Figure 3, we can see some series never quite reach 0 (indicated by the black dashed line) during the shutoff for this reason, but advertising spend is practically zero during this period. We also note the heightened amount of spending, at least for app-OS pairs in the weeks leading up to the global shutoff. This was due to GameSpace moving its spending forward, using what it would have spent during the shutoff in the weeks before. Figure 4 displays the corresponding raw, scaled category rank series around the shutoff window. Ranks are log-transformed, then demeaned and SD-scaled within each app (pooling Android and iOS) so absolute chart positions are not identifiable. Higher values indicate worse chart positions (same direction as log Rankijt in our analysis). Two additional partial shutoffs for individual apps (App5 from day 395 and App6 from days 177–239) provide supplementary context but are not our primary identification strategy. All apps were active during and after the 500 days that we observed. 7 6 5 4 le d 3 S c a 2 1 0 | 70 | 75 | 80 | 85 | 90 | | 95 | 100 | 105 | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | | | | | Date | | | | | | | App1 (Android) | | App2 (iOS) | App4 (Android) | | App5 (iOS) | | | | | App1 (iOS) | | App3 (Android) | App4 (iOS) | | App6 (Android) | | | | | App2 (Android) | | App3 (iOS) | App5 (Android) | | App6 (iOS) | | | | Scaled Figure 3: | Spend during | Shutoff. | Scaled | spend on | days 67 | through | 105 for all | of the 12 | | different app-OS | combinations | we analyze | in this | study around | the | shutoff | period, highlighted | in | gray shading. 60 70 80 90 100 Date 4 3 2 1 0 1 2 R a n k Android 60 70 80 90 100 Date iOS App1 App2 App3 App4 App5 App6 Figure 4: Scaled Category Rank during Shutoff. Demeaned and SD-scaled log Rank on days 55 through 105 around the global shutoff (gray shading). Left: Android; right: iOS. Series are demeaned and scaled within each app across both platforms so absolute chart positions are not identifiable. Higher values indicate worse chart positions; the dashed line marks each app’s sample mean. We employed two main empirical methods to analyze the paid ad shutoff experiments across GameSpace’s six different apps: an event study and fixed effects panel regressions. We describe the details of each method and the corresponding model specifications below. In our event study estimation, we treat the global ad shutoff as the event of interest and estimate outcome changes in a local window around the shutoff date (Hausman and Rapson, 2018). Our event date is the start of the global shutoff on day 75. While ad spend reactivation theoretically serves as an additional event, we excluded it from our analysis since reactivation was gradual (Figure 3). We estimate local constant and local linear specifications: log(Yijt) = δDijt + αij + ωt + ϵijt (1) log(Yijt) = δDijt + γ1(t − c) + γ2Dijt(t − c) + αij + ωt + ϵijt (2) where Yijt denotes organic, paid, or total installs for app i on OS j at date t; Dijt is an indicator equal to one after the shutoff cutoff c; αij are app-OS fixed effects; ωt are day-of week fixed effects; and ϵijt is the error term. We log-transform install counts to interpret δ as an approximate proportional change, which is appropriate given substantial heterogeneity in app scale. The coefficient δ captures the discontinuity in installs at the shutoff. Poisson robustness checks appear in the Web Appendix. Because time is the running variable, calendar date fixed effects would be collinear with Dijt during the estimation window. For Equation 2, (t− c) is the difference in days between date t and cutoff c. The slope coefficients γ1 and γ2 are not the focus of our analysis. We purposefully limited ourselves to the simpler event study specifications since the more complex specifications run the risk of overfitting. Overfitting can be especially problematic since the decrease in spending, while rather drastic, is still not perfectly sharp, thus extending the outcome shift across multiple days. As such, higher-order polynomial terms, if used, would more likely capture the underlying shift in spending, rather than the true nonlinearity of interest. For instance, as seen in Figure 3, the two series corresponding to App3 already attained a much lower level of spend 3 days before the cutoff compared to the other apps. Though this is the most prominent example, many other apps already had declining ad spend in the days before c. Furthermore, persistence effects are a concern: a consumer may install an app today due to ad exposure yesterday or the day before. While such effects will eventually decay, they will also extend the outcome shift across multiple days. Aggregate Spend. Although the event study identifies the causal effect of a shutoff in ad spend, in practice, firms adjust spend continuously rather than shutting it off entirely. To examine how day-to-day changes in ad spend relate to installs, we used the panel regression in Equation 3: Yijt = βspendijt + πijt + ϵijt (3) Here, our main variable of interest is spendijt, the number of dollars spent on advertising app i for OS j on date t across all publishers and campaigns. For this specification, we did not log transform our variables because we expect the underlying relationship between spending and installs (organic, paid, or total), is linear rather than multiplicative. We avoided using date fixed effects as date-level shocks are likely multiplicative rather than additive because calendar shocks (e.g., weekends) are likely multiplicative across apps of different scale. Instead, we employed an app-OS-time fixed effect πijt that consists of the interactions between the app-OS fixed effects αij and day-of-week fixed effects ωt as well as the interactions between app-OS fixed effects αij and week fixed effects ψt. More formally πijt = αij × ωt + αij × ψt. Thus, the effect β captures the deviations relative to the average number of installs on an app-os-day-of-week and an app-os-week level. Exploratory platform decomposed specifications appear in Web Appendix C. To investigate the underlying mechanism of the organic lift, we formally specify the models used to test the ranking mechanism, using log Rankijt. Event Study on App Store Rankings. To test whether paid ad spend causally affects category rankings, we replicate the event study design from Equations 1 and 2 but substitute log Rankijt as the dependent variable: log Rankijt = δDijt + αij + ωt + ϵijt (4) log Rankijt = δDijt + γ1(t − c) + γ2Dijt(t − c) + αij + ωt + ϵijt (5) where Dijt is the post-shutoff indicator, αij are app-OS fixed effects, ωt are day-of-week fixed effects, and (t − c) is the running variable of days relative to the cutoff. Lagged Effects of Advertising on Store Rankings. To examine whether yesterday’s advertising spend predicts today’s store ranking, we estimate the following panel regression with lagged spend: log Rankijt = θspendij(t− 1) + πijt + ϵijt (6) where πijt = αij × ωt+αij × ψt represents the app-OS-time fixed effects defined in Equation 3. Mediation Model Specification. To evaluate whether category rankings mediate the relationship between contemporaneous spend and installs, we estimate parallel fixed-effects specifications containing both variables as predictors: Yijt = β1spendijt + β2 log Rankijt + πijt + ϵijt (7) where Yijt represents organic or paid installs, spendijt is contemporaneous spend across all ad campaigns, log Rankijt is the log-transformed app store category chart ranking, and πijt is the app-OS-time fixed effects. Under the ranking mechanism, adding store ranking to the model should absorb the association between contemporaneous spend and organic installs, while paid installs (which are driven mainly by contemporaneous ad delivery) should remain strongly associated with spend.
Showing the abstract — retrieve the full paper via the Exa API.