Season 5, Episode 2: Meta's Renaissance

[MUSIC] This week's episode of the Mobile Dev Memo Podcast is brought to you by Vib. Powered by proprietary data and machine learning, Vib is the leading streaming TV ad platform for small and medium-sized businesses looking for actionable campaign performance. Vib's newest suite of AI-enabled products, like Vib Studio, Vib AI Assistant, and Vib IQ2, are just the latest in a string of intuitive and transparent tools designed to radically democratize access to premium television advertising. Try them out for yourself at Vib.co/sina. That's Vib.co/sina. [MUSIC] The problem is that the distinction needs to be drawn between the confidence of the economists and the correctness of their analysis. [MUSIC] On October 26, 2022, at the end of Metas Q3/2022 earnings call, during which the company recorded a 4% year-over-year decline in revenue. Mark Zuckerberg made a statement that I found oddly, uncharacteristically solicitous, if perhaps even vulnerable. When asked by an analyst about the company's confidence that the various initiatives into which it was investing vast sums of money would prove worthwhile, Zuckerberg said, "There are a lot of things going on right now in the business and in the world, and so it's hard to have like a simple, we're going to do this one thing, and that's going to solve all the issues." I mean, there's macroeconomic issues. There's a lot of competition. There's ads challenges, especially coming from Apple. And then there are some of the longer-term things that we're taking on expenses because we believe that they're going to provide greater returns over time. And I think we're going to resolve each of these things over different periods of time. And I appreciate the patience. And I think that those who are patient and invest with us will end up being rewarded. End quote. Metastock had opened on October 26, 2022 at roughly $132 per share. It opened the next day at roughly $98 per share. Metastock would continue to decline reaching $90 per share on February 5, 2023, which is a level the company stock hadn't seen since October 2, 2015. And this decline took place a little over a year after Metastock had reached an all-time historical high of $380 per share. But perhaps more incredible than the stocks decline from $380 to $90 within 13 months is its subsequent rebound. By the end of October 2023, Metastock was trading at around $300 with the company reporting 25% year-over-year growth in Q3, 2023. How did the company go from a 4% year-over-year revenue decline in Q3, 2022 with Mark Zuckerberg earnestly in treating investors to remain patient to 25% revenue growth in Q3, 2023, and to an all-time stock price of over $600 today as a record this. And what catalyzed both of these inflections? In this podcast, I'll explore Meta's renaissance, the factors that led to the company's tumultuous decline in 2022 and, critically, what the company changed to invigorate growth in 2023 and 2024. Ultimately, Meta's trajectory over the past few years has been defined by its response to Apple's app tracking transparency or ATT privacy policy, with its recovery serving as a blueprint for success in the modern operating environment. In short, I believe three adaptations allowed Meta to recover from the fundamental challenges imposed on the structure of its advertising platform by ATT. First, Meta's wholesale transition of its core product experiences from a friend graph to an open graph. Second, the transition of Meta's advertising measurement tools to new data ingestion mechanisms and attribution approaches. And third, the expansion of Advantage Plus, Meta's suite of automated AI-empowered advertising campaign automation tools. These components of Meta's renaissance are instructive, not just for considering the company's future, but the prospects of all companies in the social media advertising category. As such, I think the fortunes of other social media and advertising platforms broadly will be dictated by how well they can replicate the changes that Meta implemented to its own core advertising infrastructure. This podcast is structured in three parts. First, I'll describe the modern structure of Meta's advertising platform and why it was so vulnerable to the restrictions imposed by Apple's app tracking transparency privacy policy. Second, I'll describe the changes that Meta implemented to navigate those restrictions and how they complement each other. And third, I'll offer my thoughts on how Meta can continue to bolster its Advantage Plus tool suite to deliver value to advertisers, especially through the deployment of AI tools like Generative AI for ad creative production. I'll explain why I believe Meta is uniquely positioned to deliver value to advertisers from Generative AI and present what I believe is the company's product roadmap for doing so over the next several years. The sponsor of this week's episode is Incrumental. Were you looking for an alternative to attribution and went down the route of MMM, but after a four month integration, the results were confusing and unhelpful? Have you thought of Incrumentality measurement? Incrumental is the first and only, always on, Incrumentality measurement platform that doesn't rely on user level data, planned experiments, or MMM. The platform is trusted by some of the top mobile marketers. Check out the case studies section on their website to learn why. Join the future of marketing measurement with Incrumental. Google Incrumental, the future of measurement, or use the link in this episode's description to learn more about why it should form part of your marketing tech stack. Meta operates what I've called a hub and spoke advertising platform. In this configuration, Meta serves as a hub to its advertising clients, selling inventory that it directly operates in its Facebook and Instagram apps, and receiving what are known as conversion signals from those advertising clients. In this way, Meta serves as what is known as a walled garden. In contrast with an advertising network, which brokers add sales between advertisers and publishers, thereby selling advertising inventory that it doesn't own, Meta sells its own and operated inventory directly to advertisers. The hub and spoke structure of Meta's platform stems from the fact that advertisers, or the spokes to Meta's hub, by advertising inventory from Meta, but also send data back to it. This data captures the actions that users take after clicking on an ad in Instagram or Facebook, once they've leapt one of Meta's apps. This data comprises the advertising feedback loop. What Meta learns about whether a consumer purchased or otherwise engaged with an advertiser's product after clicking on the ad for it. Historically, and I'll get to the more modern approach later in the podcast, Meta collected this data through mechanisms that it built, but that are integrated into advertisers products. Think of these mechanisms as transmitters or beacons, as they're often called. These were Meta's SDK or software development kit, which advertisers would integrate into their mobile apps, and Meta's pixel, which advertisers would integrate into their websites. Both of these beacons served and continue to serve a similar purpose, to observe the actions that users take inside the products of Meta's advertisers, and to transmit information about those actions back to Meta's advertising platform. Once Meta receives this data, it can be used to optimize advertising campaigns. These actions, be they purchases or website registrations or partial checkouts are referred to as conversions. Advertisers willfully and enthusiastically integrate these conversion beacons into their products. Advertisers gain meaningful benefits by letting Meta know when a user that received an advertisement on one of Meta's properties ends up being economically valuable. The first benefit is that Meta can credit that conversion to the campaign that delivered the user to measure campaign performance. This is called attribution. But the second and more significant benefit is that Meta can consider that user across some number of properties or characteristics, and use that knowledge to serve the advertisers ads to other users that match its profile. This is called optimization. Attribution facilitates what's known as marketing measurement, or the profitability accounting of an advertising campaign, and optimization facilitates targeting, or the improvement of audience definitions for a campaign based on advertising performance across groups of addressable users. By definition, targeting is downstream of measurement, without the ability to attribute campaign performance, and advertising platform can't optimize campaigns. Meta didn't invent this feedback loop, but it may have industrialized it. In July 2016, Meta introduced a product called App Event Optimization, or AEO, which allows advertisers to set bid prices for specific events that users can take on their websites or in their mobile apps. Meta uses its machine learning infrastructure to predict the completion of these events, thereby allowing users to be targeted on the probability of converting on advertiser properties, and exactly the ways that advertisers care about. This is called event-based conversion optimization, and it allows advertisers to manage conversion risk. I described the concept in a recent episode of the podcast called Understanding App Loven, and in a piece I wrote last year called Understanding Conversion Optimization in Digital Advertising. I'll quote from that piece. When an ad platform or network allows conversion-based optimization, it effectively tells the advertiser that it will only charge them for the successful delivery of those events. This is obviously attractive to the advertiser, which can determine how much any event is worth to them, bid up to that amount, and only pay the platform when those events transpire. This is essentially risk-free money if the advertiser feels confident that its bid pricing model is statistically reliable and rigorous. Each time the advertising platform delivers some event, the advertiser pays that platform less than what the event is worth to them. End quote. AEO is effective at allowing advertisers to target the highest-value users on the basis of conversion probabilities. In May 2017, a little less than a year after AEO was introduced, I wrote a piece titled Facebook's app event optimization tool Showcases the Power of its data in Q1 earnings. In that piece, I hypothesized that AEO may have explained the substantial increase in ARPU that Facebook, which was the company's name at the time, experienced in the US, given the ARPU divergence between the US and Canada since the introduction of AEO and the likely overrepresentation of AEO targeting in US audiences. That year, in 2017, Facebook released a tool that abstracted advertiser outcomes further, called Value Optimization, or VO. VO took value-based bidding a step further by removing the need for advertisers to bid at all. With VO campaigns, advertisers don't submit bids, but rather return on investment or return on ad spend requirements. This alleviates the need for the advertiser to derive proxy metrics for customer values, which could be obfuscated through the curse of dimensionality. Instead, advertisers would simply input and expect ROI on a one-day or seven-day basis, and let meta target audiences in pursuit of those goals. This form of targeting is attractive to advertisers for a few reasons. First, many advertisers lack the infrastructure to calculate the value of specific events. Meta has much more data than does any advertiser, and it can better determine how much, for example, an ad to cart or registration should be worth to a given advertiser than the advertiser can. But potentially more meaningfully, VO campaigns mostly obviate the need for audience level targeting. With VO, Meta can simply consider users as a function of their behavioral profiles and not through demographic features to consider their conversion probabilities. Advertisers only have so many options for targeting users on a platform like Meta's. They can either upload lists of specific users by some identifier, like an email address or mobile device ID, but more on those later, or they can target against the demographic properties that those users self-certify, or they can target against observed interests. The broader a targeting group in terms of the bid that an advertiser sets to reach it, the larger the summed distance between the average value of that group and the intrinsic or observed value of any individual. One might conceive of this as the mean squared error in a regression model. This is especially true since most users are economically worthless to any advertiser, and the average value of a group is often mostly determined by a very small minority of valuable users, which is a concept that I unpack in a piece titled Digital Advertising, Demand Rooting, and Millionaires Mall. As VO, Meta untethered targeting from these groups and simply targeted individuals based on individual conversion probabilities, with campaigns being delivered and scaled against ROI targets. I think it's important here to unpack the idea of the Millionaires Mall, which is a thought experiment I posed in the piece I just mentioned. I adapted this thought experiment from a lecture that Nassim Teleb gave at Cambridge University, and it goes like this. Imagine you enter an utterly generic shopping mall somewhere in some nondescript American city. You walk up to the information counter, and before you say anything, the person behind the counter tells you an interesting fact. The average net worth of all shoppers in the mall at that moment is $50 million. In credulous, you push back. That seems impossible. The clerk says that it is indeed unlikely, but that they know with total certainty that the average net worth in the mall is $50 million. Now, it just so happens that you're a yacht salesman and you're ecstatic. You've somehow lucked in defining yourself in a target rich environment. Everyone in the mall can afford one of your yachts on average. You set up a small stand, think of a lemonade stand, at the entrance to the food court, and wait. But hours pass, and no one expresses interest in your yachts. Why? The question relates to probability distributions. Remember, the mall is not located in Natherton, or Stanford, or Beverly Hills. So is it more likely that the net worth of the shoppers in the mall follow a normal distribution with $50 million at the mean? Or is it more likely that the shoppers net worth are clustered around the US average, or the average for that state? And one billionaire simply craved a synabon that morning and skews the average much higher. Assuming that the latter is more likely, the yacht salesman will have less luck with a lemonade stand for yachts than with hunting down the Bentley or the Maybok in the mall parking lot and leaving a personalized note under the windshield wiper. My millionaire's mall thought experiment maps to the digital advertising workflow. The internet is the millionaire's mall. The internet is so vast that millions or tens of millions of relevant customers for any given product inhabit it. But they are not concentrated in one place at one time, meaning they're not all playing one mobile game or browsing one website or even using one social media apps simultaneously. And even if a long enough time period is considered such that all the relevant people would be observed, for instance, a popular social media app over the course of a year, those people can't be reached economically if they aren't targeted specifically. An advertiser can't buy all the inventory on a social media app over the course of a year, just because every one of their potential customers would see it, since that would entail buying a lot of inventory that was exposed to irrelevant users. Workable economics of digital advertising generally require targeting, reaching relevant users through prior knowledge of their predilections for purchasing certain types of products, either observed or inferred. Meta's implementation of event and value-based algorithmic bidding automated this individual level targeting without requiring advertisers to build the machinery needed to understand how certain behavioral patterns or characteristics correlated with intent to purchase. This represented a significant milestone in digital advertising, and it dramatically altered the day-to-day work of digital marketers. In April 2013, I wrote a piece titled "The Changing Face of Mobile Marketing with Event Based Bidding." It's important to note that, even with algorithmic bidding, targeting doesn't go away. It just moves up and down the funnel into places that it can preempt or assist the delivery of algorithms. Creative is the new targeting domain. Add Creative Now Determines which demographic groups are most likely to click, install, pay, etc., via relevancy. In this new model, the marketer has essentially shifted their focus to the inputs, which are add creatives, and the signals, which are events, that produce the desired level of return on ad spend. End quote. I believe that it was this improved form of targeting and advertising optimization that accelerated the growth of the direct consumer category and invigorated the mobile app economy. For instance, App Store Developer Payouts, which serve as a proxy for overall app store revenue, increased by 31% in 2018 after VO was introduced. Metas AEO and VO tools were purpose built for the kinds of direct response advertising pursued by e-commerce retailers and mobile game developers. I'd argue that Metas' introduction of these tools had as consequential of an impact on the internet economy as industrial tailorism did for manufacturing. And certainly, it was as controversial. But event in value based bidding worked phenomenally well at pairing advertisers with relevant consumers until it didn't. On June 2020, at its annual WWDC Developers Conference, Apple unveiled a privacy policy for its iOS operating system called App Tracking Transparency, or ATT. ATT allows users to choose to expose their unique device level advertising identifiers, or IDFA's, to specific apps or not, through an opt-in prompt. When a user opts out of ATT for a given app, their device identifier is not available to that app. But crucially, the restrictions of the broader ATT policy apply to more than just the use of the IDFA. ATT stipulates that an opt-out should prevent an app from transmitting any personally identifiable information about the user, including network data like their IP address, to another party for the purposes of advertising targeting. I've written extensively about ATT on mobile def memo, and will point users to the blog to understand the details of the policy if they're unfamiliar. But broadly, ATT seemed precision design to disrupt advertising platforms like Metas, which aggregated user-level data through its advertising partners, the hubs, to its spokes. I wrote about Metas' exposure to ATT in January 2021, ahead of the rollout of ATT, in a piece titled "Facebook may take 7% revenue hit from Apple privacy changes." I'm quoting from that piece and jumping around between paragraphs. The events stream for both app and web advertisers is being broken with the ATT prompt for users that opt out of tracking. This means that Facebook will no longer have full transparency into what users do in apps or on websites once they click on an ad in Facebook Blue or in Instagram. Instead, Facebook will receive a very limited amount of interaction data from the advertised property. Facebook's ad targeting and optimization is driven primarily by the event stream it receives from apps and websites. And when that event stream is broken on iOS, ad targeting will be hindered considerably. I estimated in that piece that Facebook would face a 7% revenue headwind at the nadir of ATT's impact in a base case. That estimate proved to be conservative. My later analysis, such as for Metas' Q1 2022 results, put the impact at more than 10%. But the idea that ATT's impact would reach a low point and then recede as I proposed in that piece is important to unpack. Opposited in that article that Meta would recover signal through various means over time, ameliorating ATT's restrictions. ATT was introduced to iOS in April 2021 through iOS 14.5 after being delayed from the iOS 14 release. But ATT wasn't present on a majority of iOS devices until June 2021. Metas stock price reached its local peak shortly thereafter in September 2021, before beginning its long decline. But given the lead time on the release of ATT from when it was announced, Meta had certainly begun engineering its adaptations earlier than that. I'll speak to those shortly, but before I do, it's important to address COVID as a confounding factor in understanding the effects of ATT. The Internet Economy booms during COVID. The annual retail trade survey, or ARTS, conducted by the US Census Bureau, found that e-commerce sales increased by $244 billion or 43% year-over-year in 2020. Meta and other digital advertising platforms were obvious beneficiaries of that. And the end of COVID would similarly, by logical extension, suggest a decline in e-commerce spend that would arrest that acceleration or even lead to revenue declines for those platforms. I attempt to disentangle these concurrent effects, the simultaneous headwinds of ATT and the end of the COVID engagement and spending boom. In a piece I wrote in January 2023, titled "The App Tracking Transparency Recession." While the piece isn't specific to Meta, I argue that the combined impact of a return to pre-COVID behavioral consumption norms coupled with the impact of ATT had catalyzed a recession in the social media advertising market. And I pointed to the differences between the outcomes of the social media advertising platforms that were most exposed to ATT and other digital advertising channels that were less exposed to ATT or not exposed at all, like Amazon, to quantify the COVID effects. This wasn't a scientific study by any means, but I think it's helpful in teasing apart the consequences of both forces. But perhaps what best explains the impact of ATT on Meta is the degree to which its small business client saw advertising performance decline in its wake. It's important to note that when I speak of advertising targeting, it is mostly in the context of a practice called "Direct Response Marketing," where an advertiser exposes an ad to a consumer with the intention of prompting an immediate direct response, like a purchase. This differs fundamentally from brand advertising, in which a company exposes an ad with the intention of building awareness or affinity. This is a broad topic, and these definitions may be contentious. I attempt to delineate between these two tactics, direct response and brand marketing. In a piece I published in 2021, titled "The Parallelist Mythology of Brand Marketing for Digital Products." I've quoted extensively from that piece in previous podcast episodes, so I won't do so again. But the general concept is that brand advertising tends to be deployed by larger businesses, which makes intuitive sense because they have existing brands to strengthen and support with advertising. Direct response advertising, by total dollar spent, is comprised primarily of small and medium-sized businesses, or SMBs, which makes intuitive sense because they are relatively more of them than large brands. Meta is principally a direct response advertising platform, and my belief is that its advertising revenue is generated in the majority by SMB advertisers. I have no way of knowing this for certain. The company doesn't break out revenue by brand versus direct response in 10, but it's possible to use certain data points to develop support for that hypothesis. First, Meta often points out that its ad platform serves 10 million advertisers. Sheeranumerical logic would dictate that the majority of these are SMBs. There simply aren't many advertisers that possess a promotion-worthy brand. While this doesn't speak to the proportion of spend across brand and direct response advertising, at the very least, it supports the notion that SMBs have the dominant advertiser class on the platform. But second, and more concretely, the consequence of a large-scale pullback in brand spending on Meta can be quantified. This is known because in July 2020, more than a thousand advertisers joined a boycott of Facebook's ad platform, organized by a collection of civil rights groups. The boycott was in response to a perceived lack of vigilance by Facebook at combating misinformation in its social media products. Pathematics, a social media analytics firm, estimated that Facebook's top 100 advertisers spent 12% less over the course of the boycott than that group had the year prior. And yet, Facebook delivered 22% revenue growth in that quarter. I detail this boycott episode and the underlying economics of direct response advertising in a piece titled "The Big Economy of Small Advertisers." I'm quoting from that piece. Small businesses generally pursue conversions through advertising, not reach. And while Meta has mostly adapted to ATT, the pain and experience was a function of the company's overreliance on an identifier it couldn't control the IDFA, for use in ads personalization. Broadly speaking, small businesses strive to reach high-intent relevant users with digital advertising for the purpose of delivering purchases. Large brands, by definition, can target their products broadly. Brand advertisers mostly pay for reach with their digital advertising campaigns. All things equal, an ad platform should prefer to serve as small businesses through behavioral targeting and ads personalization over large brands that mostly don't rely on that for a few reasons. One, reach is fungable and can be found in many advertising channels and formats, making it something of a commodity. The ability to target preferences and purchase proclivities is relatively rare and thus commands a premium. Two, brand advertising spending is susceptible to brand safety concerns that can arise from current events. Three, the money multiplier of performance marketing in the compounding nature of cohort math allows for direct response advertisers to grow their budgets in direct response to efficiency. Whereas brand advertisers do not generally set budgets dynamically and revisit channel allocation much more slowly. And four, as a result of second price auction dynamics, advertisers targeting the broadest and most competitive segments of traffic only generate the delta of the second and third highest bids for the platform as incremental revenue when they outbid others for inventory. But ads targeted to smaller, more niche and better-defined audiences can generate net new revenue and targeting inventory that might otherwise have gone unsolved. If it is assumed that the largest advertisers during the boycott were predominantly brand advertisers and a pullback and spend from the largest advertisers had a mostly negligible impact on meta's growth, then what happens when direct response advertising, again, assuming that it's comprised of budgets from SMBs in the main, is disrupted. The initial impact of ATT provides helpful guidance. Evaluating the impact of privacy regulation on e-commerce firms, evidence from Apple's app tracking transparency is a title of a recently updated working paper from academics at Northwestern University, Columbia University, UCLA, the University of Maryland, and the University of Hamburg. This paper explores the repercussions of ATT from the perspective of nearly 25,000 e-commerce advertisers by comparing the performance of high exposure advertisers or those most dependent on meta as an advertising platform for revenue to low exposure advertisers. The study takes into account advertisers' budgets across iOS and Android, which isn't affected by ATT. Because both dependents on meta and share of iOS are not exogenous factors, the paper constructs a model that includes firm level and monthly fixed effects from panel data. The paper finds that the introduction of ATT was associated with, quote, a decrease in revenue of 37.1% for more meta-dependent firms relative to less meta-dependent firms. And that quote, "This effect is driven by small firms." End quote. One of the regressions which isolates small firms associates the introduction of ATT with the 68% marginal decline in the dependent variable, which is logged revenue. Another regression which doesn't isolate firms by size associates the introduction of ATT with the 21% decline in log to transactions. I've linked the paper in the show notes for those curious to learn more. To recap, the premise of this podcast follows this logical course. One, meta's platform was designed to deliver value to direct response advertisers through personalized advertising, achieved through aggregated offsite behavioral data. Two, meta's advertising revenue is contributed in the majority by direct response advertisers, and the largest segment of that group by revenue is SMBs. And three, Apple's ATT privacy policy is principally responsible for the revenue slowdown and, in some cases, decline that meta-experienced from Q1 2022 to Q1 2023 inclusive. If this premise is accepted, the next obvious question to ask is, what did meta do to turn things around? How is meta trading at an all-time high stock price, with revenue growth of nearly 20% in Q3 2024, if ATT systematically, fundamentally, and permanently compromised the integrity of meta's hub and spoke advertising platform? In other words, if ATT demolished meta's business model, why is the company thriving three years after ATT was first introduced? This is a question I field often. The answer is multifaceted, which makes it somewhat unsatisfying. But as a case study in corporate strategy and repositioning, meta's adaptation to ATT is fascinating and inspiring. Because meta navigated the challenge of ATT through a wholesale to the stud's reconstruction of its advertising platform, executed in parallel, with the total overhaul of the engagement model of its two flagship apps, Facebook and Instagram. In other words, ATT did kill Facebook. The operating model of the company that existed in 2021, across not just its revenue engine, which is the ads platform, but also its consumer-facing apps, has given way to something totally new. Facebook adapted to ATT by reinventing itself, with a corporate imperative to place artificial intelligence at the center of all of its interlocking initiatives. That reinvention has taken place on three axes, engagement, conversion measurement, and campaign optimization. The first axis of reinvention is with meta's consumer-facing products. The impediments to targeting imposed by ATT rendered ads less relevant on the platform. This would have resulted in lower ad prices and lower click-through rates. And this was the case, as I've chronicled in my quarterly analyses of meta's earnings. Metas average price per ad dropped by 8% in Q1 2022, from 6% growth the quarter before, and it was negative through Q4 2023. In July 2022, as rumors circulated that meta might transition its Facebook and Instagram products. products from a friend's graph to an open graph. I wrote a piece titled Unpacking Metas Pivot to an Open Graph and Short Form video in which I enumerated the four opportunities for advertising revenue growth available to a social media platform. I'm quoting. One, increase ad load or the ratio of ads shown to each user per session relative to organic content. Two, increase reach or the number of users that engage with the product and thus are exposed to ads. Three, increase the value generated by ads through higher quality formats or better targeting, which improves the general price paid for ad inventory through increased bids from advertisers in the ad auction. And four, increase time spent on site, which provides more opportunities for ads to be served. With its ability to aggregate user level data curtailed, Meta couldn't increase the value generated by ads. And while Meta's daily active people metric, which counts the daily, de-duped users of its consumer-facing apps, has risen more or less consistently to reach $3.29 billion in Q3 2024, the company doesn't have much of the internet-connected world left on board. Similarly, Meta has increased ad load marginally since ATT was introduced, but ad load increases tend to result in user churn and ad platforms are reluctant to institute them. That leaves time spent. I posted it in that piece that Meta was transitioning to an Open Graph with a focus on short form video content in the form of its real product, because that would keep people in its apps for longer. And Meta could compensate for a decrease in ad prices if it simply served more ads. Since ad load and reach were mostly dead ends in terms of achieving that, engagement was the most obvious place to find those incremental impressions. The number of impressions served can increase if users spend more time in the app, even if ad load, which captures the number of ads served per some unit of time, remain the same. Reels is Meta's short form video content product. It presents short form user generated video in the Instagram and Facebook feeds. Meta's transition to short form video content in its flagship apps is visible and obvious to users. As is the fact that Reels are surfaced from the entire corpus of Meta content as opposed to being exclusively sourced from user's friend graph. But what is taking place below the surface of the Reels product is potentially more meaningful for driving increased engagement. And it is a component of Meta's overarching AI narrative. Meta has invested heavily into the infrastructure needed to classify and recommend relevant video content. In March 2024, that Morgan Stanley's TMT conference, Tom Allison met his head of Facebook said the following, and I'm quoting. And so we actually created an advanced technology group. They are kind of home within the Facebook organization, but their charter is to build the best content recommendation system in the world that can power all of our recommendation products, whether it's in Facebook, whether it's in Instagram, whether it's in threads. We said, well, what would this look like in recommendation space? What if we actually had instead of these per-product recommendation models? What if we had one recommendations architecture that could power all of our recommendations products and that could leverage lots of data? And so we put Facebook Reels on this new model architecture. It used the same data as the previous model that was on GPUs, but we got roughly an 8% to 10% gain in Reels watch time. So what that told us was this new model architecture is learning from the data much more efficiently than the previous generation. Now look, all of this also requires a bunch of kind of hardware investments in planning, right? So in addition to this, we're kind of like, frankly, like reconfiguring data centers, figuring out how to wire more GPUs together. This is a big part of what's going to push model development in our generative AI work, but it's also similar to recommendations. Recommendations data is actually, in a lot of ways, a lot larger of a data set than even some of the large language models used because you're looking at all the interactions of billions of people every day. End quote. So AI is a central component of Meta's renaissance from a consumer facing angle. Reels are sourced from the entire pool of content across Facebook and Instagram. But because decomposing and classifying video content is more computationally complex than doing the same for text-based content, the company needed to invest in the infrastructure to do that at scale. According to Meta, the company's AI clusters possessed the compute power equivalent to 600,000 H100s at the end of 2024. One use case for this was building the Reels recommendation system, which should account for more time spent, given one, the broader pool of content from which to draw, given the depth of the open graph, and two, the improved ability to curate that large pool of content for user-level relevance. It's worth noting here that bringing Reels into the feed was a risk. As a new content model, Reels is suitability for delivering ads wasn't unknown, and targeting logic would need to be tuned to accommodate it. When Reels was first introduced and monetized at a lower price point than standard feed ads, an actively cannibalized feed revenue. Mark Zuckerberg noted this in the company's Q2 2022 earnings call, but affirmed the company's determination to grow Reels monetization, given its strategic importance. I'm quoting from that earnings call. Reels doesn't yet monetize at the same rate as feed or stories, so in the near term, the faster that Reels grows, the more revenue that actually displaces from higher monetizing surfaces. Now in theory, we can mitigate this short term headwind by pushing less hard on growing Reels, but that would be worse for our products and business longer term, since we're confident that Reels will grow engagement overall, and quality will eventually monetize closer to feed. In Metas Q4 2023 call, Mark Zuckerberg revealed that Reels had become a net contributor to Revenue. While the units monetization had improved over time, it still underperformed the feed unit, but the increased impressions it delivered compensated for the displaced revenue. This dynamic is apparent in the chart of Metas advertising economics that I publish every quarter. Impression growth peaked in Q2 2023 at 34%, but remains positive, while declines in average price per ad reached a near-deer in Q4 2022 at negative 22%. Flipping to growth in Q4 2023 at 2% and climbing from there, reaching 11% last quarter. So along this first axis of reinvention, engagement, the combination of Reels, the improved video classification and relevancy scoring unlocked by Metas investments into AI infrastructure and consolidated model development, and the augmented pool of content available for recommendation in the transition to an open graph. Meta has attenuated the targeting challenges imposed by ATT by simply showing more ads in more engaging core content and learning how best to pair the two. The second axis of reinvention is advertising measurement. It's important here to once again distinguish between measurement and targeting. Measurement is the process by which an advertising channel receives feedback on the ads it served, whether they resulted in a conversion or not. In September 2021, a few months after ATT reached majority scale on iOS, Facebook stated in a blog post titled Navigating Change and Improving Performance and Measurement that it believed it was undercounting the conversions its ads were delivering for advertisers on iOS by about 15%. Although, quote, "There is a broad range for individual advertisers. We believe the real world conversions like sales and app installs are higher than what is being reported from any advertisers. We are committed to helping you better measure those outcomes and improve your performance." As I stated earlier, measurement is upstream of targeting and optimization. If a platform can't measure the conversions it is delivering for advertisers, it can't adjust campaign parameters automatically on the basis of performance. This 15% gap in measurement would result in a more substantial friction on ad spend because it has knock-on consequences. When an advertiser, but especially an SMB with limited internal measurement capabilities, sees a decline in its conversions in a platform's reporting, it will naturally reduce its spend there, concluding that its performance has degraded. This could happen even if the advertiser intuitively believes that the reporting is underestimating the true volume of conversions, simply because many SMBs are wholly dependent on platform reporting for making budget decisions. In order to address the shortfall, meta invested in measurement methodologies that could help it close the gap between what it believed it was delivering, I say believed because absent the deterministic IDFA, it couldn't really know, and what it could credibly report. In order to achieve this, meta has built a portfolio of measurement tools that I believe is designed to deliver redundancy. If any given approach is deemed non-compliant by either Apple or some future law, meta has others on which it can rely. The first of these methodologies is the conversions API, or Cappy. I provide a history of meta's Cappy tool, which was launched in 2020, and a piece titled is Cappy Future Proof, but in short, Cappy was originally introduced to address a different privacy intervention by Apple, which was a change to its intelligent tracking prevention policy that limited the lifetime of first-party cookies where click IDs are detected in referral links, which applies to Metas. To sidestep this, meta built Cappy, a server-to-server transmission mechanism that allows advertisers to relay conversion data outside of the browser. Cappy requires an integration, but since it was introduced to serve web advertisers, many e-commerce platforms such as Shopify created one click opt-in mechanisms. When ATT was introduced, meta expanded Cappy to app advertisers and built what it calls the Cappy Gateway, which reduces the amount of engineering resources required for integration. Metas Cappy accepts a number of different identifiers that can be used to match a conversion to a user in its own user base. Such as the user's email address, the click ID attached to a specific ad, the user's phone number, The advertisers unique identifier for the user or the user's IP address. Meta had at one point noted that the user's IP address is a high priority input for its matching logic, as I noted in a screenshot in a piece titled Meta's AEM update and the disappearing IP address. The Cappy documentation no longer lists the IP address in its hierarchy of matching identifiers, which is what I had screenshoted, although the IP address is still noted as an identifier in the Cappy documentation. The degree to which these identifiers are used for matching would be closely held within Meta. I don't know the extent to which they are used for users that have opted out of ATT. Even without them, the expansion of Cappy should have improved Meta's measurement capabilities, that is, its ability to account for the conversions delivered by a campaign. Meta may accept some or all of the identifiers I noted above for measurement without attaching that knowledge to individual profiles. That is Meta might, and again, I don't know, and I am speculating. Attribute conversions to campaigns on the basis of individual identifiers without aggregating that data at the user level for the purpose of building behavioral profiles. Another methodology that Meta has introduced since ATT reached majority scale is aggregated event measurement, or AEM. AEM was initially described as something of a web-based alternative to Apple's SK ad network framework, which app developers can use to send campaign-level conversion data to ad channels without revealing individual user information. AEM was first announced in a pre-Christmas update to advertisers in December 2020, but support for in-app events was added later, and an overhaul of the framework was announced in May 2023. This overhaul included a setting that allows app measurement companies, known as MMPs, to receive conversion events even when a user has opted out of ATT in an app. Some MMPs refer to the setting as advanced data sharing, and it can include the user's IP address and other device settings. As it exists now, in its modern incarnation, AEM serves as a vastly more functional SK ad network, providing near real-time reporting, facilitating more instrumented campaign and event identifiers, and processing events for opted out users. Again, the extent to which this data is aggregated and utilized can't be known, but even if it simply augments what can be attributed to the campaign level, it would represent an enormous improvement over SK ad network, which has mostly been abandoned by app advertisers as a measurement methodology. The third axis of reinvention relates to campaign optimization. This is captured in Advantage Plus, met a suite of AI-empowered campaign automation tools. Advantage Plus can be thought of as the natural evolution of VO, except that it applies to all components of an advertising campaign, targeting bid setting, budget setting, and, increasingly, advertising creative production. Advantage Plus merely requires that advertisers upload creative assets and specify certain performance targets. It optimizes all other decisions in an automated way through rapid, exhaustive testing, with the ostensible goal of finding the best possible permutations of all campaign settings to maximize performance for the advertiser. Advantage Plus is what's known in advertising parlance as a black box system since it manipulates campaign settings without input from the advertiser. An Advantage Plus has counterparts at other advertising platforms. Google has Performance Max or PMax. TikTok has Smart Plus. Pinterest has Performance Plus, etc. There is important context to acknowledge when considering the development of Advantage Plus. First, it's neither a new nor novel approach. Google in 2017 merged all of its search, display, and YouTube app install campaign settings into a single new campaign type called Universal App Campaigns or UAC. I wrote about UAC at the time in a piece titled, "Understanding Google's Universal App Campaign Changes." And I'm quoting, "The general crux of most complaints about this change is that UAC removes an advertiser's ability to control targeting parameters. UAC provides reporting on creative performance and demographic resonance, but these can't be influenced for a campaign since the UAC system creates these user segments without any input from the advertiser. In other words, UAC tries a lot of things and tells the advertiser what works, but the advertiser can't tell UAC what to do." It seemed clear in the years after Facebook launched VO that it would introduce something similar. In a piece I published in 2019 titled, "What Comes Next After Facebook's VO campaign strategy," I wrote, and I'm quoting, "Given the thread that runs through Facebook's contemporary ad product releases, automated creative variation, automated value targeting, automated budget allocation, it is hard to believe that these systems won't be combined into a UAC-like black box at some point. The UAC approach is simply so much more lucrative for an ad platform. It controls where and how budget is allocated and thus could potentially, and cynically, provide advertisers with just enough performance to justify continued spend, but no more. It's obvious why the black box, totally automated ad platform paradigm, is beneficial to platform owners. It virtually guarantees maximized profits as the platform strives to deliver minimally to a goal, and it also helps small advertisers get onboarded very seamlessly. And it seems almost inevitable that Facebook would evolve its platform in this direction, because where else can it go? The Facebook UAC is just a combination of dynamic creative, automatic placements, dynamic budget optimization, and AEO with a cost cap or VO with a min row ad setting. It's fairly easy to make the leap from a campaign with these settings explicitly selected to merely a campaign." Advantage Plus serves exactly this purpose, but it applies to all campaign types, both Web and in app. The idea is that Advantage Plus optimizes campaign performance, but only to the standard demanded by the advertiser, and in this way, it maximizes ad spend. As I noted in the article I just quoted, the goal is to give the advertiser just enough performance to satisfy their requirements, but to ensure that as much budget was deployed as possible in meeting that goal. This is not to say that advertisers should be cynical about these tools. The reality is that these tools also test more exhaustively and can utilize more data than any single advertiser can. For the vast majority of advertisers, it's likely that these black box tools produce results that are far better than what the advertiser could have achieved on their own, at least over some initial timeline. But it's also true that the advertiser can't know if every dollar of their budget was deployed efficiently. The advertiser sees total campaign performance and not the marginal performance of each dollar spent. I've described this dynamic as a dilemma for advertisers. Some advertisers might incur an efficient spend when using Advantage Plus, that the otherwise wouldn't, but that waste is averaged into campaign performance that meets their standards. And the reality is that many advertisers are likely better off on net with access to these technologies than they would be if left to engineer those settings themselves. By offering end-to-end automation with Advantage Plus, Meta not only likely improves the performance of its advertising clients on net with the global average across its client base, but it also maximizes spend across the portfolio. In the expansion and development of Advantage Plus coincides with the advancements made in classification and relevancy scoring for reals. That is to say that Advantage Plus is a beneficiary of Meta's AI investments. The reality is that scoring content for relevance fundamentally optimizes for the same thing as scoring ads for relevance, which is the likelihood of engaging. Of course, the form engagement takes for each use case is different. With an ad, a useful engagement is a click and some subsequent action on the advertiser's property. With reals, a useful engagement is the continued consumption of reals content. As reals get better targeted, in the sense that they better match users' interests and preferences, so too should the ads serve by Advantage Plus. And while Advantage Plus has seen the consolidation and compounded benefits of bringing all aspects of ad deliverability into one tool suite, empowered by AI, there remains one optimization opportunity that I think will fuel Meta's growth going forward. AI enabled advertising creative production. Meta introduced its AI sandbox as part of the Advantage Plus suite in May 2023. The AI sandbox initially consisted of simple, automated image-authoration tools, text variations for ad CTAs, background generation that allowed an advertiser to create multiple variants of an ad automatically, and image outcropping, which automatically adjusted image ratios for different static image ad units. Meta expanded this offering in May 2024 with automated image variation, which changed more aspects of existing ads than just their background, and prompt-based image generation. And in October, Meta announced that it is testing a tool that creates animations from static images, and that can automatically adjust image ratios for existing video ad units. In July 2023, I published a piece titled "Generative AI for Ad Creative," five value milestones, in which I outlined what I saw as the five ascending waypoints for generative AI as applied to ad creative production. Meta has progressed through these waypoints so far in the expected order with the next natural stop being video variant production. In September 2024, Meta revealed that more than 1 million advertisers had used its generative AI tools for ad creative production. This is an astounding rate of adoption, given that it represents 10% of Meta's overall advertiser client base. And while these tools are doubtlessly valuable and helping to reduce ad creative production cost, the real value in generative AI for ad creative lies in improving performance. Just as Advantage Plus can experiment more thoroughly than most advertisers can. Generative AI for ad production can unlock patterns and themes and otherwise deploy more conceptual templates than most advertisers can. It's when generative AI creates concepts, not just variants, that these tools will showcase the power to expand budgets and grow ads-bend. The obvious destination for these tools is the real-time production of personalized ads that can maximize conversion in that moment, taking into account all-extent available information about the user, the product, the campaign, and the context on the page. It's in this deployment of generative AI that ad-creative reaches the theoretical limit with performance. And while generative AI tools are deflationary from a production cost standpoint, I've argued and I believe that they are inflationary when applied in this way to ad platforms. An advertiser has two levers for improving the delivery of its ads. The bid it is willing to pay for a conversion and the creative it exposes to users. With real-time generative ad-creative, every advertiser operates with peak, user-level efficiency in their advertising messaging. Every advertiser utilizes the best and most resonant ads for their product. That leaves the bid as the only other option for winning the advertising auction. And since meta owns its own inventory, it captures all of the upward pressure on ad prices. Note here that I'm not saying that generative ad-creative produces a conversion from every ad, only that each ad achieves its maximum potential for conversion. If meta's global conversion rate is 0.01%, meaning that every 10,000th ad results in the desired conversion, then doubling that number to 0.02% results in twice as much revenue for the platform and still means that 4,999 out of every 5,000 ads don't produce a bid against conversion outcome. And meta is uniquely positioned to capitalize on this. Even if it wanted to, meta can't share user-level preference data with other platforms, at least on iOS, Apple forbids that. And outside of Google, meta is probably the only platform with the scale of both advertising clients and users to justify the expenditure into this type of infrastructure. With more data and more compute than any other social media platform except YouTube, meta's capabilities with AI might present a formidable moat. This could have two profound effects. First, it's possible that performance marketers consolidate budget with meta even further, given the increased effectiveness of their dollars there unlocked with generative ad creative. If meta produces conversion rates at the upper limit while also relaxing the need to spend money on ad production, advertisers would likely see a performance premium there that couldn't be matched elsewhere despite the inflationary pressure on ad prices. Second, this development could fundamentally challenge the nature of brand marketing. If a brand is a set of image guidelines meant to maximize affinity within some group, how does the idea of every individual being its own targeting group change the contours of brand positioning? If meta can maximize the average responsiveness of groups while reducing group sizes ultimately to one, what is a brand but a personally tailored ad? Does brand advertising exist in a world where ad creative is generated in real time informed by all possible information to elicit the best possible response? And if it doesn't, does all brand advertising become direct response? As I noted in my 2025 predictions for mobile marketing, there is an ambiance, optimism, present, and digital advertising that strikes me as inescapable. The digital marketing sphere has already changed profoundly as a result of the application of AI. Metas renaissance is a testament to that. [MUSIC PLAYING]