Living in Walled-Gardens

Alphabet (i.e., Google), Amazon, and Meta (i.e., Facebook) will account for about 50% of all ad revenue in 2022. To be clear, in 2021 Google had $218 billion in ad revenue, Facebook $115 billion, and $31 billion: this $364 billion in ad revenue combined for 74% of digital ad spend and 47% of *all trackable ad spend* in the world. 

Living in Walled-Gardens

Tobias Konitzer

Living in Walled-Gardens

David Rothschild

Alphabet (i.e., Google), Amazon, and Meta (i.e., Facebook) will account for about 50% of all ad revenue in 2022. To be clear, in 2021 Google had $218 billion in ad revenue, Facebook $115 billion, and $31 billion: this $364 billion in ad revenue combined for 74% of digital ad spend and 47% of *all trackable ad spend* in the world. 

These big three concentrate on digital ads (both Amazon and Google also have beachheads in TV through operating systems), and digital ads have zoomed past TV in ad revenue. This digital revolution brought with it a promise of increased openness and innovation, but most critical for ad revenue: personalization. Personalization meant ads targeted to individual people, rather than a collection of demographics (e.g., David, rather than white, 35 to 49-year-old, male), paired with knowledge whether David bought the widget you are selling or not (attribution). Outside firms started to buy and build lists of people to target for various advertisers, companies sprung up to deploy ads, and test their effectiveness. It felt like the future was bright with innovation, an open exchange of information that would lead to a paradise where ads would be so targeted, so perfect that customers would actually love them. Imagine: advertisers not only get precision but can trace it to conversion, customers would get such awesome ads they would feel like they would have sought out this content organically, innovators reepings in rewards for news innovation driving this utopia, and platforms and publishers rich off of everyone’s joy.

This trend came to a halt in 2018 when Facebook banned all third-party data on its platform, in a preemptive bow to stringent new rules that allowed people to control and delete their data (see: General Data Protection Regulation in Europe and California Consumer Privacy Act in California). In tandem, Facebook built out further its look-alike capabilities: Companies could upload a “seed” list of current customers, and Facebook would find similar targets based on a mix of clues: demographics, historical engagement with various types of content, and other relevant contextual clues. The crux: Facebook had at its disposal not just revealed preference data from Facebook itself, but a whole litany of location data and global telemetry data not just of how individuals interacted with the platform, but how individuals used their cell-phones writ large, and where they went. Facebook ingested a ton of external customer data, but the extrapolated (read: broader) look alike audience can only be activated on its own platform. Data on this larger audience that is shared back to the companies is scarce (and has become more scarce over time). The operation model:  all-data-in-no-data-out: a walled garden. Out of these walled gardens, Facebook is particularly crucial to small and medium-sized business (SMB), especially with a tight geographical footprint, due its unique ability to target geo-located potential customers. 

Any mom-and-pop shop could upload a list of customers, and get in return a highly optimized list of targetable (remember: only in Facebook!) group of customers. At this point, Facebook’s look-alike algorithm works well because it has significant structured and unstructured data to work with – Facebook knows each individual user much more deeply than Google does, at least along certain dimensions. At this point, if you give the algorithm buyers of your product, it will confidently find more buyers of your product, and Facebook could price-dump such that customer acquisition cost (CAC) would become irrelevant to the buyers. Companies could afford flying blind in terms of long-term, repeated interaction, the lifetime value of the customers, and Facebook partly incentivized this behavior by pixeling technology – making it easy to track the on-site conversion, or even first purchase of this newly acquired Facebook audience, but no long-term trends.

But, problems were on the horizon for Facebook. As opposed to Google, which owns distribution channels as well as general digital infrastructure (Chrome browser, Android operating system for phones). Facebook just owns distribution channels but no infrastructure (heard of the Facebook phone, anyone?). The power behind its look-alike algorithm came from added telemetry data collected by usage pattern and intent data created on the device writ large (vs. the specific platform), but there is more to it: Pixeling technology allows Facebook to understand who buys in the moment, but cell-phone technology, allows Facebook to also understand what is bought later. You are on your phone, see a cool ad on your Insta, remember to check out the product, but go on to see other stuff. Later, you purchase the product. Mobile technology (through the Mobile Ad ID) then connects the purchase made with the ad exposure, and the algorithm learns further who buys and does not buy, which, in turn, validates the advertising pipeline and makes the targeting better. 

With Apple phasing out its version of the mobile ad ID (IDFA) in late 2020, Facebook immediately lost this ability. Now, it’s technology does not only over-index on short-term behavior, but it loses a ton of signal as to who even engages in short-term behavior good for the brand (i.e., a first purchase) and who does not. As a result, the look alike algorithm now targets folks with less precision as to whether someone will become a customer, and misses out on retargeting some folks who actually did become customers but Facebook did not see. And, as a result, for advertisers relying on Facebook: cost of customer acquisition goes up and the quality of the acquired prospective customers goes down.

In a world of higher customer acquisition cost, advertisers need to care how GOOD their customers are, and all of a sudden the over-indexing of Facebook’s pixeling technology on short-term behavior becomes a real liability. In the old days, if you paid less than your marginal cost of your cheapest product for a new customer: you were making money no matter what they bought. Now, you are paying 2 or 3 times more for each customer, and it is starting to be a big deal what your customer buys, and if they are lifelong loyalists (i.e., do they have a high lifetime value?). As a consequence, brands cannot afford to fly blindly anymore, as negative LTV:CAC ratios have become a reality.

In this new world, investing in the RIGHT customer, especially at the top of the funnel, is crucial and can be the make or break of profitability. But current customer value (LTV) prediction models predict the value of each customer by past transactions, and at that stage, no prediction can be made (because no purchase has occurred yet). Advertisers are flying blind, but now it is starting to hurt their bottom line, amounting to a huge, unsolvable problem for direct to customer companies.


Alphabet (i.e., Google), Amazon, and Meta (i.e., Facebook) will account for about 50% of all ad revenue in 2022. To be clear, in 2021 Google had $218 billion in ad revenue, Facebook $115 billion, and $31 billion: this $364 billion in ad revenue combined for 74% of digital ad spend and 47% of *all trackable ad spend* in the world. 

These big three concentrate on digital ads (both Amazon and Google also have beachheads in TV through operating systems), and digital ads have zoomed past TV in ad revenue. This digital revolution brought with it a promise of increased openness and innovation, but most critical for ad revenue: personalization. Personalization meant ads targeted to individual people, rather than a collection of demographics (e.g., David, rather than white, 35 to 49-year-old, male), paired with knowledge whether David bought the widget you are selling or not (attribution). Outside firms started to buy and build lists of people to target for various advertisers, companies sprung up to deploy ads, and test their effectiveness. It felt like the future was bright with innovation, an open exchange of information that would lead to a paradise where ads would be so targeted, so perfect that customers would actually love them. Imagine: advertisers not only get precision but can trace it to conversion, customers would get such awesome ads they would feel like they would have sought out this content organically, innovators reepings in rewards for news innovation driving this utopia, and platforms and publishers rich off of everyone’s joy.

This trend came to a halt in 2018 when Facebook banned all third-party data on its platform, in a preemptive bow to stringent new rules that allowed people to control and delete their data (see: General Data Protection Regulation in Europe and California Consumer Privacy Act in California). In tandem, Facebook built out further its look-alike capabilities: Companies could upload a “seed” list of current customers, and Facebook would find similar targets based on a mix of clues: demographics, historical engagement with various types of content, and other relevant contextual clues. The crux: Facebook had at its disposal not just revealed preference data from Facebook itself, but a whole litany of location data and global telemetry data not just of how individuals interacted with the platform, but how individuals used their cell-phones writ large, and where they went. Facebook ingested a ton of external customer data, but the extrapolated (read: broader) look alike audience can only be activated on its own platform. Data on this larger audience that is shared back to the companies is scarce (and has become more scarce over time). The operation model:  all-data-in-no-data-out: a walled garden. Out of these walled gardens, Facebook is particularly crucial to small and medium-sized business (SMB), especially with a tight geographical footprint, due its unique ability to target geo-located potential customers. 

Any mom-and-pop shop could upload a list of customers, and get in return a highly optimized list of targetable (remember: only in Facebook!) group of customers. At this point, Facebook’s look-alike algorithm works well because it has significant structured and unstructured data to work with – Facebook knows each individual user much more deeply than Google does, at least along certain dimensions. At this point, if you give the algorithm buyers of your product, it will confidently find more buyers of your product, and Facebook could price-dump such that customer acquisition cost (CAC) would become irrelevant to the buyers. Companies could afford flying blind in terms of long-term, repeated interaction, the lifetime value of the customers, and Facebook partly incentivized this behavior by pixeling technology – making it easy to track the on-site conversion, or even first purchase of this newly acquired Facebook audience, but no long-term trends.

But, problems were on the horizon for Facebook. As opposed to Google, which owns distribution channels as well as general digital infrastructure (Chrome browser, Android operating system for phones). Facebook just owns distribution channels but no infrastructure (heard of the Facebook phone, anyone?). The power behind its look-alike algorithm came from added telemetry data collected by usage pattern and intent data created on the device writ large (vs. the specific platform), but there is more to it: Pixeling technology allows Facebook to understand who buys in the moment, but cell-phone technology, allows Facebook to also understand what is bought later. You are on your phone, see a cool ad on your Insta, remember to check out the product, but go on to see other stuff. Later, you purchase the product. Mobile technology (through the Mobile Ad ID) then connects the purchase made with the ad exposure, and the algorithm learns further who buys and does not buy, which, in turn, validates the advertising pipeline and makes the targeting better. 

With Apple phasing out its version of the mobile ad ID (IDFA) in late 2020, Facebook immediately lost this ability. Now, it’s technology does not only over-index on short-term behavior, but it loses a ton of signal as to who even engages in short-term behavior good for the brand (i.e., a first purchase) and who does not. As a result, the look alike algorithm now targets folks with less precision as to whether someone will become a customer, and misses out on retargeting some folks who actually did become customers but Facebook did not see. And, as a result, for advertisers relying on Facebook: cost of customer acquisition goes up and the quality of the acquired prospective customers goes down.

In a world of higher customer acquisition cost, advertisers need to care how GOOD their customers are, and all of a sudden the over-indexing of Facebook’s pixeling technology on short-term behavior becomes a real liability. In the old days, if you paid less than your marginal cost of your cheapest product for a new customer: you were making money no matter what they bought. Now, you are paying 2 or 3 times more for each customer, and it is starting to be a big deal what your customer buys, and if they are lifelong loyalists (i.e., do they have a high lifetime value?). As a consequence, brands cannot afford to fly blindly anymore, as negative LTV:CAC ratios have become a reality.

In this new world, investing in the RIGHT customer, especially at the top of the funnel, is crucial and can be the make or break of profitability. But current customer value (LTV) prediction models predict the value of each customer by past transactions, and at that stage, no prediction can be made (because no purchase has occurred yet). Advertisers are flying blind, but now it is starting to hurt their bottom line, amounting to a huge, unsolvable problem for direct to customer companies.