The after party: Cynical resignation in Adtech's pivot to privacy

“Encrypted” identification solutions

Third-party cookie IDs have been a “key prerequisite in personalized/targeted advertising since all user-centric data is associated with [them]” (IAB Europe, 2020). The deprecation of third-party cookies is leading industry actors to seek “Universal” or “Alternative” IDs that identify users based on PII and first-party cookies and require those who identify consumers to follow certain norms that limit their ability to use collected data. In contrast to current third-party cookie identification, “Universal ID” solutions offer explicit opt-out mechanisms and potentially promise more transparency on the data being collected. The practicality of those alleged improvements, however, remains questionable.

The three primary identification architectures suggested by the industry are as follows: The Trade Desk's Unified ID 2.0, that is expected to reach 250 million people; LiveRamp's RampID, that is expected to cover more than 250 million people; and the Secure Web Addressability Network (SWAN), that is expected to include top-20 publishers and advertisers (Asim, 2021).

Trade Desk's Unified ID (UID) 2.0 aims to let sellers and buyers of online ads match encrypted user IDs in the bidstream that circulates information about impressions being sold in programmatic auctions. The proposed architecture requires users to login to participating websites via email or other PII. The login email is then encrypted and converted to a “UID2 token” that is refreshed periodically by the system. Those obfuscation mechanisms ensure that the email address cannot be traced back to the user, but do not prevent the matching of IDs to the same user by those who hold access to the raw data. Such a UID 2.0 identifier is likely to provide wide and persistent visibility of the user by the governing bodies (“UID2 Operators”) and participating bodies, like registered Demand-Side Platforms (DSPs) that match user IDs from the seller to the buyer side. Users can “opt-out” of the UID2 framework, but publishers and service providers may limit access to their products for users who decline to log in. Practically, the onus is on publishers to explain why they need an email address and how personalized advertising powers the “free and open web that users appreciate” (Titone, 2021). This “opt out” mechanism will only prevent publishers from creating UID2 tokens and block the ability of DSPs to bid on ads to show to those users.

For the SWAN solution, an individual is identified via a first-party cookie when they first visit a publisher site that has adopted the architecture (Thomson and Rescorla, 2021). Upon loading the publisher page, the user is presented with a pop-up asking for consent to show personalized advertising on the current site and other sites that have adopted the solution. Here, again, the burden is on the publishers to communicate the value exchange to the users. As part of the pop-up, the user also has the option to share their email address, which can serve as an identifier. Regardless of whether they share their email address, a first-party cookie is placed by the initially visited publisher site within the SWAN network that creates a pseudonymous identifier stored on the user's browser. The identifier is potentially based on PII, making it more deterministic than a third-party cookie identifier, and enables persistent identification of the user across all participating websites.

The LiveRamp RampID architecture is distinct from the previous two, in that it is interoperable with other ID solutions (Asim, 2021), making the identification instrument even more persistent since it can identify the user across all three ID solutions. The RampID uses user email addresses, upon login to publishers’ sites in exchange for content. Instead of placing a first-party cookie, LiveRamp matches IDs across partners (Asim, 2021). Once a RampID is created, LiveRamp provides cookie syncing across industry actors to match user IDs in DSPs, for example matching to The Trade Desk's cookies. Once a publisher's and a marketer's Ramp IDs are linked, the marketer can buy the user impression. The ability of this solution to further identify users in the ecosystem is also attached to other offline PII (phone number, address history), based on information that advertisers can match with first-, second-, and third-party data.

Each of these solutions creates an identifier that enables advertisers to track users across sites, replicating that capability enabled by third-party cookies. The cookie-less IDs rely on first-party data—users’ PII—passed to a centralized governing actor or through various page redirections (in the case of SWAN), that is then distributed to “registered actors” who can identify the user in the bid stream. Cross-site tracking and profiling is therefore enabled for “authorized partners,” who get users’ consent and conform to the contractual terms of these ID architectures. In that regard, little has changed—advertisers and ad servers are still gathering large amounts of user browsing history data for targeting purposes. Those proposed identification architectures are in fact designs of new tracking systems that come with self-regulatory controls of how the data should be handled and how the real user identity should be masked. Alarmingly, the solutions mostly use PIIs as identifiers, making user IDs more deterministic and potentially more persistent than third-party cookie IDs. Importantly, however, the new identification architectures do pose one tracking limitation: only those who are registered to use each solution will have cross-site visibility on users. By contrast, third-party cookies let any party involved in the ad auction gain visibility on individuals’ online behavior.

“Tidying” targeting through clean room solutions

There is a growing trend in adtech towards investing in technology that can exploit the first-party data that organizations own. As companies lose the ability to pair that data with other sources via third-party cookies, the digital advertising industry is turning to data clean rooms as a “privacy-preserving” method for maintaining profiling capabilities.

At a high-level, a clean room is a software environment that allows an organization to pair its first-party data with other sources, including other organizations’ first-party data and purchased third-party data from data brokers. The name highlights the objective of sanitizing cross-partner data pooling through mutually agreed upon privacy practices (IAB Technology Laboratory, 2023a). A data clean room does not allow participating parties to view joined data sets, preventing reidentification of users or the exposure of PII. Deterministic PII such as email address or a universal ID solution often serves as the identifier that can pair datasets together (Haggin, 2022; Hercher, 2023a).

The primary uses of data clean rooms take place from the buyer side and include: the maintenance of audience insights and segmentation for targeting, campaign measurement and attribution, and audience modeling and activation capabilities that can be used internally within an organization and exported for targeting across activation channels, such as display, search, or social media. These capabilities would allow for continued insight into behavioral and contextual signals for targeting, the creation of look-a-like audiences, and optimization based on user behaviors (IAB and Ipsos, 2023).

While many solutions fall under the label of data clean rooms, the different types of developers often have various goals and particular values to contribute to their customers’ digital advertising capabilities. Clean rooms developed by major platforms, such as Google's Ads Data Hub and Amazon's Marketing Cloud, are characterized as walled gardens, which allow platforms to leverage the significant amount of first-party data kept within the platform (Hercher, 2023a). These clean rooms may be bundled with these platforms’ other ad and cloud infrastructure offerings, and can be tooled in order to offer direct activation of audience segments and ad campaigns. These provide efficiencies and a powerful incentive for users to select them over alternatives. Platforms have come to define this technology as “privacy compliant” based on control over the flows of first-party data and ability to track user behavior across various locations within its own suite of products (Hercher, 2023b). Alternatively, clean rooms developed independently by parties that are not primary sites of activation allow for advertisers to store and match their data in a centralized location and export it for use elsewhere (Hercher, 2023b). This matching process carried out by non-platform parties is primarily done for the purpose of creating segments for targeting that advertisers can push to activation channels. Independently developed clean rooms highlight that it would be almost impossible to coordinate compliance across all parties involved in creating an interoperable output and have transparency into the various data flowing in and out of the clean rooms (Hercher, 2023a).

Since the practices in this type of privacy-preserving environment derive from different definitions of privacy and limitations of data usage, the employment of a data clean room does not guarantee compliance with laws or industry norms. Consequently, the IAB has been attempting to standardize elements of clean room services. It has introduced “Open Private Join and Activation” guidelines for data enrichment that occurs through the pairing of different first-party data sets (IAB Technology Laboratory, 2023b). This acknowledges that the solution is being used to match data sets using personally identifiable information, with visibility from the parties involved, to continue granular level targeting. Furthermore, this technology does not allow outside parties to see if the paired data contains information related to sensitive categories that violate privacy norms.

“Anonymous” ad attribution solutions

Attribution refers to the capability of linking ad views or clicks—called “source events”—to subsequent consumer behaviors, like a purchase or app download—called “target events.” Third-party cookies and tracking pixels currently enable this through persistent monitoring of identifiable users. These linkages let advertisers measure the effectiveness of ad campaigns and optimize accordingly to maximize ROI. In preparation for the post-third-party tracking era, Google, Meta, and Apple have publicized new attribution solutions that, according to company documents, “improve” user privacy by inserting sophisticated technologies into the mediation of these data flows (Apple, 2021; Nalpas and White, 2021; Taubeneck et al., 2022).

All three solutions promise to mask user identity during the generation and distribution of aggregated reports that link source and target events. Their designs are complex and some implementation details are vague. But, in general, they aim to reduce the ability of unauthorized actors to identify, track, and profile individual users, while still joining the records of events that are generated by a user's behaviors. They accomplish this via a suite of technical mechanisms. The Interoperable Private Attribution (IPA) system proposed by Meta and Mozilla, for instance, uses unique identifiers called “write-only match keys,” which are stored on and only readable by the local device, to record source and target events from the same user; the IPA system then generates reports by way of a multi-party computation method that disperses personal information and encryption keys across trusted servers, so that useful insights can be gleaned without allowing a single party to access complete records of user behavior; those reports aggregate data and use differential privacy techniques to obfuscate individuals within a crowd; finally, privacy budgets, which restrict actors’ abilities to query the servers, prevent adversaries from deanonymizing attribution data with brute force (Taubeneck et al., 2022). Google and Apple use some but not all of these same techniques, and their approaches lean heavily on special permissions granted to Chrome and Safari browsers as well as Apple iOS. For example, Apple's SKAdNetwork, which lets registered ad networks measure the impact of advertising on app downloads and engagement, is considered private by design and thus exempted from Apple's ATT mechanism (Apple, 2023). They also add delays to the production of attribution reports to further obfuscate users’ identities. Essentially, these companies claim that as long as the identifiers assigned to users are not accessible to third parties, and as long as the anonymous information associated with individuals is stored locally (on a device or browser) or cryptographically protected, then the data collection and analysis required to provide attribution services is privacy-compliant.

These solutions aim to maintain accurate reporting on the performance of ad campaigns while promising to conceal user identity and limit the data shared in attribution reports. Observing and influencing users, however, is still the priority; the goal remains to induce higher value consumer behaviors at lower advertising costs. While they improve certain information-security functions, like preventing data leakage, these “private” attribution solutions continue to service corporate demands for online surveillance. They use clever technological designs both to guard against adversarial abuses, and to protect the premise that advertisers, intermediaries, and platforms are all entitled to determine how effectively digital ads are performing.

“Privacy-safe” optimization

The last set of solutions examined here takes us into the black-box magic of optimization. Meta and Google are both steering advertisers toward adtech products that use machine learning to identify valuable consumer targets and optimize marketing performance, exercising near-total algorithmic discretion about where to serve ads. These products, branded respectively as Meta's Advantage+ ¹ and Google's Performance Max, are not strictly “privacy-preserving,” but they are definite responses to the challenges and opportunities encountered in adtech's pivot to privacy. “PMax isn’t a privacy product,” a trade reporter explains. “But it's part of a future digital ad model without third-party tracking” (Hercher, 2022a). Meta, on the other hand, explicitly markets Advantage+ functions as “privacy-safe” means for scaling the delivery of personalized advertisements (Meta, 2022). In particular, Meta is trying to work around the difficulties posed by Apple's ATT (Murphy and Criddle, 2023). “Having read the room,” as one observer puts it, “Meta is investing in AI to develop and deploy privacy-enhancing technologies (PETs) to underpin its ad platform” (Schiff, 2023).

Meta and Google position these products in the familiar registers of automation and optimization. In Advantage+ and Performance Max, clients specify conversion targets, the price they want to pay for those conversions, and budget caps, and these systems work to maximize each campaign's return on advertising spend. These systems are designed to compensate for the new limitations on third-party measurement and data sharing. Take Meta for example. Rather than relying just on “its conversion pixel and [software development kit] to connect impressions to conversions,” Meta ingests advertisers’ sales records by integrating the latter's databases into its servers (Hercher, 2022b). This shifts the processing of conversion records from a third-party relationship, wherein Meta observes events on advertisers’ websites or apps, to something that looks like a second-party relationship, where advertisers share the first-party sales data they collect. While ATT stops Meta from monitoring the cross-app chain of events from ad exposure to purchase behavior, “Apple can’t block a company from piping its CRM, CDP or data warehouse to Meta” (Hercher, 2022b). And where conversion data are missing, these optimization systems fill in the gaps with probabilistic modeling. By shifting responsibilities for measurement and data ownership among parties, and because Meta and Google do not report granular details to advertisers about how these black boxes target or influence individual users, this whole arrangement has a veneer of privacy.

These solutions have garnered attention for being opaque but effective (Murphy and Criddle, 2023). They also demonstrate the broader transition away from situating “ground truth” in advertising around widely accessible behavioral signals and toward the use of machine-learning models that rely on massive first-party data assets and computing resources. Meta, for instance, has reportedly “invested in dramatically expanding its computing power in order to train these more complex AI models on larger data sets” (Murphy and Criddle, 2023). Large platform companies are, in these cases, promising to replicate the optimization capabilities that marketers are used to by absorbing more adtech and data-analytic functionalities and by leaning more on inferential profiling and evaluation. Through these products, Meta and Google can tell a compelling story about why advertisers should continue to see their platforms as the best options for maximizing ROI; and they can, at an operational level, exert additional control over the distribution of ad revenue, perhaps favoring their own properties. Google and Meta may steer customers toward these products by making them prerequisites for accessing other spending channels. For example, advertisers who want to promote themselves on Google Maps now must use Performance Max (Hercher, 2022a). This not only pumps revenue through these pipes, but it also builds the training data necessary for optimization. And since Meta and Google only tell ad clients how much money was spent and how much revenue was likely produced, all the information about what placements worked and what users converted stays with the platforms. This could drive adoption through a network effect: as advertisers feed data into these systems, and as the models learn to recognize valuable prospects for certain products, those models become more attractive to other advertisers targeting the same people. In sum, platform companies are using these products to exploit and intensify information asymmetries, while presenting them as solutions that will comply with restrictions on third-party tracking.

Few details about the inner workings of these products are publicly known. What is clear, however, is that they represent a waypoint for adtech's pivot to privacy, wherein optimization is powered by inferential machine learning, third-party tracking is replaced with first-party tracking and data-sharing partnerships, and platforms internalize more computing and logistical operations within walled gardens, increasing dependency on those core adtech companies. These maneuvers are emblematic of an approach to “privacy” that we call cynical resignation.

Sanitizing surveillance

One of the means through which firms evade privacy mandates is through the development of computational techniques that enable them to obfuscate data flows and sanitize profiling activities. While these technological systems may be technically compliant with privacy regimes, they do not make any meaningful underlying change to the practices that animated privacy concerns in the first place—particularly concerns that are sensitive to issues of power. For example, significant investments in PETs enable firms to comply with requirements not to retain data that reflects individual-level tracking, while still allowing them to profit from the underlying insights that can be derived from this data. Such methods include the use of multi-party computation, which enables multiple parties to match and query data from multiple business partners without revealing individual-level information, and the use of “clean rooms” that enable ad targeting to take place within a secure environment operated by a third party, without the targeting firm gaining direct access to individual insights.

These create growth opportunities for firms that maintain a cloud infrastructure business, because they are able to sell features such as clean room services as add-on compliance features for their potential clients. This enables their clients to do such things as retain anonymized audience sets within a clean room over time for persistent tracking, and combine data from across many subsidiary platforms (Hercher, 2023c). What they do not do is meaningfully address the role of data as a key source of tech firms’ power. Indeed, some of the regulatory moves happening around PETs are being used to avert more bright line data protection measures (Veale, 2022).

Another example is the turn to inferences as the key locus of insights (Solow-Niederman, 2021). Google in 2021 announced plans to shift to using “data-driven attribution,” which substitutes industry standard “last-click” attribution methods with algorithmic models designed to infer the connection between ad impressions and user activity in the absence of the data that cookies or other identifiers provide (Srinivasan, 2020). As one trade publication describes it, “Data-driven attribution may not be more privacy-compliant than last-click [attribution], in and of itself. But in a privacy-forward environment where connecting ad impressions to online user activity is often prohibited, it will be the reliable methodology” (Hercher, 2021). Though this data is probabilistic, and does not contain personal identifiable information, it nevertheless retains the information asymmetries that characterize other models in behavioral advertising, in many ways concentrating these asymmetries even further by positioning Google's algorithmic models as the ground truth on which publishers and advertisers can evaluate the effectiveness of their placements.

Party-hopping

Maximizing information asymmetries is at the heart of the second technique that characterizes cynical privacy resignation, what we’re calling “party-hopping”: pursuing more invasive first-party data practices in lieu of disreputable third-party practices. It is a policy dodge enabled by failures to enforce a meaningful definition of privacy. Party-hopping works on the premise that data collection and processing are unproblematic so long as the number of companies accessing or owning that data can be restricted. It both dismisses data minimization and encourages consolidation. While many companies are able to engage in first-party tracking, big tech firms—which can draw on their network effects and combine streams of data drawn from subsidiaries within the platform ecosystems they control—benefit most from this transition, and several emerging industry techniques illustrate how they intend to maximize this data advantage.

One example of party-hopping is through the development of new technologies that enhance companies’ first-party tracking capabilities. The clean rooms described above offer one illustration of what this looks like in practice. While many clean rooms are designed for interoperability, those offered by Google, Amazon, and Meta function as walled gardens, as one recent industry article bemoaned: “The privacy magic isn’t in encryption or even the degree to which unbeknownst users are tracked online, it's whether the data in held in the hands of one company” (Hercher, 2023b). By drawing clients into the use of their clean room technology and integrating it with their other service offerings, cloud providers are able to keep more users operating within their walled gardens (Hercher, 2023b). Meta's development of its Meta Pixel to replace its standard conversion pixel is an additional example. The former triggers the website hosting it to set a first-party cookie on the user's browser and then share data with Meta. This acts as a workaround to get past Apple's intelligent tracking prevention, which blocks third-party cookies, enabling Meta to replicate the data flows established prior to Apple's “privacy enhancing” measures (Zawadziński and Wlosik, 2022).

Universal ID solutions work along a similar principle. By cloaking persistent tracking in the legitimating cover of a first-party consent agreement, unique identifiers and profile data can once again be activated across sites by various advertisers and intermediaries who signed up for the same global ID solutions. SWAN, for example, effectively turns every publisher within the network into a “first party,” making users eligible to be targeted by advertisers across sites of other SWAN partners, despite encryption of user-level information by site. As discussed further below, Apple's ATT also hinges on the pretense that, by definition, first-party tracking is legitimate and third-party tracking is not.

Across these examples are a set of techniques that large tech firms are best placed to exploit in environments where third-party tracking is eroded. By controlling bottlenecks, such as browsers, devices, cloud computing, and marketplaces, they have the capacity to technologically innovate around regulatory environments to preserve their data advantage. And where this fails, they have the capital to buy their way into data advantages.

Sabotage

A third model of cynical resignation takes the form of what we describe as sabotage—using “privacy” as a means of disadvantaging rivals and increasing market power. This strategy not only seeks to retain core elements of data tracking in advertising, it reinforces platforms’ dominance in the market by directly undermining external competitors’ abilities to gain insights from third-party data. Our terminology here draws on Shapiro's (2023) Veblen-influenced definition of platform sabotage as the strategic withdrawal of efficiency from other market actors. This is quite applicable to digital advertising, whose quintessential value proposition is marketing efficiency, and wherein access to personal data and addressable targeting capabilities are considered the fundamental sources of that efficiency (Turow, 2011). We call this sabotage because an offending company does not improve its products so much as withhold perceived efficiencies from the rest of the marketplace.

Apple's announcement of its App Tracking Transparency privacy tools is a case in point. ATT mandates that when apps seek to collect user information to share it with third parties, a window must show up on the Apple device asking for permission, and if the user declines the app must stop tracking the user's data. The announcement of this measure led Meta to take out full-page newspaper ads denouncing the feature as harmful to small businesses. It also led Germany's competition regulator to initiate a proceeding against Apple to review ATT's impact on other companies, specifically examining whether Apple's ability to unilaterally set rules for its app store had anticompetitive effects on the market.

This maneuver should be viewed in light of Apple's growing advertising business. As mentioned above, ATT operationalizes privacy in a self-serving way; its definition of “tracking” prohibits third-party data collection (absent opt-in consent), while Apple's own first-party data collection and personalized advertising services are exempted from such prohibitions and permissions requirements (Kollnig et al., 2022). That means advertisers can get more detailed data about the performance of their ads if they buy from Apple rather than other ad-sellers, and that apps financed by personalized advertising may become more dependent on Apple. As one analyst told the Wall Street Journal, “it's not really clear why they would withhold [performance data]…unless they were just really trying to privilege their own ad networks” (Haggin, 2021). This party-based privacy policing generates a windfall for the company. In the first six months after the launch of ATT, Apple's advertising business reportedly tripled its market share (McGee, 2021).

Similarly, Google's intention to block third-party cookies on Chrome and its subsequent implementation of Privacy Sandbox foreshadow significant downstream effects for other companies (Wakabayashi, 2021). This led to complaints by publishers and the UK Competition and Markets Authority, both of whom worried that Privacy Sandbox would have negative impacts on competition in the digital ads market. The result was a negotiated set of concessions made by Google in exchange for the CMA's agreement not to pursue an investigation. The agreement is heavily reliant on a monitoring trustee to track and evaluate whether Google has adequately complied with these commitments (Competition and Markets Authority, 2022). This is not the first time Google has used “privacy” to justify terms of service and data sharing (or withholding) that make its own adtech products more valuable to advertisers and publishers, relative to competitors (Srinivasan, 2020).

FTC Commissioner Rebecca Kelly Slaughter (2021: 21) remarked on these sorts of practices in a recent speech: “Shutting off the data spigot for others while filling your own well is the kind of anticompetitive innovation that we’re bound to see more of if this space remains unregulated.” Where privacy regulations fail to take into account these kinds of industry practices, they are likely to result in further concentration of data advantages in the hands of large tech firms who are positioned to benefit from platform sabotage.