This is a guest post by Peter Kireev, a leading expert with over a decade of experience in AdTech. He is the Co-founder and Chief Product Officer at Reliz, a product-led European company, where he leads the development of its AI-driven solutions.
In 2024, the AdTech industry is undergoing a tectonic shift: the data that digital advertising relied on for decades is disappearing. On January 4, 2024, Google began phasing out third-party cookies for 1% of global Chrome users, marking the beginning of a full deprecation planned for later in the year. At the same time, Apple’s enforcement of AppTrackingTransparency continues to limit access to IDFA on iOS devices. In parallel, global privacy legislation — including GDPR, CCPA, and Brazil’s LGPD — is increasing restrictions on how data is collected, stored, and shared.
As platforms lose access to individual identifiers, behavioral tracking, and cross-site user journeys, the fundamentals of audience targeting and performance optimization are being redefined. Deterministic targeting is becoming the exception rather than the norm. In this environment, companies, whether they want it or not, have to shift toward models that can operate with incomplete data.
Machine learning is not just supporting the transition in AdTech — it is redefining the rules of the game. Treating AI as an afterthought is no longer sufficient. What matters now is embedding statistical reasoning into how AdTech systems process and act on limited signals.
Current data regulations in AdTech
The legal landscape has become one of the strongest forces reshaping the AdTech landscape. Starting with GDPR in the EU, CCPA in California, and LGPD in Brazil, global regulation has imposed strict requirements on data transparency, storage, and consent. These rules apply to any company working with international traffic, mobile and web alike.
At the same time, Apple’s privacy policy changes have reshaped the landscape from the inside out. The introduction of App Tracking Transparency in iOS 14.5 effectively made IDFA inaccessible for the majority of iOS users, with opt-in rates stabilizing well below 20%, leaving marketers with a fraction of the deterministic data they once relied on.
The initial steps toward phasing out third-party cookies in Chrome, which began on January 4, 2024, were accompanied by a shift in Google’s strategy. Due to regulatory pressure — particularly from the UK’s Competition and Markets Authority (CMA) — the company postponed the full rollout. Google cited the need to ensure both technical and commercial readiness across the ecosystem as the reason for the delay. Despite this temporary pause, the industry’s direction remains unchanged: a transition toward a consent-based data model and a move away from traditional tracking mechanisms.
What can and cannot be collected
In today’s regulatory climate, collecting or using personal data without clear, trackable user consent is no longer a grey area — it is a liability. A growing list of signals are off-limits:
- Personally identifiable information such as name, email, and declared age;
- Behaviorally derived identifiers that could be linked to a user profile;
- Device fingerprinting — the practice of generating a unique user profile based on IP address, user-agent, screen resolution, and other passive signals — is widely regarded as non-compliant and is grounds for rejection by Apple App Store, Google Play, and major demand partners.
What is still allowed is a much narrower set of data, and it must meet both legal and platform standards:
- First-party data collected within your own product, where the user has actively consented (e.g., age voluntarily submitted during onboarding);
- Aggregate behavioral context, such as app category, inferred geolocation, or content themes;
- IP address and user-agent, which are transferred by default with every web or mobile request. These signals are legally permissible but offer only about 60% accuracy in probabilistic models. Nonetheless, they remain foundational for cookie-less targeting when no persistent identifiers are available.
As access to identity data shrinks, these baseline signals are becoming vital inputs for ad performance systems, particularly for platforms that cannot rely on cross-site identity graphs.
Why AI for behavioral data processing became essential
With cookies fading out, IDFA largely unavailable, and regulation tightening across the globe, traditional user-level segmentation has become nearly impossible. Personalization mechanisms — from retargeting to lookalike modeling — can no longer rely on deterministic IDs. In this new reality, AdTech platforms are forced to generate hypotheses based on depersonalized and partial data.
In this context, AI becomes an operational requirement. Machine learning models are trained on aggregated and anonymized inputs: clicks, impressions, and in-product behavioral events. This allows platforms to compensate for the absence of direct user identifiers by identifying behavior-based patterns.
In practice, these models are used in three core areas:
- Behavioral targeting without cookies or IDFA, using generalized contextual and technical signals (e.g., content category, scroll depth, dwell time).
- Lookalike audience generation based on behavioral similarity, not on demographic or declared profiles.
- Real-time bid optimization, where algorithms estimate the value of a given impression without access to persistent identifiers.
As of 2024, leading AdTech providers — including The Trade Desk, Criteo, and Xandr — are investing heavily in probabilistic infrastructure and real-time AI bidding systems designed to operate under signal loss and compliance constraints. These systems are no longer optional innovation layers — they are how targeting and performance now function.
A standout example from this year is AppLovin’s Axon 2.0 — a predictive machine learning–based user acquisition engine. The algorithm automatically identifies users most likely to install an app and dynamically boosts impressions for those audiences. This has allowed AppLovin clients to significantly increase CPI campaign performance without manual intervention.
Limitations and realities in 2024
Despite growing reliance on AI in data-constrained environments, these systems have hard limitations — both technical and structural.
Today’s AdTech models do not “know” who the user is. Instead, they infer patterns based on incomplete signals: IP address, user-agent, interaction metadata, or historical aggregates from anonymized cohorts. The accuracy of this approach varies widely. For example, identifying a user solely based on IP and user-agent yields only around 60% reliability, and even that assumes short time windows and minimal device variability.
These models do not predict identity—they predict likelihood. For example, if a user visits a finance app and comes from a certain set of referral domains, the model might assign them a high probability of being a subscriber-type user. But it is still a bet, not a fact. Scaling such models requires large, well-structured behavioral datasets, which many advertisers still lack.
Another practical constraint is the black box problem. A model might determine that traffic from Manchester performs better than traffic from London, but it offers no transparent logic as to why. This black-box nature makes debugging and campaign diagnosis difficult, particularly in performance-sensitive environments.
Finally, probabilistic systems are extremely sensitive to training data. The model can quickly degrade if early-stage conversions are misclassified or training sets are noisy. This limits their usefulness in low-volume campaigns or products with sparse user behavior.
To operate reliably, AI infrastructure must be embedded directly into the platform’s architecture — hardwired into the platform, listening to real-time signals, and constantly learning from what happens next.
Competitiveness in AdTech now depends on precision with incomplete data
In 2024, AdTech operates under structural limitations: identifiers are disappearing, deterministic targeting collapsing, and privacy regulation shapes product and architecture.
The companies pulling ahead are not the ones with the most data — they are the ones who make the smartest use of the least. The competitive edge now comes from turning partial signals into meaningful predictions.
This shift requires more than model accuracy. It demands a full-stack rethink: real-time pipelines that capture clean behavioural signals, in-product telemetry that feeds into live models, and feedback loops that tie ad performance back to actual user outcomes.
AdTech players who fail to evolve will find themselves stuck — unable to scale, unable to optimise. Meanwhile, those who build for uncertainty are shaping the post-cookie market on their terms.
