Cookie collapse exposes cracks in AI identity systems

The browser-based identity system started to crack in 2017, when Apple’s Safari privacy changes began weakening cookie-based identity by restricting third-party cookie tracking. Apple introduced Intelligent Tracking Prevention (ITP) in Safari in 2017, which began restricting third-party cookie tracking and was subsequently tightened in later releases, weakening the browser- and cookie-based identity system that marketers and ad-tech platforms had long used as a basic mechanism for recognizing users across sites and sessions.

That weakening became an industry-wide turning point on 2020-01-14, when Google announced that Chrome would phase out support for third-party cookies within two years. Because Chrome was central to the web’s advertising and measurement infrastructure, Google’s January 2020 announcement signaled a broader shift away from third-party cookie-based identity, weakening browser and cookie identity as a foundation for tracking and resolution and initiating a move toward alternative identity resolution methods.

Identity resolution sounds precise, but in practice it is a stitching exercise across systems that were never built as one record of truth. Tableau TUG describes marketers using “three main tools to connect the dots” — UTM parameters, cookies, and identity resolution — effectively assembling breadcrumbs from separate channels, while enterprise operations show the same fragmentation problem elsewhere: financial, banking, and customer data often sit in different systems, and according to PwC, “34% of CFOs still load accounting data into spreadsheets that call for manual adjustments,” a sign that reconstruction still depends on human patchwork.

That patchwork matters because many identity decisions are inferred from partial, temporary, or probabilistic signals rather than a real-time, unified customer ID. Tableau TUG notes that a website “will place a small file called the first party cookie in their browser” and that the cookie is “like a temporary name tag” so that “even if they leave… and come back four days later…” the system can recognize them, but cross-device linking still requires stitching sessions through common links such as logins or household clues, and when signals are blocked or disconnected, verification becomes slower, less complete, and harder to reproduce consistently.

The failure is no longer theoretical: organizations are already being called on to defend identity decisions to outsiders, and many cannot do it convincingly. As StreetInsider reported from a global Regula study of 850 fraud prevention and financial crime decision-makers, “82% of organizations have been asked to document or explain an identity-related decision to a regulator, court, or external auditor,” and the headline distilled the problem bluntly: “82% of Organizations Had to Explain an Identity Decision, and One-Third Could Provide Only Limited Evidence.”

What breaks under scrutiny is not just documentation volume but evidentiary quality and reconstructability across systems. Among organizations asked to explain an identity-related decision, “68% could produce clear, audit-grade evidence, such as logs, provenance data, or a decision trace,” while “the other 32% had only limited or indirect evidence,” and the same study found that “only 50% of organizations can fully reconstruct an identity verification decision across all contributing systems, signals, and decision logic,” while “42% can identify the main systems and signals involved but cannot fully reconstruct how the final outcome was reached.”

The cost of weak data foundations is not theoretical; it shows up as wrong decisions made at scale, then paid for again in reruns and rework. As one July 2026 analysis put it, bad data leads to “wrong decisions at scale,” forcing organizations to “pay twice the money” and often spend “up to three times as long” fixing the outcome, while the damage spreads beyond a single campaign because once teams stop trusting the data, they demand extra validation and the whole organization’s decision velocity slows.

AI does not repair that weakness; it accelerates it. Jay Schwedelson warned that AI can make “a lot of really bad decisions really fast if the data foundation is weak,” and stressed the operational burden of incomplete customer information that forces decisions before the full answer is available; the remedy, he said, is disciplined data construction, noting, “We call 25 million businesses a year and verify that they’re open and at that address… So we use sort of a AI human assist model in the data construction.” HIT Consultant similarly found that AI simply builds on incomplete or inaccurate inputs, directing budget toward the wrong channels and audiences with growing confidence.

The vendors pushing autonomous marketing are effectively conceding that automation is only as safe as the identity layer beneath it. Zeta says its platform first resolves consumer identity, then automates decisioning, then executes across channels, and CFO Chris Greiner said in February 2026 that revenue is tied to the volume of decisions made, not seats, whether those decisions are made by a human or an agent. That makes the dependency explicit: if decision volume is the business model, a unified ID and shared data inputs have to come before scaled automation, especially because weak inputs can misdirect targeting at the outset. Adstra and InterMedia found that only 23% of residential IP addresses matched their intended geographic target.

The same pattern appears in newer products, but with a warning label attached. On September 22, 2026, Taboola added an identity layer named Realize ID to Realize and said it turns fragmented identifiers into a single, persistent view of consumers who show purchase intent on the open web, yet its claimed uplift of up to 2.4x rested on undefined efficiency and signal-enriched predicted intent models. No model architecture, training window, refresh frequency, or lookback period was described, underscoring that cross-channel AI needs verified, current inputs before it can be trusted to act.