Attribution After iOS 14.5: What Actually Works in 2026

Your marketing attribution is broken. This is not a guess; it is a mathematical certainty. Since Apple’s iOS 14.5 update, the data pipelines that once fed your dashboards have run dry. The models you relied on, from last-click to multi-touch, are now operating on phantom signals. Continuing to use attribution models built for a world that no longer exists is not just ineffective; it is financial negligence.

Before: The Fragile World of Pre-iOS 14.5 Attribution

Pre-iOS 14.5 attribution was a system of tracking user behavior across the web using third-party cookies and device IDs. This allowed marketers to connect ad views to purchases, but it was notoriously inaccurate. Unlike modern causal inference, it relied on flawed models like last-click that over-credited some channels and under-credited others, leading to inefficient ad spend.

Before 2021, marketing attribution felt like a solved problem. You had a suite of models at your disposal: first-touch, last-touch, linear, time-decay, and U-shaped. Each promised a different, supposedly more accurate, view of the customer journey. Brands would debate the merits of a linear attribution model explained versus a last-click model, believing they were refining their spend. In reality, they were just choosing their preferred flavor of fiction.

These models all shared a fatal flaw: they were built on the unstable foundation of third-party cookies and device IDs.

Let's be more specific. A last-click attribution model, the default for many platforms, gives 100% of the credit to the final touchpoint before a conversion. This systematically overvalues branded search and direct traffic, channels that often capture intent created elsewhere. A customer sees your TikTok ad, remembers your brand, and searches for you on Google a week later. Last-click gives all the credit to Google, leading you to believe your search ads are far more effective than they actually are. The result: you pour more money into capturing existing demand instead of creating new demand.

A first-touch attribution model does the opposite, giving all credit to the initial touchpoint. While this can highlight channels that are good at generating awareness, it ignores the complex journey that follows. It's a blunt instrument that fails to capture the nuances of a multi-channel world.

Multi-touch attribution models, like linear, time-decay, and U-shaped, attempted to solve this by distributing credit across multiple touchpoints. A linear model gives equal credit to every touchpoint, regardless of its position in the journey. A time-decay model gives more credit to touchpoints closer to the conversion. A U-shaped model gives the most credit to the first and last touchpoints. While these models seem more sophisticated, they are still based on arbitrary rules and assumptions, not on a true understanding of what caused the conversion. They are just more complicated ways of being wrong.

Then came iOS 14.5. The AppTrackingTransparency (ATT) framework was not a minor tweak; it was an earthquake. Suddenly, the entire ecosystem of mobile attribution was rendered obsolete. The data simply vanished.

After: The Wasteland of Post-ATT Attribution

Post-ATT attribution refers to the state of marketing measurement after Apple's AppTrackingTransparency (ATT) framework rollout. This new reality is defined by a lack of user-level data, making traditional attribution models obsolete.

Google has deprecated third-party cookies in Chrome, completing the transition to a privacy-centric web. The old attribution models are not just inaccurate; they are impossible to implement as designed. Yet, many platforms and agencies continue to sell them, patching together probabilistic methods and fingerprinting in a desperate attempt to recreate the past.

This has led to a crisis of confidence. The data is a mess of conflicting reports, where each ad platform takes credit for the same conversion, creating a cycle of over-attribution and wasted budget.

This is the direct result of relying on correlation, not causation. Your ad platforms are designed to find correlations. They will always find a way to correlate their ad impressions with a sale, even if no causal link exists. This is why you see a spike in branded search conversions when you run a TikTok campaign. The TikTok ad created the demand; the search ad captured it. A last-click model gives 100% of the credit to the search ad, leading you to invest more in a channel that is not actually generating new customers.

Consider a typical Dutch fashion brand spending €50,000 a month on Meta and €20,000 on Google Ads. Meta reports a 5x ROAS, and Google reports a 7x ROAS. On paper, the brand is generating €390,000 in revenue from a €70,000 spend. But the company's total revenue only increased by €150,000. Where is the missing €240,000? It's phantom revenue, the product of overlapping attribution. Both platforms are taking credit for the same sales, and neither is telling you how much of that revenue would have come in anyway, even without the ads.

Causal inference is a statistical method that moves beyond correlation to identify true cause-and-effect relationships in marketing data.

There is a way out of this mess. It does not involve finding a better attribution model or a new tracking technology. It involves a fundamental shift in how you measure marketing effectiveness: from correlation-based attribution to causal inference.

Causal inference is a branch of statistics that deals with identifying cause-and-effect relationships. Instead of just observing that two events happen at the same time (correlation), it allows you to determine if one event actually caused the other. In marketing, this means you can finally answer the most important question: what would have happened if I had not run that ad?

This is the concept of incremental sales. Incremental sales are the sales that would not have occurred without a specific marketing activity. This is the only metric that matters. ROAS, as reported by your ad platforms, is a vanity metric. It is a blend of sales that were going to happen anyway and sales that were genuinely caused by your ads. Causal inference allows you to separate the two.

How does it work? Instead of tracking individual users, causal inference uses aggregated data and statistical techniques like regression and instrumental variables to model the causal impact of your marketing spend on your sales. For example, a simple causal model might look like this:

Sales = β0 + β1 * (Meta Spend) + β2 * (Google Spend) + β3 * (Seasonality) + ε

In this model, the coefficients (β1, β2) represent the incremental sales generated by each additional euro spent on that channel, while controlling for other factors like seasonality. This is a simplified example, but it illustrates the core principle. You are no longer trying to connect individual touchpoints; you are measuring the aggregate causal impact of your investments.

More advanced causal inference techniques, such as instrumental variables, can isolate the true effect of your marketing even in the presence of confounding factors.

How to Implement Causal Inference in Practice

Implementing causal inference involves a methodological shift from tracking to modeling, starting with centralizing all marketing and sales data. It requires embracing experimentation through lift studies and adopting a modeling approach to estimate the causal impact of marketing spend. This means shifting KPIs from platform-reported ROAS to metrics like incremental revenue and true customer acquisition cost.

Moving to a causal inference framework does not require you to abandon all of your existing tools and processes. It is a shift in mindset and methodology. Here are the key steps:

1. Centralize Your Data: The first step is to bring all of your marketing and sales data into a single place. This includes your ad spend data from all platforms (Meta, Google, TikTok, etc.), your sales data from Shopify, and any other relevant data sources, such as your CRM or email marketing platform.

2. Embrace Experimentation: The gold standard of causal inference is the randomized controlled trial, or A/B test. In marketing, this takes the form of lift studies or holdout tests. By creating a control group that is not exposed to your advertising, you can directly measure the incremental impact of your campaigns. While it is not always feasible to run a perfect experiment, the principles of experimentation should guide all of your measurement efforts. [1]

3. Adopt a Modeling Approach: Where experiments are not possible, you can use statistical models to estimate the causal impact of your marketing.

4. Shift Your KPIs: Stop obsessing over platform-reported ROAS. Start focusing on the metrics that matter: incremental revenue, customer acquisition cost (CAC), and lifetime value (LTV). These are the metrics that will tell you if your business is actually growing.

Frequently Asked Questions

What is the main problem with marketing attribution after iOS 14.5?

The main problem is the loss of user-level data.

How does causal inference solve the post-cookie attribution problem?

Causal inference solves the problem by not relying on user-level tracking. Instead, it uses aggregated data and statistical modeling to measure the causal impact of marketing activities on business outcomes.

What is the difference between correlation and causation in marketing?

Correlation means two things happen at the same time, but it does not mean one caused the other. For example, your ice cream sales and your air conditioning costs might both go up in the summer, but one does not cause the other. Causation means that a change in one variable directly causes a change in another. In marketing, focusing on causation is critical for making profitable decisions.

Why is ROAS a misleading metric?

ROAS as reported by ad platforms is misleading because it blends incremental sales with sales that would have happened anyway. It does not distinguish between generating new demand and capturing existing demand. This often leads to over-investment in channels that are good at capturing intent but poor at creating it.

What is the first step to adopting causal inference?

The first step is to centralize your data. You need to bring all of your marketing and sales data into a single place. This includes your ad spend data from all platforms, your sales data from Shopify, and any other relevant data sources. This unified view is the foundation for building a causal model of your business.

References

[1] A/B Testing - A Complete Guide to Statistical Testing [2] The Future Of Attribution Is Causation