Advertising attribution: when models deceive the budget

What has changed: the apparent precision of attribution reports

Until a few years ago, attribution reports showed imperfect data that was honest about its flaws. Today, however, ad platforms return numbers that look super rigorous. Yet behind that precision often lies a statistical estimation mechanism, not direct measurement.

The phenomenon is clearly documented by Search Engine Journal , which analyzed how modeled data is progressively replacing real signals. The main cause is the signal loss : the gradual reduction of third-party cookies, restrictions imposed by mobile operating systems, and privacy regulations have eroded the observable data base. Consequently, platforms — Google, Meta, TikTok — have developed probabilistic models to reconstruct conversion paths.

The visible result is paradoxical. Dashboards are more detailed than ever. At the same time, the percentage of conversions actually tracked directly is constantly declining. Therefore, a company reading a ROAS of 4.2 on a campaign might be looking at a projection, not real data.

The real risk for those who manage media budgets

For a marketing manager, the problem isn't just theory. It pops up in daily decisions: bumping up the budget on a channel because a model says it's the most efficient one, or cutting another because the numbers look lackluster. If these choices are based on unverified estimates, they can lead to systematically flawed budget allocations.

The data quality it is considered a strategic priority even higher than adopting new tools. In a context where AI models automatically fill tracking gaps, the risk of making decisions on shaky ground grows proportionally with campaign complexity.

Furthermore, the problem is magnified in SMEs. Large companies have data science teams capable of querying the models and validating their assumptions. Small and medium-sized businesses, on the other hand, tend to accept platform numbers as absolute truth. This exposes them to a risk of budget misallocation which can quietly and progressively erode your overall ROI.

A similar phenomenon is seen in non-human traffic: as we analyzed in our deep dive on bots and AI traffic and protecting the ads budget , distortions in campaign data can have different origins but produce the same effect: spending decisions based on unreliable signals.

Measured data vs modeled data: a practical distinction

The first concrete action is to learn how to distinguish the two categories of data within your analytics tools. Not all numbers in a report carry the same epistemic weight.

• Measured data: conversion tracked directly via pixel, tags, or server-to-server integration. The signal is observed, not inferred.
• Modeled data: conversion estimated by the platform based on historical patterns, similar behaviors or probabilistic inferences. The signal is reconstructed.

Google Ads, for example, explicitly states in the column notes when a metric includes modeled conversions. Meta does the same in its system of Conversions API with the concept of estimated conversions . However, these insights are often ignored or misunderstood by those reading reports without specific training.

Therefore, the first practical step is to activate and verify the Conversions API (or server-side equivalent) to increase the share of direct signals. This does not eliminate modeling, but reduces it and improves the quality of the remaining estimates. We at SHM Studio include this check as standard in our digital marketing projects for clients with significant media investments.

What to do now: the attribution audit as an immediate priority

An attribution audit isn't a complex project. It's a structured analysis that answers three fundamental questions.

First question: what percentage of attributed conversions are directly tracked? If the answer is below 60-70%, the data foundation is fragile. In this case, budget decisions should be made with an explicit margin of uncertainty.

Second question: are the platform's attribution models aligned with the customer's actual buying cycle? A model last-click on a product with a long sales cycle systematically underestimates initial touchpoints. Similarly, a model data-driven on low volumes produces statistically unreliable estimates.

Third question: is there an independent source of truth? Tools like Google Analytics 4, CRM systems, or platforms from marketing mix modeling can offer an external point of comparison compared to the data of the single platform. This triangulation is essential to validate estimates.

At the same time, it is worth looking at how the advertising ecosystem is evolving. The growth of new formats and channels — as we discussed when talking about ChatGPT Ads and the implications for Italian SMEs — introduces additional layers of complexity in attribution. Every new touchpoint adds potential breakage points in the tracking chain.

The role of AI models in attribution: allies or risk amplifiers?

Platforms present AI modeling as a fix for signal loss. In part, that's true. Modern probabilistic models are significantly more accurate than those from five years ago. However, they are still estimates, and estimates come with margins of error that are rarely communicated clearly to advertisers.

One of the structural problems of modern attribution is the lack of transparency on uncertainty margins. Platforms show point numbers — 1,247 conversions, ROAS 3.8 — without indicating the confidence interval of those estimates. The reader thus perceives a precision that the data does not possess.

This does not mean that AI models are useless. On the contrary, in the absence of direct signals they are often the only alternative available. It does mean, however, that they must be used with methodological awareness, not as oracles. Our activities of AI consulting applied to marketing always include a phase to calibrate expectations regarding model limitations.

Data quality as a competitive advantage

In a market where all advertisers use the same platforms and attribution models, those who invest in data quality gain a structural advantage. Not because they have higher numbers, but because they have more reliable ones.

This is especially true for campaigns on fast-evolving channels. For example, new interactive formats — like the ones analyzed in our article on TikTok voice comments and polls for brands — they generate engagement signals that platforms still struggle to translate into reliable attributions. Those who understand this limitation can interpret the data more realistically.

Similarly, the evolution of touchpoints in the buying journey — including AI assistants as analyzed in the context of Amazon Alexa Shopping AI and trust in retail — makes it increasingly complex to track the entire customer journey . Therefore, the answer is not to look for perfection in tracking, but to build a robust methodology that knows how to live with uncertainty.

Our activities of Google Ads campaign management and LinkedIn campaigns always include a check of the tracking setup before any budget optimization. Without this foundation, any spending decision remains exposed to the risk of distorted attribution.

Outlook: toward a new advertising data culture

The attribution issue won't be resolved anytime soon. Market direction — more privacy, fewer cookies, more AI — structurally trends toward an increasing share of modeled data. Therefore, the answer isn't resisting this trend, but adapting your analytics practices.

Three quick tips for the upcoming months. First, set up or double-check server-side tracking on all your main campaigns. Second, back up your platform data with at least one independent sanity check—even just comparing it with your CRM data can reveal some big gaps. Third, be honest with your team about data uncertainty: treating ROAS as an estimated range rather than a hard fact completely changes how budget decisions are made and approved.

To explore the topic of attribution and data quality further in the context of paid media and digital advertising , our team is available for a session to analyze the current setup. An initial audit allows for quick identification of weak points and priorities for intervention. You can contact us from the contact page let's chat with zero pressure.