Маркетологи должны различать измеренные данные и статистические модели, чтобы принимать верные решения

Marketers have less observable data but increasingly precise-looking reporting. At the same time, a growing portion of what appears in those reports is modeled or statistically reconstructed. My argument is that understanding the difference between what was measured and what was estimated can be the difference between making a good decision and confidently making a bad one. What ‘Signal Loss’ Costs You Signal loss doesn’t come from one source. It accumulates across consent, device changes, platform restrictions, and gaps between systems. And most people only notice it when the numbers stop making sense. The user journey has become less directly observable, and what remains depends on consent configuration, first-party systems, modeled reporting, login state, CRM integration, etc. Cross-device behavior makes this even messier. A person hears you on a podcast, later searches the brand name from a work laptop, reads two articles, gets retargeted on mobile, then comes back through direct, and converts. Okay, so, which part was discovery? Which part was persuasion? Which part was merely the last identifiable interaction? Attribution systems, though, see different disconnected fragments. Depending on the setup, the podcast may be invisible, some of the research may end up classified as direct or organic, and the retargeting interaction may receive disproportionate credit. In other words, the easiest touchpoint to measure is not necessarily the one that had the greatest influence on the decision. Attribution can tell you what it was able to connect to the conversion, but that is not always the same as reconstructing the full journey that created it. The practical consequence is budget misallocation at scale. When upper-funnel channels appear to contribute nothing, teams defund them. The decision looks data-driven even though it may simply reflect what the measurement system was capable of seeing. See also: Rethinking Audience Targeting In A Signal-Loss Era (With The R.E.M. Framework) AI Steps In: Missing Data Becomes Modeled Data Platforms have responded to these gaps with more modeled measurement. Some major platforms, like Google, have integrated machine-learning-based modeling into their measurement systems. When a direct link between interactions and a conversion can no longer be observed, these systems use patterns in observable and aggregated data to estimate some of the missing attribution. Those modeled results can then feed into reporting, attribution, bidding, and campaign optimization. That solves part of the observability problem, but introduces another, more difficult issue: Knowing when the estimates are good enough to support a decision, especially when the reporting interface presents modeled, and directly observed outcomes with similar visual confidence. My advice is to treat any metric labeled “modeled” or “estimated” in your platform reporting as directional, not definitive. Why 1 Attribution Model Is Not Enough One of the more persistent myths in attribution is that there’s a correct model waiting to be discovered. Bad news: There isn’t. Every attribution model answers slightly different questions, so the goal should not be finding the one “correct” model. I’m actually more interested in what changes when I compare them. If one channel looks important under several different approaches, that’s useful. If its contribution disappears as soon as the model changes, that’s useful too. That is why I recommend teams that want to use attribution most effectively not to pick one model and defend it but to triangulate across multiple methodologies and look for where the signals converge. Treat disagreements between them as questions worth investigating, not errors to be resolved by picking a winner. When Your Measurement Systems Disagree Here’s a scenario that plays out in marketing teams constantly. Google Analytics 4 reports 150 conversions. Plausible claims 180. The CRM shows 120 new customers. Three platforms, three realities, none of them matching. Perhaps GA4 counts purchases, CRM counts approved customers, ad platforms use view-through attribution, refunds are excluded from one system but not another, or date-of-click, and date-of-conversion reporting differ. That mismatch doesn’t automatically mean you have a data quality problem. This is simply the result of each platform using its own attribution window, conversion definition, reporting logic, and modeling assumptions. I would usually start with the system closest to the actual business outcome, such as CRM data, order records, subscription data, or another backend source, and then use analytics, and advertising platforms to understand different parts of the journey around it. That changes the question. Instead of asking which platform reports the most conversions, you start asking which outcomes actually happened, which parts of those journeys you can observe around them, and which touchpoints show up consistently across those journeys. Those business-side systems aren’t perfect attribution sources either. They can contain missing acquisition data, overwritten fields, duplicate records, or very little information about what happened before the conversion. Their value is not that they explain why someone converted, but that they give you a stronger anchor for confirming whether the business outcome actually happened. That reframe doesn’t require perfect tracking, but it requires unified data, and a willingness to accept incomplete answers. Making Decisions Without False Precision Stakeholders still want a definitive answer to a question the data can’t answer definitively: Which channel deserves the budget? And being explicit about what the data shows versus what it estimates may feel risky, but it makes the uncertainty visible rather than hiding it behind precise numbers. Teams that communicate measurement limitations clearly tend to make better decisions over time, because they’re not anchoring strategy to false precision. First-party data collection has become non-negotiable in this environment, not just for privacy compliance but for measurement quality. The more directly you can observe customer behavior through your own infrastructure, the less dependent you are on external platforms to reconstruct what happened. Server-side tracking, for example, can improve data reliability and control, but, of course, it does not magically eliminate consent gaps or recreate interactions you were never able or permitted to observe. As a result, a first-party measurement setup does not remove uncertainty, but it can shift the problem from “we don’t know what happened” to “we have a reasonable picture with known blind spots.” The job of attribution is no longer to tell us exactly what caused a conversion. It is to reduce uncertainty enough to make a better decision. The marketers who navigate this well will stop expecting their attribution stack to produce ground truth and start treating it as one input among several useful, directional, but always worth questioning. More Resources: Your AI Visibility Tracker Is Quietly Breaking Your Analytics And Your Strategy Google Went ‘Not Provided’ In 2011 And Blinded Us, ChatGPT Just Shipped Its Version How Do I Effectively Measure Campaign Success Across Multiple Platforms? – Ask A PPC Category Digital Marketing Marketing Analytics Analytics & Data