Attribution is one of those problems that every performance marketing team has wrestled with, usually reaching some version of "we know it's imperfect, so we just pick a model and stay consistent." That is a reasonable operational decision. But the specific model you pick introduces systematic biases that are not random noise. They are structural distortions that consistently overstate some channels and understate others. Staying consistent with a broken model just makes you consistently wrong.
The Core Problem With Touch-Based Attribution
Last-click, first-click, and linear attribution models all share a fundamental assumption: that purchases can be meaningfully explained by the sequence of ad interactions that preceded them. The differences between models are really just disagreements about which interaction in that sequence deserves the most credit.
Last-click assigns all revenue to the final touchpoint before conversion. First-click assigns it to the first. Linear spreads it evenly across all touchpoints. Time-decay gives more weight to recent touchpoints. Each of these models is a mechanical rule applied to a sequence that the model itself does not actually understand.
The deeper problem is that touch-based attribution models assume the counterfactual they are trying to measure. When last-click says "Google search gets credit for this sale," it is implicitly claiming that the sale would not have happened without that Google click. But that customer might have bought anyway. They might have been convinced by the Meta ad they saw three days earlier and only searched to navigate back to the product page. The Google search was a navigation step, not a persuasion event. Last-click assigns full conversion credit to the navigation step.
How Last-Click Specifically Distorts Channel ROAS
Last-click attribution systematically overstates the ROAS of channels that appear late in the purchase journey. Google branded search is the clearest example. A customer sees a TikTok video for a DTC skincare brand. They think about it for a few days. They see a Meta retargeting ad. They search for the brand name on Google and click the branded search ad. Last-click gives full credit to Google branded search.
The effect on reported ROAS is dramatic. Google branded search campaigns typically show reported ROAS of 8 to 15x under last-click attribution for DTC brands, because they are capturing navigation from customers who were already sold. The channel's contribution to the actual persuasion event was zero or minimal. But the attribution model cannot see that.
The inverse is true for channels that operate at the awareness or consideration stage. Meta prospecting, TikTok in-feed, and YouTube non-skippable ads all tend to be severely undercounted under last-click, because they rarely appear as the final click before conversion even when they were the primary persuasion driver. In last-click models, these channels appear expensive and low-ROAS. Teams cut them. Then branded search volume drops three months later because the top-of-funnel has been depleted, and the team cannot figure out why their "best-performing" Google campaigns are declining.
First-Click Has the Mirror Problem
First-click attribution overcorrects in the opposite direction. It assigns full credit to the first touchpoint, which tends to be an awareness or prospecting placement. This overstates the value of upper-funnel channels and undercounts the contribution of conversion-stage channels like retargeting and branded search.
Neither model is wrong in the sense of being maliciously designed. Both are wrong in the same structural way: they take a binary credit-assignment approach to a problem that is fundamentally about probabilistic influence. A customer's purchase decision involves multiple touchpoints with different levels of causal contribution, and no touch-based attribution model can measure that causal contribution directly.
Linear and Time-Decay Models Are Less Wrong, Not Right
Linear attribution distributes credit equally across all touchpoints, which at least avoids the extreme distortions of last-click and first-click. Time-decay gives more weight to recent interactions, which reflects the intuition that the purchase decision was more proximal to recent ads. Both are improvements over the binary models.
But they still cannot answer the fundamental question: would this customer have purchased if they had not seen this specific ad? That is an incrementality question, not an attribution question. Touch-based models of any variety cannot answer it, because they measure what actually happened, not what would have happened in the counterfactual where one of the touchpoints was removed.
This matters practically because two channels might both appear in the conversion path with equal frequency but have very different incremental contributions. Google branded search appears in the path for almost every conversion on a DTC brand with any awareness, but its incremental contribution approaches zero for customers who were already searching for the brand by name. Meta prospecting appears in the path for a smaller percentage of conversions, but its incremental contribution is high because without that touchpoint, many of those customers would never have entered the funnel at all.
What a Better Attribution Approach Needs
The attribution problem does not have a clean single solution, but there are approaches that reduce systematic bias significantly. The most practically useful are media mix modeling (MMM) and holdout-based incrementality testing. They work differently but address complementary aspects of the same problem.
MMM uses aggregate-level regression to estimate the marginal contribution of each channel's spending to total revenue, accounting for external factors like seasonality and organic trends. It does not rely on individual touchpoint data and is therefore not subject to the cookie and tracking limitations that affect touch-based models. Its limitation is granularity: MMM gives you channel-level estimates, not campaign or ad-set level insight.
Incrementality testing, by contrast, measures the causal effect of a channel by running controlled experiments where some portion of the audience is held out from seeing ads. The measured difference in conversion rates between exposed and held-out groups is the incremental lift attributable to the channel. This is the most direct measurement of what you actually want to know.
The practical limitation of incrementality testing for smaller DTC brands is that it requires statistical power, which requires volume. Running a valid holdout experiment on a channel spending $5,000 per week is difficult because the sample sizes are too small to reach significance without running the test for several weeks, during which creative and audience conditions may shift.
The Implication for Forecasting
At Flyweel, we build channel-level ROAS forecasting models that are trained on historical realized ROAS, not platform-reported ROAS. This is not a small distinction. Platform-reported ROAS reflects whatever attribution model the platform uses (which is almost always favorable to itself). Realized ROAS, measured as actual revenue divided by actual ad spend, removes the attribution model from the signal entirely.
The forecasting models we train are channel-specific, because the patterns that predict Google search ROAS over the next 7 to 14 days are structurally different from the patterns that predict Meta prospecting ROAS or TikTok ROAS. A single cross-channel model would be averaging across signal patterns that have different underlying drivers and different lag structures. Separate models per channel let us capture those differences.
We are not claiming this solves the attribution problem. Attribution and forecasting are different questions. But grounding the forecasts in revenue-side data rather than platform-reported attribution numbers means the forecasts are not inheriting the systematic biases that touch-based models introduce.
The point of this article is not that attribution models are useless. Used carefully, they provide useful operational guidance. The point is that if you are using last-click or first-click ROAS as your primary signal for budget allocation decisions, you are systematically redirecting spend toward channels that appear at the end of the conversion path and away from channels that do the actual persuasion work. Over time, that allocation pattern depletes the very channels that make your converting channels possible.