Attribution is the accounting layer of digital advertising. It decides which ad, click, or impression gets the credit for a sale. The trouble is, most attribution models rest on assumptions that fall apart the moment you look at them closely. They aren’t measurement instruments. They’re allocation rules—and the rules are often made up.
Marketers talk about attribution like it’s a settled science. It’s not. The dominant models—last-click, first-click, linear, time-decay, even many of the so-called data-driven approaches—are all variations on the same broken idea: that a user’s path to conversion can be sliced cleanly and handed out to individual touchpoints. That idea ignores how people actually make decisions. It ignores the mess of multi-device browsing, the weight of offline conversations, and the plain fact that not every ad exposure does anything at all.
The Last-Click Default and Its Distortions
Last-click attribution is still the most common model because it’s the easiest to turn on. It hands 100% of the credit to whatever touchpoint happened right before the conversion. That creates a relentless bias toward bottom-of-funnel channels—branded search, retargeting, affiliate links—while starving the channels that introduced the brand in the first place.
Take a standard e-commerce path. Someone sees a display ad for a new running shoe brand. No click. A week later, they search “best lightweight trainers,” read a review site that mentions the brand, and still don’t click. Two days after that, they Google the brand name directly, click a paid search ad, and buy. Under last-click, the branded paid search ad gets all the credit. The display ad and the review site get nothing. The marketing team then shifts budget away from display and toward branded search, which is really just harvesting demand that display helped build. Over time, the top of the funnel thins out, branded search volume drops, and the team wonders why performance is slipping.
This isn’t a thought experiment. It’s the predictable output of a model that mistakes correlation for causation. The last click is often the easiest action, not the most influential one. Attribution models that ignore this aren’t just imprecise—they actively steer budgets in the wrong direction.
Multi-Touch Attribution: More Complex, Same Blind Spots
Multi-touch models try to spread the credit around. Linear attribution gives equal weight to every interaction. Time-decay leans toward the interactions closest to the conversion. Position-based models hand most of the credit to the first and last touch, with the remainder split across the middle. These are heuristics, not insights. They’re guesses dressed in percentages.
The deeper problem is that these models still operate on a single, logged-in user journey. They assume the sequence of tracked touchpoints tells the whole story. It rarely does. Someone might see a display ad on their phone during a commute, research on a work laptop, and finally buy on a home tablet. If those devices aren’t connected—and they often aren’t—the model sees three separate users, not one. The attribution gets fragmented across phantom individuals.
Even when cross-device identity resolution works, it usually depends on deterministic matching (like a login) or probabilistic signals (IP address, device fingerprint). Deterministic matching is accurate but covers a tiny slice of users. Probabilistic matching is broader but introduces error. The result is an attribution dataset that’s incomplete at best and systematically skewed at worst. Users who log in aren’t representative of all users. They’re more engaged, more loyal, and more likely to convert regardless of the ad they saw.
Data-Driven Attribution: The Black Box That Still Needs Light
Google’s Data-Driven Attribution (DDA) and similar offerings from other platforms are sold as a step beyond heuristic models. They use machine learning to analyze converting and non-converting paths and assign credit based on statistical patterns. That sounds rigorous, but the limitations are substantial.
First, DDA is walled-garden attribution. It only sees touchpoints inside the platform’s ecosystem—Google Ads, Display & Video 360, Campaign Manager. If you run ads on Facebook, TikTok, or through a direct publisher deal, those touchpoints are invisible to Google’s model. The model isn’t measuring the true contribution of Google channels relative to everything else. It’s measuring their contribution relative to each other, inside a closed system. That inflates the apparent value of Google channels because the model can’t account for the influence of non-Google touchpoints that may have actually driven the conversion.
Second, DDA is a relative model, not an absolute one. It tells you that, within the observed data, certain channels tend to show up more often in converting paths. It doesn’t tell you whether those channels caused the conversion. A billboard, a friend’s recommendation, or a podcast mention could be the real driver, and the Google touchpoints are just correlated—users who are already interested tend to click Google ads. The model confuses interest with influence.
Third, DDA needs a minimum volume of conversions to work, often 600 conversions in 30 days for Google Ads. Smaller advertisers are locked out. Even for larger advertisers, the model’s output can be unstable, shifting credit allocations month to month based on noise in the data. That makes budget planning a headache. A channel that looks highly valuable one month may look mediocre the next, not because its true effectiveness changed, but because the model’s training data fluctuated.
Incrementality: The Test That Attribution Models Fail
The only way to know if an ad caused a conversion is to measure incrementality. Incrementality testing asks a straightforward question: did this ad exposure make a conversion more likely than it would have been without the exposure? That requires a control group—a set of users who are similar to the exposed group but who didn’t see the ad. The difference in conversion rates between the exposed and control groups is the incremental lift.
Attribution models don’t do this. They look only at exposed users and try to infer causality from sequence. That’s a basic category error. Attribution is a counting exercise. Incrementality is a causal measurement. They answer different questions, but the industry routinely treats attribution output as if it were incremental truth.
When incrementality tests are run, they frequently contradict attribution models. A 2019 study by the advertising effectiveness firm NCSolutions found that, on average, only 38% of attributed sales were actually incremental. The rest would have happened anyway. For a typical campaign, attribution models overstate impact by a factor of more than 2.5x. The channels that look best in attribution are often the ones with the lowest incrementality—because they reach users who were already going to convert.
Running proper incrementality tests is operationally hard. It requires the ability to hold out a randomized control group from ad exposure, which many platforms don’t natively support. Facebook’s lift tests and Google’s conversion lift experiments exist, but they’re limited to those platforms and need significant scale. For cross-channel incrementality, advertisers have to build custom infrastructure or use third-party measurement partners. That’s expensive and technically demanding, so most don’t do it. They lean on attribution models instead and make decisions based on numbers that are largely fictional.
The Hidden Cost of Attribution-Driven Optimization
When teams optimize toward attribution signals rather than incremental impact, they systematically defund the channels that create new demand. This isn’t a minor edge case. It’s a structural bias that compounds over time.
Brand advertising—video, audio, high-impact display, sponsorships—is especially vulnerable. These channels rarely get direct click attribution because users don’t click and immediately buy. Instead, brand advertising works by raising the probability that a user will later search for the brand, click a retargeting ad, or recognize the product in a store. In an attribution model, the credit goes to the search ad or the retargeting click. The brand campaign that made those actions possible gets zero credit. The optimization algorithm then recommends cutting brand spend and increasing performance spend. Short-term ROAS improves. Long-term demand generation collapses.
This dynamic is well-documented. Les Binet and Peter Field’s analysis of the IPA Databank, covering decades of advertising effectiveness data, shows that brand-building campaigns drive long-term growth but underperform on short-term attribution metrics. Performance campaigns show the opposite pattern. An attribution-only optimization framework will always favor the short-term, attributable activity at the expense of the long-term, harder-to-measure activity. The business slowly eats its own seed corn.
What a Better Approach Looks Like
The fix isn’t to hunt for a more sophisticated attribution model. It’s to stop using attribution as the primary decision-making framework for budget allocation. Attribution has a role—it can help diagnose funnel blockages, identify which creative messages resonate at which stages, and provide directional signals for tactical adjustments. But it shouldn’t be the basis for answering the question “how much should I spend on this channel?”
For budget allocation, the framework should be incrementality-first. That means:
- Run regular incrementality tests on major channels, using platform-native tools where available and third-party measurement where not. Accept that these tests are imperfect—they have statistical noise, they require scale, and they can’t be run continuously on every channel. But they provide a ground truth that attribution cannot.
- Use Marketing Mix Modeling (MMM) as a complement. MMM uses aggregate time-series data—weekly spend by channel, weekly sales, seasonality, pricing, competitor activity—to estimate the contribution of each channel over long time horizons. MMM captures the indirect effects that attribution misses, like the impact of brand advertising on branded search volume. Modern MMM approaches, using Bayesian methods and higher-frequency data, are more agile than the old annual models and can provide ongoing strategic guidance.
- Calibrate attribution to incrementality. If an incrementality test shows that display’s true contribution is 2x what attribution says, adjust the display attribution weights accordingly. This isn’t perfect—the calibration factor will vary over time and across campaigns—but it’s better than using raw, uncalibrated attribution numbers.
- Set guardrails based on business logic, not just attribution ROAS. Maintain minimum spend levels on brand-building channels even when attribution says they underperform. Treat these as infrastructure investments, not variable costs to be optimized in real time.
This approach is messier than simply following the numbers in a dashboard. It requires judgment, experimentation, and tolerance for uncertainty. But that’s the actual nature of advertising measurement. The clean numbers in the attribution report are a fiction. The messier reality—that we can’t perfectly measure everything, that some channels work in ways we can’t easily track, that the best decision is often a reasoned bet rather than a calculated optimum—is the truth. Operating on the truth, even when it’s uncomfortable, produces better outcomes than operating on a precise-looking lie.
FAQ
Why is last-click attribution still so common if it’s flawed?
Last-click persists because it’s simple, free, and built into every ad platform by default. It requires no extra setup, no statistical knowledge, and no cross-platform coordination. For performance marketing teams judged on short-term ROAS, last-click often shows the highest numbers because it credits the channels closest to the purchase—channels that are already capturing demand rather than creating it. Switching to a more accurate model can make ROAS look worse in the short term, which creates an organizational disincentive to change, even when the long-term business impact would be positive.
Can’t multi-touch attribution solve the last-click problem?
Multi-touch attribution distributes credit more evenly, but it doesn’t solve the fundamental issue: it still only sees tracked, digital touchpoints and still confuses correlation with causation. Giving 30% credit to a display ad and 70% to a search ad is less extreme than 0% and 100%, but it’s still an arbitrary allocation if the display ad was the true driver and the search ad was just a navigational convenience. Multi-touch models are heuristics, not measurements. They feel more sophisticated, but they’re built on the same flawed data and the same flawed logic.
What’s the difference between attribution and incrementality?
Attribution asks: “Among the tracked touchpoints this user had, which ones get credit for the conversion?” Incrementality asks: “Did the ad exposure itself cause conversions that would not have happened otherwise?” Attribution is a division of credit among observed interactions. Incrementality is a causal measurement that requires a control group of unexposed users. Attribution can be calculated from standard campaign data. Incrementality requires an experiment. The two often produce very different answers, and incrementality is the one that actually measures advertising effectiveness.
How can smaller advertisers measure incrementality without big budgets?
Smaller advertisers can start with platform-native lift tests. Facebook and Google both offer conversion lift studies that are free to run, though they require minimum spend thresholds. For cross-channel measurement, a lightweight approach is to use geographic holdout tests: pause advertising in a randomly selected set of regions and compare sales trends in those regions to regions where advertising continued. This requires enough geographic diversity and sales volume to detect a signal, but it’s far cheaper than full third-party measurement. The key principle is to run some form of test, even if imperfect, rather than relying solely on attribution.
