Most digital advertising is measured wrong. Not a little wrong—fundamentally wrong. The models we use to decide which channel gets credit for a sale are built on assumptions that barely survive contact with reality. And because those models steer billions of dollars in ad spend, the error isn’t just academic. It inflates budgets, warps strategy, and quietly rewards channels that look brilliant on a dashboard but do almost nothing to drive real demand.
Let’s walk through why attribution is broken, what the common models actually do, and how to think about measurement if you care more about accuracy than a flattering report.
The Core Illusion: We Can Isolate Cause and Effect
Attribution models start with a seductive idea: that we can trace a conversion back through every touchpoint and figure out which one “caused” it. The reality is messier. Someone sees a display ad while reading the news, ignores it, later searches for the brand on Google, clicks a paid search ad, leaves, gets retargeted on Facebook, still doesn’t click, and finally types the URL directly into their browser and buys. Which of those touchpoints gets the credit? The model picks one—or splits it—based on a rule someone invented in a conference room.
The truth is, we can’t isolate cause and effect from a log of exposures. We can only see what happened before the sale. That’s correlation, not causation. And in complex, multi-channel environments, correlation is a weak substitute for understanding.
Last-Click: The Default That Won’t Die
Last-click attribution is the cockroach of measurement models: simple, resilient, and almost impossible to kill. It gives 100% of the credit to the final touchpoint before conversion. If someone clicked a branded paid search ad and then bought, paid search gets the sale. Everything that happened before—the display ads, the social content, the email nurture sequence—gets zero.
Why does it survive? Because it’s easy. Every platform can report last-click numbers without any heavy lifting. And because it makes bottom-funnel channels look like heroes. Branded paid search, in particular, becomes the star of every report. The problem, of course, is that people searching for your brand name were already looking for you. The ad didn’t create the demand; it just intercepted it. Last-click attribution confuses interception with creation.
Multi-Touch Models: More Sophisticated, Same Blindness
Multi-touch attribution (MTA) was supposed to fix this. Instead of giving all the credit to the last click, MTA spreads it across several touchpoints. Linear models split it evenly. Time-decay models give more weight to interactions closer to the conversion. U-shaped models assign 40% to the first touch, 40% to the last, and sprinkle the remaining 20% across the middle.
These feel fairer. They acknowledge that an early awareness campaign might have planted the seed. But they introduce a new problem: they treat every logged touchpoint as if it actually mattered. A display ad that loaded below the fold and was never seen gets the same weight as a deliberate search click. An auto-played video running in a muted, background tab counts as an “interaction.” MTA doesn’t solve the causality problem. It just spreads the error across more line items.
View-Through Conversions: Credit Laundering at Scale
If you want to see attribution at its most dishonest, look at view-through conversions. Here’s how it works: a demand-side platform (DSP) serves a display ad to someone who was already likely to buy—maybe they visited your site last week, or they’re in a high-intent audience segment. The ad appears. The person doesn’t click. Days later, they convert through direct traffic or organic search. The DSP, using its own attribution window (often 30 days), claims that conversion as “view-through.”
This isn’t measurement. It’s credit laundering. The platform is taking conversions that would have happened anyway and stamping its name on them. And because most marketers don’t deduplicate across platforms, the same sale gets claimed by Facebook, Google, and the DSP simultaneously. Add up all the platform-reported conversions and you’ll often get a number larger than your actual revenue.
Data-Driven Attribution: Google’s Black Box
Google’s answer to the flaws in rules-based models is data-driven attribution (DDA). It uses machine learning to analyze converting and non-converting paths, then algorithmically assigns credit based on which touchpoints appear to make a statistical difference.
That sounds like progress. But DDA has a structural conflict of interest: it’s built by Google, trained on Google’s data, and it decides how much credit Google’s own channels receive. Google has never published the full methodology. The model is a black box, and independent analyses consistently show it shifting credit toward Google’s paid channels—branded search and YouTube—while reducing credit for organic and direct sources. That’s a convenient outcome for the company selling the ads.
The Incrementality Gap
Every attribution model shares a fatal flaw: it measures correlation, not causation. It looks at what touchpoints were present before a conversion and assigns credit based on presence. It doesn’t measure what would have happened without those touchpoints.
Incrementality testing tries to answer that question directly. You run a controlled experiment: one group sees the ad, a matched group doesn’t, and you measure the difference in conversion rate. That difference is the true incremental effect. Everything else is noise.
When companies actually run incrementality tests on their digital channels, the results are often sobering. Display and video campaigns that looked fantastic in attribution models frequently show near-zero incremental lift. Retargeting campaigns that appeared to drive huge volumes turn out to be capturing demand that was already inbound. Branded paid search—the golden child of last-click—often shows minimal incrementality because those clicks are just intercepting organic traffic that would have converted anyway.
Why the Industry Sticks with Broken Models
If incrementality testing is more accurate, why isn’t it the standard? The answer is structural. Attribution models are cheap, automated, and flattering. They produce numbers that make everyone look competent. Incrementality testing is expensive, slow, and often produces numbers that make campaigns look wasteful.
Ad platforms have zero incentive to promote incrementality. Their business depends on advertisers believing their ads work. If every campaign were subjected to rigorous geo-experiments or randomized controlled trials, a significant chunk of digital ad spend would disappear. The platforms know this, so they invest heavily in attribution tools that make their inventory look effective while quietly discouraging independent measurement.
Agencies face a similar conflict. Their fees are often tied to media spend. If incrementality testing reveals that half the budget is wasted, the agency’s revenue drops. There’s a soft but persistent pressure to keep using models that justify the current spending level.
What Actually Works: A Practical Framework
So if attribution models are unreliable and incrementality testing is resource-intensive, what should a marketing team actually do? Here’s a framework that doesn’t require a PhD or a seven-figure testing budget.
1. Separate Reporting from Decision-Making
Use attribution models for directional reporting, not for budget allocation. It’s fine to look at last-click or multi-touch numbers to understand user paths, but don’t let those numbers directly control spend. The moment a model’s output becomes a target, the model gets gamed.
2. Run Cheap Incrementality Tests Where You Can
You don’t need a full geo-experiment for every channel. Start with the biggest line items. For branded paid search, pause it in a few geographic regions for two weeks and measure the impact on total conversions (not just paid search conversions). For retargeting, split your audience into a control group that receives no retargeting and compare. These tests are imperfect but far better than trusting an attribution model.
3. Deduplicate Across Platforms
At minimum, use a single source of truth for conversions—your own backend data, not platform-reported numbers. If you’re spending on Facebook, Google, and a DSP, all three will claim credit for overlapping conversions. Centralize conversion tracking so you can see the real total and identify double-counting.
4. Evaluate Channels by Their Role, Not Their Score
Some channels create demand. Some capture existing demand. Some assist. Attribution models collapse these distinct roles into a single number. Instead, map your channels onto a simple framework: demand generation (net-new awareness and interest), demand capture (converting people already in-market), and demand assistance (supporting the path without being the primary driver). Judge each channel by whether it performs its role, not by a blended attribution score.
5. Watch for the View-Through Trap
If a platform reports view-through conversions, treat those numbers as fiction until proven otherwise. Compare view-through claims against a holdout group or a period when the campaign was off. If the view-through volume doesn’t drop when the campaign stops, those conversions were never caused by the ads.
The Bottom Line
Attribution models are not measurement tools. They are allocation mechanisms built on assumptions that benefit the platforms selling the ads. The more you treat them as truth, the more money you’ll waste on channels that look effective but aren’t.
The alternative isn’t perfect. Incrementality testing has its own challenges—sample size requirements, seasonality effects, contamination between test and control groups. But it at least tries to answer the right question: did this ad cause a change in behavior that wouldn’t have happened otherwise?
Until the industry aligns incentives around that question, attribution will remain what it is today: a convenient fiction that costs advertisers billions.
Frequently Asked Questions
Why is last-click attribution still so common if it’s flawed?
Last-click persists because it’s simple to implement, easy to understand, and makes bottom-funnel channels like branded search look highly effective. Platforms default to it, and teams running those channels have little incentive to switch to a model that would reduce their reported performance. Changing attribution models often means redistributing credit—and budgets—which creates internal friction.
What’s the difference between multi-touch attribution and incrementality testing?
Multi-touch attribution divides credit for a conversion among the touchpoints a user encountered. It’s based on correlation: if a touchpoint was present, it gets some credit. Incrementality testing measures causation by comparing a group exposed to an ad against a control group that wasn’t. The difference in conversion rates between the two groups is the true incremental effect. MTA tells you what happened; incrementality tells you what would have happened anyway.
How can a small marketing team test incrementality without a big budget?
Start with simple on/off tests for your largest channels. Pause a campaign in select geographic regions or for a specific audience segment while keeping everything else constant. Measure the impact on total conversions, not just conversions attributed to that channel. Even a rough test with imperfect controls will give you more useful information than trusting an attribution model blindly.
Are view-through conversions ever legitimate?
In most cases, no. View-through conversions credit an impression for a conversion that happened through another channel, often days or weeks later. The vast majority of these conversions would have occurred without the impression. The only scenario where view-through measurement might be defensible is when the ad itself contains no clickable element and the conversion path is direct—for example, a billboard or a non-clickable video ad where the user later visits the site by typing the URL. Even then, controlled testing is necessary to separate real influence from coincidence.