Attribution modeling is marketing measurement's most widely used tool and its most misunderstood one. Every major ad platform includes attribution reporting. Most analytics tools default to last-click. And almost all of it, used without understanding its limitations, will lead you to wrong conclusions about which channels are working.
That is not an argument against attribution. It is an argument for understanding what attribution actually measures, and what it cannot measure, no matter how sophisticated the model.
Attribution modeling connects ad exposure to conversion outcomes across the customer journey
What Attribution Models Are Actually Doing
Attribution models assign credit for a conversion to the marketing touchpoints that appeared in the customer's journey before that conversion occurred. They divide credit among the channels and interactions that were present, using different rules for different models.
What attribution models do not do: they do not measure causality. They do not tell you whether any of those touchpoints caused the conversion. They tell you which touchpoints were present and assign them credit according to a set of rules that were chosen before the model ran.
A user who saw a YouTube ad, received an email, clicked a retargeting ad, and then searched for your brand name to buy, how do you divide credit among those four touchpoints? Every attribution model answers this differently. None of them can tell you which touchpoint, if any, actually caused the conversion. The customer might have bought regardless. The attribution model has no way to know.
This is not a problem that better models solve. Data-driven attribution, the most sophisticated common model, is still fundamentally answering the same question as last-click: among the paths that converted, which touchpoints appeared more often? That tells you about correlation between touchpoint presence and conversion. It does not tell you whether the touchpoints caused the conversion.
Every Model Applied to the Same Journey
To make the differences between attribution models concrete, take a single customer journey and apply each model. This journey is not hypothetical, it is a typical e-commerce path:
A customer sees a Facebook prospecting ad on a Tuesday. Three days later, they click a YouTube video ad for your brand. On Saturday, they receive an email with a promotion and open it but do not click. The following Monday, they search for your brand name on Google and click a paid search ad. They buy. Order value: $120.
Four touchpoints: Facebook (view), YouTube (click), Email (open), Google Search (click).
Last-click attribution: Google Search gets $120. The other three touchpoints get $0.
First-click attribution: Facebook gets $120. The other three touchpoints get $0.
Linear attribution: Each touchpoint gets $30. (If you count the Facebook view as a touchpoint, some models do, some do not.)
Time-decay attribution: More recent touchpoints get proportionally more credit. Google Search gets roughly $50-60, Email gets $30-35, YouTube gets $20-25, Facebook gets $10-15.
Position-based (U-shaped) attribution: Facebook (first touch) and Google Search (last touch) each get 40% of credit ($48 each). YouTube and Email split the remaining 20% ($12 each).
Data-driven attribution (DDA): Algorithmic, uses your actual conversion path data to estimate the contribution of each touchpoint by comparing paths that converted with paths that did not. The output varies by account. One important caveat: Google's DDA is trained on Google-observable data. It cannot see touchpoints outside Google's ecosystem. Channels like Facebook, TikTok, and email are underrepresented or absent in Google DDA's training data, which means their contributions are systematically underweighted.
Look at that range. Facebook gets $0 under last-click and $120 under first-click. Google Search gets $120 under last-click and $48 under position-based. The "right" answer for any given touchpoint varies by more than 10x depending solely on which model you use. And none of these numbers reflect whether the touchpoint actually caused the conversion.
The Four Fundamental Problems with Attribution
Understanding why attribution has structural limitations, not just model choice limitations, matters for using it correctly.
1. It Only Sees What It Can Track
Cross-device journeys are largely invisible to attribution. If a customer saw your YouTube ad on a smart TV, read your blog post on a work laptop, and then bought on a personal phone, the most your attribution tool can connect is the last two interactions at best, and only if the user was logged in or identified via email match.
Incognito browsing, ad blockers, cookie expiration, and device switching all create gaps. Offline behavior, in-store purchases, phone calls, conversations with colleagues, is completely outside attribution's view. Studies consistently show that 20-40% of the customer journey is invisible to digital tracking tools. That number is rising as privacy protections strengthen.
Attribution is not a model of the customer journey. It is a model of the trackable portion of the customer journey. The trackable portion systematically over-represents bottom-of-funnel, digital, last-touch interactions.
2. Platforms Grade Their Own Homework
Meta's attribution shows Meta ads performed well. Google's attribution shows Google ads performed well. TikTok's attribution shows TikTok ads performed well. This is not coincidence, each platform's attribution window and logic is set by the platform itself, an entity with a direct financial interest in showing high performance.
This is not fraud. It is a structural incentive problem that produces predictable systematic bias. Platforms set attribution windows that maximize the conversions they can claim credit for. Meta's default 7-day click / 1-day view window claims credit for any conversion that happens within seven days of a click or one day of an ad view, even if the user converted via another channel, even if they would have converted without any ad.
View-through attribution is particularly aggressive: Meta counts a conversion as attributed if the user was shown your ad in their feed and converted within one day, whether or not they ever clicked it. This dramatically inflates Meta's attributed conversion count, particularly for retargeting audiences with high organic purchase intent.
3. It Measures Correlation, Not Causality
A user who was already intending to buy your product will naturally encounter many marketing touchpoints before converting, they are actively searching, reading reviews, comparing options. Attribution gives credit to those touchpoints for the conversion. But the touchpoints did not cause the conversion; the prior intent caused it. The touchpoints happened to be present.
This is the fundamental problem with all rule-based attribution models, and it is why incrementality testing provides something attribution cannot: actual causal evidence from an actual experiment, not inferred credit assignment based on correlation.
4. Cookie Deprecation Is Making This Worse
Third-party cookie deprecation, iOS 14's App Tracking Transparency, and browser privacy restrictions have already reduced the coverage and accuracy of digital attribution significantly. Meta estimates that iOS 14 caused it to underreport iOS web conversions by roughly 15% immediately after the change. The actual impact on attribution accuracy varies by business type, but the direction is consistent: tracking coverage is shrinking, gaps are growing, and the degree to which attribution reflects reality is declining.
This trajectory will continue. The infrastructure that made deterministic multi-touch attribution possible in the early 2010s, third-party cookies, cross-app tracking, device fingerprinting, is being systematically restricted. Attribution tools are adapting with modeled data and statistical inference, which helps, but also introduces new assumptions and error sources.
What Attribution Is Actually Useful For
Given these limitations, why use attribution at all? Because it is useful when applied to the right questions.
Intra-channel optimization: Within a single platform, attribution signals are valuable for comparing campaigns, creatives, audiences, and objectives against each other. If you are choosing between two Meta prospecting audiences, both operating under the same attribution model and the same platform biases, the relative comparison is meaningful even if the absolute numbers are inflated. Use Meta's attribution to optimize Meta campaigns. Use Google's attribution to optimize Google campaigns. Do not use either to compare Meta against Google.
Customer journey sequencing: Understanding which channels appear early versus late in converting journeys, regardless of how you assign credit, can reveal insights about how customers discover you, what content they engage with, and what channels are doing awareness versus conversion work. Multi-touch attribution path data is useful for journey analysis even if the credit numbers are unreliable.
Qualitative directional signals: Attribution is a leading indicator. If a new channel you are testing shows zero attributed conversions across 30 days and millions of impressions, something is wrong, either the tracking is broken, the audience targeting is off, or the channel genuinely is not working. Attribution as a sanity check on new channel performance is reasonable, as long as you are not making budget allocation decisions based on the attributed ROAS.
Calibration baseline: When you run an incrementality test and discover Facebook retargeting's true iROAS is 1.4x versus an attributed ROAS of 4.8x, you now have a calibration factor (roughly 0.29x, multiply attributed ROAS by 0.29 to get an approximate iROAS estimate). You can apply that calibration factor to other Facebook retargeting campaigns you have not yet tested. Attribution data becomes more useful when it is calibrated against actual causal evidence.
How to Combine Attribution with Incrementality Data
The most practical use of attribution in a mature measurement stack is as a signal to calibrate, not a number to trust directly.
Here is the workflow. Run an incrementality test on a channel. Calculate the ratio of iROAS to attributed ROAS, this is your calibration factor. Apply that factor across other campaigns in the same channel with similar audience types. When you run the next test, refine the calibration.
Over 12-18 months of testing, you build a calibration library: a set of adjustment factors that translates your attribution data into approximate iROAS estimates across channels and audience types. This is not as accurate as running an incrementality test on every campaign, but it gives you far better guidance than using attributed ROAS at face value.
The calibrated attribution data feeds into your MMM as an additional validation layer. If your MMM says YouTube has 3x ROI and your calibrated attribution (based on a geo experiment) says 2.8x iROAS, you have high confidence in that number. If the two diverge significantly, investigate before acting.
Attribution alone is an unreliable guide. Attribution calibrated against incrementality data is a useful input to a measurement system that can actually support confident budget decisions.