Attribution is dead. It was dead on arrival.

Marketer: I’d like to return this attribution model.

Vendor: Return it? Whatever for?

Marketer: It’s dead.

Vendor: Dead? No, no. It’s just suffering from a little signal loss.

Marketer: Signal loss? It hasn’t produced a credible causal answer in twenty years.

Vendor: It’s a sophisticated model.

Marketer: I spent three million dollars on paid media, and it told me Facebook, Google, email, affiliate, and branded search were each responsible for roughly four million dollars of revenue.

Vendor: Excellent performance.

Marketer: My company only made five million dollars.

Vendor: Cross-channel synergy.

Marketer: They all claimed the same customers.

Vendor: Multi-touch.

Marketer: Then Apple removed tracking.

Vendor: Modeled conversions.

Marketer: Third-party cookies disappeared.

Vendor: Data-driven attribution.

Marketer: Half the journey happens inside platforms I can’t see.

Vendor: Advanced machine learning.

Marketer: The platforms are estimating their own performance.

Vendor: Privacy-safe measurement.

Marketer: Google is grading Google. Meta is grading Meta. Everyone somehow gets an A.

Vendor: Strong ROAS.

Marketer: I turned one of the channels off and sales didn’t move.

Vendor: Delayed halo effects.

Marketer: I turned it off for three months.

Vendor: Brand lift.

Marketer: Nothing happened.

Vendor: Omnichannel influence.

Marketer: It. Is. Dead.

Vendor: I prefer to say the methodology is evolving.

And that, more or less, is how we have spent the last two decades talking about attribution, with deepest apologies to Monty Python.

Every time reality presents another reason to question whether attribution can actually tell us what caused a business outcome, the marketing industry invents another layer of methodology to keep the idea alive. First touch was too simplistic, so we tried last touch. Last touch was obviously flawed, so we built multi-touch. Tracking deteriorated, so we added modeling. Platforms became increasingly opaque, so we accepted their estimates. When the numbers still disagreed, we bought software to reconcile them.

At no point did we stop and ask the more uncomfortable question: What if the attribution model isn’t malfunctioning? What if the premise itself is dead?

We need to stop trying to fix attribution. Not improve it, not add another touchpoint, not layer on another model, and certainly not sprinkle AI over it and pretend the underlying problem has gone away. The fundamental issue with attribution is not that we have failed to find the right methodology. It is that attribution has been answering the wrong question from the beginning.

For years, marketers have obsessed over some version of the same question: Which channel gets credit for this conversion? Was it the first touch, the last touch, the Facebook ad in the middle, the branded search click five minutes before purchase, or some mathematically weighted combination of all of them? Or was it the damn billboard?

We created first-touch models, last-touch models, linear models, time-decay models, position-based models, multi-touch models, data-driven models, and increasingly sophisticated statistical approaches, all in pursuit of a more accurate way to distribute credit.

But there is a much more important question that has to come first: Would the customer have converted anyway?

If the answer is yes, then there may be nothing meaningful to attribute in the first place. That is the inconvenient truth underneath modern marketing measurement. We have spent years refining systems for assigning credit before proving that the marketing activity itself caused the outcome.

Attribution confuses being present with being responsible

Imagine I am already planning to buy a pair of running shoes (I’m not. I only run from bears). I research them for a week, read reviews, visit the brand’s website twice, and eventually search for the company by name. A paid search ad appears, I click it, and then I purchase.

Paid search now gets the conversion.

Depending on the attribution methodology, it may get all of the credit, some of the credit, or a weighted portion of the credit alongside other channels I encountered during the journey. From the dashboard’s perspective, the system has worked beautifully. A customer clicked an ad and then bought something, so revenue can be associated with that marketing interaction.

Except for one small problem: I was probably going to buy the shoes anyway.

The paid search ad did not necessarily create the sale. It may simply have been present at the end of a decision-making process that was already well underway. The platform observed the conversion happening near an ad and inferred value from that proximity. But observing that one event preceded another does not establish that the first event caused the second.

This is where attribution has always struggled. It is exceptionally good at reconstructing what happened before a conversion, but much less capable of proving why the conversion happened. It tells us that an interaction occurred, that a customer passed through a channel, or that an ad was somewhere in the path to purchase. What it does not reliably tell us is whether that marketing interaction changed the outcome.

That distinction matters because return on ad spend (ROAS) is supposed to answer a business question: If I spend another dollar here, will I create additional revenue? Attributed ROAS often answers a much easier question instead: How much revenue can I associate with customers who encountered this marketing?

Those numbers can look remarkably similar on a dashboard, but economically they are worlds apart.

Then we handed the grader the answer key

The attribution problem became even more problematic with the rise of walled gardens because the companies selling the media increasingly became the companies measuring the media.

Think about the arrangement for a moment. You give a platform money. The platform serves the ads, observes some portion of customer behavior, determines which conversions should be associated with those ads, reports the resulting ROAS, and then recommends how much more money you should spend on the platform. We have somehow normalized a system in which the seller is deeply involved in determining the effectiveness of what it sold.

To be fair, these platforms are operating in an increasingly difficult measurement environment. Signal loss is real, privacy restrictions are real, and the ability to directly connect ad exposure to customer behavior has eroded.

Google, for example, explicitly describes using conversion modeling to estimate conversions when direct links between ad interactions and conversions cannot be observed because of privacy or technical limitations. Those modeled conversions are then included alongside observed conversions in reporting.

That may make the platform’s measurement more complete, but it does not make the measurement independent and certainly not accurate. The same company selling the media is still participating in determining the value of the media. Put less politely, we let the platforms grade their own homework and then use those grades to decide next semester’s tuition.

Privacy did not kill attribution. It exposed it.

It has become fashionable to say that attribution broke because of privacy. Apple restricted cross-app tracking, an all-out war was launched on third-party cookies, regulators constrained the collection and movement of personal data, and consent requirements made customer journeys harder to stitch together. All of that is true.

But privacy did not create attribution’s fundamental flaw. It merely exposed it.

Even in the golden age of cookies, when marketers could follow users around the internet with astonishing precision, we still were not reliably answering the causal question. We were simply better at connecting dots. A customer saw Ad A, clicked Ad B, received Email C, searched for the brand, and eventually purchased. That sequence could be reconstructed in impressive detail, and the resulting customer journey looked wonderfully scientific.

But even a perfectly reconstructed customer journey does not tell you what would have happened if none of those marketing interactions had occurred.

That is the counterfactual, and without a credible counterfactual, we cannot confidently establish incremental impact. We can observe a path with exquisite precision and still misunderstand what actually caused the customer to buy.

So we invented more attribution

When last-touch attribution began to look obviously simplistic, the industry did not question the premise behind attribution. Instead, we created more attribution.

First touch promised to recognize where the journey began. Linear attribution spread credit across the entire journey. Time-decay attribution gave more weight to recent interactions. Position-based models emphasized the beginning and end. Multi-touch attribution attempted to account for every relevant interaction, and data-driven attribution introduced increasingly sophisticated algorithms intended to distribute credit based on observed patterns.

Each new approach sounded more advanced than the one before it, but sophistication does not necessarily solve a flawed premise. We became better at dividing the credit without first proving that marketing deserved the credit.

We were debating who deserved the assists before establishing whether marketing had actually scored the goal.

This is not a semantic distinction. It determines where billions of dollars in marketing investment move. Even the advertising industry now acknowledges how strained the measurement ecosystem has become, with privacy regulation, signal loss, fragmented data environments, and platform-embedded optimization all making it harder to confidently connect media exposure to outcomes.

At some point, we have to entertain the possibility that the problem is not that we need a better attribution model. The problem may be that attribution is simply the wrong place to start.

Ask the $10 million question before the $10,000 question

Most marketing organizations begin measurement by asking which campaign performed best, which channel had the highest ROAS, which audience converted, which creative won, or which keyword generated the most revenue. These are perfectly reasonable questions, but they are second-order questions.

Before asking any of them, there is a more fundamental question: Does this channel actually create incremental business value?

Not whether conversions appear in the channel. Not whether revenue is attributed to it. Not whether the platform reports a 7.4x ROAS. The question is whether the business would materially change if that marketing activity did not happen.

If I turn the channel off, what happens? If I withhold the campaign from an otherwise comparable group, do sales decline? Do I acquire fewer customers? Does average order value change? Does retention move? Or does the business continue on almost exactly the same trajectory?

That is the question incrementality is designed to answer. Incrementality looks at the difference between what happened and what would have happened anyway. A well-designed randomized experiment can compare a group exposed to a marketing intervention with an equivalent group that is not exposed, helping isolate the causal impact of the marketing itself.

That is profoundly different from assigning credit. It shifts the question from “Which channel gets the conversion?” to “Did this activity actually change customer behavior?”

That is the question marketing measurement should have been asking all along.

Attribution and incrementality can tell completely different stories

Consider two hypothetical channels.

Channel A reports $1 million in spend and $8 million in attributed revenue, producing an 8x ROAS. Channel B reports the same $1 million in spend but only $3 million in attributed revenue, producing a 3x ROAS.

If you are optimizing purely on attribution, the decision looks obvious: move more money to Channel A.

Now imagine that you run a controlled incrementality test. The test reveals that Channel A generated only $500,000 in revenue that would not otherwise have occurred. The majority of customers exposed to that channel were already extremely likely to purchase, and the channel was simply very effective at appearing near the end of the buying process.

Channel B, on the other hand, generated $2 million in genuinely incremental revenue.

Now which channel is actually doing more work for your business?

The attribution dashboard told you Channel A. Causal measurement told you Channel B.

This is why attribution-based ROAS can be so dangerous. It can reward channels for harvesting demand rather than creating it. In fact, the better a platform becomes at identifying people who are already likely to convert, the more impressive its attributed ROAS can appear.

That may represent excellent targeting. It does not necessarily represent excellent incremental marketing.

This does not mean every model goes in the trash

None of this means observational measurement is useless. Customer journey data can still help us understand how people interact with a brand. Attribution can help teams understand common paths to purchase and operational patterns. Marketing mix modeling can help analyze aggregate relationships between spend, seasonality, macroeconomic factors, and business outcomes.

The mistake is asking these methodologies to prove something they cannot prove on their own.

Measurement methods should be used for the questions they are actually capable of answering. Attribution can tell you where a conversion appeared to come from. MMM can help identify macro-level relationships across channels and time. Journey analytics can show how customers interact with touchpoints.

But none of those things should automatically be confused with causal proof.

Incrementality is different because it explicitly asks what happened because of the marketing that would not otherwise have happened. It is centered on a counterfactual, and that is the foundation required to make a causal claim.

The distinction is simple but important: attribution tells you who was nearby when the conversion happened. Incrementality tells you whether marketing changed the outcome.

The measurement hierarchy needs to change

Marketing measurement should happen in a different order.

The first question is does it work? Does this channel, campaign, or intervention create incremental business value? That should be established before we spend time optimizing the details.

The second question is for whom does it work? Which customers, segments, products, markets, or situations actually respond incrementally to the intervention?

Only then should we move to what specifically works? Creative, message, offer, timing, frequency, channel, audience, and sequencing all matter, but they matter much more once we know we are optimizing something that actually changes behavior.

Finally, we can ask where should the next dollar go? At that point, budget allocation becomes much more intelligent because we are comparing the incremental value of different investments rather than comparing the amount of revenue each system has managed to claim.

Notice what is missing from the top of that hierarchy: who gets credit?

That is because credit is fundamentally an accounting question.

Growth is a causal question.

This matters even more in the age of AI

There is another reason we need to break our addiction to attribution now: AI is about to dramatically increase the number of marketing decisions companies make.

More audiences will be created. More creative will be generated. More offers will be tested. More messages will be personalized. More budget decisions will happen automatically, across more channels, at a speed no human marketing team could possibly manage manually.

If the learning system underneath all of that activity is optimized around attribution, we have a serious problem.

AI will become extraordinarily efficient at maximizing whatever signal we give it. If we give it attributed conversions, it will learn how to maximize attributed conversions. That is not necessarily the same thing as maximizing incremental business outcomes.

Would you really choose to optimize towards ad clicks rather than conversions?

We could very easily build the world’s most sophisticated machine for finding customers who were already going to convert and then taking credit for the conversion.

The opportunity is much bigger than that.

Imagine a marketing system that continuously learns that a certain message causes a specific type of customer to purchase, while another offer does nothing. Imagine it learns that one channel increases acquisition but has no impact on retention, while another changes long-term customer value. Imagine it discovers that one intervention is highly effective for new customers but simply cannibalizes purchases among loyal ones.

Now the system is not simply getting better at predicting outcomes. It is getting better at creating them.

That is the difference between correlation and causal intelligence, and it is where marketing measurement has to go next.

Attribution was dead on arrival

Privacy did not kill attribution. Walled gardens did not kill attribution. Cookie deprecation did not kill attribution.

They simply made its limitations harder to ignore.

Attribution was compromised from the moment we decided that observing a customer interaction meant we could determine how much credit that interaction deserved. For years, we responded to that flaw by inventing increasingly elaborate ways to divide the credit, and an entire ecosystem of technology vendors emerged to help us do it.

But businesses do not actually need better ways to distribute credit. They need better ways to make decisions.

Should I spend money here? Did that investment change customer behavior? What happened because of my marketing that would not have happened without it? Where should I invest the next dollar?

Those are the questions that matter.

Before you ask which channel deserves the credit, prove there was something worth taking credit for, because, yes, this parrot is dead.