Marketing Attribution Limitations: Why Tracking Cannot Tell You Why Customers Buy

Marketing attribution has become very good at describing the parts of a customer journey that happen inside systems we can observe. We can see a paid-search click before a purchase, a social interaction two weeks earlier, an email opened in between and perhaps the campaign associated with the first identifiable website visit. More sophisticated attribution models can distribute credit across those interactions rather than assigning the entire sale to whichever channel happened to appear last.

All of that can be useful. It helps marketers understand observable journeys, identify where customers interact with the business and diagnose how different parts of the digital system work together. The problem begins when that behavioural record is treated as an explanation of why the customer bought.

Those are different questions.

Attribution can tell us something about how a customer reached the point of purchase. It cannot tell us, from tracking alone, why the customer wanted the product, why the brand entered their consideration set, why it was preferred to the alternatives or what finally gave them enough confidence to buy.

That information often exists outside the tracked journey altogether.

A customer may have known the brand for years before becoming identifiable in an analytics system. They may have developed a preference through previous experience, advertising, recommendations, reputation, physical availability or repeated exposure long before they opened Google and searched for the product. By the time the measurable journey starts, much of the commercially important work may already have happened.

This is why I think companies often overinvest in tracking the path while underinvesting in understanding the choice.

The measurable journey usually starts too late

Think about a high-consideration category such as cars.

If you had to buy a car tomorrow, you probably already know which brands you would consider. You might open Google to compare models, prices or dealers, but the search did not create the entire consideration set at the moment you typed the query.

Some of those brands have been accumulating mental availability for years. You have seen them on the road, heard people talk about them, experienced them yourself, seen advertising or formed opinions about quality, design, reliability and status.

Once the active purchase journey begins, those existing preferences influence everything that follows.

The customer searches one brand rather than another. They spend longer on certain product pages. They visit a dealer. They click a retargeting advert. Eventually they search the company name and convert.

Digital attribution sees the later stages extremely clearly because those stages leave identifiable traces. The earlier formation of preference is much harder to observe.

That creates a systematic imbalance in measurement. Activity close to the transaction is easier to measure and therefore tends to receive more credit, while the influences that made the customer interested in the first place often disappear from the reporting system.

This is particularly dangerous when attribution is used for budget allocation. A channel close to the sale can appear exceptionally efficient because it repeatedly appears in journeys involving people who were already highly likely to buy. The reporting system then gives the final observable interaction more strategic importance than it necessarily deserves.

The issue is not that the click did not happen. It is that observing the click tells us very little about what created the customer's underlying preference.

More touchpoints do not solve the problem

The obvious response is to track more of the journey.

Instead of observing five interactions, perhaps we can observe fifty. We record impressions, video views, website sessions, emails, product pages, comparison pages, retargeting impressions and searches before the eventual conversion. We can then build a much richer customer path and assign revenue across it.

This improves our description of observable behaviour, but it does not necessarily improve our understanding of customer motivation.

Two buyers can follow almost identical digital journeys for completely different reasons.

One may choose the company because they have trusted the brand for twenty years. Another may have received a strong recommendation from somebody they know. Another may have chosen because of product design, availability, customer service, price or an experience in a physical store.

The tracked behaviour can look similar even though the purchase logic is different.

That matters because marketers are often trying to make strategic decisions from the journey data. They want to know which messages matter, which aspects of the brand create preference, why customers reject competitors and what is stopping people from buying.

Those questions cannot be answered simply by adding more touchpoints.

A more complete behavioural record remains a behavioural record.

This is also why I separate the limitations of attribution from the limitations of incrementality. Attribution is weak at establishing causality because it observes what happened rather than the counterfactual. It is also weak at explaining motivation because much of the customer's reasoning is never present in the tracking data in the first place.

Incrementality can improve the causal question by deliberately changing an intervention and measuring the resulting effect. It can tell us whether a media change generated additional sales under the conditions tested. But even that does not automatically tell us why customers responded, which message changed their perception or why one brand was preferred over another.

That still requires another type of evidence.

If the question is why customers buy, ask the customer

This sounds almost too obvious, but many businesses invest heavily in attribution, MMM and incrementality before they have developed a serious system for collecting buyer evidence.

I do not mean adding a single dropdown at checkout asking, "How did you hear about us?"

That can provide some information, but it often reproduces the same problem as digital attribution. The customer chooses Google, Instagram, TV or Friend, and the organisation creates another channel attribution report.

The more useful questions are about the decision itself.

What first made the brand worth considering? Which alternatives did the customer seriously evaluate? What mattered most when choosing between them? Was there anything that nearly prevented the purchase? What created confidence? Had they known the company before entering the market? Why did they begin looking now?

Those answers can reveal influences that will never appear in a clickstream.

Someone may report that they had wanted the product for a year after seeing it repeatedly used by people they respected. Another may say that reputation mattered more than price. Someone else may explain that a recommendation got the brand into consideration, but a store visit ultimately created enough confidence to purchase.

The important point is not to treat the customer's answer as perfect causal truth. People forget things, simplify complex decisions and sometimes rationalise choices after the fact. Asking customers to identify the single advertising channel that "caused" their purchase simply gives them a causal inference problem they are not equipped to solve either.

Buyer evidence is useful because it reveals stated motivations, consideration, objections and influences that are invisible to behavioural tracking.

It answers a different question.

Treat qualitative evidence as data, not anecdotes

The weakness of qualitative research often comes from how it is used rather than from the evidence itself.

A marketing team runs ten interviews, selects three memorable quotes and builds a narrative around them. That is not the standard I mean when I talk about buyer evidence.

The evidence becomes much more useful when it is collected and analysed systematically.

Post-purchase surveys can capture what mattered to customers who eventually bought. Non-buyer and loss surveys can show what prevented other prospects from doing the same. Sales calls contain recurring objections, competitor comparisons and reasons for urgency. Reviews contain unprompted language about the features, experiences and benefits people consider important enough to mention publicly.

The objective is to classify recurring themes and examine their distribution rather than celebrate individual anecdotes.

If hundreds of customers repeatedly mention trust, that is more useful than one particularly articulate quote about trust. If customers with high order values disproportionately mention one decision factor, that deserves attention. If people who chose a competitor repeatedly describe the same objection, that may identify a commercial problem that no attribution model will reveal.

Free-text responses are especially valuable because they avoid forcing every customer into a predefined channel category. They allow unexpected reasons to emerge.

Modern language models make analysing large quantities of this material much easier, but the underlying discipline remains the same. Categories need to be defined carefully, classifications should be auditable and somebody who understands the business still needs to inspect whether the interpretation makes sense.

At scale, qualitative research stops being a collection of stories and becomes another structured source of evidence.

The disagreement with attribution is often useful

Once buyer evidence exists, I want to compare it with what the attribution system says.

The objective is not to decide which source is correct and discard the other. The disagreement often tells us more than agreement would.

Suppose paid search receives the majority of attributed conversions, while customer research shows that most buyers already knew exactly which company they wanted before searching.

That does not prove paid search has no value. It does suggest that its role may be more heavily weighted towards capturing or facilitating existing demand than the attribution report implies.

Now imagine customers repeatedly mention a particular creator or campaign as the first reason they became interested in the product, while digital attribution gives that activity very little conversion credit. Again, the survey has not established an incremental ROI for the campaign. It has identified an influence that the tracked journey may not represent well.

That creates a hypothesis worth investigating.

The same principle works in the opposite direction. Customers may frequently report that television influenced them while an incrementality experiment finds little detectable sales effect from a recent TV increase. The correct response is not automatically to believe the survey or the experiment.

Perhaps people are remembering advertising that happened before the test. Perhaps the experiment was underpowered. Perhaps TV affects a subset of customers but the aggregate effect is small. Perhaps customers overstate the importance of television because it is memorable.

The disagreement tells us where to look next.

A good measurement system should create better questions rather than force every source of evidence into one number.

Buyer evidence can improve quantitative measurement

This is where qualitative evidence becomes much more than a customer-insight exercise.

It can help determine what should be measured and tested quantitatively.

If buyers consistently say they knew the brand long before entering the active purchase journey, that should affect assumptions about measurement windows and the likely role of upper-funnel marketing.

If store experience appears repeatedly in purchase decisions, store-locator behaviour may deserve more attention in mid-funnel analysis.

If one product benefit consistently drives preference, creative analysis can examine whether advertising that communicates that benefit produces different behavioural or experimental outcomes.

If a particular influence appears repeatedly in customer research but is almost absent from digital measurement, that creates an obvious area for investigation.

The qualitative evidence does not replace the quantitative work. It improves the questions we ask of it.

This is one of the reasons I prefer building marketing measurement in layers. Start with the evidence closest to the question. If the question is why people choose the brand, buyer evidence is more direct than trying to infer motivation from a media coefficient.

Observed behavioural data can then tell us whether the patterns implied by the research appear in what customers actually do. Incrementality can test important causal hypotheses when the decision requires stronger evidence. MMM can then help place those results inside the wider marketing system.

The order matters because each layer constrains the next.

Why another model is often the wrong first answer

There are companies with advanced marketing measurement stacks that still cannot explain, with any confidence, why customers choose them.

They know which channels received the conversion. They know which campaigns touched the customer. They may have a sophisticated MMM estimating channel contribution and an incrementality programme testing major media investments.

But ask what customers value, which competitors they seriously considered, what almost stopped the purchase or why the brand entered consideration in the first place, and the evidence can become surprisingly thin.

That is a measurement gap.

It cannot automatically be solved by buying another model because the missing information is not necessarily hidden somewhere inside the transactional data waiting for a cleverer algorithm to recover it.

Sometimes the information has never been collected.

This is where the principle of using the least assumption-heavy method capable of answering the question becomes useful.

If you want to know whether a marketing intervention caused additional revenue, design a causal test.

If you want to understand how the wider mix relates to business outcomes, an MMM may be appropriate.

If you want to understand why customers say they chose you, ask them and analyse the responses properly.

Sophistication should come from choosing the right evidence for the question, not from choosing the most complicated methodology available.

Attribution is useful when it stays within its boundary

None of this is an argument for abandoning attribution.

Digital attribution remains useful for understanding observable customer paths, diagnosing campaign and tracking problems, analysing landing-page behaviour and helping teams operate marketing channels.

Its limitation appears when the output is expected to explain more than the underlying data contains.

The fact that a customer converted after clicking paid search is useful information. It does not tell us how much of the underlying demand paid search created. It does not tell us why the customer preferred the company. It does not reveal the full set of influences that existed before the measurable journey began.

Those questions require different evidence.

Attribution can describe the route. Buyer evidence can help explain choice. Behavioural evidence can show how customers progress. Incrementality can test the causal effect of a defined intervention. MMM can help estimate relationships across the wider marketing system.

No single one of those perspectives is sufficient on its own.

The mistake is pretending that because one of them produces a precise number, it must contain the entire answer.

Growth Dynamics uses buyer-reported evidence alongside observed demand, behavioural data, execution analysis, incrementality and MMM because customer choice cannot be reconstructed completely from tracking data.

Knowing how somebody reached the checkout is useful.

Understanding why they wanted to buy from you requires different evidence.