In our first post in this series, we described the three pillars of classical customer research — the sample, the survey, the retrospective report — and the pressures that are reshaping all three at once. This post examines the first of the four methodological shifts at the heart of that reorganization: the move from probability-sampled inquiry to population-level behavioral observation.

It is a shift that has been building for two decades and is now reaching the point where it changes how serious organizations approach customer understanding. Where it applies, it changes both what can be known and how confidently it can be known. Where it does not apply, the older methods remain — and we will come back to that boundary at the end.

The logic of the sample, and where it came from

The probability sample is one of the great methodological inventions of the twentieth century. Its central insight is that you do not need to ask everyone to know about everyone. If you select respondents in a way that gives every member of a population a known, non-zero chance of being selected, you can quantify the uncertainty around any estimate you compute from them. The sample of a few thousand becomes a statement about the population of millions, with a margin of error you can defend.

This was the foundation on which the modern market research industry was built. Survey panels, telephone polling, mall intercepts, online recruitment — every method that came after assumed that the route to understanding the customer ran through a carefully drawn subset of them. For most of the twentieth century, this assumption was correct because there was no alternative. The only way to know what customers thought, preferred, or intended was to ask some of them in a way that allowed inference to the rest.

What changed

Two things changed, and the combination is what matters.

First, the practical foundation of probability sampling has degraded. Telephone response rates, which were around 30 percent in the late 1990s, are now in single digits. Online panels increasingly recruit from a pool of professional respondents whose representativeness of the target population is contested. The cost per usable response has risen sharply, while the willingness of the population to respond has fallen. Probability sampling did not stop working in theory; it became increasingly difficult to do well in practice.

Second, and more importantly, an alternative emerged that was simply not available before. When customers interact with digital systems — websites, apps, point-of-sale terminals, support channels, loyalty programs — they leave continuous records of what they actually do. These records are not a sample. They are, for many categories of customer behavior, the entire population.

A retailer with five million loyalty members does not need to sample to know what those members purchased last quarter. The data covers all of them. A media platform with twenty million monthly users does not need to survey to understand which articles were read, how long, in what order. A bank does not need to ask its customers how often they use mobile banking; the logs answer the question completely and continuously.

This is what we mean by the shift from sample to trace: from a methodologically constructed subset to an exhaustive behavioral record.

What trace data is good at

Where it is available, trace data has properties the survey could never match.

It is continuous, not periodic. You do not wait for the next wave of fieldwork to see what changed; the data updates with each customer interaction.

It is behavioral, not declared. People are notoriously unreliable narrators of their own behavior — they overestimate how much they exercise, underestimate how much they spend, and misremember which brands they bought. Trace data records what actually happened. The gap between stated and revealed preference, which has occupied consumer researchers for a century, becomes operationally manageable when the revealed side is directly observable.

It is at scale, not sampled. The methodological worry about whether the sample represents the population dissolves when the data is the population. There is no margin of error in the classical sense, because no inference from a subset is being made.

And it is disaggregable. Trace data can be filtered by segment, channel, time period, and customer characteristic without requiring a new study. Surveys had to be designed with anticipated cuts in mind; if a new question emerged, you went back into the field. With trace data, most new questions can be answered by querying what is already there.

What trace data is not good at

This is the part where the shift gets more interesting. Trace data is excellent at answering questions about behavior, but it is silent on the questions surveys were originally designed to address.

Trace data tells you that a customer purchased a competitor's product. It does not tell you why. Trace data tells you that a customer canceled a subscription. It does not tell you what frustrated them — only that they stopped paying. Trace data tells you that a customer abandoned a cart at the payment step. It does not tell you whether it was the price, the shipping cost, the form length, or something else entirely.

For these why questions, the older methods remain necessary in some form. But what has changed is how those methods are used. In a trace-data world, qualitative inquiry becomes targeted: instead of broad surveys to a general sample, you talk in depth with the specific customers whose trace behavior raised the question. The qualitative interview becomes more like a diagnostic instrument than a discovery instrument.

And — this is where the second shift of our series enters, which we will treat in the next post — much of what surveys used to be the only way to get can now be approximated by analyzing what customers already say in reviews, support interactions, social posts, and other unstructured text. The hard line between behavioral observation and stated preference is becoming softer than it used to be.

What this means for the customer research function

Three implications for organizations that operate customer research as a discipline.

The first is what to ask. With trace data covering a large share of behavioral questions, the survey budget can — and should — be reallocated toward questions that genuinely require asking: motivations, perceptions, future intentions, sensitive topics. The survey was over-applied for a century because it was the only general-purpose tool available. Now that more specialized tools exist for behavioral questions, the survey can be used where it is actually the best instrument.

The second is how to combine sources. Trace data alone produces a description of what is happening; it does not produce a diagnosis of why or a recommendation about what to do. The customer-research function that thrives in this environment is the one that builds disciplined workflows for combining trace data, targeted qualitative inquiry, and other instruments into integrated views. The skill that matters is not running surveys; it is orchestrating evidence.

The third is what to instrument. Trace data exists only where systems have been built to capture it. Organizations that have invested in event-level logging, identity resolution, and clean data infrastructure have access to a class of customer understanding that organizations without that infrastructure simply do not. The methodological shift from sample to trace is, at its base, also a shift in what the customer-research function depends on the data team to deliver.

Where this leaves the sample

The probability sample is not disappearing. For populations that do not generate trace data — non-customers, potential customers, customers of competitors — sampling remains the only access route. For questions that are sensitive in ways that behavior does not reveal — political attitudes, brand perceptions, future purchase intent at the category level — surveys remain necessary. For regulatory contexts that require attested data of a particular kind, the survey remains the instrument that produces it.

But the sample's role has changed. It has gone from being the method of customer research to being one of several methods, each appropriate to a different class of question. That is a significant demotion from the central position the sample held for the better part of the twentieth century. It is also, on balance, a healthier place for the method to be: used where it is actually best, rather than used because nothing better was available.

In the next post in this series, we will turn to the second of the four shifts — from asking to reading — and examine how the analysis of unstructured customer expression is changing what surveys are even needed for.