In the first post in this series, we described the three pillars of classical customer research and the pressures reshaping all three. In the second post, we examined the first methodological shift — from sample to trace — and the move from probability-sampled inquiry to population-level behavioral observation. This post takes up the second shift, which addresses a class of questions that trace data cannot answer.

When customers behave in unexpected ways, trace data shows what they did but not why. For the why questions, the survey has been the canonical instrument for a hundred years. But something has changed about the supply of customer language that makes a different approach viable for many of the questions surveys used to handle. Customers express themselves, unprompted, at a scale and in a form that — until very recently — could not be analyzed except by hand. That is no longer true.

The economics of unstructured customer expression

Customers have always talked about products, brands, and categories. They complained to friends, wrote letters, called support lines, posted on forums, reviewed on retailers' sites. The volume of this unprompted expression is enormous. A large consumer brand may receive millions of mentions per year across reviews, social platforms, and customer-service channels. A mid-sized retailer accumulates hundreds of thousands of product reviews and support transcripts. A bank logs every customer-service call.

For most of the past century, this material was either invisible to the firm or, where visible, prohibitively expensive to analyze at scale. Reading a thousand reviews to understand recurring themes required a trained analyst working through them one at a time. Coding ten thousand was a project. Coding a hundred thousand was effectively impossible, which meant that the unstructured expression of most customers — beyond the few thousand whose reviews were sampled — never made it into the firm's understanding of itself.

Large language models have changed the economics of this analysis by roughly two orders of magnitude. What used to take a coding team weeks now takes a well-prompted system hours. More importantly, the analysis can be reproduced, re-run with different questions, and applied to entire archives rather than samples of them. The constraint that made unstructured customer expression a marginal source — the human cost of reading at scale — has largely fallen away.

What this changes

The implications are not subtle. For a substantial class of questions that organizations have historically answered by surveying customers, they can now answer by reading what customers already said.

What do customers complain about most? They have told you, repeatedly, in support tickets and reviews. What features matter to them in a category? They have written about it in forum discussions and competitor reviews. How does a recent campaign land? They are responding to it on social channels in real time, with a frankness that surveys rarely elicit. How does customer perception of a brand evolve over months and years? The archive of unstructured expression contains the answer if you have the means to query it.

This is what the shift from asking to reading means. It is the recognition that for many questions, the customer has already answered — not in response to a designed instrument, but in the natural course of having opinions about things. The role of the customer-research function is no longer primarily to construct instruments that elicit responses; it is, increasingly, to construct systems that read responses that were given without prompting.

What "reading" actually means

It is worth being specific about what large language models can do with unstructured customer text, because the popular framing tends to either oversell or underspecify.

A well-built system can do at least four things competently.

It can identify themes in a corpus of text: clustering similar complaints, surfacing the dominant praise patterns, distinguishing genuine product issues from generic dissatisfaction. This is what topic modeling used to do, only with substantially better semantic understanding and much less manual tuning.

It can track sentiment over time, including granular sentiment about specific product attributes rather than overall brand-level sentiment. A retailer can now see not just that reviews of a particular product have grown more negative but that the negativity concentrates around fit issues that emerged after a manufacturing change.

It can answer specific questions posed against the corpus. "What are the top three reasons customers cite for canceling subscriptions?" was, until recently, a question that required a dedicated study. It is now a question you can ask of your support-ticket archive and receive a defensible answer in minutes.

And it can integrate text with other data sources — linking review themes to product attributes, sentiment to campaign timing, complaint clusters to specific customer segments. The reading does not happen in isolation; it becomes one source within a connected evidence base.

What it does not mean

A few cautions, because the gap between what these systems can do and what they are sometimes claimed to do is substantial.

Reading what customers already said tells you about the customers who said something. It does not tell you about the customers who said nothing — and the silent customers are often the majority. For categories where review behavior is concentrated in particular demographics or particular kinds of experiences, the reading approach inherits that selection bias. A retailer learning from reviews learns disproportionately from customers who had either a very good or a very bad experience; the modal customer, whose experience was unremarkable, leaves no trace.

Reading also produces patterns, not explanations. The system can tell you that complaints about delivery speed have risen 40 percent over the past quarter. It cannot tell you whether this reflects a real degradation in delivery, a change in customer expectations, a competitor's new offering raising the bar, or a shift in which customers are reviewing. The interpretation of patterns is still a human task.

And reading is not a substitute for asking when the question is genuinely forward-looking. Customers can only express themselves about things they have experienced or considered. Questions about hypothetical products, novel category extensions, or unfamiliar value propositions cannot be answered by reading historical text. For those, you still need either to ask, to test, or — as we will see in the next post — to simulate.

The hybrid that is emerging

What we are watching, in customer-research functions that have begun to integrate large language models into their practice, is the emergence of a hybrid model.

The default for most questions has shifted. Instead of starting with "what survey should we run?", the default question is becoming "what do we already know — from trace data, from existing customer expression, from prior research — and what genuine gaps remain?" Surveys, when commissioned, are commissioned to address the residual gaps rather than to produce the primary evidence.

This changes the cadence of customer research substantially. The classical pattern was a sequence of discrete studies, each requiring weeks of design and fielding. The emerging pattern is a continuous reading capability against archives that update daily, supplemented by targeted asking where the reading runs out.

It also changes what the customer-research function looks like internally. The skills that matter most in a reading-centric model are different from those of survey design: prompting and querying large-language-model systems against text corpora, integrating text sources with behavioral and transactional data, distinguishing robust signal from selection-biased noise, and connecting findings to operational decisions. These are not the survey-methodology skills the field has trained for. Functions that recognize the shift are reconfiguring their teams accordingly; functions that do not will find themselves competing on instruments whose advantages have eroded.

What this means for surveys

The survey is not going away, for the same reasons the probability sample is not going away. There are questions reading cannot answer, populations that do not generate enough unstructured expression to be readable, and contexts where direct asking remains the only valid method.

But the survey's center of gravity is moving. From a default instrument applied broadly because nothing better existed, it is becoming a targeted instrument applied to questions for which it is uniquely suited. That is, again, a healthier place for the method to be — used where it is best, rather than used by default.

In the next post in this series, we will turn to the third shift — from asking to simulating — and examine what happens when customer response can be anticipated through synthetic audiences rather than measured through real fieldwork.