Just a few years ago, marketing rested on a fairly predictable pattern: market research took months and cost a fortune, customers found a brand mainly through Google search, and service quality was measured by a handful of well-established indicators. Today every one of those elements is undergoing a rapid overhaul. Generative artificial intelligence is compressing consumer research from months into days, reshaping how customers find companies online in the first place, and simultaneously forcing managers to think more rigorously about which customer data actually matters and whose voice is worth listening to first. This is not one revolution but several running in parallel — and this article looks at four of them, drawing on the latest research and the experience of marketing practitioners.
1. Market research: from months to days
The classic market research process — defining the problem, designing the study, sampling, collecting data, analysis, conclusions — has looked much the same for decades, taking anywhere from several weeks to several months and costing tens to hundreds of thousands of dollars. By the time the results land on a decision-maker's desk, the market has often already moved on. Neeraj Arora, Ishita Chakraborty, and Yohei Nishimura of the Wisconsin School of Business, drawing on their own study published in the Journal of Marketing and on conversations with marketing leaders, show that large language models (LLMs) are starting to fundamentally change this calculation across the research and insights industry, currently valued at $153 billion.
Digital twins of consumers
The most spectacular application is the use of so-called digital twins — synthetic, data-driven personas capable of simulating decision processes, shifting preferences, and reactions to marketing stimuli. Instead of waiting to collect new data, analysts can test product concepts, assortment decisions, or campaign reactions in a virtual environment before the company incurs real costs.
A good example is the collaboration between the research firm Evidenza and a German ICT-sector company that wanted to test whether B2B customers would trust it on cybersecurity and the storage of sensitive data in the cloud. The team built synthetic samples of decision-makers based on the demographic and behavioral profiles of real respondents, then asked an LLM to adopt these personas and answer the research questions. Comparing the results with a traditional study on real people showed correlations of 0.75–0.88 — strong enough to conclude that synthetic samples produce directionally accurate findings, and at a fraction of the time and cost of a classic study.
The authors emphasize that LLM-generated data can even be deeper and more insightful than human responses — models are not constrained by time or reluctance to elaborate. Digital twins work particularly well where reaching real respondents is difficult or expensive — for example, among physicians or senior executives, who rarely have time to take part in surveys. What's more, synthetic answers can be used before a survey ever reaches real people, to test its design, remove weak questions, and sometimes even avoid running the full study altogether — if the simulation results are already unambiguous.
AI as interviewer, evaluator, and probe
The second area of transformation is qualitative data collection. LLMs can conduct interviews while performing three functions at once: asking questions according to a script, scoring the quality of a respondent's answer on a 1–100 scale, and, when the score falls below a threshold, asking for clarification. The Wisconsin researchers found that data from AI-moderated interviews preserves the meaning of human answers, and that independent human raters even scored them higher for depth and insight than traditionally conducted interviews. An AI moderator has no geographic or time constraints and can run thousands of conversations simultaneously at a fraction of the cost of classic in-depth interviews — though it still lags behind an experienced researcher at reading tone of voice, body language, or subtle non-verbal cues.
Fast analysis of unstructured data — reviews, customer-service transcripts, social media posts — is also enormously valuable. As the managing director of C+R Research noted, AI tools can process and synthesize what would take days of manual work in a matter of hours. Companies such as Voxpopme, which offer analysis of multimodal qualitative data (video, audio, text), report cost reductions of 30–50% on research projects as a result, along with up to a 60-fold increase in analysis speed.
RAG as the connective tissue between data silos
A standalone LLM, stripped of company context, can nonetheless be unreliable — it tends to generate answers that are too homogeneous and lack the logical coherence typical of human thinking. The solution is retrieval-augmented generation (RAG), which feeds the model with data from internal sources — historical qualitative research, CRM systems, survey results. RAG can become a kind of connective tissue linking scattered insight streams that, in most organizations, still don't "talk" to one another. It is not a perfect solution, though — as the knowledge base grows, retrieval precision and speed decline, and if the system pulls incomplete or context-free data, the quality of the generated insights deteriorates sharply. As the VP of analytics at Procter & Gamble observes, a substantial share of organizational knowledge still exists only in presentations, on slides, or in people's heads — and AI has no way to "read" that.
Humans still hold the wheel
Across all of these applications, the human-in-the-loop principle remains essential. Defining the research problem and designing the study are stages dominated by human market intuition, experience, and budget realities — AI can support them at best in a supporting role. Only at the execution and analysis stage do language models become a genuine "collaborator." As one insights leader at Microsoft put it, generative AI radically speeds up data synthesis and allows work at a previously unattainable scale, but it is still the human who decides how to frame the right questions and translate raw data into business-relevant conclusions.
The authors don't shy away from the risks, either. LLMs inherit the gender, racial, and cultural biases present in their training data, which demands constant oversight. The ease of generating reports creates a temptation to run research "quickly," at the expense of rigor. There are signals that the number of junior marketing roles is shrinking — and it is precisely those roles that, for years, served as the natural training ground where young specialists learned analytical thinking. Finally, digital-twin technology can be abused by respondents who use LLMs to mass-produce fake but credible-sounding answers in online surveys, purely to collect participation incentives faster. The proposed answer to these risks is rigorous analytical oversight, systematic validation of results, and a "test and learn" approach rather than uncritical full-scale AI rollout.
2. Customer experience: fewer metrics, more meaning
Where market research suffered from a shortage of data, customer experience (CX) faces the opposite problem today — an excess of metrics that nobody can meaningfully use. Charles H. Patti, Maria M. van Dessel, and Steven W. Hartley, drawing on consulting work with a group of 14 subscription-service companies, show how the runaway proliferation of CX metrics has paralyzed managers' ability to make decisions.
According to Gartner, large enterprises track more than 50 CX metrics on average, and some as many as 200, scattered across departments and channels — call center, chat, email, website, physical locations. The companies studied by the authors used more than a hundred such measures combined. The problem is that customers today receive a satisfaction survey minutes after every brand interaction, yet the sheer volume of data collected — operational, perceptual, and financial — does not automatically translate into better business decisions. Many metrics are collected simply because "that's how the industry does it," with no reflection on whether they actually help improve the customer experience.
Reduction through regression analysis
The authors proposed a simple but effective tool: regression analysis, which identifies which metrics most strongly — and most weakly — explain variance in key business outcomes such as churn, contact volume, or Net Promoter Score. Among the subscription companies studied, of the 13 metrics tracked in the call-center channel, a single one — the share of calls redirected to the IVR system — alone explained most of the variance in NPS. Meanwhile, service level (the percentage of calls answered within 30 seconds) and the call-transfer rate turned out to be weak predictors of all the variables analyzed — making them strong candidates for elimination.
The result? A real reduction of roughly 15% in the number of metrics tracked in the call-center channel alone, and with it, savings in human and financial resources and less "survey fatigue" among customers, which in any case undermines the reliability of the data collected. Interestingly, in each of the 14 companies studied, managers asked the same question after seeing the results: why didn't we do this sooner? The process also had a side effect — it broke entrenched "we've always done it this way" habits of thought and made it harder to keep tacking on new metrics unreflectively in the future.
Mapping metrics to the customer journey
A second, complementary tool is mapping CX metrics onto the stages of the customer journey — from first contact with the product category, through selection and purchase, to usage, potential churn, and win-back attempts. The authors ran such an exercise in the spirit of design thinking with a group of new customers, mapping the onboarding stage in detail — information search, vendor verification, service activation, first billing, first service request.
The result was instructive: many financial-services firms rely almost exclusively on NPS, ignoring the pre-purchase stage entirely, where trust or brand-position indicators would be far more useful. Similarly, many organizations measure NPS only at the very end of the onboarding process, even though the same metric applied earlier would enable real-time intervention — before a negative first impression turns into churn. Measurement gaps of this kind usually stem from an organization's excessive attachment to one or two favorite metrics, rather than from a lack of data as such.
The authors also point to a cultural factor: organizations oriented primarily around profitability tend to gravitate toward operational and cost-saving metrics, while customer-centric companies invest in understanding the emotions and motivations behind purchase decisions — which, over the longer run, still translates into sales results. In the age of AI, this conclusion takes on added significance: it is precisely clean, sensibly mapped CX data that becomes the raw material language models can later work with effectively — junk, redundant metrics make any automation of analysis harder.
3. Visibility online: a new game for customer attention
The third front of change concerns perhaps the most fundamental function of marketing — whether a customer finds the brand at all. Michael Pettiette and Kimberly A. Whitler describe how generative-AI-based search is upending the mechanisms that digital marketing has relied on for twenty years.
The examples they cite are sobering. One of the largest fitness chains in the US, with an enormous search-marketing budget, was stunned to discover that in AI search results it was outranked by a small, local company from Houston. An executive in the financial industry observed that a customer looking for the best offers used ChatGPT instead of Google — and his company, despite being a market leader with the largest SEO investment, did not appear in the AI's recommendations at all, losing out to a far smaller player. Years of investment in traditional search rankings turn out to carry no automatic weight in the new environment.
Search without clicks
The key phenomenon is so-called zero-click search — a situation where the consumer never has to click a link at all, because the AI platform immediately provides a direct, synthesized answer. This applies not only to chatbots such as ChatGPT, Perplexity, or Gemini, but also to Google's own AI Overview feature, which increasingly eliminates the need to browse further results. Search-engine traffic is declining steadily, and the rising popularity of generative AI tools is only accelerating that trend.
The Information Search Marketing (ISM) framework
Because the industry lacked a shared vocabulary to describe this shift, the authors propose the Information Search Marketing (ISM) framework, covering four pillars: classic SEO (search engine optimization) and SEM (paid search), alongside the new GEO — generative engine optimization, meaning adapting content so it appears in AI-generated answers — and GEM — generative engine marketing, the advertising that accompanies those answers. GEO relates to GEM the way SEO relates to SEM, and ISM is the umbrella concept tying both worlds together.
The authors set out five principles for success in this new environment, framed as the 5A model:
Authority. While traditional search rewards mainly the number of inbound links and domain-quality signals, AI platforms favor high-quality citations and expert opinion, where quality clearly outweighs quantity.
Answers. Classic search engines serve up a list of pages to browse, while AI aims for a single, precise answer (e.g., naming one specific company), so FAQ-style content and semantic alignment matter more than exact keyword matching.
Arrangement. SEO favored "less is more" and avoided repetition; GEO is the opposite: AI models prefer complete, extensive, well-structured content, even if it is partly repetitive.
Attribution. Clicks are no longer a reliable measure of effectiveness, and attribution analytics for GEO and GEM are still in their infancy, forcing companies to build entirely new measurement models.
Agnostic alignment. Because the number of competing AI platforms keeps growing and no clear market leader has yet emerged, companies should not commit to a single tool but should continuously test and monitor the whole ecosystem.
For a sense of scale: Google still accounts for roughly 90% of the traditional search market, while in the AI segment ChatGPT currently holds close to 60% share, followed closely by Microsoft Copilot, Perplexity, Google Gemini, Claude, and Grok — though the situation remains exceptionally fluid.
Three concrete steps for companies
The authors recommend three actions. First, reallocating budgets and resources toward ISM — given that the search industry as a whole generates roughly $350 billion in annual revenue, the scale of this reallocation could be enormous. Second, choosing the right people and partners — teams or agencies that understand the full ISM domain, not just its familiar, classic slice. Third, investing in measurement and attribution: building a GEO analytics layer that tracks brand presence in AI citations and share of answers across a defined set of queries, using first-party tools (e.g., Google Analytics 4) to analyze AI-originated sessions, consistently applying UTM parameters, and ultimately building attribution models that triangulate GEO signals with branded search and direct traffic, supported by quasi-experimental testing.
Importantly, this shift can be just as favorable for small players as it is dangerous for large ones. Brands that have built their position for years through sheer advertising-budget scale may lose their edge to smaller, more agile companies that better understand the new logic of "being cited" rather than "being clicked."
4. Whose voice to listen to when expanding into new markets
The fourth thread concerns a question that seems unrelated to AI technology itself but is central to understanding how the approach to customers is changing overall: whose voice a company should listen to first when entering a new market. Nataliya Langburd Wright of Columbia Business School, drawing on data from more than a thousand technology startups, shows that the choice of first customers — the so-called beachhead market — is a far more strategic decision than is usually assumed.
Managers face a dilemma: learn from well-known users in the home market, who are easier to understand, or target users in the destination market right away, who better represent future customers. The author calls these two competing values legibility and transferability of signals. Familiar users communicate in ways that are easy to interpret — shared language, cultural norms, context — but their preferences may carry no weight in the destination market. Foreign users send signals that are harder to read, but far more representative of future demand.
Two factors that determine the choice
According to the author, choosing the right strategy depends on two variables. The first is the degree of similarity in customer preferences across markets — low-fragmentation categories such as SaaS software or productivity tools tend to have a global, homogeneous base of needs, while highly fragmented categories such as language learning, food, or industrial automation show large cultural differences. The second variable is the homogeneity of the home market — the more linguistically and culturally uniform the market a company knows well, the easier it is to correctly read signals from local customers; in highly diverse markets, such as India with more than a dozen dominant languages, even "well-known" local users can be harder to understand than one might expect.
Combining the two dimensions yields a simple decision rule: high similarity between markets plus a homogeneous home market argues for starting with local, familiar users; low similarity plus a diverse home market argues for going straight to users in the destination market. The author illustrates this with examples: Australia's Canva, operating in the low-fragmentation category of creative tools, successfully built its product around local, linguistically homogeneous users before launching its global expansion. Ukraine's Grammarly, by contrast, targeted English-speaking students and professionals abroad from the very start, despite being founded in Ukraine — because the language-learning industry is highly fragmented culturally, and local users would have provided signals that were legible but poorly transferable.
Wright proposes a four-step decision process: precisely defining the target market, estimating the similarity of preferences between markets, assessing the homogeneity of the home market, and combining both dimensions into a concrete recommendation. It's a reminder that even in the age of synthetic respondents and AI-moderated interviews, marketing's fundamental question remains the same: whose signals to listen to, and how much to trust what the first customers say.
What ties these four stories together
At first glance, LLM-powered market research, CX metric reduction, optimization for generative search, and the choice of a beachhead market are four separate topics. But they share a common denominator: artificial intelligence does not replace strategic thinking about the customer — it raises the stakes and accelerates the pace at which that thinking must be applied.
Language models drastically cut the cost and time of gathering insights, but only where a company has well-organized, meaningful input data — hence the growing importance of practices like trimming redundant CX metrics and mapping them onto the real customer journey. The same models are rewriting the rules of search, shifting attention from "being clicked" to "being cited" by AI — which demands new competencies and new metrics that are only now taking shape. And at the very foundation lies a question AI will not answer on a manager's behalf: which customers are actually worth learning from, so that conclusions — human or synthetic — hold up in the destination market.
The same warning recurs across all four areas: technology delivers scale and speed, but responsibility for quality, context, and interpretation stays with people. Companies that understand this, and treat AI as a powerful but supervised collaborator rather than an automatic substitute for strategic thinking, will gain a real advantage. Those that treat AI as a shortcut around the hard work of understanding the customer risk the same fate as the large fitness chain that was stunned to find itself losing AI search results to a small company from Houston.
Sources. This article was written based on four sources: Neeraj Arora, Ishita Chakraborty, Yohei Nishimura — "AI-Powered Market Research"; Charles H. Patti, Maria M. van Dessel, Steven W. Hartley — "A Smarter Approach to Measuring Customer Experience"; Michael Pettiette, Kimberly A. Whitler — "Will Customers Find Your Brand Today? Strategies for AI-Based Search"; Nataliya Langburd Wright — "The Best Customers to Study When Expanding Into a New Market."
