Week 10 · Marketing Research & Analytics
Davood Wadi, PhD
Multivariate analysis provides the evidence on which segmentation, positioning, and product design decisions are based.
Segmentation divides a market into groups of customers with similar needs or behavior. Positioning determines how a brand is perceived relative to competitors in the minds of target customers.
Both decisions depend on data that involve many variables at once, such as ratings of brands on numerous attributes.
Factor analysis reduces a large set of attributes to a small number of underlying dimensions, on which perceptual maps are based.
Cluster analysis groups customers into segments.
Conjoint analysis measures how customers trade off product attributes against one another and against price.
Consumers are often asked to rate brands on many attributes, many of which are correlated with one another.
Factor analysis identifies the smaller number of dimensions that underlie these ratings.
Perceptual maps display the positions of brands on these dimensions.
Factor analysis is an interdependence technique: it examines the relationships among a set of variables, without distinguishing dependent from independent variables.
It identifies factors, which are underlying dimensions that account for the correlations among the observed variables.
For example, ratings of banks on “friendly staff,” “helpful advice,” and “courteous service” may reflect a single factor of service quality, while ratings on “low fees,” “competitive interest rates,” and “no hidden charges” may reflect a factor of value.
A factor loading is the correlation between a variable and a factor. Each factor is interpreted and named according to the variables that load highly on it.
The number of factors is commonly determined by retaining factors with an eigenvalue greater than one, by examining a scree plot, and by the interpretability of the solution.
Factor analysis is also used to assess the convergent and discriminant validity of multi-item scales, introduced in Week 4.
A smartphone survey asks for ratings on eight attributes. The factor analysis gives eigenvalues of 3.1, 2.2, 0.8, 0.6 and smaller. On the first factor, battery life, durability and repair cost load highly; on the second, camera quality, screen color and design load highly. Which conclusion follows?
A perceptual map is a graphical representation of consumers’ perceptions of competing brands on two or more dimensions.
The dimensions may be derived by factor analysis of attribute ratings, or by multidimensional scaling of judgments of the similarity between brands.
Brands positioned close together are perceived as similar, and are therefore likely to compete directly.
An ideal point represents the combination of attributes most preferred by a segment of consumers. Brands closer to the ideal point are expected to be preferred by that segment.
Perceptual maps represent perceptions, which may differ from the objective characteristics of the brands.
The proximity of two brands indicates the degree of perceived similarity, and therefore the intensity of competition between them.
The distance between a brand and a segment’s ideal point indicates how well the brand’s perceived position matches the preferences of that segment.
An unoccupied region of the map may indicate a positioning opportunity, but only if a sufficient number of consumers have an ideal point in that region.
A brand may be repositioned by changing the product itself, or by changing consumers’ perceptions of it through communication.
A perceptual map describes perceptions at the time of measurement, and is updated as the market changes.
A perceptual map of sports drinks shows brands and the ideal points of three segments. What does each observation indicate?
Proximity indicates perceived similarity, and therefore the intensity of competition between the two brands.
The region is unoccupied, and a sufficient number of consumers have their ideal point there.
A short distance to a segment's ideal point indicates that the brand's perceived position matches that segment's preferences.
An unoccupied region is an opportunity only if enough consumers have their ideal point in it.
The distance to each ideal point shows how well the brand's perceived position matches each segment; here the match is weak.
Brands positioned close together are perceived as similar and compete directly.
Which concept does this case show?
Segments can be defined in advance, on the basis of a single variable such as age or region, or derived from the data.
Data-driven segmentation groups customers according to their similarity on several variables at once, such as needs, attitudes, and purchase behavior.
Cluster analysis is the principal method of data-driven segmentation.
Cluster analysis classifies objects into groups, called clusters, so that objects within a cluster are similar to one another and different from objects in other clusters.
Similarity is measured by a distance measure, most commonly the Euclidean distance between the objects’ values on the clustering variables.
Variables measured on different scales are standardized before clustering, so that variables with larger numerical ranges do not dominate the distance measure.
Hierarchical methods combine objects into progressively larger clusters, and display the result in a tree diagram called a dendrogram.
Nonhierarchical methods, such as k-means clustering, assign objects to a number of clusters specified in advance, and are suited to large datasets.
The researcher first selects the clustering variables, according to the marketing research problem, such as the benefits customers seek from a product.
The researcher then standardizes the variables where necessary.
The researcher then selects a distance measure and a clustering method.
The researcher then determines the number of clusters, according to statistical criteria and the interpretability of the solution.
The researcher then profiles each cluster on variables not used in the clustering, such as demographics and media use, so that the segment can be identified and reached.
Finally, the researcher assesses the stability of the solution, for example by repeating the analysis on separate halves of the sample.
A bank segments its credit card holders. Put the steps of the clustering procedure in order.
Tap a step, then the number where it belongs.
Segments produced by cluster analysis are evaluated before they are used in targeting decisions.
Measurable: the size and characteristics of the segment can be estimated.
Substantial: the segment is large or profitable enough to justify a distinct marketing program.
Accessible: the segment can be reached through identifiable channels and media.
Differentiable: the segment responds differently from other segments to marketing actions.
Actionable: the firm has the resources to design and implement a program for the segment.
Cluster analysis produces clusters even when the data contain no meaningful group structure. A solution that fails these criteria, or is unstable across samples, is not used for targeting.
Each segment found by cluster analysis fails one evaluation criterion. Which one?
The segment is too small to justify a distinct marketing program.
A segment that does not respond differently to marketing actions need not be treated separately.
The segment cannot be reached through identifiable channels and media.
The firm lacks the resources to design and implement a program for the segment.
The size and characteristics of the segment cannot be estimated.
Which concept does this case show?
Discussion: A fitness chain holds data on the visit frequency, class attendance, time of visits, and survey-reported goals of 50,000 members. Which variables should be used to cluster the members, which should be used to profile the clusters, and how should the resulting segments be evaluated before a targeting decision is taken?
Customers rarely find a product that is best on every attribute, and must therefore make trade-offs.
Asking customers directly how important each attribute is tends to produce answers in which every attribute appears important.
Conjoint analysis infers the importance of attributes from the choices customers make among complete products.
Conjoint analysis treats a product as a bundle of attributes, each with a set of levels.
For example, a laptop may be described by brand, with three levels, battery life, with levels of 8, 12, and 16 hours, and price, with levels of 900, 1,100, and 1,300 dollars.
Respondents evaluate a series of product profiles, each combining one level of every attribute.
In choice-based conjoint, the most widely used form, respondents repeatedly choose one profile from a small set, often with the option to choose none.
The analysis estimates a part-worth utility for each level of each attribute, representing its contribution to the overall preference for a product.
Attributes are selected because they are relevant to customers’ choices and actionable for the firm, often on the basis of prior qualitative research.
Levels are realistic, cover the range the firm is considering, and are unambiguous to respondents.
The number of attributes is limited, typically to six or fewer, since respondents simplify their choices when profiles become too complex.
Because the number of possible profiles grows rapidly with the number of attributes and levels, a fractional factorial design presents a subset of profiles from which all part-worths can still be estimated.
Within each attribute, a higher part-worth indicates a more preferred level.
The relative importance of an attribute is the range of its part-worths divided by the sum of the ranges across all attributes.
For example, if the ranges of the part-worths are 1.0 for brand, 2.0 for battery life, and 3.0 for price, the relative importance of price is 3.0 divided by 6.0, or 50 percent. Battery life accounts for 33 percent and brand for 17 percent.
Relative importance depends on the levels included in the study. A wider range of prices increases the apparent importance of price.
Part-worths can be estimated for each respondent, which allows segments to be formed on the basis of their preferences.
Headphone part-worth ranges are 0.6 for brand, 1.8 for noise cancelling and 1.6 for price, with prices tested from $150 to $250. What is the relative importance of noise cancelling, and how would testing prices from $100 to $400 change the importance of price?
Because price is included as an attribute, part-worths can be converted into monetary values.
The willingness to pay for an improvement in an attribute is the gain in utility from the improvement divided by the utility of one dollar, derived from the price part-worths.
For example, if an increase in battery life from 8 to 12 hours adds 1.2 units of utility, and each dollar of price reduces utility by 0.04 units, the willingness to pay for the improvement is 1.2 divided by 0.04, or 30 dollars.
Market simulators use the estimated part-worths to predict the shares of preference for competing products under alternative scenarios, such as a new product launch or a competitor’s price change.
Because conjoint choices are hypothetical, they may overstate willingness to pay. Estimates are therefore compared with market data where possible.
A conjoint design can be administered to an AI agent in the same form as to a consumer: the agent chooses among profiles whose attributes vary according to an experimental design.
The part-worths estimated from the agent’s choices show how much each attribute level contributes to the agent’s selection.
For example, a sponsorship label can be included as an attribute, so that its negative part-worth measures the penalty the agent applies to paid listings.
The part-worths can be compared across conditions of the agent’s instructions. In hotel-booking choice experiments, agents told they served the booking platform penalized sponsored listings less than agents told they served the traveler (Wadi & Ma, 2026c).
The wording of the label can be varied as an attribute level. The label “Sponsored” reduced choice of the paid listing more than the label “Promoted” (Wadi & Ma, 2026c).
Because AI agents show position effects, the order of the profiles in each choice task is randomized (Wadi & Ma, 2026a).
A retailer gives a shopping agent choice tasks of four profiles varying in price, rating, delivery time and an "Ad" label. The agent is told in one condition that it serves the shopper, in another that it serves the retailer. Which approach is sound?
Traditional segmentation relies on structured variables, such as survey ratings and purchase records.
Much of the information about customers’ needs is contained in unstructured text, such as open-ended survey responses, reviews, and service conversations.
Language models allow such text to be represented numerically, so that it can be used as a basis for segmentation.
An embedding is a numerical representation of a piece of text, produced by a language model, in which texts with similar meanings receive similar values.
For example, the responses “I need something quick before work” and “breakfast has to take five minutes” receive embeddings that are close to each other, although they share almost no words.
Embeddings can represent open-ended responses, reviews, product descriptions, and sequences of customer behavior.
The individual dimensions of an embedding have no direct interpretation, unlike the factors of a factor analysis.
Customers are clustered on the embeddings of the texts they have written, using the clustering methods introduced in this week, such as k-means.
The resulting segments are based on the needs and concerns customers express in their own words, rather than on attributes selected in advance by the researcher.
The approach can be applied to very large volumes of text, and can be combined with behavioral data in the same analysis.
Because the dimensions are not interpretable, each cluster is interpreted by reading a sample of the texts assigned to it.
The results depend on the language model used and on the preparation of the texts, and are subject to the same stability tests and segment evaluation criteria as any other cluster solution.
A persona is a narrative description of a typical member of a segment, with a name, a context, goals, and frustrations, used to communicate the segment to managers and designers.
Language models can be used to generate personas from the profile of each cluster, and some tools allow managers to converse with a persona as a simulated customer.
A generated persona may contain details that are not supported by the segment data, such as an occupation, a family situation, or a quotation, and these details may reflect stereotypes.
A persona that answers questions in conversation is a synthetic respondent, and is subject to the limitations introduced in Week 5.
Every characteristic of a persona is therefore traceable to the data describing the segment, and the persona is used to communicate evidence, not as a source of evidence.
A language model wrote a persona from the profile of a commuter segment: 68% are aged 25–34, 81% travel by train, 59% buy coffee on the way, and the main theme of their answers is reliability. Sort each line of the persona.
Tap a card, then tap the group it belongs to.
Discussion: A product team uses an AI tool to create a conversational persona for each of five customer segments, and begins to test new feature ideas by asking the personas for their reactions. What does this practice offer the team, which risks does it introduce, and which evidence should replace the personas’ answers before a feature is built?
Factor analysis reduces many correlated attributes to a few underlying dimensions, and perceptual maps display brands and ideal points on those dimensions.
Cluster analysis groups customers by their similarity on several variables, and the resulting segments are profiled, tested for stability, and evaluated before use.
Conjoint analysis estimates part-worth utilities from choices among product profiles, from which attribute importance and willingness to pay are derived.
Relative importance and willingness to pay depend on the attributes and levels included in the study, and hypothetical choices may overstate willingness to pay.
Embeddings allow customers to be segmented on the needs they express in unstructured text, with each cluster interpreted by reading its texts.
AI-generated personas communicate segment data, and every characteristic they contain must be supported by that data.
Conjoint designs administered to AI agents estimate the part-worths agents assign to attributes such as price and sponsorship, and these part-worths can be compared across the instructions the agents receive.