Segmentation, Targeting and Personalization

Introduction

Segmentation, targeting and positioning (STP) is a central framework in marketing management. Segmentation divides a heterogeneous market into groups with relatively similar needs or behaviours. Targeting evaluates those groups and selects the audiences the organisation can serve effectively. Positioning defines the value proposition and associations the firm seeks to establish in the minds of selected customers. Personalization extends this logic by adapting content, offers, products, prices, service and communication to an individual or context.

Artificial intelligence (AI) has expanded the scale and speed of STP. Clustering can identify naturally occurring customer groups; propensity models can estimate the likelihood of a future action; look-alike models can find prospects resembling valuable existing customers; and recommender systems can select a next-best product, message or action. These tools can improve relevance and resource allocation, but they can also reproduce social inequality, expose sensitive information, manipulate vulnerable people or deny opportunities unfairly.

The managerial challenge is therefore not simply to maximise prediction accuracy. It is to make useful, explainable and proportionate decisions while preserving customer autonomy, privacy, equal treatment and organisational accountability.

Learning Objectives

By the end of this chapter, students should be able to:

1. Explain segmentation, targeting, positioning and personalization as related but distinct activities.
2. Select suitable variables and methods for AI-supported market segmentation.
3. Interpret clustering, propensity scores, look-alike audiences and recommender systems.
4. Evaluate targeting and next-best-action models using commercial and ethical criteria.
5. Design safeguards against harmful discrimination, privacy violations and manipulative personalization.
6. Translate analytical results into a coherent marketing decision and value proposition.

1. From STP to AI-Enabled Decision Making

Traditional segmentation may use geographic, demographic, psychographic and behavioural variables. Contemporary segmentation commonly combines first-party data—such as transactions, loyalty activity and consented interactions—with contextual information, such as device, time, location or browsing environment. Data should be collected for legitimate purposes, documented and retained only as long as necessary.

The STP process can be represented as a decision chain:

1. Define the market and decision problem. A firm might ask which customers are most likely to renew, which prospects need education, or which customers should receive a retention intervention.
2. Construct meaningful variables. Examples include purchase frequency, monetary value, product category breadth, engagement recency and service history.
3. Discover or predict. Unsupervised methods discover groups; supervised methods predict outcomes; recommendation methods rank alternatives.
4. Assess strategic value. A statistically distinct group is not automatically attractive. Size, growth, accessibility, profitability, competitive intensity and fit with brand purpose matter.
5. Design positioning and action. The organisation translates evidence into a differentiated value proposition and an appropriate experience.
6. Measure outcomes and revise. Models and segments must be monitored because customer behaviour, markets and data quality change.

Segmentation describes similarity; targeting is a resource-allocation decision; positioning is a market meaning decision; personalization is an execution mechanism. Confusing these concepts may lead an organisation to personalise irrelevant or harmful offers very efficiently.

2. Segmentation Variables and Data Quality

A useful segmentation variable should be relevant to the decision, sufficiently reliable, available at the required time and lawful to use. Common variable families include:

Data preparation is often more important than algorithm selection. Analysts should check missingness, duplicate records, inconsistent identifiers, outliers, label errors and sampling bias. A variable can be a proxy for a protected characteristic even when the protected characteristic is not explicitly recorded. For example, postcode, language, device type or browsing pattern may correlate with ethnicity, disability or income.

A strong data dictionary records each variable’s meaning, source, collection date, lawful basis, permitted use, retention period and known limitations. Feature selection should be driven by the marketing decision rather than by the availability of data.

3. Clustering for Market Segmentation

Clustering is an unsupervised learning approach that groups observations so that members of the same group are more similar according to selected variables than members of different groups. It is useful when the organisation does not have a predefined outcome label.

3.1 K-means clustering

In k-means clustering, the analyst chooses the number of clusters, k, and assigns each observation to the nearest cluster centre. The usual objective is to minimise within-cluster squared Euclidean distance:

Here, J is the objective function; Cⱼ is cluster j; xᵢ is the vector of variables for customer i; and μⱼ is the mean vector of cluster j. The method alternates between assigning observations and recalculating means until assignments stabilise or improvement becomes negligible.

K-means assumes that numeric variables are meaningfully represented by distances and that cluster centres based on means are appropriate. Variables should normally be standardised when measured on different scales. A common standardisation is:

where x is an observation, μ is the sample mean and σ is the sample standard deviation. The resulting z-score is dimensionless. Standardisation is appropriate when scale differences would otherwise dominate distance calculations, but it can reduce the influence of substantively important units.

K-means is sensitive to outliers, initial centres and the selected value of k. Alternatives include hierarchical clustering, density-based methods and model-based clustering. Categorical data may require suitable distance measures or algorithms designed for mixed data.

3.2 Choosing and interpreting clusters

The number of clusters should not be selected solely by an automated metric. The analyst can combine statistical diagnostics, managerial interpretability, stability across samples and actionability. The silhouette coefficient, for example, compares within-cluster cohesion with separation from the nearest alternative cluster. Higher values generally indicate better separation, but no metric proves that a segment is strategically meaningful.

A cluster profile should report size, distinguishing characteristics, needs, expected value, reachability, preferred channels, possible harms and a proposed proposition. Labels such as “budget-conscious planners” are more useful than labels that imply fixed personality types. Segments are probabilistic summaries, not natural or permanent identities.

Worked example: interpreting a simple clustering result

Suppose a retailer standardises three variables—purchase frequency, average order value and digital engagement—and obtains three clusters. Cluster A contains 20% of customers, purchases frequently, has high average order value and responds to product education. Cluster B contains 50%, purchases occasionally and is price-sensitive. Cluster C contains 30%, has low engagement and high service contacts.

A simplistic response would target Cluster A with expensive promotions because it appears valuable. A more careful interpretation asks whether Cluster A is already loyal and needs retention rather than discounting, whether Cluster B can be grown profitably, and whether Cluster C requires service recovery before promotional communication. The cluster output informs strategy but does not determine it.

4. Propensity Models and Value-Based Targeting

A propensity model estimates the probability that a customer will perform a specified action within a defined period, such as purchase, churn, respond to an email or adopt a subscription. The target variable must be clearly defined: “purchase within 30 days” is more useful than “likely to be interested.”

A logistic model for a binary outcome can be written as:

Yᵢ is the outcome for customer i, pᵢ is the predicted probability, xᵢⱼ is predictor j for customer i, β₀ is the intercept, βⱼ are estimated coefficients, and e is the base of the natural logarithm. The model produces a value between 0 and 1. It assumes that the log-odds are linearly related to the predictors unless nonlinear terms or other models are introduced.

A high propensity is not necessarily a high incremental opportunity. Customers predicted to purchase may do so without marketing contact. Therefore, targeting should ideally use uplift or incremental-response modelling, which estimates the difference between outcomes under treatment and no treatment:

T=1 denotes exposure to a marketing treatment and T=0 denotes a suitable control condition. Uplift requires experimental or carefully designed causal data. Correlation-based propensity scores should not be interpreted as causal effects.

A simple expected incremental value rule is:

EVᵢ is expected incremental value for customer i; Uᵢ is estimated incremental probability; Mᵢ is contribution margin if the action succeeds; and Cᵢ is the cost of contact and fulfilment. The units of Uᵢ are probability, Mᵢ and Cᵢ are monetary units, so EVᵢ is monetary. This rule assumes the estimates are calibrated and that relevant costs and benefits have been included. It should not be used to justify actions that violate law, consent or fairness requirements.

5. Look-Alike Audiences

A look-alike audience begins with a seed group, such as customers with high retention, and identifies non-customers who resemble that group across selected attributes. The method may use classification, nearest-neighbour similarity or representation learning. It is common in digital advertising because it expands reach beyond known customers.

The central question is not merely “Who resembles our best customers?” but “Which similarities are legitimate, useful and non-discriminatory?” A seed group may reflect historical exclusion, unequal purchasing power or biased acquisition practices. If the algorithm learns those patterns, the expanded audience may systematically omit disadvantaged groups.

Good practice includes defining the seed outcome precisely, excluding sensitive or unjustified proxy variables, testing reach across relevant groups, using frequency and exposure controls, and validating the quality of leads rather than assuming resemblance equals value. Look-alike modelling should not be used to infer sensitive traits or to deny access to essential products, employment, housing, education or financial services.

6. Recommender Systems and Next-Best Action

A recommender system ranks items, content or actions for a user or situation. Common approaches include:

• Content-based recommendation: recommends items whose attributes resemble items previously engaged with.
• Collaborative filtering: uses patterns of interaction among users and items.
• Hybrid systems: combine content, collaborative and contextual information.
• Context-aware systems: incorporate time, device, location or stage in the customer journey.

A recommendation score can be understood conceptually as an estimate of expected relevance, value or utility. The ranking objective must be explicit. Optimising clicks alone may increase sensational content, repetitive exposure or low-quality purchases. A broader objective may include relevance, margin, customer satisfaction, diversity, novelty, complaint risk and long-term retention.

The next-best action is not necessarily the next-best advertisement. It may be a service explanation, product comparison, reminder, human escalation, apology or decision to do nothing. Action policies should include eligibility rules, consent checks, contact-pressure limits, exclusion rules and a fallback when confidence is low.

Cold-start problems occur when a new customer or item has little interaction history. Solutions include contextual information, explicit preferences, carefully designed onboarding and exploration. However, exploration should be proportionate and should not expose people to unreasonable risk.

7. Evaluation: Predictive, Commercial and Customer Metrics

Model evaluation should distinguish technical performance from business and social outcomes.

For a binary classifier, precision is the proportion of predicted positives that are actually positive, while recall is the proportion of actual positives that are identified. Their usefulness depends on decision costs. A model with excellent ranking performance may still produce poor campaigns if the treatment is ineffective or if the threshold is chosen without considering capacity and harm.

Randomised holdout tests are often preferable for measuring incremental impact. A simple estimated treatment effect is:

where Ŷ_T is the average outcome in the treatment group and Ŷ_C is the average outcome in the control group. The estimate is interpretable as causal only when treatment assignment and measurement satisfy appropriate experimental assumptions, including no systematic treatment-selection bias and consistent outcome measurement.

8. Ethical, Legal and Governance Considerations

AI-supported targeting creates several ethical risks:

1. Discrimination and unequal opportunity. A model may give different access to offers, credit, insurance, employment or information based on protected characteristics or proxies.
2. Privacy intrusion. Combining datasets can create detailed profiles that customers did not reasonably expect.
3. Manipulation. Personalization can exploit anxiety, addiction, financial distress or limited decision-making capacity.
4. Opacity. Customers and staff may not understand why a message or offer was shown.
5. Autonomy and consent. Persistent tracking can make refusal difficult or meaningless.
6. Feedback loops. Recommendations can reinforce previous behaviour, reducing exposure to alternatives and making historical patterns appear natural.
7. Security and misuse. Profiles can be leaked, repurposed or accessed by unauthorised parties.

Governance should be integrated into the marketing process rather than added after deployment. A practical control framework includes:

• Define the purpose and prohibited uses before collecting data.
• Apply data minimisation and obtain valid, understandable consent where required.
• Separate sensitive attributes from operational features while retaining protected-group information, where lawful and appropriate, for fairness auditing.
• Test group-level selection, error, calibration and outcome differences.
• Review proxy variables and document reasons for retaining them.
• Use human review for high-impact decisions and provide escalation routes.
• Give customers meaningful controls, explanations and opt-out mechanisms.
• Monitor drift, complaints, disparate impact and unintended outcomes.
• Maintain model cards, data lineage, version histories and incident procedures.

Legal requirements vary by jurisdiction. Depending on the application, relevant principles may include purpose limitation, transparency, data protection, consumer protection, anti-discrimination, profiling restrictions and requirements for human oversight. Compliance is a baseline, not a complete ethical analysis.

9. Illustrative Case: Personalization for a Subscription Retailer

Background

Imagine a fictional online retailer, Northstar Home, selling household products through a subscription service. It has three million registered customers and wishes to reduce cancellations while avoiding excessive discounting.

Problem

The marketing team proposes a churn-propensity model and a personalised retention offer. Early analysis shows that customers in areas with unreliable delivery infrastructure are more likely to cancel. A purely predictive campaign would send increasingly large discounts to customers predicted to churn.

Analysis

The team separates three questions: who is likely to cancel, what causes cancellation, and which intervention creates incremental retention. It combines service records, delivery reliability, subscription tenure, product usage and customer feedback. It excludes inferred health information and does not use postcode as a direct targeting rule. A randomised test compares three permitted actions: service-recovery information, a flexible delivery option and a small incentive.

The team also tests whether exposure and outcomes differ by region, language preference and accessibility needs. It discovers that some customers with low digital engagement were underrepresented in the training data and were receiving fewer useful interventions.

Approach and solution

Northstar deploys a two-stage policy. First, customers with delivery-related problems receive transparent service recovery and an opportunity to change delivery preferences. Second, customers with high estimated incremental response may receive a modest incentive. Customers with low confidence are routed to a general support message rather than a highly personalised claim. Contact frequency is capped, and customers can opt out.

Outcome

The service-recovery intervention reduces cancellations more effectively than blanket discounting and produces lower cost per retained customer. The firm also improves delivery communication. The result is not simply a better model; it is a better decision design that addresses a structural cause rather than treating cancellation as an individual propensity.

Learning points

• Prediction is not explanation or causation.
• Incremental testing is preferable to targeting solely on predicted probability.
• Fairness analysis may reveal data coverage problems, not only biased algorithms.
• The best next action may be service improvement rather than promotion.
• Personalization should preserve choice and avoid exploiting vulnerability.

10. Implementation Blueprint for Marketing Managers

A responsible implementation can follow these stages:

1. Decision charter: specify the business objective, customer benefit, time horizon, affected groups and prohibited uses.
2. Data and consent review: document sources, permissions, quality, retention and sensitive proxies.
3. Baseline strategy: define what happens without AI so that incremental value can be measured.
4. Model selection: use the simplest method that meets the decision need and can be explained to stakeholders.
5. Validation: assess predictive accuracy, calibration, stability, subgroup performance and business value.
6. Experimentation: use holdouts or controlled tests where feasible.
7. Policy design: establish eligibility, frequency caps, escalation, opt-out and “do nothing” rules.
8. Deployment: begin with limited exposure and human oversight.
9. Monitoring: review drift, complaints, opt-outs, fairness indicators, profitability and customer outcomes.
10. Retirement or revision: remove models that no longer serve the stated purpose or create unacceptable harm.

Employability requires more than operating software. Marketing professionals must be able to brief data scientists, challenge weak assumptions, interpret uncertainty, explain decisions to managers, collaborate with legal and compliance teams, and communicate with customers in plain language.

Contemporary Developments and Critical Perspective

Generative AI is increasingly used to create segment-specific copy, images and conversational experiences. This can improve speed but increases the risk of inconsistent claims, stereotyping and large-scale experimentation without adequate oversight. Privacy-enhancing technologies, federated learning and on-device processing may reduce the need to centralise personal data, although they do not automatically guarantee fairness or lawful use.

Representation learning can identify complex behavioural patterns, but more complexity may reduce interpretability. Causal machine learning and uplift modelling are becoming more relevant as firms move from “who is likely to act?” to “who will act because of this intervention?” There is also growing interest in fairness-aware ranking, diversity constraints and responsible exploration in recommender systems.

A critical perspective asks who benefits from personalization, who bears its risks and which customer needs cannot be solved by better prediction. Segmentation is a managerial construction, not a discovery of fixed human categories. Responsible marketing therefore combines quantitative evidence with qualitative research, customer participation, social impact assessment and institutional accountability.

Chapter Summary

Segmentation divides markets into meaningful groups; targeting selects groups and allocates resources; positioning establishes a differentiated meaning; and personalization adapts execution to individuals or contexts. Clustering is useful for discovering patterns, while propensity models estimate the likelihood of defined outcomes. Look-alike audiences expand reach from a seed group, and recommender systems rank products, content or next-best actions.

These techniques require clear objectives, suitable variables, careful data preparation and evaluation beyond predictive accuracy. Propensity is not causation, and high likelihood is not the same as high incremental value. Holdout experiments and uplift approaches help identify whether an intervention creates additional outcomes. Responsible implementation requires privacy protection, consent, transparency, human oversight, fairness testing, contact limits and mechanisms for correction and appeal.

The central professional principle is that AI should support better customer and organisational decisions, not merely more efficient persuasion. A successful system is accurate enough, commercially valuable, explainable, proportionate and fair.

Key Terms

Artificial intelligence: Computational systems that perform tasks associated with human intelligence, such as prediction, classification or recommendation.

Clustering: Unsupervised grouping of observations according to similarity.

Data minimisation: Collecting and using no more personal data than is necessary for a defined purpose.

Incremental impact: The additional outcome caused by a marketing intervention compared with an appropriate alternative or control.

Look-alike audience: Prospects selected because they resemble a defined seed group.

Next-best action: The most appropriate permitted action for a customer or context at a particular time.

Personalization: Adaptation of marketing content, offers, products or experiences to an individual or context.

Positioning: The intended meaning and value proposition associated with a brand in a target market.

Propensity score: An estimated probability that an individual will perform a specified action within a defined period.

Recommender system: A system that ranks products, content or actions for users.

Segmentation: Division of a market into groups with relatively similar characteristics or needs.

Targeting: Evaluation and selection of market segments for organisational attention.

Uplift: The estimated difference between an outcome under treatment and an outcome under no treatment.

Calibration: Agreement between predicted probabilities and observed frequencies.

Cold-start problem: Difficulty making reliable recommendations when a user or item has little interaction history.

Proxy variable: A variable that indirectly represents a sensitive or protected characteristic.

Fairness audit: Systematic examination of whether a model or policy produces unjustified differences among groups.

Feedback loop: A process in which model outputs influence future data, reinforcing existing patterns.

Review Questions

Short Answer Questions

1. Distinguish segmentation, targeting, positioning and personalization.
2. Why should variables often be standardised before k-means clustering?
3. What does a propensity model estimate?
4. Why is a high propensity score not necessarily evidence of incremental impact?
5. Identify two ethical risks associated with look-alike audiences.

Long Answer Questions

1. Explain the k-means method, including its objective, assumptions and limitations.
2. Compare content-based, collaborative and hybrid recommender systems.
3. Discuss how a marketing manager should evaluate an AI targeting model beyond predictive accuracy.
4. Explain how privacy, consent and fairness should influence personalization design.
5. Describe a responsible implementation process for an AI-supported STP programme.

Analytical / Application Questions

1. A retailer has a cluster of frequent, high-value customers and a cluster of infrequent, price-sensitive customers. Design different positioning and next-best-action strategies for each, and identify possible risks.
2. A churn model performs well overall but poorly for customers with limited digital interaction. Diagnose possible causes and propose corrective actions.
3. Design a controlled test to determine whether a personalised retention offer creates incremental value. Specify the treatment, control, outcome measure and ethical safeguards.

Multiple Choice Questions with Answers

Question 1

Which activity primarily involves selecting market groups for organisational attention?

A. Positioning

B. Targeting

C. Clustering

D. Personalization

Correct Answer: B. Targeting

Question 2

What is the main purpose of standardising variables before distance-based clustering?

A. To remove all missing values

B. To make every customer identical

C. To prevent variables with larger scales from dominating distance

D. To guarantee the correct number of clusters

Correct Answer: C. To prevent variables with larger scales from dominating distance

Question 3

A propensity model estimates:

A. The causal effect of every possible campaign

B. The probability of a specified future action

C. The legal acceptability of a campaign

D. The exact lifetime profit of a customer

Correct Answer: B. The probability of a specified future action

Question 4

Which expression represents uplift for a customer profile x?

A. P(Y=1) + P(T=1)

B. P(Y=1 | T=1,x) − P(Y=1 | T=0,x)

C. P(T=1 | Y=1,x) × P(Y=0 | x)

D. P(Y=1 | T=0,x) − P(Y=1 | T=1,x) + 1

Correct Answer: B. P(Y=1 | T=1,x) − P(Y=1 | T=0,x)

Question 5

A look-alike audience is normally created by finding prospects who:

A. Have opted out of all communication

B. Are randomly selected from the population

C. Resemble a defined seed group on selected attributes

D. Have identical purchase histories

Correct Answer: C. Resemble a defined seed group on selected attributes

Question 6

Which recommender approach uses interaction patterns among users and items?

A. Collaborative filtering

B. Geographic segmentation

C. Rule-based positioning

D. Service recovery

Correct Answer: A. Collaborative filtering

Question 7

Why can optimising recommender systems only for clicks be problematic?

A. Clicks are never measurable

B. It may promote low-quality, repetitive or sensational content

C. It prevents any ranking of items

D. It automatically eliminates personalization

Correct Answer: B. It may promote low-quality, repetitive or sensational content

Question 8

In the expression EV_i = U_i × M_i − C_i, what does C_i represent?

A. The customer’s cluster number

B. The model’s confidence interval

C. The cost of contact and fulfilment

D. The customer’s consent status

Correct Answer: C. The cost of contact and fulfilment

Question 9

Which method is generally strongest for estimating whether a campaign caused an additional outcome?

A. A randomised treatment-control test

B. A descriptive demographic profile

C. A cluster label alone

D. A look-alike similarity score alone

Correct Answer: A. A randomised treatment-control test

Question 10

Which statement best reflects responsible AI-supported targeting?

A. Protected characteristics should always be hidden and never audited

B. The most accurate model should always be deployed

C. High predicted likelihood automatically justifies contact

D. Models should be evaluated for value, fairness, privacy and customer impact

Correct Answer: D. Models should be evaluated for value, fairness, privacy and customer impact