AI and Algorithmic Performance Marketing

Introduction

Performance marketing has traditionally relied on rules, historical reporting, segmentation and manual optimisation. Contemporary platforms increasingly use machine learning (ML) and artificial intelligence (AI) to decide whom to target, which message to show, when to bid, how much to bid and how to allocate budget across channels. This transformation is not simply a matter of replacing human decisions with automated ones. It changes the object of management: marketers must now design data-generating systems, define optimisation objectives, evaluate model quality and govern automated decisions.

AI in performance marketing includes predictive bidding, recommendation engines, propensity and churn models, predictive audiences, generative creative systems and conversational interfaces. These applications can improve relevance and efficiency, but they may also amplify biased data, optimise misleading proxies, conceal causal uncertainty and create brand, legal or financial risks.

The central managerial question is therefore not “Can AI automate this task?” but “Under what conditions should this task be automated, with what evidence, safeguards and human accountability?”

Learning Objectives

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

1. Explain the role of machine learning in bidding, recommendations, propensity modelling and churn prediction.
2. Distinguish prediction, optimisation, experimentation and causal inference.
3. Evaluate predictive audiences and algorithmic targeting critically.
4. Assess generative creative and conversational AI in performance marketing.
5. Interpret model-quality measures, calibration, lift and business value.
6. Design appropriate human oversight, monitoring and escalation controls.
7. Identify privacy, fairness, explainability, security and automation risks.
8. Develop an AI-enabled performance-marketing operating model.

1. From Rule-Based Optimisation to Learning Systems

A rule-based campaign might bid more for users in a selected location or reduce bids after a specified time. An ML system estimates outcomes from many variables and updates decisions as new data arrive. Typical inputs include device, context, past interactions, conversion history, product attributes, price, creative features and channel conditions.

A useful distinction is:

• Prediction: estimating an unknown outcome, such as conversion probability.
• Optimisation: selecting an action, such as a bid or recommendation, to maximise an objective.
• Experimentation: deliberately varying treatment to estimate incremental effects.
• Causal inference: estimating what would have happened under an alternative action.

These are related but not interchangeable. A user who converts after seeing an advertisement is not necessarily converted because of the advertisement. The user may have converted anyway. A highly predictive model can therefore be a poor targeting model if it identifies existing intent rather than advertising incrementality.

In a simplified decision setting, an advertiser chooses action a from a set of possible actions A. The expected value of an action can be represented as:

Here, x represents observed information, y is a possible outcome, P(y|x,a) is the estimated probability of that outcome given the context and action, V(y,a) is its value and C(a) is the cost. The formula assumes that the relevant outcomes and costs can be specified and that the estimated probabilities are sufficiently reliable. In practice, the value function may include revenue, margin, customer lifetime value, strategic value and compliance constraints.

2. Machine-Learning Bidding

2.1 The bidding problem

In real-time advertising auctions, a platform may have milliseconds to estimate the probability that an impression will produce a desired event. The system then combines predicted value, bid constraints, auction conditions and delivery objectives.

A simplified value-based bid is:

where p(c|x) is the predicted probability of conversion, V(c) is the monetary value of the conversion, and k is a scaling factor reflecting budget, target return, competition and platform-specific auction mechanics. This is a conceptual relationship, not a universal platform formula. Actual systems may use second-price or generalized second-price mechanisms, value rules, pacing, quality scores, constraints and private proprietary models.

If the predicted conversion probability is 0.04, the conversion value is £80 and the scaling factor is 0.5, the conceptual bid is:

0.04 × £80 × 0.5 = £1.60.

The result is expressed in currency per impression under the stated assumptions. It is not necessarily the final amount paid.

2.2 Learning objectives and feedback loops

Bidding models may optimise conversions, revenue, value, return on advertising spend (ROAS), cost per acquisition or a proxy such as clicks. Each objective produces different behaviour. Optimising clicks can favour curiosity-inducing creative, while optimising revenue may favour high-value customers and neglect acquisition of new customers. Optimising short-term conversion can underinvest in brand formation or retention.

Automated bidding also creates feedback loops. Impressions are more frequently delivered to users predicted to respond. The resulting data overrepresent those users, making the model increasingly confident about the same audience. This is sometimes called exposure bias or selection bias. Exploration—showing some alternatives to less-observed users—can be necessary for learning, although it introduces cost and risk.

2.3 Evaluation

A bidding system should be evaluated at several levels:

• Prediction: calibration, discrimination and stability.
• Auction: delivery, win rate, cost and pacing.
• Business: incremental conversions, margin, customer quality and lifetime value.
• Causal: controlled lift compared with a suitable holdout or experiment.

A model may have strong area under the receiver operating characteristic curve (AUC) but poor calibration. AUC measures ranking performance; it does not guarantee that a predicted probability of 0.20 corresponds to an actual event rate of approximately 20%.

For a binary outcome, a common calibration measure is the Brier score:

Here, n is the number of observations, pᵢ is the predicted probability for observation i and yᵢ is the observed outcome, coded 0 or 1. Lower values are better. The measure is sensitive to both calibration and, to some extent, discrimination. It should not be interpreted without considering class prevalence and the business decision.

3. Recommendation Engines and Propensity Models

3.1 Recommendation systems

Recommendation systems select products, content, offers or messages for an individual or context. Two broad approaches are:

• Content-based recommendation: matches item characteristics with a user’s previous interests.
• Collaborative filtering: uses patterns of interaction among users and items.

Modern systems often combine these methods with deep learning, contextual features, business constraints and real-time signals. They may optimise click-through rate, purchase probability, watch time, margin or long-term engagement.

The main risk is objective narrowing. A recommendation that maximises immediate clicks may promote sensational or low-margin content. A system should therefore incorporate constraints such as inventory, profitability, diversity, frequency, suitability and customer experience.

3.2 Propensity modelling

A propensity model estimates the likelihood of an event, such as purchase, subscription, response or app installation. A logistic model for a binary outcome is commonly expressed as:

Y is the binary outcome, x₁,…,xₖ are predictors, β₀ is the intercept, β₁,…,βₖ are estimated coefficients and e is the base of the natural logarithm. The model assumes a specified functional form and requires suitable data, meaningful predictors and an outcome definition. Coefficients describe changes in log-odds, not direct percentage-point changes in probability.

Propensity is not the same as persuasion. High-propensity users may already be likely to buy. For advertising allocation, a treatment-effect or uplift model may be more relevant. Conceptually, individual treatment effect is:

where T=1 denotes exposure to treatment and T=0 denotes a control condition. In observational data, this difference requires strong assumptions or statistical adjustment; randomised experiments provide more credible causal estimates.

3.3 Churn models

A churn model estimates the likelihood that a customer will discontinue, lapse or become inactive during a defined period. It requires a precise churn definition, such as “no paid transaction for 90 days,” and a prediction window that does not leak future information.

Useful variables may include recency, frequency, monetary value, service usage, complaints, failed payments and engagement trends. A retention intervention should be evaluated for incremental retention and profitability, not merely model accuracy. Offering a discount to a customer who would have stayed wastes margin, while failing to intervene with a high-risk, high-value customer destroys value.

4. Predictive Audiences and Customer Value

A predictive audience is a group generated by a model rather than solely by explicit demographic or behavioural rules. Examples include likely purchasers, high-value prospects, likely responders, likely churners and users similar to existing customers.

Audience construction involves four decisions:

1. Outcome definition: what counts as success?
2. Observation and prediction windows: which information is available at decision time?
3. Eligibility and exclusion: who must not be targeted or must be suppressed?
4. Refresh and governance: how often is the audience recalculated and audited?

Customer lifetime value (CLV) can be represented in a simplified discounted form:

Here, mₜ is contribution margin in period t, cₜ is service or retention cost, r is the discount rate per period, T is the modelling horizon and CAC is customer acquisition cost. This formula assumes that future margin and cost estimates are available and that discounting is appropriate. It is not a complete financial valuation where uncertainty, retention probability and varying cash flows are ignored.

Predictive audiences can improve efficiency, but they can also reproduce historic exclusion. If past campaigns underinvested in a group, the model may infer that the group is low value. Sensitive attributes may be absent yet indirectly encoded through location, language, device, browsing patterns or purchasing history. Governance must therefore examine outcomes across relevant groups, not only input variables.

5. Generative AI for Performance Creative

Generative AI can produce text, images, video, audio, variants, product descriptions and creative concepts. In performance marketing it is particularly useful for rapid variation, localisation, structured experimentation and adaptation to audience or placement requirements.

A responsible creative workflow includes:

1. Define brand, legal, accessibility and claims requirements.
2. Supply approved product facts and reference material.
3. Generate multiple candidate variants.
4. Review for factual accuracy, copyright, stereotypes, unsafe content and disclosure requirements.
5. Test variants using controlled methods where possible.
6. Monitor post-launch performance and user feedback.

Generative systems may hallucinate facts, reproduce copyrighted styles, create synthetic testimonials or produce culturally inappropriate imagery. High click-through rate does not validate a misleading claim. Human review remains essential for regulated categories, health or financial claims, comparative advertising, political communication and sensitive audiences.

Creative optimisation should also avoid confusing correlation with causation. If a creative variant performs better, the difference may result from audience composition, placement, seasonality or auction conditions. Randomised creative experiments or carefully controlled tests provide stronger evidence than simple before-and-after comparisons.

6. Conversational AI in the Customer Journey

Conversational AI can qualify leads, answer product questions, recommend products, recover abandoned carts and support service operations. Its performance should be measured beyond conversation volume. Relevant outcomes include qualified leads, resolution rate, conversion, customer satisfaction, escalation rate, complaint rate and cost per resolved interaction.

A conversational system should clearly communicate its nature when required, avoid pretending to be a human, preserve an audit trail and provide an accessible route to a human agent. It should not invent policy, pricing or eligibility information. Retrieval from approved knowledge sources, constrained response patterns and confidence-based escalation can reduce risk.

The most important design principle is bounded autonomy: the system may answer low-risk questions automatically but must seek confirmation or transfer to a human before irreversible, high-value or sensitive actions.

7. Model Quality, Explainability and Measurement

7.1 Data quality and leakage

Model quality begins with data provenance. Marketers should document collection purpose, consent status, event definitions, timestamp logic, missing values, identity resolution and retention periods. Target leakage occurs when a predictor contains information that would not have been available at the moment of decision. Leakage can produce impressive offline results and poor live performance.

Training, validation and test data must respect time. Random splitting is often inappropriate for campaigns with seasonality or changing auction conditions. Time-based validation better reflects deployment, although it cannot eliminate all distribution shifts.

7.2 Metrics

Different objectives require different metrics:

Accuracy alone is inadequate when errors have unequal costs. A false positive may waste media spend, whereas a false negative may lose a high-value customer. Thresholds should therefore reflect decision costs and operational capacity.

7.3 Explainability

Explainability has several meanings. A global explanation describes general model behaviour; a local explanation describes a particular prediction; a process explanation documents data, governance and decision rights. Feature importance is not necessarily causation, and post-hoc explanations can be unstable. Simpler models may be preferable where decisions are consequential and performance differences are small.

8. Human Oversight and Automation Risk

Automation risk arises when a system acts at scale faster than people can detect errors. Common risks include:

• Objective risk: optimising a proxy rather than the true business goal.
• Data drift: the relationship between inputs and outcomes changes.
• Concept drift: customer behaviour or market conditions change.
• Feedback loops: model decisions alter the data used for future learning.
• Automation bias: staff accept machine recommendations without scrutiny.
• Concentration risk: one platform or model controls too much expenditure.
• Adversarial risk: fraudsters manipulate signals or generate fake activity.
• Reputational risk: inappropriate targeting or creative spreads rapidly.

A practical control framework contains four layers:

1. Preventive controls: permissions, exclusions, constraints, approved data and spending limits.
2. Detective controls: anomaly detection, drift monitoring, audits and outcome comparisons.
3. Corrective controls: pause mechanisms, rollback, retraining and budget reallocation.
4. Accountability controls: named owners, documentation, review boards and incident records.

Human involvement should be proportionate to impact. Low-risk copy variation may be automatically generated and sampled. A targeting decision involving children, health information or financial vulnerability requires stronger review and possibly prohibition.

9. Ethical, Legal and Governance Considerations

AI performance marketing engages privacy, consumer protection, equality, intellectual property, platform policy and sector-specific regulation. Compliance cannot be reduced to obtaining consent. Organisations must consider purpose limitation, data minimisation, security, retention, transparency and the rights of individuals under applicable law.

Important governance questions include:

• What data were used, and were they collected for this purpose?
• Can the organisation explain the decision in understandable terms?
• Could the system systematically disadvantage a group?
• Is the claim generated by AI accurate and substantiated?
• Is a person able to challenge or reverse a consequential decision?
• Who is responsible when a platform’s proprietary model causes harm?

Fairness is context-dependent. Equal selection rates, equal error rates and equal outcomes cannot generally all be achieved simultaneously when base rates differ. Therefore, teams must define the relevant harm, affected groups and acceptable trade-offs rather than applying a single universal fairness metric.

10. Applied Case: Illustrative Retail Retention Programme

Background

An online retailer has 1.2 million active customers and declining repeat purchase rates. It plans to use a churn model to target customers with personalised offers.

Problem

The initial model predicts 90-day churn with high ranking performance. However, the marketing team discovers that most targeted customers receive discounts even when they would have purchased without intervention. The programme increases reported retention but reduces margin.

Analysis

The model predicts risk, not treatment responsiveness. The team creates a randomised holdout within each risk band. For each customer, incremental retention is estimated by comparing the treatment and control groups. The team also separates high-value customers from low-margin customers and checks whether discount delivery differs systematically across regions and customer groups.

Approach

The revised programme uses:

• a churn-risk model to identify customers requiring attention;
• an uplift model and experiment to estimate likely incremental response;
• CLV and contribution margin to prioritise profitable interventions;
• contact-frequency limits and exclusion rules;
• human approval for unusually large incentives;
• weekly monitoring of calibration, drift, margin and group outcomes.

Suppose the treatment group retains 2,400 of 10,000 customers and the control group retains 2,200 of 10,000. The estimated absolute incremental retention is:

(2,400/10,000) − (2,200/10,000) = 0.020,

or 2 percentage points. The relative lift is 2,400 ÷ 2,200 − 1 ≈ 9.1%, but the absolute lift is usually more useful for planning capacity and financial value. These estimates are meaningful only if assignment was random or the adjustment strategy plausibly addresses confounding.

Outcome

The retailer reduces unnecessary discounts, improves contribution margin and retains a smaller but more responsive audience. The model becomes one component of a controlled decision system rather than the sole decision-maker.

Learning Points

1. Churn prediction does not establish that retention activity will work.
2. Incrementality and margin are more valuable than retention volume alone.
3. Holdouts reveal whether automation creates additional value.
4. Governance must monitor distributional effects as well as aggregate performance.

11. An Implementation Framework for Organisations

An organisation should move through the following sequence:

Define the decision

Specify the action, outcome, time horizon, constraints, cost of errors and accountable owner.

Establish data and consent controls

Document sources, permissions, identity rules, retention, access and prohibited uses.

Select the appropriate modelling approach

Use prediction for prioritisation, causal experimentation for incrementality and optimisation for allocating constrained resources. Do not use a predictive model to answer a causal question without appropriate design.

Establish a baseline

Compare the AI system with current rules, expert judgement and a simple benchmark. Automation is justified by measurable improvement, not novelty.

Test safely

Use offline validation, time-based testing, simulation where appropriate, limited pilots and randomised holdouts. Define stop conditions before launch.

Monitor continuously

Track performance, calibration, drift, spend, fairness, complaints, fraud, creative quality and business value. Monitoring should include alerts and assigned responders.

Review and retire

Models require version control, documented changes, periodic revalidation and retirement when data or objectives change.

Contemporary Developments

Recent developments include multimodal foundation models, real-time generative creative, privacy-preserving measurement, clean rooms, contextual targeting, synthetic data, reinforcement-learning allocation and agentic marketing workflows. These developments may improve scale, but they also increase opacity and operational complexity. Privacy-enhancing technologies can reduce direct data sharing but do not automatically guarantee valid measurement or fairness. Agentic systems require especially clear permissions because they may plan and execute sequences of actions across platforms.

The strategic advantage will therefore come less from possessing a generic AI tool and more from integrating reliable first-party data, sound experimentation, distinctive creative assets, disciplined governance and skilled human judgement.

Chapter Summary

AI and algorithmic performance marketing uses machine learning to predict outcomes, recommend actions, allocate bids, identify audiences, generate creative and conduct conversations. Prediction, optimisation, experimentation and causal inference must be distinguished: a model that predicts conversion does not prove advertising caused conversion.

Effective systems align objectives with commercial value, use calibrated and appropriately validated models, monitor feedback loops and evaluate incremental outcomes. Propensity and churn models should be supplemented by uplift analysis when the managerial question concerns who will respond to treatment. Recommendation and bidding systems require constraints for profitability, customer experience, fairness and safety.

Generative and conversational AI can improve speed and personalisation but require factual, legal, brand and human review. Model quality involves data integrity, leakage prevention, calibration, drift monitoring and business measurement. Explainability should be matched to decision impact. Human oversight, bounded autonomy, auditability, rollback mechanisms and accountable ownership are essential because automated errors can scale rapidly.

Key Terms

Algorithmic performance marketing: Use of computational models to optimise measurable marketing outcomes.

Artificial intelligence: Systems performing tasks associated with human cognitive capabilities, such as prediction or language generation.

Calibration: Agreement between predicted probabilities and observed event frequencies.

Causal inference: Estimation of the effect of an action compared with a credible alternative.

Concept drift: Change in the relationship between predictors and outcomes.

Conversational AI: AI system that interacts with users through natural-language dialogue.

Data leakage: Use of information unavailable at the time a prediction would be made.

Generative AI: Models that produce new text, images, audio, video or other content.

Incrementality: Outcomes caused by marketing beyond what would have occurred otherwise.

Machine learning: Methods that learn patterns from data to make predictions or decisions.

Model drift: Decline in model validity caused by changing data or environments.

Propensity model: Model estimating the likelihood of an event.

Predictive audience: Audience defined using modelled future behaviour or value.

Recommendation engine: System selecting content, products or offers for users.

Uplift model: Model estimating differential response caused by treatment.

Churn: Customer discontinuation or lapse under a defined operational rule.

Explainability: Ability to communicate how or why a model produces decisions.

Automation bias: Uncritical reliance on machine-generated recommendations.

Review Questions

Short Answer Questions

1. Distinguish prediction from causal inference in performance marketing.
2. What is calibration, and why does it matter in automated bidding?
3. Explain why a churn model does not necessarily identify customers who will respond to a retention offer.
4. Identify two risks associated with generative AI creative.
5. What is bounded autonomy in conversational AI?

Long Answer Questions

1. Explain the design and evaluation of an ML-based bidding system, including objectives, feedback loops and safeguards.
2. Compare content-based recommendation, collaborative filtering and propensity modelling in marketing applications.
3. Discuss how predictive audiences can improve efficiency while reproducing historic exclusion.
4. Evaluate the role of explainability, monitoring and human oversight in high-scale automation.
5. Design an organisational governance framework for generative and conversational AI in advertising.

Analytical / Application Questions

1. A retailer’s conversion model has excellent AUC but poor calibration. Explain the likely operational consequences and recommend corrective action.
2. A campaign reports a 12% increase in conversions after an AI bidding change. Design an experiment to determine whether the increase is incremental.
3. A recommendation engine increases clicks but decreases margin and increases complaints. Analyse the objective failure and propose a revised optimisation and governance approach.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes prediction from causal inference?

A. Prediction estimates an outcome, while causal inference estimates the effect of an action relative to an alternative.

B. Prediction always requires randomised experiments, while causal inference never does.

C. Prediction measures profit, while causal inference measures clicks.

D. Prediction is qualitative, while causal inference is always descriptive.

Correct Answer: A. Prediction estimates an outcome, while causal inference estimates the effect of an action relative to an alternative.

Question 2

In the conceptual bidding relationship Bid = p(c|x) × V(c) × k, what does p(c|x) represent?

A. The auction clearing price

B. The predicted probability of conversion

C. The customer acquisition cost

D. The platform’s commission

Correct Answer: B. The predicted probability of conversion

Question 3

What does a lower Brier score generally indicate?

A. More accurate probability forecasts under the chosen outcome coding

B. Higher media reach regardless of model quality

C. Higher advertising frequency

D. Greater creative variety

Correct Answer: A. More accurate probability forecasts under the chosen outcome coding

Question 4

Why can a high-propensity audience be a poor audience for incremental advertising?

A. It may contain users who would have converted without advertising.

B. It necessarily contains users with low purchase intent.

C. Propensity models cannot use behavioural data.

D. High-propensity audiences are always prohibited by privacy law.

Correct Answer: A. It may contain users who would have converted without advertising.

Question 5

Which method is most directly suited to estimating incremental retention from an offer?

A. A randomised treatment and control experiment

B. Counting all post-campaign purchases

C. Selecting users with the highest predicted churn only

D. Comparing this month’s sales with a different product category

Correct Answer: A. A randomised treatment and control experiment

Question 6

What is target leakage?

A. A model’s failure to produce any prediction

B. Use of information unavailable at the time the decision is made

C. A customer receiving too few advertisements

D. A platform charging an incorrect tax rate

Correct Answer: B. Use of information unavailable at the time the decision is made

Question 7

Which is an appropriate safeguard for a conversational AI system handling high-risk requests?

A. Allowing the system to act without logging

B. Removing all human escalation routes

C. Requiring confirmation or escalation before irreversible actions

D. Maximising conversation length regardless of outcome

Correct Answer: C. Requiring confirmation or escalation before irreversible actions

Question 8

A treatment group retains 2,400 of 10,000 customers and a control group retains 2,200 of 10,000. What is the absolute incremental retention?

A. 0.2 percentage points

B. 2 percentage points

C. 9.1 percentage points

D. 20 percentage points

Correct Answer: B. 2 percentage points

Question 9

Which situation is an example of automation bias?

A. A manager tests a model against a simple benchmark.

B. An analyst rejects a model after finding data leakage.

C. A team accepts an algorithmic recommendation without reviewing obvious anomalies.

D. A marketer documents model ownership and escalation procedures.

Correct Answer: C. A team accepts an algorithmic recommendation without reviewing obvious anomalies.

Question 10

What is the principal purpose of a holdout group in performance marketing?

A. To increase exposure for the most active customers

B. To estimate what would have happened without the intervention

C. To guarantee that every user sees identical creative

D. To replace all model validation procedures

Correct Answer: B. To estimate what would have happened without the intervention