Marketing Analytics and Evidence-Based Decisions

Introduction

Digital entrepreneurship generates an unusually rich stream of behavioural and financial data: impressions, clicks, sessions, product views, registrations, purchases, renewals, support contacts and cancellations. Data abundance, however, does not automatically produce managerial insight. Analytics creates value only when data are converted into valid measures, interpreted in context, and connected to decisions.

Marketing analytics is the systematic use of data, statistical reasoning and business knowledge to understand marketing performance and improve decisions. Evidence-based marketing extends this idea by requiring managers to distinguish correlation from causation, evaluate uncertainty, and allocate resources according to demonstrated incremental value rather than attractive but potentially misleading metrics.

For an entrepreneurial firm, the central question is not simply “How many people clicked?” It is: “Which activities create additional, profitable and strategically valuable customer behaviour, for which segments, over what time period, and at what risk?”

Learning objectives

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

1. Define marketing objectives, KPIs, metric hierarchies and dashboards.
2. Analyse funnels and cohorts while recognising denominator, timing and selection problems.
3. Explain attribution and distinguish attributed performance from incremental impact.
4. Calculate campaign ROI, customer economics and break-even acquisition thresholds.
5. Apply forecasting, experimentation and decision rules to entrepreneurial marketing.
6. Evaluate data quality, privacy, ethical and governance risks.
7. Translate analysis into defensible managerial action.

1. From objectives to KPIs and dashboards

A metric is a quantified observation; a KPI is a metric explicitly linked to an important objective and decision. The distinction prevents “metric collecting” from being confused with performance management. A start-up whose objective is profitable retention might monitor renewal rate, contribution margin per retained customer and payback period. Impressions may be useful diagnostic information, but they are not necessarily KPIs for that objective.

A useful hierarchy has four levels:

A KPI should have a clear definition, owner, time window, population, data source and decision threshold. “Conversion rate” is incomplete unless the organisation specifies whether the denominator is impressions, sessions, visitors, leads or qualified leads, and whether conversions are counted within one day, seven days or another window.

Core equations

Conversion rate for a defined population is:

The numerator and denominator must refer to the same period, population and event definition. If one person can convert several times, the analyst must specify whether the unit is people, orders or events.

Customer acquisition cost (CAC) is commonly calculated as:

The cost boundary must be stated. A narrow media CAC excludes salaries and technology; a fully loaded CAC includes them. Neither is universally correct, but mixing definitions makes comparisons invalid.

A dashboard is a decision interface, not merely a collection of charts. An effective dashboard normally contains: current value, target or benchmark, trend, segmentation, uncertainty or data-quality status, and the action owner. It should permit drill-down from outcome to diagnosis while avoiding excessive visual complexity.

A dashboard can be organised around a KPI tree. For example, revenue can be decomposed as:

This identity is useful only when the definitions are compatible. A subscription business may instead use active customers × monthly revenue per active customer. The tree helps managers ask whether a revenue change resulted from acquisition, retention, frequency or price/mix.

2. Measurement architecture and data quality

Analytics begins before modelling. A measurement plan should specify events, properties, identities, timestamps, consent status, source systems and retention rules. Typical events include ad_impression, landing_page_view, signup, activation, purchase, renewal and cancelled_subscription.

Important quality dimensions include:

• Validity: the event represents the intended behaviour.
• Completeness: required events and fields are recorded.
• Consistency: definitions agree across platforms.
• Uniqueness: duplicated events do not inflate counts.
• Timeliness: data arrive soon enough for the decision.
• Lineage: the organisation can trace a number to its source and transformations.

Identity resolution is especially difficult. A person may appear as an anonymous browser, a mobile-device user and a logged-in customer. Combining these records can improve analysis but can also create false matches. Analysts should document deterministic matching, probabilistic matching and unresolved identities separately.

A major analytical error is denominator drift. Suppose 1,000 visitors produce 50 purchases in January, a 5% visitor conversion rate. If February reports 50 purchases from 800 qualified checkout users, its reported rate is 6.25%, but this does not prove performance improved: the population changed. Every KPI should therefore be accompanied by its denominator and eligibility rule.

3. Funnels and customer journeys

A funnel represents sequential stages through which a defined population progresses. A digital commerce funnel might be impression → landing-page visit → product view → cart → checkout → purchase. A B2B funnel could be lead → marketing-qualified lead → sales-qualified lead → opportunity → contract.

Stage conversion from stage i to stage j is:

where N_i is the number entering stage i and N_j is the number reaching stage j during a specified observation period. If stages are strictly sequential and the same cohort is followed, overall conversion is:

where p_i is the conditional conversion rate at each step. This multiplication is not valid when stage counts come from different populations or overlapping time windows.

Worked example: funnel diagnosis

An online course receives 20,000 landing-page visits. Of these, 2,400 begin registration, 1,200 complete registration, and 180 purchase.

Landing-to-registration conversion is 2,400 ÷ 20,000 = 12%. Registration completion is 1,200 ÷ 2,400 = 50%. Purchase conversion among completed registrations is 180 ÷ 1,200 = 15%. Overall landing-to-purchase conversion is 180 ÷ 20,000 = 0.9%, which also equals 12% × 50% × 15%.

The largest relative loss occurs at registration completion. Yet a manager should not automatically optimise that stage: improving a low-quality registration source may increase completions without increasing profitable purchases. Funnel analysis must therefore connect intermediate metrics to downstream value.

Funnels also have limitations. Customers can enter at different stages, repeat actions, move backwards, use multiple devices or convert after long delays. For complex journeys, path analysis, state-transition models or cohort analysis may represent reality better than a single linear funnel.

4. Cohort analysis and retention

A cohort is a group sharing a defined starting event or characteristic, such as customers acquired in the same week, users whose first purchase occurred in the same month, or subscribers entering through a particular channel. Cohort analysis separates customer age from calendar time.

Retention for cohort c at age t may be expressed as:

The definition of “active” must be explicit. It may mean a login, paid subscription, completed order or a usage threshold. Revenue retention and logo retention are not interchangeable.

A retention table might show:

The improving early retention of later cohorts may indicate better onboarding, improved product-market fit or a change in acquisition mix. It may also reflect incomplete observation: March cannot yet be compared with January at month three. This is right-censoring and must not be interpreted as superior long-term retention.

For subscription businesses, churn rate over a period can be defined as:

The population at risk excludes customers who joined after the period began if the purpose is to measure cancellation among opening customers. Different churn definitions produce different results; annualised churn estimates require assumptions about stability and are not automatically equal to monthly churn × 12.

5. Attribution: useful description, weak causation

Attribution assigns credit for a conversion to marketing touchpoints. Common rules include:

• Last-touch: all credit goes to the final recorded touchpoint.
• First-touch: all credit goes to the first touchpoint.
• Linear: credit is divided equally across recorded touches.
• Time-decay: later touches receive greater credit.
• Position-based: larger shares go to first and last touches.
• Algorithmic or data-driven: credit is estimated from observed patterns.

Attribution answers a descriptive question: which recorded contacts preceded conversions under a chosen rule? It does not necessarily answer the causal question: what would have happened without the marketing activity?

Attribution is vulnerable to missing identifiers, cookie restrictions, cross-device behaviour, view-through inflation, self-selection and correlated campaigns. Retargeting illustrates the problem: people who are already likely to buy may receive more retargeting exposure. Last-touch attribution can then reward retargeting for conversions it did not cause.

Attribution remains operationally useful for campaign tagging, journey diagnosis and budget administration, but managers should not interpret attributed revenue as incremental revenue without stronger evidence.

6. Incrementality and causal decision-making

Incrementality is the additional outcome caused by an intervention relative to a credible counterfactual. In its simplest form:

The second quantity is unobservable for the same unit at the same time, which is the fundamental problem of causal inference. Experiments approximate the counterfactual through randomisation.

In a randomised holdout test, eligible units are assigned to treatment or control. An incremental conversion rate is:

where p_T is the treatment-group conversion rate and p_C is the control-group conversion rate. Incremental conversions can be estimated as N_T × ∆p when N_T is the number of eligible treatment units and rates are measured over the same period.

Relative lift is:

This measure is undefined when p_C = 0 and can appear very large when the baseline is small. Absolute percentage-point change should also be reported.

Randomisation balances observed and unobserved characteristics in expectation. In practice, analysts must check assignment integrity, contamination, non-compliance, sample-ratio mismatch, seasonality and statistical uncertainty. A pre-registered primary outcome reduces selective reporting.

When individual randomisation is impossible, geo experiments, matched-market tests, interrupted time-series designs and difference-in-differences may be used. Difference-in-differences compares changes over time between treatment and comparison groups:

The key assumption is parallel trends: without treatment, the two groups would have followed comparable trajectories. This assumption should be examined using pre-treatment data, not merely asserted.

7. Campaign economics, ROI and unit economics

Return on investment should be defined with a consistent numerator and denominator. A simple campaign ROI is:

Contribution profit is revenue minus variable costs attributable to the incremental activity. If only attributed revenue is available, the result is attributed ROI, not incremental ROI.

Return on ad spend (ROAS) is:

ROAS is not profit. A campaign can have high ROAS but destroy value if gross margin, fulfilment, refunds or service costs are high.

Contribution margin per order is:

If a customer generates several orders, customer contribution over a time horizon is the sum of contribution margins minus variable retention costs. A simplified gross-margin lifetime value for a stable subscription is:

This steady-state approximation assumes constant revenue, margin and churn, and an effectively long horizon; it ignores discounting, acquisition mix, expansion, contraction and finite customer life. A more general discounted formulation is:

Here T is the horizon, t is the period, and r is the per-period discount rate. Probabilities of survival can be incorporated into expected contribution_t. The units of LTV must match CAC—for example, dollars per customer.

A campaign is economically viable when incremental customer contribution exceeds incremental acquisition cost, subject to cash-flow timing and risk. Payback period is the time required for cumulative contribution from acquired customers to recover acquisition cost. Short payback is particularly important for cash-constrained ventures.

Worked example: incremental campaign decision

A firm spends $12,000 on a controlled campaign. The treatment group contains 10,000 eligible prospects and converts at 4.0%; the control group contains 10,000 and converts at 3.0%. The incremental rate is 1 percentage point, producing an estimated 100 incremental customers (10,000 × 0.01). Each incremental customer generates $70 contribution profit over the measured horizon. Incremental contribution is therefore $7,000. Incremental ROI is ($7,000 − $12,000) ÷ $12,000 = −41.7%. The campaign should not be scaled on these measured economics, even though attributed revenue could look attractive.

8. Forecasting for entrepreneurial decisions

Forecasting estimates future outcomes using historical data, assumptions and scenario logic. It is not a promise and should be presented with uncertainty.

A naïve forecast uses the latest observation:

where Y_t is the most recent actual value. A moving-average forecast uses the mean of the previous k observations:

A weighted moving average assigns weights w_i summing to 1:

These methods are transparent but may perform poorly under structural change, promotions or strong seasonality. Exponential smoothing gives greater weight to recent observations:

where 0 ≤ α ≤ 1. A high α reacts quickly but is noisy; a low α is smoother but slower to detect change.

Forecast accuracy should be assessed on holdout periods, not only on the data used to fit a model. Mean absolute error is:

where n is the number of forecast periods. MAE retains the original unit, such as customers or dollars. Percentage errors can be unstable when actual values approach zero. Forecasts should include base, upside and downside scenarios, with assumptions about traffic, conversion, price, retention, capacity and spend.

9. Experimentation as a management capability

A sound experiment begins with a decision, not a tool. The process is:

1. State the business question and primary outcome.
2. Formulate a testable hypothesis.
3. Define the eligible population, treatment and control.
4. Select the unit of randomisation and prevent contamination.
5. Establish duration, sample-size logic and stopping rules.
6. Instrument outcomes and guardrail metrics.
7. Analyse according to the pre-specified plan.
8. Decide whether to stop, iterate, scale or reject.

Primary metrics should capture value; guardrails detect harm. For example, a checkout redesign may target purchase conversion while monitoring refunds, support contacts, payment failures and margin. A statistically detectable change may still be economically trivial, while a commercially important change may be uncertain with a small sample.

Multiple testing creates false-positive risk when many variants, segments and outcomes are inspected. Organisations should control this through pre-specification, correction procedures where appropriate, replication and disciplined interpretation. “No significant result” does not prove no effect; confidence intervals or credible intervals should be examined for the range of plausible effects.

Ethically, experiments must not expose customers to unreasonable harm, discriminatory treatment or deceptive practices. Consent, transparency and the ability to withdraw may be required depending on jurisdiction, data type and intervention.

10. Integrated illustrative case: LumaLearn

Background and problem

LumaLearn, an illustrative subscription-based language-learning venture, acquires users through paid social media, search and referrals. Its dashboard reports strong growth: 40,000 new registrations in a quarter and a reported paid-media ROAS of 3.2. Cash, however, is tightening, and management wants to decide whether to increase advertising spend.

Analysis

The analytics team first standardises definitions. A new customer is a person making a first paid subscription; CAC includes media and directly allocated agency fees; contribution excludes fixed engineering salaries but includes payment, support and content-delivery costs. The team discovers that the ROAS dashboard uses last-touch revenue, while referral conversions are often recorded after a retargeting impression.

A funnel review shows registration conversion rising, but activation—the completion of three lessons in the first week—falling from 46% to 32%. Three-month cohort retention is 24% for the earlier cohort and 17% for the latest paid-social cohort. The latest cohort has incomplete month-three observation, so the comparison is flagged as provisional rather than treated as final.

Management then runs a geo-level holdout test for paid social. Treatment regions receive normal advertising; control regions receive a reduced advertising schedule. After adjusting for pre-period differences and checking pre-treatment trends, the estimated incremental subscription rate is 0.6 percentage points. The observed lift is smaller than last-touch attribution implied.

Suppose the test estimates 600 incremental subscribers. Average three-month contribution per incremental subscriber is $18, and campaign cost is $8,000. Incremental contribution is $10,800, so incremental ROI is ($10,800 − $8,000) ÷ $8,000 = 35%. The campaign is economically positive over three months, but its margin of safety is modest. The team therefore models retention uncertainty and cash payback rather than scaling without limits.

Approach and decision

LumaLearn reallocates part of the budget from poorly retained paid social to referral incentives, improves first-week onboarding, and repeats the holdout test after the product change. The dashboard is redesigned to show incremental contribution, cohort retention, activation and data-quality warnings alongside attributed metrics.

Learning points

1. Growth in registrations is not equivalent to growth in profitable customers.
2. Attribution can guide diagnosis but overstate causation.
3. Cohorts reveal quality and timing differences hidden by aggregate averages.
4. Incrementality testing changes budget decisions from credit assignment to causal value.
5. A decision should incorporate uncertainty, payback and strategic constraints, not only a point estimate.

11. Governance, privacy and responsible analytics

Marketing analytics often processes personal data. Governance should address purpose limitation, lawful basis or consent where required, data minimisation, access controls, retention periods, deletion requests, vendor oversight and breach response. Sensitive attributes should not be used to produce discriminatory targeting or differential access to essential opportunities.

Consent banners alone do not make every use ethical. Organisations should explain material profiling, avoid dark patterns, document automated decisions and provide human review where appropriate. Data should be aggregated or pseudonymised when individual-level detail is unnecessary. Pseudonymisation reduces exposure but does not make data anonymous if re-identification remains reasonably possible.

AI systems can assist with forecasting, segmentation, creative testing and anomaly detection. They can also amplify biased historical data, leak confidential information, generate fabricated explanations or optimise a proxy that conflicts with customer welfare. Responsible deployment requires validated data, human accountability, audit logs, access restrictions, monitoring for drift and clear separation between exploratory suggestions and approved decisions.

12. Best-practice decision protocol

An evidence-based marketing decision should document:

• the decision and economic objective;
• KPI definitions, denominators and time horizon;
• data sources, exclusions and known quality limitations;
• descriptive evidence and causal evidence separately;
• assumptions, uncertainty and alternative explanations;
• customer, legal and ethical risks;
• the action threshold and accountable owner;
• a follow-up measurement plan.

The discipline is iterative. Measurement identifies a problem, diagnosis proposes an explanation, experimentation tests an intervention, and financial analysis determines whether the result merits scaling. The organisation then updates its assumptions and repeats the cycle.

Chapter Summary

Marketing analytics connects data to decisions through carefully defined objectives, KPIs and measurement systems. Dashboards should expose outcomes, diagnostics, trends, targets and data quality rather than encourage vanity-metric optimisation. Funnels identify stage losses, while cohort analysis separates customer age, acquisition time and retention quality.

Attribution assigns descriptive credit but does not establish incremental causation. Holdouts and other quasi-experimental designs provide stronger evidence about what marketing caused. Campaign evaluation must distinguish ROAS from profit and attributed performance from incremental contribution. CAC, contribution margin, LTV and payback are useful only when their boundaries, time horizons and assumptions are compatible.

Forecasts support planning but require validation, scenario analysis and explicit uncertainty. Experiments should begin with a decision, use primary and guardrail metrics, and address contamination, multiple testing and ethical risk. Finally, privacy, fairness, data lineage and AI governance are integral to credible analytics. Evidence-based management is therefore not the worship of numbers; it is disciplined reasoning under uncertainty.

Key Terms

KPI: A metric explicitly linked to an important organisational objective and decision.

Dashboard: A decision-oriented interface presenting performance measures, trends and diagnostics.

Conversion rate: The proportion of a defined eligible population completing a specified action.

Funnel: A sequence of stages through which users or customers progress.

Cohort: A group sharing a defined starting event or characteristic.

Retention: The proportion of a cohort remaining active at a specified age.

Churn: The rate at which customers cancel or become inactive during a defined period.

Attribution: Assignment of conversion credit to marketing touchpoints.

Incrementality: The additional outcome caused by an intervention relative to a counterfactual.

Control group: Eligible units not receiving the tested intervention.

Lift: The change in outcome associated with treatment, often relative or absolute.

CAC: Acquisition cost per newly acquired customer under a stated cost boundary.

ROAS: Attributed advertising revenue divided by advertising spend.

Contribution margin: Revenue remaining after specified variable costs.

LTV: Expected contribution from a customer over a stated future horizon.

Payback period: Time required for cumulative contribution to recover acquisition cost.

Forecast: An estimate of a future value based on data and assumptions.

Guardrail metric: A measure monitored to detect unintended harm or deterioration.

Data lineage: The traceable origin and transformation history of a data element.

Denominator drift: A change in the population used to calculate a rate, making comparisons misleading.

Review Questions

Short Answer Questions

1. Distinguish a KPI from a general metric and give one example of each.
2. Why can last-touch attribution overstate the value of retargeting?
3. Define a cohort and explain why cohort age matters in retention analysis.
4. State two assumptions behind the simplified subscription LTV approximation.
5. What is a guardrail metric, and why is it useful in experimentation?

Long Answer Questions

1. Explain how an entrepreneurial firm can design a KPI hierarchy from strategic objectives to operational measures.
2. Compare attribution analysis with incrementality analysis, including their managerial uses and limitations.
3. Discuss how funnel and cohort analyses can be combined to diagnose declining customer economics.
4. Explain the requirements for a credible marketing experiment and the threats to validity that should be monitored.
5. Evaluate the roles of privacy, fairness, data lineage and human oversight in responsible marketing analytics.

Analytical / Application Questions

1. A campaign produces 5,000 treatment prospects with a 6% conversion rate and 5,000 control prospects with a 4% rate. Each incremental customer yields $40 contribution and campaign cost is $3,000. Calculate the incremental customers, incremental contribution and incremental ROI. Should the campaign be scaled solely on this evidence?
2. A subscription cohort has 1,200 customers at entry, 540 active after one month and 300 active after three months. Calculate month-one and month-three retention, then identify one reason why the figures should not be compared with a newer cohort lacking three-month observations.
3. A manager reports ROAS of 4.0 but has no control group, and gross margin is 25% before support and refund costs. Design a measurement plan that would determine whether the campaign creates incremental profit.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes a KPI from a general metric?

A. A KPI is always a financial measure.

B. A KPI is explicitly linked to an important objective and decision.

C. A KPI is collected only from paid media.

D. A KPI cannot be displayed on a dashboard.

Correct Answer: B. A KPI is explicitly linked to an important objective and decision.

Question 2

A funnel has 20,000 visits and 180 purchases. What is the visit-to-purchase conversion rate?

A. 0.09%

B. 0.9%

C. 9%

D. 11.1%

Correct Answer: B. 0.9%

Question 3

Why is a newer cohort's apparently higher retention potentially misleading when it has not reached the same age as an older cohort?

A. It creates survivorship bias in advertising cost.

B. It reflects right-censoring and incomplete observation.

C. It guarantees that the newer cohort is more profitable.

D. It changes the definition of CAC.

Correct Answer: B. It reflects right-censoring and incomplete observation.

Question 4

What does incrementality attempt to estimate?

A. All revenue recorded by an advertising platform.

B. The number of touchpoints before conversion.

C. The additional outcome caused by an intervention.

D. The average order value of control customers.

Correct Answer: C. The additional outcome caused by an intervention.

Question 5

Treatment conversion is 5% and control conversion is 4%. What is the absolute lift?

A. 0.25 percentage points

B. 1 percentage point

C. 20 percentage points

D. 25 percentage points

Correct Answer: B. 1 percentage point

Question 6

Which measure is revenue attributed to advertising divided by advertising spend?

A. CAC

B. LTV

C. ROAS

D. Payback period

Correct Answer: C. ROAS

Question 7

A campaign costs $10,000 and generates $14,000 of incremental contribution. What is incremental ROI?

A. 4%

B. 40%

C. 71.4%

D. 140%

Correct Answer: B. 40%

Question 8

In a difference-in-differences analysis, the parallel-trends assumption means that:

A. Treatment and control must have identical outcomes after treatment.

B. Without treatment, the groups would have followed comparable trends.

C. Both groups must receive the same advertising exposure.

D. The treatment effect must be constant for every customer.

Correct Answer: B. Without treatment, the groups would have followed comparable trends.

Question 9

Which forecasting method gives the form F_t+1 = αY_t + (1 − α)F_t?

A. Naïve forecasting

B. Simple moving average

C. Exponential smoothing

D. Difference-in-differences

Correct Answer: C. Exponential smoothing

Question 10

Which practice most directly reduces the risk of confusing attributed performance with causal impact?

A. Increasing the number of dashboard colours.

B. Using a randomised holdout or credible comparison group.

C. Reporting only the highest-converting segment.

D. Replacing contribution profit with impressions.

Correct Answer: B. Using a randomised holdout or credible comparison group.