Ethics, Privacy, Governance and Future Measurement

Introduction

Campaign measurement is not merely a technical exercise in counting impressions, clicks, conversions and revenue. It is also an exercise in deciding what may be observed, who may be represented, whose interests are protected, and which forms of evidence are acceptable for managerial action. A measurement system can be statistically sophisticated and still be unethical, unlawful, insecure or socially harmful.

Contemporary advertising operates through extensive data infrastructures: websites and applications, connected television, mobile identifiers, loyalty programmes, data brokers, retail media platforms, customer relationship management systems and automated bidding algorithms. These infrastructures create opportunities for more relevant communication and better resource allocation, but they also create risks of surveillance, discrimination, manipulation, data leakage, opaque decision-making and unreliable attribution.

Ethical campaign analytics therefore requires a broader definition of performance. A campaign should be evaluated not only by financial return, but also by legitimacy, fairness, safety, accountability and resilience. This chapter examines the principles and practical systems needed to achieve that balance.

Learning Objectives

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

1. Explain consent, purpose limitation, data minimisation, fairness and accountability in advertising measurement.
2. Distinguish legitimate measurement from excessive surveillance and discuss algorithmic bias.
3. Evaluate ad fraud, brand-safety and data-security risks.
4. Design a measurement-governance framework with clear ownership, controls and escalation procedures.
5. Compare privacy-enhancing technologies, data clean rooms and emerging cross-media standards.
6. Assess campaign performance using both financial and responsible-measurement criteria.

1. The Ethical Foundations of Measurement

1.1 Data is not ethically neutral

Data reflects choices about what is recorded, how categories are defined and which behaviours are treated as valuable. For example, a conversion model trained on historical purchases may allocate more media to affluent neighbourhoods because those areas have generated more observed sales. If historical exposure, income, language or access to credit is uneven, the model may reproduce that inequality while appearing mathematically accurate.

Measurement decisions can affect real opportunities. Advertising may influence access to employment, housing, education, insurance, healthcare or political information. Consequently, ethical review is essential when targeting or optimisation may produce material effects on individuals or groups.

1.2 Core principles

These principles may conflict. Highly granular data can improve short-term attribution while increasing privacy risk. Strong anonymisation can reduce risk but also reduce analytical utility. Ethical governance is therefore a process of justified trade-offs, not a claim that all risk can be eliminated.

2. Consent, Identity and the Problem of Surveillance

2.1 Meaningful consent

Consent is more than the presence of a banner or a preselected checkbox. Meaningful consent is generally expected to be informed, specific, freely given and revocable. A person should understand the categories of data collected, purposes of processing, major recipient categories and consequences of refusing consent.

A useful operational distinction is:

• Notice: telling people what the organisation does.
• Choice: allowing people to accept, reject or modify purposes.
• Control: ensuring the choice is technically respected across systems.
• Proof: retaining evidence of when and how the choice was obtained.

Consent-management failures often occur after collection. A platform may receive a rejection but continue to send identifiers to an analytics vendor, or a customer may withdraw consent while historical profiles remain active in activation systems. Governance must therefore connect consent records to collection, processing, activation, suppression and deletion workflows.

2.2 Lawful basis and jurisdiction

Regulatory requirements differ by jurisdiction and by the nature of the data. Depending on the applicable legal framework, organisations may rely on consent, contract, legal obligation, legitimate interests or another recognised basis. This is a legal determination requiring qualified advice; analytics teams should not assume that a business purpose automatically authorises processing.

Important regulatory themes include data-subject rights, transparency, children's privacy, sensitive or special-category data, cross-border transfers, breach notification, automated decision-making and vendor accountability. Examples of major regulatory regimes include the European Union's General Data Protection Regulation, the ePrivacy framework, the California Consumer Privacy Act and related state privacy laws, Brazil's Lei Geral de Proteção de Dados and sector-specific rules in many countries. These instruments change over time, and compliance requires current legal review rather than reliance on historical summaries.

2.3 Surveillance and contextual alternatives

Surveillance becomes a concern when observation is persistent, difficult to avoid, poorly understood or disproportionate to the advertising purpose. Cross-site tracking, device graphs, inferred sensitive attributes and indefinite retention can create profiles that people cannot reasonably inspect or correct.

A privacy-respecting measurement strategy often moves from individual-level observation toward:

1. contextual targeting based on the content or environment rather than a person;
2. aggregated reach, frequency and conversion reporting;
3. first-party data collected through a transparent relationship;
4. experiments and geo-based tests that do not require extensive personal histories; and
5. short retention periods and strict separation between identity and analysis.

The objective is not to prohibit all personalisation. It is to ensure that the information intensity of a system is proportionate to a legitimate purpose.

3. Fairness and Algorithmic Bias

3.1 Sources of bias

Bias may enter a measurement system through:

• Representation bias: important groups are missing or underrepresented in the data.
• Measurement bias: proxies such as device type or postcode poorly represent the construct of interest.
• Historical bias: past decisions reflect unequal social conditions.
• Selection bias: only exposed, logged-in or converting users are observed.
• Label bias: outcomes such as “high value” encode organisational assumptions.
• Optimisation bias: a platform maximises a narrow metric that conflicts with social or legal goals.

Automated advertising systems may also use protected characteristics indirectly. A model need not receive a person's ethnicity or disability status to produce disparate outcomes; correlated variables can act as proxies.

3.2 Fairness assessment

No single fairness metric is universally correct. Teams should identify the relevant harm, protected groups, decision context and acceptable trade-offs. Common assessments include:

• Demographic parity: selection rates are similar across groups.
• Equal opportunity: true-positive rates are similar across groups.
• Calibration: a predicted probability has comparable meaning across groups.

These criteria can conflict, especially when base rates differ. A fairness review should therefore combine quantitative testing with domain expertise, qualitative consultation and legal analysis. Small sample sizes should not be used to make confident claims about group differences.

A basic disparity ratio for a selection rate can be written as:

Here, p_A and p_B are the proportions selected or reached in groups A and B. The ratio is dimensionless and is meaningful only when the groups, outcome definition, observation window and decision rule are comparable. It is a diagnostic indicator, not proof of discrimination or fairness.

3.3 Bias mitigation

Mitigation may occur before, during or after modelling:

• improve representation and data quality;
• remove unjustified proxy variables;
• constrain targeting categories;
• use fairness-aware optimisation where appropriate;
• monitor delivery and outcomes by relevant groups;
• establish human review for high-impact decisions; and
• suspend models when unexplained disparities exceed agreed thresholds.

It is dangerous to treat fairness as a one-time model audit. Campaign objectives, inventory, audiences and social conditions change. Continuous monitoring and documented remediation are required.

4. Ad Fraud, Invalid Traffic and Measurement Integrity

Ad fraud is the deliberate generation of false or non-human advertising activity for financial gain. Invalid traffic may include bots, automated browsing, deceptive placements, domain spoofing, click farms, incentivised activity that does not represent genuine interest, malware-driven traffic and fabricated conversions.

Fraud distorts both media cost and analytical conclusions. If fraudulent impressions are counted as reach, frequency estimates are wrong. If fraudulent clicks enter an attribution model, budget may shift toward low-quality sources. If fake conversions are credited to a campaign, return on investment is overstated.

A conventional financial measure is:

ROI is dimensionless and is valid only when revenue and cost are measured on a comparable basis and when “incremental” means revenue and cost caused by the campaign relative to a credible counterfactual. Fraud can contaminate either term and can also create false confidence in attribution.

Controls include independent verification, supply-path analysis, ads.txt or equivalent authorised-seller mechanisms, app-ads.txt, sellers.json, traffic-quality monitoring, conversion deduplication, frequency controls, payment anomaly review and contractual audit rights. No single vendor's fraud score should be treated as absolute truth. Different vendors use different definitions, detection methods and observation capabilities.

5. Brand Safety and Suitability

Brand safety concerns the risk that advertising appears beside illegal, extremist, hateful, sexually exploitative, dangerous or otherwise damaging content. Brand suitability is broader and concerns whether an environment is appropriate for a particular brand, audience and campaign objective.

Keyword blocking alone is inadequate. A word may be harmless in one context and damaging in another; conversely, harmful content may use language that simple filters do not detect. Effective controls combine:

• content classification and semantic analysis;
• publisher and supply-chain due diligence;
• exclusion and inclusion lists;
• human escalation for ambiguous cases;
• contractual remedies and make-goods;
• ongoing incident monitoring; and
• documented risk appetite by market and campaign.

Overblocking can harm legitimate publishers and reduce access to diverse viewpoints. Underblocking can expose audiences and brands to serious harm. Governance should therefore record why a category is restricted and assess both false positives and false negatives.

6. Data Security and Responsible Data Handling

Privacy and security are related but distinct. Privacy concerns appropriate collection and use; security concerns protection against unauthorised access, disclosure, alteration or destruction. A dataset may be securely stored but collected without a valid purpose, or lawfully collected but poorly protected.

A practical security architecture includes:

1. Data inventory: identify sources, fields, purposes, owners, locations and recipients.
2. Classification: distinguish public, internal, confidential, personal and sensitive data.
3. Access control: apply least privilege, role-based access and periodic access reviews.
4. Protection: use encryption in transit and at rest, key management and secure secrets handling.
5. Pseudonymisation: separate direct identifiers from analytical records, recognising that pseudonymous data may remain personal data.
6. Retention control: delete or irreversibly aggregate data when the purpose ends.
7. Monitoring: log access, detect unusual exports and test incident response.
8. Supplier management: assess processors, sub-processors, certifications, breach duties and deletion obligations.

Data retention should be justified rather than indefinite. If a campaign needs weekly aggregate sales for optimisation, retaining every individual event for years is difficult to defend. Backups, replicated warehouses and vendor copies must be included in deletion planning.

7. Privacy-Enhancing Technologies

Privacy-enhancing technologies, or PETs, seek to preserve analytical value while reducing exposure of personal information. They do not automatically make a processing activity lawful or ethical; their protection depends on implementation, threat models and governance.

7.1 Aggregation and de-identification

Aggregation reports results for groups rather than individuals. De-identification removes or transforms identifiers, but re-identification may remain possible when datasets are linked or when groups are small. Suppression, minimum cell sizes and controlled outputs reduce risk.

7.2 Differential privacy

Differential privacy provides a formal way to limit how much the output of an analysis changes when one person's record is added or removed. A mechanism M is commonly described as satisfying ε-differential privacy when, for neighbouring datasets D and D′ and every possible output set S:

Here, ε is the privacy-loss parameter, D and D′ differ in one individual's record, and smaller ε generally indicates stronger privacy for the same mechanism. The guarantee depends on the defined neighbouring relationship, the mechanism, composition across repeated queries and protection of the implementation. Differential privacy does not guarantee that a result is accurate, unbiased or legally permissible.

7.3 Federated learning and secure computation

Federated learning trains a model across distributed data locations without centralising raw records, although model updates can still leak information if not protected. Secure multiparty computation allows parties to compute certain functions over protected inputs. Homomorphic encryption permits computation on encrypted data but may involve significant computational cost. These technologies are valuable where raw data cannot or should not be pooled, but they require specialist threat analysis and operational maturity.

7.4 Data clean rooms

A data clean room is a controlled environment in which two or more parties compare or analyse datasets under restricted access and output rules. Typical applications include measuring overlap between a retailer's purchasers and a media platform's exposed audience, or estimating campaign outcomes without exchanging raw identifiers.

A credible clean room should specify:

• permitted data fields and purposes;
• identity-resolution method and error rates;
• minimum aggregation thresholds;
• query approval and logging;
• prohibited joins and re-identification controls;
• retention and deletion rules;
• independent audit rights; and
• responsibility for errors and regulatory requests.

Clean rooms can reduce direct data sharing but may centralise power in the platform operating the environment. They can also limit independent verification if the platform controls identity graphs, methodology and reporting. Buyers should require methodological documentation, uncertainty estimates and the ability to compare results with experiments or alternative sources.

8. Measurement Governance

Measurement governance is the system of roles, policies, standards, controls and escalation mechanisms through which an organisation decides what measurement is allowed and how results may be used.

8.1 Governance operating model

A strong operating model assigns:

• an executive sponsor accountable for risk and value;
• a privacy or legal function responsible for regulatory interpretation;
• data stewards responsible for definitions, quality and access;
• analytics leaders responsible for methods and uncertainty;
• information-security specialists responsible for technical controls;
• media and brand-safety owners responsible for supply and placement risk; and
• an independent review or audit function.

The model should distinguish ownership from execution. An agency may execute a measurement plan, but the advertiser remains responsible for understanding what data is shared on its behalf and what claims are made to decision-makers.

8.2 Measurement register and approval gates

Every significant measurement initiative should have a register containing purpose, data sources, legal basis, affected populations, vendors, retention period, outputs, risks, controls, owner and review date. Approval gates can be placed at:

1. idea and purpose definition;
2. data acquisition and vendor onboarding;
3. model or dashboard development;
4. activation or optimisation;
5. publication of results; and
6. retirement and deletion.

A model card or measurement card should document the population, outcome, training data, exclusions, assumptions, known limitations, validation results, fairness checks and intended use. This prevents an attribution model designed for directional optimisation from being presented as audited financial causality.

8.3 Evidence hierarchy

A governance system should classify claims according to evidence strength. Randomised experiments generally provide stronger causal evidence than observational last-touch attribution, although experiments may be expensive, imperfect or infeasible. A practical hierarchy is:

• descriptive reporting: what happened;
• diagnostic analysis: what patterns accompanied it;
• predictive modelling: what may happen;
• causal inference: what changed because of the campaign; and
• decision analysis: what should be done given objectives, constraints and uncertainty.

The stronger the business consequence, the stronger the required evidence and review. A small bid adjustment may rely on directional evidence; a major budget reallocation or public claim requires greater validation.

9. A Responsible Measurement Scorecard

Campaign performance should be reported through a balanced scorecard rather than a single return number.

The scorecard should report confidence intervals, data coverage and material limitations where possible. A result based on 70% observable conversions should not be presented with the same certainty as a result based on a complete, validated transaction ledger.

10. Emerging Cross-Media Standards and Measurement

Cross-media measurement seeks to estimate reach, frequency and outcomes across television, streaming video, audio, display, search, social and retail media without double counting. The central difficulty is that each channel has different identifiers, log structures, definitions of exposure and reporting incentives.

Emerging approaches increasingly emphasise:

• common taxonomies for impressions, viewability and exposure;
• interoperable event and audience definitions;
• independent accreditation and auditing;
• privacy-preserving identity and clean-room collaboration;
• deduplicated reach and frequency estimates;
• incrementality testing rather than exposure-only attribution; and
• transparent uncertainty and methodology disclosure.

No standard eliminates the need for judgment. “Reach” may mean served impressions, viewable impressions, authenticated people, estimated households or attention-qualified exposure. A cross-media report is meaningful only when its population, geography, time period, deduplication method and exposure definition are explicit.

Illustrative Case: The Northstar Retail Launch

Background

Northstar, a fictional national retailer, launched a six-week campaign across connected television, social video, search and retailer-owned media. Its agency proposed combining customer email addresses, publisher identifiers and purchase data in a clean room to calculate return on advertising spend and build lookalike audiences.

Problem

The initial plan had four weaknesses. Consent records were stored separately from the customer data warehouse. The proposed audience model used postcode and browsing behaviour as value predictors, creating a risk of socioeconomic discrimination. The platform reported conversions using different attribution windows, and the campaign included low-cost video inventory with unusually high click rates. Brand-safety exclusions were based only on keywords.

Analysis

The governance committee mapped data flows and found that an opt-out suppression file was updated weekly, while activation lists were refreshed daily. It also found that the campaign's reported revenue was observational rather than incremental. A traffic-quality audit identified a large discrepancy between ad-server impressions and platform-reported video views. A fairness review showed materially lower delivery in several language communities, but the team could not determine whether this reflected legitimate audience strategy or proxy discrimination.

Approach and solution

Northstar adopted the following controls:

1. consent status was synchronised before activation and withdrawal was propagated within a defined service level;
2. the clean room limited analysis to hashed, access-controlled records and aggregate cells above a minimum threshold;
3. postcode was removed from the value model until a documented fairness assessment was completed;
4. an incrementality test with geographically matched markets was introduced;
5. suspicious inventory was paused pending independent verification;
6. semantic brand-safety classification and human review replaced keyword blocking alone; and
7. the board report separated observed attributed revenue from estimated incremental revenue and disclosed uncertainty.

For the financial analysis, the team defined incremental ROI as:

Here, ΔRevenue is the estimated difference in revenue between treatment and control markets attributable to the campaign, subject to the test assumptions. If estimated incremental revenue were $1,200,000 and campaign cost were $800,000, then:

The calculation is valid only if the treatment-control design provides a credible counterfactual and the revenue estimate is expressed on the same time, market and accounting basis as cost.

Outcome

Northstar reported a lower short-term return than the platform's original attribution report, but the result was more defensible. It reduced exposure to questionable inventory, improved consent compliance, documented audience limitations and created a repeatable approval process. Management accepted the lower reported number because it had greater evidential quality and lower regulatory and reputational risk.

Learning points

The case illustrates that governance is not an obstacle placed after analytics. It changes the design of data collection, experimentation, optimisation and reporting. It also shows why ethical measurement can improve managerial decisions: unreliable or unlawfully obtained evidence is not a genuine performance advantage.

Challenges and Limitations

First, privacy-preserving methods often reduce granularity. Small-cell suppression can make niche audiences difficult to evaluate, while differential privacy introduces a privacy-accuracy trade-off. Second, fairness metrics may conflict, and a technically balanced delivery pattern may still be substantively unfair if the objective itself is inappropriate. Third, clean rooms can create vendor dependence and limited transparency. Fourth, regulatory obligations vary across jurisdictions and may apply differently to advertisers, agencies, publishers and processors. Fifth, measurement systems can produce a false appearance of objectivity when definitions and assumptions remain hidden.

There are also organisational limitations. Governance can become bureaucratic if every low-risk dashboard receives the same review as a high-impact profiling system. Conversely, rapid experimentation can bypass controls if responsibility is unclear. A proportionate, risk-based approach is therefore essential.

Best-Practice Checklist

Before approving a campaign measurement system, ask:

1. What precise decision will the data support?
2. Is the collection proportionate to that purpose?
3. What is the legal basis, and is consent genuinely meaningful where required?
4. Can individuals exercise access, correction, objection, deletion or opt-out rights?
5. Which groups could experience unequal delivery or outcomes?
6. What controls address fraud, unsafe placements and supply-chain opacity?
7. Who owns security, methodology, vendor performance and incident response?
8. Are the reported results descriptive, predictive or causal?
9. What uncertainty, missingness and data-quality limitations must be disclosed?
10. When will the data, model and access permissions be reviewed or retired?

Chapter Summary

Ethical campaign measurement combines performance analysis with respect for autonomy, fairness, security, safety and accountability. Consent must be meaningful and technically enforceable; data collection should be limited to legitimate purposes; and surveillance risks should be reduced through contextual, aggregated and experimental approaches. Algorithmic bias can arise from data, labels, proxies and optimisation objectives, requiring continuous assessment rather than a single audit.

Ad fraud and brand-safety failures compromise both financial results and public trust. Security requires inventories, least-privilege access, encryption, retention controls and supplier oversight. Privacy-enhancing technologies, including aggregation, differential privacy, federated learning and clean rooms, can reduce exposure but do not replace legal and ethical governance. A mature governance system assigns owners, uses approval gates, documents methods, distinguishes causal evidence from attribution and reports uncertainty. Future cross-media measurement will depend on interoperable standards, independent verification and privacy-preserving deduplication.

Key Terms

Consent: A voluntary, informed and specific agreement to a defined processing activity.

Purpose limitation: Restricting data use to specified, legitimate purposes.

Data minimisation: Collecting and retaining no more personal data than necessary.

Pseudonymisation: Separating or transforming identifiers while recognising that re-identification may remain possible.

Surveillance: Persistent or extensive observation of people or behaviour, especially where control and transparency are limited.

Algorithmic bias: Systematic error or disadvantage produced through data, design or deployment of an algorithmic system.

Demographic parity: A fairness criterion based on similar selection rates across groups.

Differential privacy: A formal privacy guarantee limiting the effect of one person's record on an output.

Privacy-enhancing technology: A technology designed to reduce data exposure while preserving useful processing.

Data clean room: A controlled environment for analysing datasets without freely exchanging raw records.

Ad fraud: Deliberate manipulation of advertising activity for financial gain.

Invalid traffic: Traffic that is non-human, deceptive, duplicated or otherwise unsuitable as evidence of genuine advertising activity.

Brand safety: Protection against advertising appearing in harmful or damaging environments.

Brand suitability: The degree to which an environment fits a particular brand's risk tolerance and audience.

Incrementality: The change caused by a campaign relative to a credible counterfactual.

Measurement governance: Roles, policies, controls and processes governing analytical data and claims.

Data steward: The person responsible for data quality, definitions, access and lifecycle management.

Clean-room output threshold: A minimum group size or aggregation rule intended to reduce re-identification risk.

Cross-media measurement: Estimation of reach, frequency and outcomes across multiple media channels.

Accountability: The duty to take responsibility for decisions and demonstrate compliance and responsible practice.

Review Questions

Short Answer Questions

1. Why is consent more than displaying a privacy banner?
2. Identify four sources of algorithmic bias in campaign analytics.
3. Distinguish privacy from information security.
4. What is the purpose of a data clean room?
5. Why should attributed revenue be distinguished from incremental revenue?

Long Answer Questions

1. Explain how purpose limitation, data minimisation and retention controls can be applied to a cross-channel campaign.
2. Critically evaluate the strengths and limitations of differential privacy for advertising measurement.
3. Discuss how ad fraud can distort both campaign ROI and managerial decision-making.
4. Design a measurement-governance operating model for a multinational advertiser.
5. Explain why fairness metrics may conflict and propose a responsible bias-monitoring process.

Analytical / Application Questions

1. A platform reports high conversion rates from a new video publisher, but independent verification finds unusual click patterns and large discrepancies in impressions. Diagnose the risks and recommend an investigation and budget decision.
2. A clean-room analysis shows strong campaign performance, but the platform will not disclose identity-resolution error rates or query rules. Evaluate whether the result should guide budget allocation and specify additional evidence required.
3. A retailer's model delivers substantially fewer advertisements to a language minority after using postcode and browsing variables. Develop a structured assessment covering possible causes, fairness tests, mitigation and governance escalation.

Multiple Choice Questions with Answers

Question 1

Which principle requires an organisation to collect only the personal data reasonably necessary for a specified purpose?

A. Data minimisation

B. Purpose expansion

C. Frequency capping

D. Calibration

Correct Answer: A. Data minimisation

Question 2

What is the principal purpose of a credible incrementality test?

A. To maximise the number of attributed conversions

B. To estimate what would have happened without the campaign

C. To remove all uncertainty from revenue forecasts

D. To replace consent management

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

Question 3

Which statement best describes a data clean room?

A. A public database containing unencrypted customer records

B. A system that automatically guarantees legal compliance

C. A controlled environment for analysing datasets under restricted access and output rules

D. A tool used only to block unsafe keywords

Correct Answer: C. A controlled environment for analysing datasets under restricted access and output rules

Question 4

In differential privacy, what does a smaller ε generally indicate, assuming the same mechanism and definitions?

A. Stronger privacy protection

B. Higher guaranteed campaign revenue

C. Perfectly unbiased data

D. Complete removal of sampling error

Correct Answer: A. Stronger privacy protection

Question 5

Which practice most directly addresses ad fraud?

A. Using longer attribution windows for every channel

B. Independent traffic verification and authorised-seller checks

C. Removing all brand-safety exclusions

D. Increasing the number of audience segments

Correct Answer: B. Independent traffic verification and authorised-seller checks

Question 6

Why can postcode function as a problematic variable in an advertising model?

A. It is always randomly assigned

B. It may act as a proxy for socioeconomic or protected characteristics

C. It prevents all forms of measurement

D. It guarantees demographic parity

Correct Answer: B. It may act as a proxy for socioeconomic or protected characteristics

Question 7

Which statement correctly distinguishes privacy from security?

A. Privacy concerns appropriate collection and use; security concerns protection against unauthorised access or alteration

B. Privacy and security are identical terms

C. Security determines whether a purpose is ethically justified

D. Privacy applies only to encrypted data

Correct Answer: A. Privacy concerns appropriate collection and use; security concerns protection against unauthorised access or alteration

Question 8

A campaign has incremental revenue of $900,000 and cost of $600,000. Using ROI = (Incremental Revenue − Incremental Cost) ÷ Incremental Cost, what is ROI?

A. 25%

B. 50%

C. 66.7%

D. 150%

Correct Answer: B. 50%

Question 9

Which evidence claim is generally stronger than last-touch attribution for estimating causal campaign impact?

A. A randomised or well-designed controlled experiment

B. A larger dashboard colour palette

C. A longer list of descriptive metrics

D. A platform's unverified conversion count

Correct Answer: A. A randomised or well-designed controlled experiment

Question 10

Why is keyword blocking alone insufficient for brand-safety governance?

A. Keywords are never present in digital content

B. Context can make the same word safe or harmful, and harmful content may evade keyword filters

C. Keyword systems always suppress too little inventory

D. Brand safety concerns only campaign cost

Correct Answer: B. Context can make the same word safe or harmful, and harmful content may evade keyword filters