Analytics, Bidding and Campaign Optimisation

Introduction

Programmatic advertising is not optimised by a single metric or a single algorithm. It is an ongoing decision system in which advertisers evaluate audiences, inventory, bids, budgets, creative, timing and measurement evidence. The central managerial task is to determine where the next unit of expenditure is likely to create the greatest incremental business value.

A campaign may generate inexpensive impressions but weak outcomes, excellent conversion rates but insufficient scale, or strong reported performance that is largely attributable to retargeting and would have occurred without advertising. Effective optimisation therefore combines funnel analysis, cohort analysis, placement and audience diagnosis, supply-path evaluation, controlled experimentation and marginal-return reasoning.

This chapter treats optimisation as a disciplined analytical process rather than as continual manual adjustment. The objective is not simply to reduce cost per acquisition (CPA), but to improve incremental outcomes subject to budget, reach, brand-safety, privacy and delivery constraints.

Learning objectives

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

1. Diagnose programmatic performance across the funnel.
2. Compare audiences, placements, cohorts and supply paths using appropriate denominators.
3. Allocate budgets and manage pacing under delivery and performance constraints.
4. Apply frequency management and bid optimisation without confusing correlation with causation.
5. Design and interpret A/B tests and incrementality studies.
6. Use marginal-return reasoning to make defensible optimisation decisions.
7. Recognise measurement, privacy, fraud, brand-safety and governance limitations.

1. The optimisation problem

A programmatic campaign receives opportunities to bid on impressions. For each opportunity, the decision system estimates value and compares it with the expected cost of winning. The value may be a purchase, qualified lead, store visit, attention event or a longer-term customer relationship.

At campaign level, optimisation can be represented conceptually as:

Here, B is spend, R is reach, F is frequency, S represents supply quality, and V is business value. This is a conceptual optimisation statement, not a claim that every platform exposes or solves this exact equation. In practice, platforms use predictive models, auctions, constraints and feedback signals.

A useful optimisation hierarchy is:

Optimisation should proceed from business objectives to operational settings. Changing a bid before verifying conversion tracking or supply quality is often an expensive diagnostic error.

2. Measurement foundations and funnel diagnosis

The funnel is a sequence of measurable events. A simplified display funnel is:

Impressions → viewable impressions → clicks or engaged visits → conversions → revenue or profit.

Common metrics include:

For example, if spend is $12,000 and conversions are 300, then:

If attributed revenue is $36,000, then:

These calculations assume that spend, conversions and revenue refer to the same attribution window, population and reporting period. A platform-reported CPA can be misleading if view-through conversions, duplicated events, refunds or cross-device identity errors are not understood.

Diagnosing leakage

A funnel diagnosis asks where the largest loss or quality problem occurs:

• Delivery leakage: low win rate, underdelivery or excessive auction loss.
• Exposure leakage: poor viewability, short exposure time or unsuitable placement.
• Engagement leakage: weak creative relevance, slow landing page or incorrect audience promise.
• Conversion leakage: landing-page friction, weak offer, poor traffic quality or tracking failure.
• Value leakage: conversions occur but have low margin, low retention or high cancellation.

A high CTR with low conversion rate may indicate accidental clicks, misleading creative or a poor landing page. A low CTR with high conversion rate may indicate a narrow but valuable audience. Optimisation must therefore use the complete path rather than automatically rewarding the highest CTR.

3. Cohort, audience and placement analysis

Cohort analysis

A cohort is a group sharing a meaningful characteristic, such as first-exposure week, acquisition date, device, geography or customer status. Cohort analysis is especially useful for distinguishing immediate response from later value.

For cohort c:

and, where customer value is available:

The notation is intentionally expressed in words where necessary: lifetime value and acquisition cost must be defined consistently. LTV estimates should include the same customer population, observation period and margin assumptions. A cohort acquired cheaply may be unattractive if it churns rapidly.

A post-graduate analyst should compare cohorts by maturity. A cohort acquired yesterday has had less time to convert and retain than one acquired six months ago. Comparing immature and mature cohorts without adjustment creates survivorship and censoring problems.

Audience analysis

Audiences may be first-party, contextual, publisher-declared, platform-modeled or third-party. Analysis should examine:

• scale and unique reach;
• incremental conversions;
• frequency distribution;
• conversion lag;
• new versus existing customers;
• marginal CPA as spend expands;
• privacy and consent status;
• overlap with other audience groups.

Audience labels are not equivalent to audience value. A “high-intent” segment may perform well because it contains users already close to purchase. If the campaign merely captures existing demand, reported efficiency may overstate incrementality.

Placement analysis

Placement-level reports identify domains, apps, content categories, ad positions and formats associated with performance. However, small samples are unstable. A placement with two conversions from 20 clicks should not automatically be promoted over one with 200 conversions and slightly lower rate.

Useful placement checks include:

1. Is the inventory viewable and brand safe?
2. Are impressions human and valid?
3. Is the traffic consistent with the intended environment?
4. Does the placement produce new customers or mainly retarget existing ones?
5. Are results stable across time, device and creative?
6. Is the supply path unnecessarily long or duplicated?

Placement optimisation should use minimum-volume rules, confidence intervals where appropriate, and business-quality outcomes—not only apparent rate differences.

4. Supply-path analysis

The supply path is the sequence through which an impression request travels: publisher or app, supply-side platform (SSP), exchange, reseller, verification provider and demand-side platform (DSP). Supply-path optimisation (SPO) seeks efficient, transparent and high-quality routes.

Important dimensions include:

• Media cost: clearing price and fees.
• Path length: number of intermediaries and resellers.
• Auction duplication: multiple routes to the same impression.
• Transparency: accurate sellers.json, app-ads.txt and schain information.
• Quality: viewability, invalid-traffic rate, brand safety and attention.
• Access: unique inventory or merely redundant access.
• Outcome: conversions and incremental value per dollar.

A cheaper path is not automatically superior. It may have lower viewability, weaker identity quality or less valuable inventory. Conversely, a premium path may justify its cost if it produces better outcomes or unique reach.

A practical supply-path scorecard should compare paths using matched inventory and consistent attribution. Analysts should avoid comparing one path dominated by mobile video with another dominated by desktop display without controlling for format and audience.

5. Budget allocation and pacing

Budget allocation

Suppose a campaign has a fixed budget and several line items. A basic allocation process is:

1. Establish minimum spend required for learning or contractual delivery.
2. Estimate performance by line item using sufficiently mature data.
3. Protect strategic reach, prospecting and testing budgets.
4. Shift discretionary spend toward higher incremental return.
5. Reassess after a defined learning period.

Average CPA is not enough for allocation. The relevant quantity is often the marginal CPA: the cost of the next group of conversions generated by additional spend.

The symbol ∆ means change between two spend levels. If spend rises from $20,000 to $24,000 and conversions rise from 500 to 580, then:

This is distinct from the new average CPA of $24,000 ÷ 580 ≈ $41.38. The marginal CPA is the more relevant figure when deciding whether to add the next $4,000, assuming the comparison is causal and the two periods are comparable.

Pacing

Pacing distributes budget over time. A simple cumulative pacing ratio is:

A ratio below 1 indicates underspending against the schedule; above 1 indicates overspending. This measure assumes the planned schedule is appropriate and should be interpreted alongside inventory availability and performance.

Pacing strategies include:

• Even pacing: approximately equal delivery over time.
• Front-loaded pacing: heavier early delivery, useful for learning or launches.
• Performance pacing: spend varies with predicted marginal return.
• Daypart pacing: delivery is concentrated in valuable time periods.

Aggressive underpacing can prevent learning; aggressive overpacing can exhaust audiences, inflate frequency and leave insufficient budget for later high-value periods. Sudden bid changes can also destabilise an algorithm already in a learning phase.

6. Frequency management

Frequency is the number of exposures delivered to a user or household within a defined period. Average frequency is:

If 2,000,000 impressions reach 500,000 identified users, average frequency is 4. This average conceals the distribution: some users may see one impression while others see twenty.

Frequency management should therefore examine frequency bands, such as 1, 2–3, 4–6 and 7+. The optimal level depends on objective, creative wear-out, buying cycle and audience. Retargeting often tolerates higher frequency than broad prospecting, but excessive exposure can create irritation, wasted spend and negative brand effects.

Frequency caps are imperfect because identity resolution is incomplete across browsers, devices and authenticated states. A cap should specify the counted event, population, time window and measurement method. Advertisers should also coordinate caps across campaigns to reduce cross-campaign duplication.

7. Bid optimisation

A bid expresses the maximum economic value an advertiser is prepared to exchange for an opportunity, subject to platform mechanics. In a simplified expected-value model, a bid may be informed by:

Here, p(a) is the predicted probability that impression opportunity a produces the defined outcome, v(a) is the value of that outcome, and c is the expected cost. This is a decision aid, not a universal bidding formula. It assumes probability calibration, a correctly defined value, and comparable cost and outcome units.

For a purchase campaign, value may be contribution margin rather than gross revenue. If the advertiser pays on a CPM basis, the expected value must be expressed per impression or converted into a consistent thousand-impression basis. Unit inconsistency is a common source of erroneous bidding logic.

Bidding strategies include:

• Fixed or rule-based bids: transparent and useful when data are limited.
• Goal-based automated bidding: optimises toward CPA, ROAS or value.
• Value-based bidding: assigns different values to outcomes.
• First-price auction adjustment: bids account for the need to pay the clearing price rather than a second price.
• Floor-aware bidding: considers publisher or exchange price floors.

Bid optimisation must monitor win rate, clearing price, spend, conversion quality and marginal return. Increasing bids can raise win rate while lowering efficiency because the campaign begins buying more marginal opportunities.

8. Experimentation and causal measurement

A/B testing changes one planned factor—such as creative, landing page, bid strategy or frequency policy—while maintaining a credible comparison. Random assignment is the foundation of causal interpretation.

For a binary outcome, the observed conversion rate is:

where x is the number of conversions and n is the number of eligible users or opportunities. A difference in observed rates does not prove causality unless assignment, exposure and measurement are appropriately controlled.

A simple difference in rates is:

where T denotes treatment and C denotes control. Relative lift is:

These measures assume a clearly defined denominator and comparable observation windows. Analysts should consider confidence intervals, statistical power, multiple testing, contamination, noncompliance and practical significance.

In programmatic advertising, geo experiments, audience holdouts and public-service announcements may be more credible than platform attribution alone. A holdout group estimates what would have happened without advertising, although imperfect randomisation, spillover and limited sample size remain concerns.

9. Marginal-return reasoning

Optimisation is fundamentally about the next decision. Average performance describes what has already been purchased; marginal performance predicts whether additional investment is worthwhile.

If the incremental contribution margin from additional conversions exceeds incremental spend, expansion may be economically rational. A simplified decision rule is:

This assumes that contribution margin includes variable product and service costs and that the measured increment is causal. If brand effects, retention or long-term value matter, they should be represented explicitly rather than hidden in an arbitrary target CPA.

Diminishing returns are common: the highest-intent users are reached first, and additional spend accesses less responsive users or more expensive inventory. A campaign should not be scaled solely because it meets an average CPA target; it should be scaled while the marginal CPA remains acceptable and strategic constraints are respected.

10. An illustrative case: Northstar Home Insurance

Background

Northstar, a fictional regional insurer, used display and online video to generate quote requests. Its DSP reported a CPA of $38, below the internal target of $45. However, completed policies were not increasing proportionally.

Problem

The marketing team suspected that reported conversions were concentrated among existing website visitors and that frequency was excessive. Spend was distributed across three supply paths and four audience groups.

Analysis

The team reconciled DSP logs, analytics data, CRM records and policy systems. It found:

• retargeting produced 62% of attributed quote requests but only 29% of new policies;
• prospecting had a higher reported CPA but generated more new customers;
• one reseller path had low CPM but poor viewability and high invalid traffic;
• users exposed more than ten times in seven days had declining incremental policy completion;
• the final conversion event had inconsistent mobile tracking.

The team then created cohorts by first exposure week and separated new from existing customers. It used a 10% geographic holdout for prospecting, retained a smaller retargeting holdout, and standardised conversion definitions.

Approach and solution

Northstar applied the following changes:

1. Removed the lowest-quality reseller route after verification.
2. Reduced retargeting frequency to five impressions per user per seven days.
3. Shifted 20% of discretionary budget to prospecting, but only after confirming supply quality.
4. Optimised toward completed policy applications rather than quote starts.
5. Introduced creative rotation and tested two value propositions.
6. Evaluated scale using marginal completed policies, not platform-reported quotes.

The campaign was allowed a learning period before further changes. The team documented each intervention, its expected mechanism and its evaluation window.

Outcome

In the illustrative evaluation, reported quote CPA increased from $38 to $42, but the cost per incremental completed policy fell from $118 to $96. The result demonstrates why a superficially worse platform metric can accompany better business performance. The outcome should not be generalised without considering the test design, seasonality and measurement quality.

Learning points

The case illustrates five principles: define the business outcome; distinguish new demand from captured demand; evaluate marginal rather than average performance; control frequency and supply quality; and change one interpretable set of conditions at a time.

11. Governance, ethics and limitations

Optimisation uses personal, behavioural and inferred data, creating legal and ethical responsibilities. Key controls include consent management, purpose limitation, data minimisation, retention controls, access governance and clear vendor accountability. The applicable legal framework depends on jurisdiction and may include privacy, consumer-protection, online safety and sector-specific regulation.

Important risks include:

• discriminatory exclusion or differential pricing;
• opaque audience classifications;
• excessive retargeting and manipulation;
• use of sensitive data or proxy variables;
• invalid traffic and made-for-advertising inventory;
• brand adjacency and harmful content;
• measurement practices that overclaim causality;
• automated bidding that cannot be adequately audited.

AI-based optimisation can improve prediction and scale, but it may amplify biased historical data, optimise a flawed proxy or make decisions that practitioners cannot explain. Human oversight should include model objectives, training data governance, drift monitoring, incident procedures and the ability to pause delivery. Privacy-enhancing technologies, clean rooms and aggregated measurement may reduce exposure of user-level data, but they do not eliminate governance obligations.

12. Professional optimisation workflow

A robust workflow is:

1. Validate measurement: reconcile spend, impressions, clicks, conversions and revenue.
2. Define the decision: specify whether the objective is scale, efficiency, reach, profit or learning.
3. Diagnose the funnel: identify the largest quality or volume constraint.
4. Segment carefully: compare cohorts, audiences, placements and supply paths with adequate volume.
5. Check marginal economics: estimate the effect of additional or reduced spend.
6. Form a hypothesis: state what will change and why.
7. Test where feasible: use randomisation or a credible holdout.
8. Implement controlled changes: avoid simultaneous unexplained interventions.
9. Monitor guardrails: frequency, viewability, fraud, brand safety, privacy and delivery.
10. Document and learn: preserve decision logs and update assumptions.

This workflow turns optimisation into an organisational learning process rather than a sequence of reactive platform adjustments.

Chapter Summary

Programmatic campaign optimisation requires analysis across business value, funnel stages, audiences, cohorts, placements and supply paths. Metrics such as CTR, CPA and ROAS are useful only when their denominators, attribution windows and business meanings are explicit. Budget decisions should consider marginal CPA and incremental contribution rather than relying solely on averages. Pacing controls when budget is spent; frequency management controls repeated exposure; and bid optimisation allocates auction participation according to predicted value and cost.

A/B tests and holdouts improve causal understanding, while supply-path and quality analysis reduce waste and risk. The strongest practice combines quantitative evidence with governance: privacy, fairness, transparency, invalid-traffic control, brand safety and human oversight. Optimisation is therefore both a technical and managerial discipline.

Key Terms

Funnel: Sequence from exposure to engagement, conversion and business value.

Cohort: Group sharing a defined characteristic or time of entry.

Placement: Specific publisher, app, position, format or content location.

Supply path: Route an impression follows through publishers and intermediaries.

Supply-path optimisation: Evaluation and selection of efficient, transparent inventory routes.

Pacing: Control of expenditure and delivery over time.

Frequency: Number of exposures delivered to a user or defined audience.

Frequency cap: Rule limiting exposures within a stated time window.

Marginal CPA: Additional spend divided by additional conversions.

Incrementality: Outcomes caused by advertising beyond what would otherwise occur.

Holdout: Control group deliberately withheld from advertising.

Attribution: Rule assigning credit for an outcome to advertising contacts.

Viewability: Whether an impression meets a specified opportunity-to-see standard.

Invalid traffic: Non-human or otherwise illegitimate advertising activity.

Value-based bidding: Bidding that uses predicted economic value rather than only event volume.

Learning phase: Period in which an optimisation system gathers data and adjusts predictions.

ROAS: Attributed revenue divided by advertising spend.

CPA: Advertising spend divided by attributed conversions.

Auction duplication: Multiple buying routes offering access to the same impression.

Conversion lag: Time between advertising exposure and conversion.

Review Questions

Short Answer Questions

1. Why can a low CPA fail to indicate strong business performance?
2. Define marginal CPA and distinguish it from average CPA.
3. What is the purpose of a frequency cap?
4. Name two dimensions used in supply-path analysis.
5. Why is random assignment important in an A/B test?

Long Answer Questions

1. Explain how an analyst would diagnose leakage across a programmatic advertising funnel.
2. Discuss the differences between cohort, audience and placement analysis, including appropriate cautions.
3. Explain how pacing and frequency management interact in a large campaign.
4. Evaluate the strengths and limitations of automated value-based bidding.
5. Describe a rigorous workflow for reallocating budget among prospecting and retargeting campaigns.

Analytical / Application Questions

1. A line item increases spend from $30,000 to $35,000 and conversions from 600 to 680. Calculate its marginal CPA and explain whether it should be expanded without further evidence.
2. A retargeting campaign has excellent reported ROAS but produces few new customers. Design a test to estimate its incremental value.
3. Two supply paths have identical CPMs, but one has higher viewability and unique reach while the other has more conversions in a small sample. Explain how you would decide between them.

Multiple Choice Questions with Answers

Question 1

Which metric measures the cost of the additional conversions generated by additional spend?

A. Average CPM

B. Marginal CPA

C. Average frequency

D. Viewability rate

Correct Answer: B. Marginal CPA

Question 2

If spend is $12,000 and conversions are 300, what is CPA?

A. $4

B. $30

C. $40

D. $AI

Correct Answer: C. $40

Question 3

What is the principal purpose of a holdout group?

A. To increase frequency

B. To estimate outcomes without advertising

C. To reduce the clearing price

D. To guarantee higher CTR

Correct Answer: B. To estimate outcomes without advertising

Question 4

Which finding most directly indicates exposure leakage?

A. Low landing-page conversion rate

B. Poor viewability

C. High contribution margin

D. Accurate conversion tracking

Correct Answer: B. Poor viewability

Question 5

Average frequency is calculated as:

A. Spend ÷ conversions

B. Clicks ÷ impressions

C. Impressions ÷ unique reach

D. Revenue ÷ spend

Correct Answer: C. Impressions ÷ unique reach

Question 6

Why can a placement with two conversions from 20 clicks be misleading?

A. It necessarily has invalid traffic

B. Its small sample may be unstable

C. It has no measurable CTR

D. It cannot be viewable

Correct Answer: B. Its small sample may be unstable

Question 7

In the simplified expected-value model, p(a) represents:

A. Predicted probability of the defined outcome

B. Publisher price floor

C. Pacing ratio

D. Number of placements

Correct Answer: A. Predicted probability of the defined outcome

Question 8

A pacing ratio above 1 generally indicates that:

A. Actual cumulative spend exceeds planned cumulative spend

B. No impressions have been won

C. Frequency is zero

D. The campaign has no conversions

Correct Answer: A. Actual cumulative spend exceeds planned cumulative spend

Question 9

Which practice most improves causal interpretation of an A/B test?

A. Changing several variables simultaneously

B. Using random assignment and a comparable control

C. Selecting only successful users

D. Removing the denominator

Correct Answer: B. Using random assignment and a comparable control

Question 10

Why should marginal return be considered when scaling a campaign?

A. Additional spend may reach less responsive or more expensive opportunities

B. Average CPA always rises exactly linearly

C. Frequency becomes irrelevant at scale

D. Attribution becomes automatically causal

Correct Answer: A. Additional spend may reach less responsive or more expensive opportunities