How to Build a Marketing Optimisation Operating Model

Marketing optimisation is the organised process of changing marketing decisions in response to evidence. Its purpose is not to generate more reports, launch more tests or make every metric rise. Its purpose is to improve the organisation’s choices about programmes, channels, offers, customer outcomes and resources.

That makes optimisation different from strategy, planning and execution:

• Strategy chooses the customers, markets and sources of advantage the organisation will pursue.
• Planning converts those choices into budgets, programmes, capacities and commitments.
• Execution delivers the planned work.
• Optimisation reviews evidence and changes a decision that remains changeable.

A strategy can be sound while a programme underperforms. A campaign can be executed correctly while its underlying assumption is wrong. A channel can report efficient attributed returns while capturing demand that another activity created. Optimisation exists to detect those differences and decide what should happen next.

The operating model below treats marketing as a portfolio of decisions rather than a list of channels. It applies to acquisition, retention, pricing and offers, customer experience, conversion, operating efficiency and measurement. It does not replace a marketing measurement framework, customer-journey mapping or controlled experimentation. It uses those disciplines when the decision requires them.

The optimisation unit is a decision

“Improve paid social” is not an optimisation brief. It does not identify what can change, whose outcome matters or how a result would alter resource allocation.

A usable optimisation unit names a decision:

Should we reduce prospecting spend in audience A, maintain it in audience B and move the released budget into retention activity C for the next planning period?

That statement exposes the choices, affected populations, opportunity cost and next action. It can be connected to evidence.

For every initiative, record the following:

• Decision. The choice that can change.
• Current state. What is happening now.
• Proposed change. The intervention being considered.
• Customer affected. The eligible or exposed population.
• Expected mechanism. Why the change might affect an outcome.
• Metric. The primary outcome used for the decision.
• Guardrail. An outcome that must not deteriorate beyond an agreed level.
• Evidence quality. What is known, how it was learned and how uncertain it remains.
• Owner. The person with authority to make the next decision.
• Cost. Cash, labour, opportunity and transition cost.
• Reversibility. How easily the change can be undone.
• Evaluation design. The method used to judge the change.
• Stop condition. The condition requiring pause or termination.
• Next decision. Continue, revise, scale, stop or reallocate.

The record prevents a common failure: a team evaluates an activity, produces an interesting result and discovers that nobody agreed what decision the result was meant to change.

The 16-step marketing optimisation cycle

1. Define the organisation’s objectives and constraints

Begin with outcomes the organisation values and limits it must respect. Objectives may include incremental contribution, sustainable acquisition, retained revenue, customer access, service quality or reduced operating cost. Constraints may include cash, capacity, inventory, regulation, consent, brand commitments, accessibility, technical dependencies and risk tolerance.

Do not convert every objective into a single financial number. Commercial, customer and operational outcomes can conflict, and some constraints are non-negotiable. HM Treasury’s appraisal guidance similarly separates the objectives being pursued from the costs, benefits and risks of the available options, and states that it does not itself set objectives or make the decision. (1)

2. Translate objectives into decisions

Ask what someone could do differently. “Increase retention” is an objective. Decisions include whether to change an onboarding sequence, reduce contact frequency, revise an offer, fund service improvements or alter acquisition mix because a channel attracts customers with poor later retention.

A decision should identify its owner, time horizon and feasible alternatives. A metric without an attached decision is a monitoring signal, not an optimisation system.

3. Inventory current programmes and interventions

Create a portfolio view of active work. Include not only campaigns but also offers, lifecycle programmes, audience rules, channel budgets, customer-experience changes, conversion work, operational processes and measurement improvements.

Record the commitment attached to each item: spend, labour, contractual lock-in, customer exposure and dependencies. This exposes work that persists because it is habitual rather than because it remains the best use of resources.

4. Establish baselines

A baseline describes the current state before a change. It should include the outcome level, eligible population, time window, costs, operating effort, known seasonality, data limitations and material external changes.

Baselines are not automatically counterfactuals. Last month’s performance does not prove what would have happened this month without an intervention. Advertising returns can be difficult to estimate precisely even in very large field experiments, because individual purchase outcomes vary so much and observational comparisons carry selection bias, so uncertainty should be reported rather than hidden behind a point estimate. (2)

I spent years watching brands celebrate their app numbers. Conversion, order value, repeat rate, all comfortably ahead of the website, every time. It is the wrong comparison. People buy from more than thirty brands a year and cannot keep thirty apps on a phone, so the ones who installed yours were already loyal before they downloaded anything. The app did not create that behaviour, it collected it. That is a commercial observation rather than a legal one, but it changed how I read every before-and-after chart I was handed afterwards.

5. Identify performance, customer and operational problems

Search across three dimensions:

• Performance: contribution, acquisition quality, retention, reach, demand, conversion or another outcome is below an agreed expectation.
• Customer: people face unnecessary effort, exclusion, confusion, poor service, inappropriate targeting or harm.
• Operational: cycle time, manual work, error, data latency, fragmented ownership or technical failure prevents effective delivery.

A high-performing programme can still create a customer or operational problem. A low-performing programme may be correctly executed but aimed at an unproductive decision. Keep these diagnoses separate.

6. Distinguish diagnosis from proposed solution

“Customers need more reminders” is a solution disguised as a diagnosis. The observed problem may be that eligible customers do not complete renewal, but possible causes include lack of value, technical failure, unclear terms, timing, affordability or a deliberate decision not to renew.

Write the problem without the proposed fix. For a defined digital task, use the separate guide to conversion diagnosis and hypothesis development. Journey-wide tasks and service failures belong in customer-journey mapping.

7. Estimate opportunity and uncertainty

Estimate the plausible value of resolving the problem, not the maximum imaginable upside. Consider:

• affected population and frequency;
• contribution rather than revenue alone;
• customer benefit or avoided harm;
• operating cost and capacity released;
• implementation and transition cost;
• confidence that the problem is real;
• confidence in the proposed mechanism;
• time until evidence will be available;
• what must be displaced to do the work.

Use ranges where uncertainty is material. Appraisal guidance treats optimism bias and uncertainty as matters to be stated and adjusted for rather than concealed inside a single figure. (1) The objective is not false precision; it is to compare opportunities honestly.

8. Prioritise the optimisation portfolio

The portfolio should distinguish at least the following areas, each with its own characteristic decision:

• Acquisition optimisation. Which audiences, propositions or sources should receive the next unit of spend?
• Retention optimisation. Which customer problem or lifecycle intervention justifies investment?
• Pricing or offer optimisation. Which price, package, discount or eligibility rule should change?
• Channel allocation. Where should resources move across demand creation, capture and service channels?
• Customer-experience improvement. Which experience failure should be resolved first?
• Conversion improvement. Which defined digital task warrants diagnosis and testing?
• Operational-efficiency improvement. Which process change reduces cost or delay without degrading outcomes?
• Measurement improvement. Which uncertainty is important enough to justify better data or evaluation?

These areas conflict. A conversion increase may reduce contribution. Lower reported CAC may come from capturing people who were already likely to buy, while later demand weakens. More contact can lift immediate sales and increase opt-outs. A faster application can create more inappropriate completions. A channel can appear efficient because it intercepts existing demand rather than creating it. Large field experiments at one online marketplace found no measurable short-term return from brand-keyword search advertising and negative average returns on the non-brand campaigns tested, alongside positive effects concentrated among new and infrequent users, which is a direct illustration of how observational attribution can overstate causal value for some populations. (3)

The campaigns clients asked me for most were rarely the ones the numbers argued for. They wanted the tab title that changes when you switch away from the page, the social proof badge, the personalised homepage slider. Those things present beautifully in an internal review. One tab-title campaign came back with a conversion uplift I was formally entitled to report and never once believed. The slider work was not broken, it simply consumed months of two teams’ time to land somewhere around an ordinary campaign. My point is commercial rather than legal: internal marketability is a real force in prioritisation, and it is not the same force as evidence.

A scoring framework can make judgement visible, but it cannot make the judgement objective. Record assumptions and allow decision owners to override a score only with a documented reason.

9. Assign decision owners

The owner is not necessarily the analyst, channel manager or person building the change. The owner must have authority to choose among the defined alternatives and accept the consequences.

Define who:

• proposes the change;
• owns the commercial outcome;
• represents customer and legal constraints;
• approves technical deployment;
• interprets the evidence;
• makes the final continue, revise, scale or stop decision.

Shared input is useful. Shared accountability without a final decision right is usually delay.

10. Choose an evaluation method

Match the method to the question, risk and feasible evidence. The current Magenta Book describes evaluation as something to be designed proportionately around the intervention, the questions being asked and the intended use of the findings, rather than as a single prescribed method. (4)

Possible methods include:

• descriptive monitoring for known, reversible operational adjustments;
• usability or qualitative research for comprehension and experience questions;
• controlled experiments for the effect of a defined change on an eligible population;
• incrementality tests for whether marketing activity caused additional outcomes;
• marketing-mix modelling for aggregate allocation questions, conditional on its assumptions;
• quasi-experimental or time-series designs where randomisation is infeasible;
• process evaluation where the immediate question is whether delivery occurred as intended.

A French beauty retail chain I worked with sent its loyal segment a discount thirty days after every purchase. For five months the reporting was excellent: opens, clicks and conversions all healthy. Then we split the audience and held a group back with no discount at all. Opens held. Clicks were lower. Conversions differed by about two per cent. The programme had been paying margin to people who were going to buy anyway, and no amount of dashboard health would have revealed that without the control group. That is a commercial observation, and I should disclose that I now run a business in the rewards and loyalty space.

The marketing measurement framework should define the evidence architecture. The incrementality guide owns detailed causal measurement of marketing activity. The CRO experimentation guide owns randomisation, sample size, inference, stopping and scaling for digital experiments.

11. Implement a controlled change

“Controlled” does not always mean randomised. It means the organisation knows what changed, when, for whom, by whose authority and with which protections.

Prefer changes that are observable and reversible. Version the intervention. Preserve the previous state. Record concurrent activity that could alter interpretation. Use staged release where technical or customer risk is material.

Online experimentation research has long reported that plausible changes can produce surprising or adverse results, and that statistical power, sample size, randomisation and instrumentation all have to be addressed before a result can be trusted, which is why deployment and evaluation should be designed before the change is exposed broadly. (5)

12. Monitor commercial, customer and operational outcomes

The primary outcome should answer the decision. Guardrails should detect unacceptable trade-offs.

Track as appropriate:

• incremental contribution;
• customer outcome;
• cost;
• operating effort;
• time to decision;
• uncertainty reduced;
• action taken;
• guardrail effects;
• resources reallocated;
• learning reused;
• stopped work.

Do not use the number of initiatives, experiments or dashboard views as the definition of success. Those are activity measures.

Metric interpretation is itself a source of risk. Changes in telemetry, exposure, denominators, user mix or logging can move a metric without the underlying customer outcome changing. Industry experimentation research catalogues recurring interpretation errors of exactly this kind and recommends diagnostic and guardrail measures because a single headline metric can be read wrongly. (6)

13. Continue, revise, scale or stop

Agree these decision states before reviewing the result:

• Continue: maintain the current change within its tested or evaluated scope.
• Revise: retain the problem but alter the intervention or mechanism.
• Scale: expand exposure after checking capacity, external validity and guardrails.
• Stop: end the change or the initiative because it is harmful, uneconomic, unsupported or no longer important.

An inconclusive result is not permission to declare success. It may imply that the effect is too small to distinguish, the measurement is weak, implementation failed or the question is not worth further cost.

14. Record what was learned

Store more than “winner” and “loser”. Record:

• the original decision and alternatives;
• evidence available before the change;
• the expected mechanism;
• implementation fidelity;
• observed outcomes and uncertainty;
• guardrail changes;
• scope and exclusions;
• what the result does not establish;
• the next decision.

Learning is reusable only when another team can understand the context in which it was produced.

15. Reallocate resources

Optimisation is incomplete until a decision changes. Released money, capacity or attention should have a named destination. Otherwise stopped work quietly returns or the saving disappears into an unexamined budget.

Reallocation should consider the marginal opportunity, not only historical average performance. It should also preserve a balance between exploiting established activity and exploring uncertain opportunities. Formal models of organisational learning show how a system can become efficient in the short run by exploiting what it already knows while undermining its own long-run adaptation if exploration is crowded out. (7)

For acquisition economics, use the appropriate CAC definition rather than a single blended figure. The CAC methodology guide distinguishes operational, fully loaded, incremental and attributed views. For the economic balance between new and existing customers, use the acquisition-versus-retention framework.

16. Review the portfolio

Portfolio review asks whether the organisation is working on the right decisions, not whether every initiative is on schedule.

Use multiple cadences:

• Operational review: active incidents, guardrails, delivery and reversible adjustments.
• Decision review: evidence sufficient for continue, revise, scale or stop.
• Portfolio review: opportunity cost, balance across objectives and resource movement.
• Strategic review: whether objectives, constraints or the portfolio structure itself have changed.

The cadence should reflect risk and decision speed. A live technical failure cannot wait for a quarterly meeting. A long-term brand or retention decision should not be judged by a few days of response data.

Governance and incentives

A workable optimisation system needs rules that survive pressure for a positive result.

Decision rights

Publish who can launch, pause, stop and scale a change. Give customer, legal, accessibility, security and operational owners explicit guardrail authority where appropriate.

Stopping rules

Stop conditions can include customer harm, legal non-compliance, accessibility regression, technical failure, contribution deterioration, capacity breach or evidence that the problem is too small to justify further work.

Documentation

Require a decision record before implementation and a learning record after evaluation. The burden should be proportionate, but material changes should not depend on disappearing slide decks or individual memory.

Incentives

Do not reward teams for test win rate, number of initiatives or budget retained. Those incentives encourage weak hypotheses, selective reporting and resistance to stopping work.

Reward:

• important uncertainties resolved;
• decisions changed by evidence;
• customer and commercial guardrails protected;
• low-value work stopped;
• resources moved to stronger opportunities;
• learning reused by another team;
• errors found before broad exposure.

Ethical and legal constraints

Ethical practice is not an optimisation tactic whose value depends on a return claim. Consent, fairness, transparency, accessibility and the avoidance of manipulation belong in the operating model as obligations and constraints, and several of them are legal duties rather than preferences.

In the United Kingdom, the Digital Markets, Competition and Consumers Act 2024 prohibits misleading actions, misleading omissions, aggressive practices and contraventions of professional diligence where they are likely to cause the average consumer to take a different transactional decision, and treats a defined list of practices as unfair in all circumstances, including falsely stating that a product or particular terms are available only for a limited time in order to force an immediate decision. (8) Where a trader could reasonably foresee that a particular group is especially vulnerable to a practice, the assessment is made against the average member of that group rather than the general average consumer. (8) The Competition and Markets Authority has published guidance explaining how it reads those provisions, while stating that the guidance illustrates how the law may apply and is not a definitive interpretation of it. (9)

None of that prescribes an optimisation method, a research technique or a target outcome. It constrains the means by which a result is obtained. A change that lifts a headline metric by hiding a cost, manufacturing urgency or obstructing a decision is not a successful optimisation with an ethical caveat attached. It is a change the organisation may not make. Use the ethical digital-marketing checklist for campaign-level review.

Technology should follow the decision and evidence requirements. Use the martech stack audit when the problem concerns duplicated capability, cost, integration or ownership. Do not purchase an optimisation platform before defining the decision it must improve.

What a mature optimisation system produces

A mature system does not promise continuous upward movement. Markets change, measurements fail and interventions have trade-offs. Its advantage is more disciplined adaptation.

At any point, leaders should be able to see:

• which decisions are open;
• what evidence each requires;
• which customer and commercial constraints apply;
• where resources are committed;
• what has been continued, revised, scaled or stopped;
• which uncertainties were reduced;
• which learning has been reused;
• where resources moved as a result.

That is marketing optimisation: not permanent activity, but a governed process for changing decisions when the evidence justifies change.

Frequently asked questions

References

1. HM Treasury. The Green Book (2026). UK central government appraisal guidance. Reference PU3608; ISBN 978-1-918417-12-8; 2026 edition. GOV.UK publication record first published 18 April 2013; current edition page updated 5 February 2026. Not marked draft, under review or superseded; the parent publication page records that previous editions are withdrawn. Not UK law and not regulator guidance: it is central government appraisal guidance, mandatory only within the governmental circumstances the document describes. https://www.gov.uk/government/publications/the-green-book-appraisal-and-evaluation-in-central-government/the-green-book-2026

2. Randall A. Lewis and Justin M. Rao. The Unfavorable Economics of Measuring the Returns to Advertising. Peer-reviewed research article, The Quarterly Journal of Economics, volume 130, issue 4, pages 1941 to 1973. DOI 10.1093/qje/qjv023. Published 6 July 2015; issue dated November 2015. No last-updated date shown and no correction, erratum or retraction notice displayed on the publisher record. Interested-party disclosure: the publicly accessible publisher record displays no author affiliation or funding statement, and the subscriber-only published PDF and its acknowledgements were not independently verified, so no disclosure is asserted here either way. Not UK law. https://academic.oup.com/qje/article-abstract/130/4/1941/1914592

3. Thomas Blake, Chris Nosko and Steven Tadelis. Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment. Peer-reviewed research article, Econometrica, volume 83, issue 1, pages 155 to 174. DOI 10.3982/ECTA12423. Issue dated January 2015; first published online 18 February 2015. No correction, erratum or retraction notice displayed. Interested-party disclosure: all three authors are listed with an eBay Research Labs affiliation and the experiments were conducted at that company. Not UK law. https://onlinelibrary.wiley.com/doi/abs/10.3982/ECTA12423

4. HM Treasury and the Evaluation Task Force. Magenta Book: Central Government guidance on evaluation (HTML). Non-statutory UK central government evaluation guidance. No reference or version number printed on the document. Document dated May 2026; GOV.UK publication record first published 27 April 2011; HTML version added and page updated 15 May 2026. Not marked draft, under review or superseded. Not UK law and not regulator guidance. https://www.gov.uk/government/publications/the-magenta-book/magenta-book-central-government-guidance-on-evaluation-html

5. Ron Kohavi, Roger Longbotham, Dan Sommerfield and Randal M. Henne. Controlled experiments on the web: survey and practical guide. Peer-reviewed review article and practical guide, Data Mining and Knowledge Discovery, volume 18, issue 1, pages 140 to 181. DOI 10.1007/s10618-008-0114-1. Published online 30 July 2008; issue dated February 2009. No correction, erratum or retraction notice displayed. Interested-party disclosure: all four authors are listed with a Microsoft affiliation and the article draws substantially on that company’s experimentation practice. Not UK law. https://link.springer.com/article/10.1007/s10618-008-0114-1

6. Pavel Dmitriev, Somit Gupta, Dong Woo Kim and Garnet Vaz. A Dirty Dozen: Twelve Common Metric Interpretation Pitfalls in Online Controlled Experiments. Peer-reviewed conference paper, KDD ‘17: Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 1427 to 1436. DOI 10.1145/3097983.3098024. Published 13 August 2017. No correction, erratum or retraction notice displayed in the retrievable publisher record. Interested-party disclosure: the authors are listed with a Microsoft Corporation affiliation and the paper reports lessons from that company’s experimentation platform. Not UK law. https://dl.acm.org/doi/10.1145/3097983.3098024

7. James G. March. Exploration and Exploitation in Organizational Learning. Peer-reviewed research article and formal modelling study, Organization Science, volume 2, issue 1, pages 71 to 87. DOI 10.1287/orsc.2.1.71. Published online 1 February 1991; issue dated February 1991. No correction, erratum or retraction notice displayed. No interested-party affiliation or funding statement identified on the publisher record. Not UK law. https://pubsonline.informs.org/doi/10.1287/orsc.2.1.71

8. Parliament of the United Kingdom. Digital Markets, Competition and Consumers Act 2024. Primary legislation, 2024 chapter 13. Royal Assent 24 May 2024. Sections 225 to 230 and 245 to 247, and Schedule 20, came into force on 6 April 2025 under regulation 2 of the Digital Markets, Competition and Consumers Act 2024 (Commencement No. 2) Regulations 2025, SI 2025/272. The consolidated text states that it is up to date with changes known to be in force on or before 28 July 2026. Section 260, concerning subscription cancellation, is not yet commenced and is not relied on here. UK law. https://www.legislation.gov.uk/ukpga/2024/13

9. Competition and Markets Authority. Unfair commercial practices. Non-statutory UK regulator guidance, reference CMA207. Published 4 April 2025; last updated 18 November 2025. Final; the preceding consultation draft of December 2024 is superseded. The guidance states that it illustrates how the law may apply, that its examples are not exhaustive, and that it is not a substitute for or a definitive interpretation of the law. UK regulator guidance, not UK law. https://www.gov.uk/government/publications/unfair-commercial-practices-cma207

Position as at 29 July 2026. Government appraisal and evaluation guidance, regulator guidance and the commencement position of individual statutory provisions all change, and several of the sources above are revised periodically or carry errata. Check the issuing body’s current page before relying on any normative statement in this article.