Beyond ROAS: When Marketing Mix Modelling Makes Sense for Australian Businesses

ROAS Answers a Narrow Question

Platform ROAS divides attributed conversion value by spend. It is useful for operating campaigns inside a platform, but it does not reveal the full incremental effect of marketing.

The same sale may be claimed by several systems. Brand search may capture demand created elsewhere. Offline media, promotions, distribution, seasonality and competitor activity can change sales without leaving a click trail.

Marketing mix modelling uses aggregate time-series data to estimate relationships between marketing inputs and a business outcome. It can help answer a different question: how should the next budget be distributed under uncertainty?

When MMM Is Worth Considering

MMM becomes more credible and commercially useful when the organisation has:

• consistent weekly or daily outcome data over a meaningful history
• material spend across several channels
• enough changes in spend to distinguish patterns
• non-digital or privacy-constrained activity that click attribution misses
• recurring budget-allocation decisions
• people who can maintain and interpret the model

There is no universal minimum spend or row count. Suitability depends on variation, signal strength, granularity and the decision being made.

When It Is Probably the Wrong First Project

Delay MMM when:

• conversions are very sparse
• tracking definitions changed repeatedly
• one channel represents almost all spend
• budgets barely varied
• finance and marketing totals do not reconcile
• the business wants a model to produce certainty it cannot support

For a small service business generating a handful of leads each month, fixing CRM outcomes and running targeted experiments is usually more valuable.

Define the Decision First

Do not begin with “build an MMM”. Begin with a decision:

• How should the next quarterly budget be split?
• What is the likely effect of reducing paid social by 20%?
• How much demand appears to persist after a campaign ends?
• Are upper-funnel channels contributing beyond tracked clicks?

The decision determines the required time grain, geography, channels and outcome.

Assemble the Data

A practical dataset may include:

Outcome: orders, revenue, qualified leads or contribution margin.

Media: spend, impressions or reach by channel and period.

Commercial controls: price, promotion, stock, distribution and sales capacity.

External controls: holidays, seasonality, macroeconomic indicators and major events where relevant.

Use Australian calendar and market effects appropriate to the business. State boundaries or media-market splits can be useful only if the underlying data is reliable at that level.

Create a data dictionary with source, owner, definition, timezone, currency, tax treatment and known breaks.

Reconcile Before Modelling

Compare media totals with invoices and platform finance views. Compare the outcome with finance or CRM records. Investigate missing periods, duplicated imports and sudden definition changes.

Document:

• GA4 or CRM migrations
• consent or tracking changes
• store openings and closures
• product launches
• major price changes
• COVID-era or other structural breaks

A sophisticated model cannot recover information that was never measured consistently.

What Meridian Adds

Google's Meridian is an open-source Bayesian MMM framework. Google has also announced tools such as Scenario Planner and integration-related workflows with Google Analytics 360.

These developments lower some implementation friction and make scenario exploration more accessible. They do not remove the need to choose priors, transformations, controls, validation and sensible aggregation.

Treat vendor-provided data and priors as inputs with assumptions, not privileged truth.

Read Ranges, Not Point Estimates

MMM outputs are uncertain. A channel contribution estimate should be presented with credible intervals or ranges, diagnostics and sensitivity analysis.

A planning discussion should sound like:

“Under these assumptions, reallocating $100,000 from channel A to channel B is estimated to improve the outcome within this range, but the result is sensitive to the treatment of the promotion period.”

It should not sound like:

“The model proved channel B will return exactly $4.73 for every dollar.”

Combine MMM with Experiments

Use geo tests, holdouts, incrementality tests or controlled budget changes where practical. Experiments can inform model assumptions and test whether a recommended reallocation behaves as expected.

The strongest measurement system triangulates:

• MMM for aggregate allocation
• experiments for causal questions
• platform data for campaign operation
• GA4 for identifiable site journeys
• CRM and finance for business outcomes

No single method covers every question.

A Readiness Checklist

Answer yes or no:

• We have at least a meaningful continuous history for the planned time grain.
• Marketing, CRM and finance definitions are documented.
• Spend varied enough to produce information.
• We can model a stable outcome tied to the business.
• The decision is large enough to justify the work.
• We have an owner for refreshes and interpretation.
• Leadership accepts ranges rather than guaranteed answers.
• We can run at least some validation or experiments.

Several “no” answers indicate that data foundations should come first.

Official Sources

• Google: Meridian and Google Analytics 360
• Google: Meridian Scenario Planner
• Google: Meridian open-source documentation