Marketing Measurement Framework: Start With What You Can Observe

Marketing measurement discussions often begin with the methodology rather than the decision.

A company wants to understand where its next marketing dollar should go, and within a few minutes the conversation has moved to whether it needs an MMM, an incrementality programme, better attribution or another reporting layer. The choice of measurement technique becomes the problem to solve before anyone has established what the business actually needs to know or how much evidence would be sufficient to make the decision responsibly.

I think the sequence should be reversed.

The starting point should be the allocation question. What decision is being made? How much money is involved? What information already exists? Which uncertainties actually matter to that decision, and what is the least assumption-heavy form of evidence capable of reducing them?

Sometimes the answer will be an experiment. Sometimes it will be marketing mix modelling. In many cases, however, a business has not yet extracted the value available from evidence it can observe much more directly.

It may already be possible to see whether preference is strengthening relative to competitors, what customers say influenced their choice, where meaningful intent is appearing before revenue arrives, and where media investment is being lost through poor execution. None of those observations provides a complete causal answer, but they can answer a surprising number of practical marketing questions without requiring the business to introduce assumptions that the decision does not yet need.

This is the principle behind the Growth Dynamics measurement framework. We start with observed evidence, move to designed causal evidence when the decision requires a counterfactual, and use modelled evidence when the business needs to understand the wider system.

The progression is deliberate. As the questions become harder, the assumption load increases.

Why starting with the model creates unnecessary complexity

Marketing measurement has become increasingly model-led. Attribution systems attempt to reconstruct customer journeys, experiment platforms promise causal answers, and MMM providers offer a decomposition of the entire media portfolio. Each method has a legitimate role, but problems arise when a business purchases the methodology before defining the question it needs the methodology to answer.

An MMM can be commissioned before anyone has properly understood the underlying demand signals. An incrementality test can be designed before the company has examined whether the treatment itself is being delivered effectively. Attribution can be refined continuously while nobody has systematically investigated why customers chose the brand in the first place.

The result can be technically sophisticated measurement that is poorly matched to the business decision.

This is particularly common because different methods produce very different levels of apparent precision. A descriptive observation might tell us that paid brand clicks increased while branded organic clicks declined. An MMM might return a channel ROI of 2.47 with a response curve and an optimised budget recommendation.

The second answer looks more advanced, but that does not automatically make it more useful. If the immediate question is whether paid brand search is cannibalising organic demand, the simpler evidence may be much closer to the decision.

The relevant standard should therefore be whether the method answers the question reliably enough, not whether it represents the most sophisticated technique available.

That distinction is also useful when dealing with uncertainty. Observed evidence usually carries fewer assumptions but answers narrower questions. Designed experiments introduce a counterfactual and provide stronger causal evidence, but depend on treatment design, measurement windows and transportability. MMM can connect the whole marketing system, but it requires substantially more assumptions about lag, saturation, priors, baseline demand and the relationships between correlated variables.

The progression should reflect the difficulty of the decision.

1. Growth quality: is preference actually increasing?

Before attempting to attribute growth to individual channels, I want to know whether the business appears to be creating more preference in the first place.

Revenue alone is not enough to answer that question. A company can grow because the category is expanding while losing relative preference to competitors. It can maintain revenue temporarily while underlying brand demand weakens. Conversely, a business can strengthen its competitive position before that improvement is fully visible in final sales.

This is where indicators such as Share of Search, branded organic search, direct demand and baseline revenue become useful.

I particularly like Share of Search because it introduces competitive context. Absolute branded search can increase because the entire category is growing. If the company's share of category search is also increasing, the evidence that relative preference is strengthening becomes much more interesting.

Branded organic search also deserves to be separated from paid brand traffic. A rise in paid brand clicks can simply reflect a decision to buy more of the demand that already exists. Organic branded demand is a cleaner observation of people actively looking for the company without the additional decision to pay for the click.

These measures should not be turned into causal claims they cannot support. If Share of Search increases following a television campaign, the movement does not prove that television caused it. Other marketing and business changes may have occurred during the same period.

The purpose of this layer is different. It establishes whether the underlying pool of people willing to seek out the brand appears to be strengthening and whether the business is successfully converting that preference into revenue.

That matters because performance marketing cannot indefinitely compensate for a stagnant demand pool. Increasing spend against the same group of in-market customers eventually drives the business into diminishing returns. Costs rise, platforms compete more aggressively for increasingly marginal opportunities, and attributed conversions can continue to look respectable long after incremental growth has weakened.

Understanding the quality of growth therefore gives the rest of the measurement system an important reference point. Before deciding which channel deserves credit, establish whether preference itself is moving.

2. Buyer evidence: what do customers say drove the decision?

Behavioural tracking is much better at showing how somebody moved through an observable customer journey than explaining why they made the purchase.

A customer may search the brand, visit several pages, return through paid search and eventually convert. Those interactions can be reconstructed relatively accurately, but they do not tell us why the company entered the customer's consideration set, why it was preferred to competitors or what ultimately gave the buyer enough confidence to proceed.

Much of that information exists outside the tracked journey.

The buyer may have known the company for years. They may have received a recommendation, experienced the product previously, visited a physical store or associated the brand with a particular quality that no analytics event can represent.

If the business question concerns customer choice, the first source of evidence should therefore include the customer.

Post-purchase surveys, non-buyer research, sales-call transcripts, reviews and open-text feedback can reveal recurring reasons for choosing or rejecting the company. The useful unit of analysis is not the isolated anecdote but the distribution of themes across enough customers to identify meaningful patterns.

This also requires better questions than a checkout dropdown asking how somebody heard about the brand. Channel self-report can be useful, but it often reduces rich purchase decisions to another attribution table.

I am more interested in what made the company worth considering, which alternatives the buyer seriously evaluated, which attributes mattered during the decision, what nearly prevented the sale and whether the customer already knew the brand before entering the market.

These responses do not provide causal proof. Customers cannot observe what they would have done in a counterfactual world any more than an attribution platform can. They can misremember or simplify the reasons behind complex decisions.

But the evidence answers a question behavioural tracking cannot answer directly.

It also creates better hypotheses for the later measurement layers. If buyers repeatedly describe store availability as important, store-locator behaviour becomes a more meaningful intermediate signal. If trust repeatedly emerges as a purchase driver, the business has context for interpreting brand activity that may look weak in short-term attribution.

The value of buyer evidence is not that it replaces quantitative measurement. It fills an information gap quantitative tracking often cannot see.

3. Traffic quality and mid-funnel behaviour: where is demand moving?

Final revenue can be a very slow signal.

This becomes particularly problematic in high-consideration categories where several weeks or months can pass between initial marketing activity and purchase. If marketing teams wait for mature revenue before evaluating every campaign, they are effectively operating without feedback for large parts of the customer journey.

Platforms solve this by providing faster metrics, but many of those metrics describe activity inside the platform rather than meaningful progress towards a commercial outcome.

Clicks, impressions and generic engagement are easy to produce. The more useful question is whether customers are displaying behaviours associated with genuine purchase intent.

The exact behaviours depend on the business. They might include visiting pricing information, using a store locator, configuring a product, adding something to a basket, beginning a quote or repeatedly returning to commercially important parts of the website.

Before interpreting those signals, however, the traffic itself needs to be trustworthy. Bot activity, poor inventory and low-quality traffic can make downstream metrics look more active without representing more real customers. Traffic cleaning is therefore not a cosmetic analytics exercise. It determines whether the behavioural evidence is worth interpreting at all.

Once that foundation exists, mid-funnel measurement can provide fast and relatively traceable feedback. We can inspect whether one campaign consistently generates more meaningful behaviour than another, whether certain traffic sources produce high volumes with almost no progression, or whether a change in marketing activity corresponds with visible movement deeper in the journey.

This remains descriptive evidence. If customers exposed to a campaign visit pricing more frequently, that association does not prove the campaign caused their eventual purchase.

Its value lies in helping the business understand where demand appears to be moving before the final commercial outcome arrives. That can be enough for many operational decisions and can also tell us which larger questions deserve stronger causal testing.

4. Execution economics: what is happening inside the media investment?

Channel-level reporting hides a large amount of variation.

A paid media channel is made up of campaigns, audiences, bidding decisions, placements, creative, objectives and landing experiences. Treating the channel as a single intervention can therefore hide both strong and weak economics underneath an acceptable average.

This distinction is important because a channel can be incrementally valuable while still containing substantial waste. The reverse is also possible: a channel can appear disappointing because execution quality is poor rather than because the underlying media opportunity has disappeared.

Before making the broader channel decision, I want to understand what was actually delivered.

Campaign and account mechanics can reveal spend being directed towards the wrong objectives, irrelevant queries or overlapping audiences. Placement analysis can expose inventory consuming meaningful budget while producing little commercially useful behaviour. Creative analysis can show whether a small number of assets are responsible for most of the useful response while a weak tail continues receiving investment.

These should not be viewed as a hierarchy. Creative, placement and campaign mechanics interact with one another.

Strong creative shown in poor inventory may never receive useful attention. Good inventory carrying ineffective advertising can mean the business is efficiently distributing something that does not persuade anyone. Both can be compromised by account mechanics that optimise towards the wrong event or direct spend into the wrong audience.

Creative is particularly important because it challenges the idea that a channel is a stable treatment.

A two-year MMM coefficient for Meta, for example, is averaging across a historical portfolio of creative that may have changed continuously during the modelling period. Audiences, placements, objectives and platform behaviour may also have changed.

The coefficient can still be useful, but it describes the economics of the historical portfolio. It should not be interpreted as an intrinsic property of the channel that automatically transfers to a different creative treatment next year.

Execution analysis gives the business a more accurate picture of what the money bought before it moves to the harder question of what the channel caused overall.

5. Incrementality: what changed because of the intervention?

Some allocation decisions eventually require causal evidence.

If a business is considering removing several million pounds from a channel, descriptive patterns may not provide enough confidence. The relevant question becomes what changed because marketing was deliberately increased, reduced or removed.

This is where incrementality testing becomes useful.

Its advantage is that variation is created deliberately rather than inferred entirely from observational history. A well-designed experiment can therefore provide a stronger counterfactual than attribution or descriptive analysis.

The interpretation still requires discipline.

Incrementality is often reduced to whether the revenue would have happened anyway, but the allocation question is usually more complicated. We need to understand the size of the effect, the treatment that was actually delivered, the uncertainty around the estimate, the conditions under which the effect occurred and how far the result can be applied beyond the tested population.

It is also not a yes-or-no property of a channel.

The first portion of spend can be strongly incremental while later spend moves into saturation and produces very little additional value. This is particularly obvious in areas such as brand search, where a core level of paid coverage may protect valuable demand in a competitive auction while additional spend increasingly buys clicks from customers who would have reached the company organically.

The same principle applies more broadly. Finding that a channel creates some incremental revenue does not establish that its current budget is optimal.

Geographic experiments introduce another important distinction between identifying an effect locally and transporting that effect elsewhere. Two regions can have similar historical revenue while differing in brand strength, competition, distribution, saturation or customer mix. A causal effect identified in the test markets is evidence about those markets under the treatment delivered. Applying the same ROI nationally introduces another assumption.

This is why the earlier observational layers remain useful even after an experiment has been run. They help us understand whether treatment delivery varied, whether the test markets resemble the wider business and what mechanisms might explain heterogeneous responses.

Designed evidence is stronger because it creates a counterfactual. It is not assumption-free.

6. MMM: how does the wider system fit together?

Marketing mix modelling sits at the top of this framework because it attempts to answer the broadest question.

A strong MMM can connect several media channels, online and offline activity, business controls and historical outcomes within one economic framework. It can estimate diminishing returns, incorporate lagged effects and help leadership compare large allocation scenarios across the portfolio.

For a sufficiently large advertiser, this can be extremely useful.

The difficulty is that MMM is being asked to estimate effects from observational history, and the strength of the answer depends on the amount of signal contained in that history.

Large channels that vary materially over time and create relatively immediate effects are easier to identify than small channels whose impact is weak relative to normal revenue variation. Long-term effects are harder again because the signal is spread over a longer period during which many other things are changing.

As the signal gets weaker, modelling decisions matter more.

The choice of priors, lag structures, controls, baseline specification and saturation functions can increasingly influence the resulting decomposition. The model can represent a six-month advertising effect mathematically, but its ability to represent that effect does not establish that the historical data uniquely identified it.

This is why I would not want an MMM to be the first serious measurement exercise undertaken by the business.

If the company already understands how preference is changing, what buyers say drove their choice, where meaningful behavioural intent is appearing, how the media treatment has changed and what major experiments have found, the model operates in a much richer evidence environment.

Those earlier layers can inform assumptions and challenge conclusions.

If the MMM says search is responsible for a large share of growth while branded demand has been increasing strongly and customers report knowing the company long before they searched, the decomposition deserves scrutiny. If a channel's estimated historical ROI remains strong while its creative treatment has changed radically, the forecast should reflect that change. If an independent experiment produces a substantially different estimate, the disagreement needs to be understood rather than forced away through calibration.

MMM becomes more valuable when it is part of the measurement system rather than the measurement system itself.

Matching the evidence to the decision

The framework is deliberately not a funnel in which every business must complete six steps before making any decision.

The appropriate stopping point depends on the question.

If a programmatic placement is consuming substantial spend while producing obviously poor-quality traffic, the business may have enough evidence to act without commissioning an experiment.

If customer research consistently shows that buyers do not understand the product proposition, there is little reason to wait for an MMM before changing the messaging.

If leadership is deciding whether to remove £10 million from television, the evidence standard should be substantially higher. Descriptive movement in Share of Search or mid-funnel behaviour can inform the decision, but a well-designed experiment or credible MMM may be justified.

The size and irreversibility of the allocation should determine how much uncertainty the business is willing to accept.

This approach also prevents methods being used outside their useful boundary. Attribution remains valuable for understanding observable journeys without being treated as proof of causality. Qualitative buyer evidence informs customer choice without being converted into a channel ROI. Mid-funnel evidence provides faster behavioural feedback without pretending that every association is incremental. Experiments provide local causal evidence without automatically becoming national truths. MMM provides a strategic system view without being asked to measure every small execution change.

The point is not to minimise methodological sophistication. It is to introduce sophistication when it earns its place in the decision.

A measurement system should make uncertainty more visible, not less

One of the reasons companies invest in advanced measurement is the desire for certainty.

Unfortunately, the hardest marketing allocation questions are often precisely the questions where certainty is least available.

Small channel effects can disappear into normal business noise. Long-term brand effects require assumptions about what happened between media investment and later revenue. Experiments can identify effects that do not transport cleanly to the rest of the market. Historical channel coefficients describe treatments that may no longer exist because the creative, audience and platform have changed.

A useful measurement system should not hide these limitations behind a single ROI.

It should show where the evidence is strong, where the conclusion depends heavily on assumptions and where additional learning would materially change the decision.

That is what allows leadership to distinguish between an allocation that can be made confidently today and one that should first be tested.

The objective is therefore not to eliminate uncertainty. It is to make the relationship between evidence, assumptions and the size of the decision explicit.

That is a much better foundation for capital allocation than asking one methodology to produce a definitive answer to every marketing question.

Start with what you can observe

The principle behind the framework is simple.

Use observable evidence for the questions it can answer. Introduce designed causal evidence when the decision requires a counterfactual. Use modelling when the business needs to connect the wider system and the underlying data contains enough information to support the inference.

This creates a chain of evidence rather than a hierarchy of fashionable methodologies.

Growth quality tells us whether preference appears to be strengthening. Buyer evidence helps explain customer choice. Mid-funnel behaviour shows where meaningful intent is moving before revenue matures. Execution economics reveals what the media investment actually contains. Incrementality tests important causal hypotheses. MMM helps connect those findings across the wider portfolio.

Each layer adds information, but it also introduces different assumptions and limitations. Understanding both is what allows the business to decide how far it needs to go.

Growth Dynamics uses this framework to produce an evidenced allocation decision rather than another measurement output. Depending on the question, the answer may be to invest, divest or hold, supported by an explicit view of the evidence, confidence and assumptions behind the recommendation.

The purpose of marketing measurement is not to prove that we have measured everything.

It is to understand enough to make the next allocation decision better.