An acquisition cost should be explainable

At lePERMISLIBRE, one of the questions I worked on sounded simple: how much does it cost to acquire a learner, using their first hour of driving as the reference event?

To answer it, we linked each learner to the touchpoints observed along their journey, assigned a cost to those interactions, then compared several attribution models.

This work illustrates a principle I apply to my data projects: an acquisition cost should be explainable. Starting with the number, we should be able to trace it back to the spend, interactions and calculation rules that produced it. That matters when the number changes, or when two methods produce different results.

Defining what counts as an acquisition

We chose the first hour of driving as the acquisition event. This gives the metric a specific meaning: account creation, a first purchase and a first hour of driving are three distinct steps in the journey, and each can have its own acquisition cost. Until we know which step is being used, we cannot interpret the number.

The scope of the spend needs the same care. Here, I am discussing advertising campaign costs and how they are allocated. A metric that also included sales costs, external services or salaries would measure something else.

These choices need to remain visible to the people using the number. They determine which comparisons are possible and which questions the metric can answer.

Assigning a cost to each touchpoint

We had two ways to assign a monetary value to the interactions.

The first assigns each click the campaign’s average CPC for that day. Average CPC is the total cost of clicks divided by their number, as counted by the advertising platform. It is an average, so it can differ from the price paid for any particular click.

The second divides the campaign’s spend for the day by the number of clicks observed in Google Analytics.

Only the denominator changes, but that is enough to change how spend is allocated across the observed interactions.

Two calculations for the same campaign

Consider a simplified, fictional example. These are not lePERMISLIBRE’s figures. For one campaign, on one day:

• €1,000 in spend;
• 500 clicks counted by the advertising platform;
• 400 clicks observed in Google Analytics.

We assume that the two counts cover a comparable scope, and that the 400 clicks include all the observations used in the calculation, whether or not they lead to an acquisition.

With the first method, each of the 400 observed clicks is assigned €2, for a total of €800. The remaining €200 is not linked to any observation.

With the second, the full €1,000 is allocated across the 400 clicks, at €2.50 per click.

The difference between €2 and €2.50 comes entirely from the allocation rule. Both are assigned average values, and neither method gives the actual price of each click. The second simply allocates all the spend to the observations available.

That distinction needs to be visible in the results. A team using an amount needs to know how it was calculated and what it represents.

Understanding why an amount changes

This choice has a very concrete consequence. Keep the €1,000 in spend and the 500 clicks counted by the platform, but suppose Google Analytics now observes only 250 clicks. With the second method, the amount assigned to each click rises to €4.

Neither the spend nor the platform’s click count has changed. The increase comes entirely from the number of observations across which the spend is allocated.

That is what I expect from an explainable metric: it should let us identify what changed before drawing any conclusions about the campaign.

The difference between the two click counts then needs its own diagnosis. The numbers alone do not reveal its cause. The calculation rule does, however, immediately explain how that difference affects the amounts. Helping teams ask the right question when a result changes is, for me, part of the data work itself.

Separating interaction costs from attribution

The project also compared several attribution models, which form a separate part of the reasoning.

Assigning a cost to a touchpoint means assigning it a share of the spend. An attribution model determines how credit for a conversion is distributed across the interactions in the journey.

We can therefore keep the same journeys and costs, change only the attribution rule, and see the number of conversions credited to each campaign change. The point is to observe what each rule brings out, with explicit assumptions. When a conclusion depends heavily on the model chosen, that dependence needs to be visible in the discussion.

A third question remains: how many acquisitions did a campaign actually cause? That is the question of incrementality. It requires a method capable of measuring a causal effect, such as an experiment. Distributing credit across touchpoints is not enough to establish it.

Including spend that does not lead to an acquisition

One final distinction: the cost of a learner’s journey and the overall acquisition cost are two different measures.

Adding up the cost of the interactions observed before their first hour of driving gives us the cost of their journey, under the conventions chosen. An overall metric also needs to include spend on journeys that do not lead to an acquisition. Otherwise, it leaves out part of what had to be spent to acquire the learners who did convert.

Moving from an individual journey to an aggregate metric therefore requires us to define the spend included, the population studied and the periods compared. We also need to account for the time that can pass between a touchpoint and the first hour of driving.

What I want a data project to make possible

On a topic like this, I want a team to be able to start with a metric and find everything needed to interpret it:

• the event that defines an acquisition;
• the touchpoints actually observed;
• the spend included and the rule used to allocate it;
• the model used to attribute conversions;
• the amounts and journeys left out of the calculation.

The level of detail needs to suit each person. Someone deciding how to allocate a budget needs to understand what the number covers. Someone maintaining the system needs to be able to find the rules and data that produced it.

That is the connection I try to build between business needs and engineering: making it possible to examine a change, discuss an assumption and revise a rule with a clear understanding of the consequences.

Before making a decision based on an acquisition cost, I want to be able to explain it.

If you are looking to connect your marketing spend, customer journeys and business events, write to me about the number your team currently finds difficult to interpret.