What is an attribution model, and which one should a service business use?

A model is a bookkeeping rule, not a measurement of cause

Every attribution model answers the same narrow question: a sale happened, several marketing touches preceded it, how should the credit be divided? Nothing in that question establishes that any touch caused the sale. Change the model and revenue moves between channels while the actual business results stay exactly the same.

That distinction matters because people spend real money on model output. Switching from last touch to first touch can double the apparent value of your top-of-funnel channels overnight. No customer behaved differently. Only the arithmetic changed. Keeping that clear is the first job of an honest marketing intelligence setup.

The models you will actually be offered

• Last touch. All credit to the final identified click or call. Simple and stable, but biased toward branded search and remarketing.
• Last non-direct touch. The same rule, but it skips sessions labeled direct, which are usually mislabeled anyway. A better default than plain last touch.
• First touch. All credit to the earliest interaction you can see. Useful for judging discovery, useless for judging closing efficiency.
• Linear, time decay, position based. Credit split across touches by a fixed formula. Different faces, same arbitrariness: the weights are chosen, not measured.
• Data-driven. The platform fits weights from observed conversion paths. Better logic, but it only sees what happens on its own surfaces.

Why short cycles let service businesses stay simple

A burst pipe, a dead air conditioner and a garage door off its track share a property: the research phase runs minutes to days, not months. Most journeys carry one or two identified touches, so the model has almost nothing to divide. When the median path length is one, last touch and linear produce nearly identical reports.

The exception is planned, high-ticket work such as system replacements and remodels. Those journeys run for weeks and genuinely cross several channels. If a meaningful share of your revenue is planned work, split the reporting by job type before you argue about models.

Pick the model that would change a decision

Choose a model you can defend, then hold it still. Comparing this month's last-touch report against last quarter's data-driven report is not analysis, it is noise. Print the model name on the report so nobody has to guess which convention produced the numbers.

Then ask the only question that matters: would a different model change what you fund next month? If not, the debate is recreational. If yes, the answer is usually a holdout test rather than another model, because joining ad platforms to completed jobs tells you more than reweighting the same clicks ever will.