Bell Statistics

Media measurement and attribution

Whether your advertising caused anything, and the three families of method — geo experiments, mixed modelling and attribution — that claim to tell you.

13 terms

Advertising measurement is the hardest applied causal problem most companies have, because the thing you want to know — what would have happened without the spend — is never observed, and every convenient number available in a dashboard is a correlation dressed as an answer. The terms here separate the ones that support a budget decision from the ones that do not.

Incrementality is the quantity that matters: the conversions that would not have happened anyway. A platform reporting conversions it attributed to itself is not measuring that, and the gap between the two is routinely a factor of two or more on retargeting and branded search. Establishing it takes an experiment — usually a geo experiment — or a model calibrated against one.

Marketing mix modelling is the other family, and its two hardest parameters have entries of their own: adstock, because advertising seen today converts next week, and diminishing returns, because the tenth impression is worth less than the first. Both are shapes fitted from data, both are weakly identified, and both are why an MMM calibrated by experiment beats one that is not.

See our geo testing and marketing mix modelling work.

Terms in this group

  • Adstock

    Advertising does not stop working the week it stops running — and this is how models say so.

  • Attribution window

    A dial that changes every channel's reported performance — and platforms do not set it the same way.

  • Budget scaling

    How much to move, and what the next pound buys — which is never what the last pound averaged.

  • Conversion lift study

    A real randomised holdout, run by the platform being measured — genuine evidence with a conflict of interest attached.

  • Diminishing returns

    The tenth million does less than the first — and why average ROAS is the wrong number to budget on.

  • Ghost ads

    Log the ad you would have shown instead of showing it — exposure-matched control, and no wasted spend.

  • Incremental CPA

    Cost per conversion you actually caused — the number bids should be set against, and rarely are.

  • Incrementality

    The conversions that would not have happened anyway — and the gap between that and what platforms report.

  • iROAS

    Return on spend counting only what the advertising caused — routinely a fraction of the platform's number.

  • Last-click attribution

    All the credit to the last touch — reproducible, universally understood, and wrong in a predictable direction.

  • Marketing mix modelling

    One regression across every channel, built on aggregate data — no tracking, and strong assumptions.

  • MMM calibration

    Anchor the model to an experiment — the step that turns a plausible MMM into a trustworthy one.

  • Multi-touch attribution

    Credit spread across the path — better than last-click, and still describing correlation rather than cause.

Geo Testing at Bell Statistics

We measure what advertising actually caused, with geo experiments and mixed models that are calibrated against them rather than against a dashboard. See how we work.

Other topics

Back to the glossary