Which half of the budget works?

It is a statistical way of explaining what drove your sales. Instead of following individual users, it looks at weekly history — what you spent on each channel, what you charged, what you promoted, what stock you had, what season it was — and estimates how much each factor contributed. Because it works on aggregate data it needs no cookies and no user identity, which is why it survived the privacy changes that broke click-based tracking. It also covers the channels attribution cannot see at all: television, radio, print, sponsorship, retail media and brand campaigns.

They answer different questions and both are useful. Attribution follows measurable journeys and tells you which touchpoints preceded a conversion, which makes it good for day-to-day optimisation inside a channel. Modelling works top-down on aggregate history and tells you what each channel contributed overall, including the ones with no click at all, which makes it the right tool for setting budgets. Neither is the truth on its own, and experiments are what settle the disagreements between them. Our attribution team builds the bottom-up side.

Ideally two to three years of weekly data across every channel, plus price, promotion and stock history. One year can work for a business with strong variation in spend, but a single seasonal cycle makes it hard to separate Christmas from a campaign. If the history is not there we will tell you at the feasibility stage rather than modelling anyway: geo experiments and holdouts give a cleaner answer sooner, cost less, and become the calibration evidence for a model later. Starting to collect a consistent dataset now is the other half of that advice.

Open-source Bayesian frameworks, chiefly Google's Meridian and Meta's Robyn, with custom modelling where a business does not fit either. We prefer them because they are free, documented and fully inspectable: you can re-run the model, hand it to another team, or have a third party audit our assumptions. Paid platforms are a reasonable choice when you want continuous refreshes and no internal data science, and we will help you evaluate them — but a model you cannot see inside is difficult to defend in a budget meeting.

It is a fair question and the reason we insist on inspectable models and independent calibration. The frameworks themselves are open source and the mathematics is public, but any model's answer moves with its priors and its inputs, so we set priors from your own experiments rather than from vendor benchmarks, keep spend definitions consistent across channels, and validate with geo tests that no platform runs for us. We also publish the sensitivity: if a channel's estimated return swings when one assumption changes, you will see it in the readout rather than discover it later.

Accurate enough to plan with, and never precise enough to quote to two decimal places. We report credible intervals rather than single numbers, so a channel's return arrives as a range with a most-likely value, and we state which channels the data can and cannot separate — two channels that always moved together are genuinely hard to untangle. Validation against holdout periods and live experiments is what turns a plausible model into a usable one, which is why the calibration calendar is part of the service rather than an upsell.

Quarterly for most businesses, monthly where spend is large or the market moves quickly. That cadence matters more than people expect: a model refreshed once a year arrives after the decisions it should have informed, which is precisely why so many modelling projects are remembered as expensive slides. We automate the data pipeline during the first build specifically so refreshes are cheap, and each refresh folds in the results of any experiments run since, so the model gets better rather than merely newer.

It can give you a defensible starting point and the shape of the trade-offs. The model produces response curves per channel showing where extra spend starts delivering less, and the scenario planner lets your team test allocations and see projected outcomes with their uncertainty. What it cannot do is predict a competitor's launch or a price change you have not told it about, so we treat the output as a plan to be tested rather than an instruction. Where the recommendation involves a big move, we usually propose a geo test first.

Yes, with two adjustments. The outcome variable moves from same-week orders to pipeline or closed revenue, and the model has to handle the lag between marketing and money explicitly rather than assuming a weekly response. That needs CRM data joined to the media history, which our CRM analytics team usually helps assemble. The result is often more valuable than for ecommerce, because in long cycles click attribution is at its least reliable and brand investment is at its hardest to defend.

Less than you might fear, and the right people rather than many. We need someone who can get us spend history and access to the data warehouse, someone commercial who knows the promotions, price changes and stock problems of the last two years, and the budget holder who will act on the result. That last one is the important one: modelling projects that skip the decision-maker produce reports nobody uses. Beyond a kick-off, a mid-point review and the readout, we keep the demands on your team small.

Ask three questions. Will you own the model, the code and the data pipeline, or rent an output? How will the model be validated against live experiments rather than only against its own history? And what decision-support comes with it — a scenario planner and a quarterly readout, or a PDF? That last one separates the field: the Harvard Business Review Analytic Services research found nearly 40% of marketers struggle to connect model outputs to real decisions. Ask to see a redacted readout, because that shows how an agency thinks.

The feasibility assessment is a small fixed fee and sometimes ends with us recommending experiments instead of a model, which is a cheaper outcome for you. The first build is fixed-scope and quoted up front, covering the dataset, pipeline, model, validation, scenario planner and readout. After that, refreshes and the experiment programme are a smaller recurring fee, since the expensive part — assembling and automating the data — is already done. As a rule of thumb, modelling earns its cost once media spend is large enough that a few percent of reallocation matters.