Working out where the next advertising dollar should go

What this actually solves

The owner opens the search advertising account and it reports one hundred and twenty conversions. The social platform reports ninety-five. The email tool claims sixty. The shop recorded one hundred and eighty orders in total. Add the platforms together and they have sold ninety-five things that never existed.

This is where the confusion becomes expensive. Faced with three sources that cannot all be right, the temptation is to pick the most flattering one, or to stop believing any of them and allocate budget by feel. Both are costly, and the second is worse than it sounds — the budget still moves every month, it just moves for reasons nobody can defend afterwards.

What is being purchased is not a model. It is one number to spend against, and an honest account of what that number can and cannot tell you. The hardest and most valuable part of this work is not the analysis, it is explaining the limits of the analysis without destroying your confidence in your own data. A report that says here is the answer is easy to write and easy to act on wrongly. A report that says here is what this model assumes, here is where it will mislead you, and here is the one figure that is hard to distort — that is what changes spending decisions safely.

What we do, and in what order

1. Assess whether the tracking can support this at all. Campaign tagging coverage, duplicate tags, test transactions in the totals, payment domains showing as referrers, purchase events arriving without a value. You get a written statement of what can and cannot currently be answered, before any modelling starts. Three days to two weeks, depending on what is found.
2. Design and apply a campaign tagging scheme. One consistent convention across paid channels, email, affiliates, print codes and offline references, applied everywhere rather than mostly. It is dull work and nothing downstream is reliable without it. Three to five days.
3. Bring spend, behaviour and actual orders into one table. Platform spend at campaign level, site behaviour, and the real orders from the system that takes the money — joined on dates and identifiers that have been checked rather than assumed. The order system is the anchor, because it is the only source that knows what was genuinely sold. One to two weeks.
4. Run the comparison and write the limits. The same period through several models, side by side, with a plain statement under each one of what it assumes and where it will mislead. The variance between models is presented as a finding in its own right, because a channel that looks strong under every model and one that looks strong under only one deserve very different treatment. About a week.
5. Define the reporting set. A small number of figures, each attached to the specific question it answers, including the blended revenue-to-spend ratio as the anchor. Built so it can be produced every month without reanalysis. Three to five days.
6. Design a holdout test, if you are willing to run one. Which channel, which region or audience, for how long, what will be measured, and how large the effect needs to be before it means anything. Designing it is quick. Running it takes commercial nerve, and declining to run it is a legitimate choice as long as that decision is made knowingly. Three to five days to design.

What moves the price

How many channels, and how much is being spent through them. More channels means more overlapping journeys to reconcile. Larger budgets mean the conclusions carry more consequence, which raises the standard of evidence required before anything is recommended.

Whether tracking has to be repaired first. This is the most common reason an attribution project costs more than expected, and it is why the assessment comes before the quote for the modelling.

Whether offline sales are in scope, and whether you want a report or a running system. A one-off analysis that answers this quarter's question is a smaller commitment than a maintained monthly reporting system that keeps answering it, and both are legitimate purchases depending on how often the decision recurs.

What usually goes wrong

A confident wrong recommendation. The report is opened, one channel looks unprofitable, and the advice is to switch it off. It is switched off, and the following month the search numbers fall too — because a substantial share of those searches were people who had seen the paused channel first. The channel was not unprofitable, it was upstream. Recommendations from attribution models should be framed as hypotheses to test, and only a holdout test justifies removing a channel outright.

The model is built on inputs nobody checked. Missing campaign tags on email traffic push it into direct. A duplicated tag inflates one side of the join. Staff test orders sit inside the conversion totals. The modelling then distributes credit with great precision across numbers that were wrong before it started, and precision is exactly what makes the output persuasive.

The data-driven model is treated as the truth. It is the best-reasoned option available and it is also a system whose workings you cannot fully inspect or reproduce. It is a good default for day-to-day optimisation and a poor sole basis for a decision that reallocates a large budget, because when someone senior asks why a channel lost its credit, no defensible answer exists. Pair it with a model you can explain, and with the blended figure that does not depend on attribution at all.