On one of my first calls of September I asked a marketing operations leader why her company had stopped trusting its own attribution reporting. Her answer was about arithmetic.

More pipeline attributed to marketing than the whole company generated that quarter. Her tool was reporting past the size of the business.

She had been receiving that report for two years. Nobody had run the two reports below.

If you are on Salesforce you can do this in the tab you already have open, and you will know the answer before you reach the next section.

If the first number is larger than the second, you do not have a contribution number. You have a sum of overlapping reports, and you can prove it in the next sixty seconds.

The fix is inside the report you just built. Campaign Influence is a junction object, and it already stores a percentage and a revenue share against each campaign for each opportunity. Swap Opportunity Amount out of your summary and sum the revenue share column instead. That total cannot exceed the opportunity, because it is a split of it.

On HubSpot the objects are named differently and the test is identical. Sum the attributed amounts, sum the deal amounts over the same window, and see which is bigger.

Here is why the two totals differ. One quarter, five opportunities, three campaigns.

Actual pipeline created: $600,000.

Now pull each campaign's influence report. Every campaign shows the full amount of every opportunity it touched, because it did touch it.

$760,000 against $600,000 of pipeline. One-hundred and twenty seven per cent, and not one of those three campaign reports is wrong. Opportunity 1 is counted three times because three campaigns each correctly claim to have touched it.

Now split each opportunity once, evenly across the campaigns that touched it.

$480,000 of $600,000. Eighty per cent, and it reconciles, because $480,000 is simply the sum of the four opportunities that had any marketing touch at all. Opportunity 4 is the remaining $120,000.

Three words are doing separate jobs here and most reports use them interchangeably.

Her report was adding the first column and labelling it the second.

Export one row per opportunity for the quarter. Six columns.

Four formulas underneath it.

That version credits any opportunity with at least one recorded touch, and it is the one to build first because it cannot double count.

If you publish a weighted model, export one row per touchpoint instead of one row per opportunity, put the model's own revenue figure in column F, and add this SUMIF check beside it:

That has to equal the opportunity's amount, every time. Higher and the model is double counting. Lower, if you did not intend a partial model, and touches are missing. Sort the column and look at the top and bottom twenty rows before you look at anything else.

Column D is the one people skip. Filter it to whatever finance means by pipeline before you calculate anything, because a contribution percentage computed over a different deal population than finance uses is not a disagreement about marketing, it is a disagreement about the company.

The two headline figures are the ones from the call on the first of September. The rest is filled in so the parts are visible, and your own specifics go where mine are.

Two lines in there earn the whole page. The total sits above marketing's number in finance's own definition, so the arithmetic can be checked without leaving the page. And the allocated figure is a subset by construction, every row traceable to an ID inside that total, instead of a parallel calculation that happens to come out smaller.

Most contribution reports get discounted for the arithmetic. The good ones get discounted for pretending the model is a measurement, and that discount is deserved.

Go back to the allocation table. The webinar got $33,333 of opportunity 1. Ask why it was a third.

I picked it, by choosing to divide evenly. First touch would have given the webinar all $100,000. Marketo Measure’s W-shaped model would have given it thirty per cent, because that model assigns thirty to first touch, thirty to lead creation, thirty to opportunity creation and splits the last ten across everything in between. Each of those is a rule someone wrote down. None of them is an estimate of anything.

Run it on the five opportunities and you can watch the effect. Same data, same touches, one dropdown changed from linear to first touch.

The ranking held here, because five opportunities is a very small room. Across a dozen channels and a real quarter it often does not. A twenty-two per cent swing in the webinar's number came from a settings change, not from anything the webinar did. Run that comparison on your own data before you present anything. If your channel ranking flips between two standard models, your budget is being set by a dropdown.

This is where the method has to change. We're no longer just making the allocation more precise; we're trying to measure contribution from the data.

Weights should be derived from data instead of asserted. One way to do that is the removal effect: take all your observed paths, remove one channel from those paths, and see how much the conversion rate falls.

For example, if your paths convert at 10% with Paid Search present and 8% when Paid Search is removed, the removal effect is 20%. Do that for each channel and you have weights based on what changes in the data rather than a rule someone picked in a settings menu.

Shapley values take a different approach, measuring each channel's marginal contribution across different possible orderings of touches.

These methods are more defensible than simply declaring that first touch gets 40% and last touch gets 40%. But they still don't prove that the channel caused the outcome.

Both need conditions your current data may not meet.

You need enough converting paths for the transition probabilities to be stable, which means thousands, not the forty deals in a good quarter. You need touch capture that is consistent across channels, which is exactly what check 01 above tells you that you do not have. And the Markov version assumes a constant transition probability, so what it describes is an average customer who may not exist inside any of your actual segments.

Even done properly, that is still correlation. A channel with a high removal effect may be where people who were already going to buy happen to show up.

Which is why the last step is experimental, and it is the only one that answers the question finance is actually asking.

Hold a channel out from a random or geographic slice of the market and measure the difference in outcomes. Or build a marketing mix model and calibrate it against those experiments.

Google is unusually candid about what that costs you in certainty. Their Meridian documentation states that the required assumption, conditional exchangeability, means your control set "includes all confounding variables" and "excludes any mediator variables", that "in reality, the exchangeability assumption is never perfectly met", and that "there is no absolute best solution to MMM, which follows from the fundamental principles of causal inference from observational data."

The line worth carrying into your next vendor conversation is the other one they publish: "goodness of fit metrics don't give a complete picture of how good a model is for causal inference, and it is likely that the best model for causal inference is different from the best model for prediction."

A model that fits your history beautifully is not thereby a model that knows what caused it. Anyone showing you an R-squared as proof of incrementality is showing you the wrong number.

There are really three levels here, and most companies end up arguing about the last one before they've properly dealt with the first.

First, reconcile the numbers. Then estimate contribution. Then run experiments to understand causality.

I've been building these models for twenty years, and I've never seen one where marketing accounts for all of the pipeline. There's always something that sits outside it.

That's what I've seen across the engagements I've worked on. I wouldn't treat that as a benchmark for what your number should be. But if your remainder is very small, I'd want to go back and look at the definitions before I believed it.

Be clear about which stage you're actually at.

Finance can work with a reconciled number that admits its weights are assumptions. Give them a number they can pull apart, and they'll argue with the model. Give them one they can't, and they'll quietly start arguing with you.

If someone on your team owns attribution or builds the finance pack, send them the comparison at the top of this issue. They can run it before your next review, and the answer comes back either reassuring or urgent.

The Moneyball CMO goes out weekly to marketing leaders who have been disappointed by at least one attribution tool. Every issue takes one measurement problem I have watched a customer hit, and hands over the report, the formula or the field to check.

Read by teams at Zscaler, Glean, Abnormal Security and Ncontracts. Subscribe at https://the-moneyball-cmo.beehiiv.com/subscribe.