Open three tabs: your ad platforms, your analytics, your back office. Ask each how many sales you made last month. You will get three different numbers, and the gap between them will be larger than the change you were trying to measure.
This is not a bug in your setup. It is what happens when three systems with different definitions, different windows and different access to data all try to describe the same thing.
Why they disagree
The differences are structural, so they will not be fixed by better tagging.
- Attribution windows. Platforms count a conversion within their own window, often seven days after a click. Your analytics may use a different window entirely. Both are choices, not facts.
- View-through. Some platforms count people who saw an ad and did not click. Whether that is real influence or coincidence depends entirely on how much of your audience was going to buy anyway.
- Every platform claims the same sale. Google and Meta both counted it. Neither is lying, and adding them together is meaningless.
- Modelled conversions. Where consent or cookies are missing, platforms estimate. Estimates are reasonable in aggregate and unreliable for any specific segment.
- Consent. If a visitor declines tracking, the sale still happens. It just does not appear in the same places.
Pick one source of truth, and it is not a platform
The single most useful decision you can make is to nominate one number as the truth, and to accept that everything else is an instrument rather than a scoreboard.
That number should be revenue in your own back office. It is the one system with no incentive to over-report, and the only one that knows about refunds, cancellations, failed payments and the customers who never pay their second invoice.
Once that is settled, the platform numbers stop being a source of argument and become what they actually are: optimisation signal. The algorithm needs conversion events fed back quickly in order to learn. It does not need to agree with your accountant.
Optimise on platform data. Report on your own.
The number that actually answers the question
Most attribution arguments are really one question wearing a disguise: if we spend another ten thousand, what do we get?
Attribution models cannot answer that, because they describe the past by dividing credit for sales that already happened. A few methods can:
Holdout tests
Turn a channel off for a defined group or region, keep everything else steady, and compare. Expensive and slightly frightening, and the cleanest evidence available.
Geo tests
Run a channel in some regions and not others. Well suited to national campaigns and much easier to sell internally than a full pause, because the downside is bounded.
Spend-response curves
Change budget deliberately in steps, and record what happens to marginal cost per acquisition rather than the blended figure. Blended cost per acquisition always looks fine while the last euro is losing money.
A monthly rhythm that works
- Reconcile first. Platform conversions against back office revenue. Not to make them match, but to know the size and direction of the gap.
- Steer on the blended number. Total spend against total revenue is crude and honest, and it cannot be gamed by attribution settings.
- Use platform data inside the channel. Which campaign, which creative, which audience. It is reliable for relative comparisons within one platform.
- Test incrementality on the channels you doubt. Usually retargeting, brand search and native. Anywhere the audience was likely to arrive anyway.
The short version
Stop trying to make the numbers agree. Decide which one you steer on, use the others for what they are good at, and reserve the real budget questions for tests that can actually answer them.
Want this applied to your own accounts? Book an audit; the first read costs nothing.