Each platform answers a different question
One counts anyone who saw an ad within a month. Another counts only clicks within a week. A third models the ones it cannot see at all.
None of these is wrong on its own terms. They simply cannot be added together, any more than you can add a distance to a temperature.
Pick one source of truth and make it boring
Your own record of orders is the only number that has to balance. Everything else becomes a hint about where to spend, not a claim about what happened.
Once the internal number leads, the platform reports become useful again: relative movements within one platform are meaningful even when absolute totals are not.
Watch the shape, not the digits
The question worth answering weekly is whether the total is moving, not which channel deserves the credit for a given order.
The cheapest honest test is to turn something off. If total revenue does not move, that channel was reporting on demand you already had.
- Total orders against total spend, weekly
- One channel paused, then observed for a full cycle
- New customers separated from returning ones
- The lag between first contact and payment
Retargeting flatters itself
Advertising to people who already visited will always look efficient. It is harvesting demand something else created, and it takes the credit because it was last in line.
That does not make it useless. It makes it a closing cost rather than an acquisition cost, and it should be judged as one.
How to run a holdout without wasting a quarter
A holdout is the only test that answers the question directly: what happens to total revenue if this spend stops? Everything else is inference.
Run it on one channel, for at least one full purchase cycle, and decide the success measure before you start. Stopping early because the first week looked bad is how holdouts produce confident wrong conclusions.
- One channel at a time, never several
- A full purchase cycle, not a calendar week
- The measure agreed and written before the pause
- Total revenue watched, not the paused channel's own report
Why the mismatch gets worse over time
Tracking has become less complete every year, and it will continue to. Browsers restrict what can be observed, people decline consent, and platforms fill the gap with modelling.
This means the direction of travel is towards fewer observed conversions and more estimated ones. Anyone whose budget decisions depend on precise attribution is building on ground that is still moving.