How incremental attribution works
Standard attribution answers a bookkeeping question: which ad interaction came before this conversion. Incremental attribution answers a much harder one: would this conversion have happened without the ad at all.
The only reliable way to answer that is comparison. You withhold the advertising from part of the market — a random slice of the audience, a group of regions, a defined period — keep everything else the same, and compare what happens in the two groups. The difference is what the advertising caused. Meta packages this as conversion lift and split tests; the same logic works by hand when you switch a campaign off in one area and watch sales there against a comparable area, which is what a holdout test does.
Where a platform offers incremental optimisation as a setting, it is being told to bid for the conversions its models judge would not otherwise have occurred, rather than for whichever conversions are easiest to claim. Check what your own account actually offers, because these options are renamed and revised regularly.
Why incremental attribution matters
Reported conversions and caused conversions are different quantities, and the gap between them is widest exactly where advertisers spend most confidently. Retargeting is the clearest case: an ad shown to somebody who already has your product in their basket will be credited with a great many sales, most of which were coming anyway. The reported return looks superb and the incremental return can be slight.
Brand search and campaigns that lean heavily on view-through conversions share the problem, since the ad may have been on screen for a moment and done nothing else.
What this changes is budget. Allocate on reported return and money flows towards the campaigns best at claiming credit rather than the ones creating demand. Over time the account looks efficient while the business stops growing, which is one of the more frustrating patterns to inherit.
Where it goes wrong
The commonest error is running a test while changing other things. A lift test only means something if the two groups differ in one respect, so altering creative, budget or the landing page part-way through destroys the comparison.
The second is testing too small. Where the market or the conversion volume is small — and in Nepal many accounts are small — the difference you are trying to detect can be smaller than ordinary weekly variation, and the test returns nothing usable. An honest “we cannot tell yet” is worth more than a result read into randomness.
The third is expecting the incremental figure to reconcile with Ads Manager. It will not, and it is not meant to. Putting the two side by side in a report without explaining that invites an argument nobody wins.
How to act on it
Start where the doubt is greatest. Retargeting, broad brand campaigns and anything with a large share of view-based credit are where reported and caused results diverge most, so test those before testing prospecting that is plainly reaching new people.
Keep the test boring. One variable, a period long enough to cover your normal sales cycle, no other changes, and a decision rule agreed before it starts — otherwise the result gets read to suit whoever wanted the answer.
Then change how you decide. Use reported attribution for day-to-day steering, because it is fast and always available. Use incrementality to set the shape of the budget: which channels get more, which get less, and which have been flattering themselves. If a formal test is out of reach, the crude version still works — turn a campaign off for a defined period and see whether the business notices. That single exercise has changed more budget decisions than any dashboard I have built.