How to test whether a client's ads caused the sales.

Every platform reports conversions it believes it caused. None of them can tell you what would have happened if the ads had not run, because it never observed that.

That gap matters most for channels aimed at people who already know the brand. An ad shown to someone who was going to buy anyway gets credited with the sale, and the report is not wrong about what it saw. It is answering a different question from the one a budget decision needs.

Incrementality testing answers the budget question directly: turn the spend off for a comparable group and see what changes.

You split the client's audience into two groups that should behave the same, show ads to one and not the other, and compare outcomes. Whatever difference appears is the part the ads caused.

The split has to be something you can actually hold apart. Geography is the usual answer, because two sets of regions can be compared and a person cannot be in both. Splitting by audience inside a platform works where the platform supports it, but be sure the two groups are not being reached by the client's other channels differently.

For an agency, the hard part is often the conversation before the test. A holdout means the client deliberately stops advertising to part of their market, so agree in writing before it starts how large the held-out group is, how long it runs, and what result would change the budget.

Four mistakes, and the first two invalidate the test entirely.

• The groups were not comparable

  Two sets of regions with different seasonality, different store coverage or different competition are not a test. Check that they tracked each other before the test as well as during it.

• Something else changed at the same time

  A price change, a promotion, a press mention or another channel shifting its spend will all show up as your result. Freeze what you can and record what you cannot.

• It ran for too short a time

  The test has to run longer than the time it normally takes someone to buy, or you are measuring the beginning of an effect and calling it the whole thing.

• The result was read as a single number

  A test gives a range, not a figure. Reporting the midpoint as though it were exact is how a test that showed almost nothing becomes a slide claiming a precise uplift.

The useful output is not a number to put in a deck. It is a correction factor for the reporting you already have: now you know roughly how much of a channel's reported conversions were incremental, and you can read next month's report with that in mind.

Re-test occasionally rather than once. The answer changes when the client's audience size, creative or competition changes, and a factor from two years ago is a guess wearing a measurement's clothes.

A holdout test needs clean spend and outcome data by region and by day, in one currency, from before the test as well as during it. That is a warehouse question rather than a platform question, because no single platform holds both the spend and the outcome.

SightShark lands both, for every client, in a warehouse you own and keeps the history, so a test you run next quarter can be compared against a baseline that was not assembled after the fact.