There is a quiet change that affects a lot of reports and rarely comes up in the meeting. Part of the conversions you see in an ad account are not observed events. They are modelled estimates.
That happens because some people do not consent to measurement cookies, and others use a browser or device where conventional tracking does not work. The platforms fill the resulting gap with an estimate.
Because an estimate beats a zero. Without it the system would conclude that an entire segment does not convert, and would move spend away from it. Modelling is what prevents that.
In three places.
Daily variance is larger. An estimated number does not settle as fast as an observed one. Reacting to daily data means reacting to noise.
Segments are not equal. Where more people decline consent, the estimated share is bigger. So comparing two campaigns is not always a like for like comparison.
Your account and your own system will not match. They do not need to. But if you expect them to match, you will treat real data as broken.
Get the consent handling right. Not only as a legal tick box, but for measurement reasons too. If the consent state is not passed correctly, you either lose more data than you should or collect data you are not allowed to.
Watch your own revenue as well. The ad account tells you direction, the books tell you fact. If the two diverge persistently, you are optimising for the account rather than the company.
Decide on a longer window. At weekly or fortnightly granularity the effect of estimation is smaller than at daily.
Send more from the server side. What you send from your own server has fewer things that can interfere with it than measurement running in the browser.
This is not about the data having gone bad. It is that part of the data is now a probability rather than a count. Knowing that, you decide well with it. Not knowing it, you will rebuild a campaign because of two weeks of variance.
More posts
Smart Bidding will not correct a measurement mistake. It will act on it faster and with more of your money. Here is how to see what the system actually sees.
If the system only sees that a form arrived, it will bring you forms. Here is how to give it back the information about which ones became business.
Automation is good at volume and speed. It is bad at judgement. A simple split that decides what you should let go of.