Attribution models and effective windows

An attribution model is a rule for assigning credit. When someone sees three ads, clicks one, and buys four days later, the model decides which ad gets the sale. Change the rule and the same underlying events produce different reported winners, which is why two systems watching identical traffic can disagree completely.

YieldBI applies incremental attribution models with defined effective windows rather than one default rule across every campaign.

Why the default is often wrong for your funnel

Meta’s default attribution is a single setting applied to every campaign in the account. That is convenient and rarely correct for all of them at once.

A short window on a considered purchase drops conversions that genuinely came from the ad, making prospecting look worse than it is and pushing budget toward retargeting. A long window on an impulse purchase does the opposite: it sweeps in sales that would have happened regardless, and the campaign that gets the credit did not earn it.

Neither error announces itself. Both quietly move budget in the wrong direction.

Effective windows

An effective window is the period after an ad interaction during which a conversion is still credited to that ad. Two things set it:

• Click windows count conversions from people who clicked. This is the stronger signal.
• View-through windows count conversions from people who only saw the ad. Weaker, and the most common source of inflated reported performance. See attribution window and view-through, and Meta’s own documentation of how its attribution works.

Choose the window from your actual time-to-purchase, not from a benchmark. If most orders land within two days of the click, a seven-day window is mostly adding noise. If your median is eleven days, a seven-day window is systematically hiding the campaigns that work.

Why your numbers will never match Meta’s exactly

They are counting different things, and both are internally consistent:

• Meta credits on the ad interaction date; your store credits on the order date. Around a spike, those land in different days.
• Meta includes modeled conversions where direct observation is not possible; your database only holds observed orders.
• Meta cannot see conversions that finish offline unless you send them back.

Reconciling to the row is not a realistic goal. Pick the system you make decisions with, understand what it counts, and use the other as a sanity check on direction.

Where this shows up

• Pixel and conversion events feed the raw signal, so a gap in the Conversions API setup limits every model built on top.
• SKAdNetwork covers iOS app campaigns, where device-level tracking is unavailable and results arrive aggregated and delayed.
• Ad-level revenue visibility and audience discovery insights are reported using the model and window you configured, so the daily action list matches the model you trust rather than a platform default nobody chose.

What attribution still cannot answer

Attribution assigns credit among touchpoints that already exist. It cannot tell you what would have happened without the ad at all. For that you need a holdout: see incrementality testing and geo holdout testing.