Carryover effect

The relationship to adstock trips people up in vendor meetings, so here is the clean version. Carryover is the phenomenon; adstock is the most common way to model it, a decaying transformation of spend.¹ Meridian's documentation treats carryover as a first-class modeling choice with its own parameters and priors.² When a vendor says "we model carryover with Weibull adstock, max lag eight weeks", they have told you the phenomenon, the mechanism, and the window; that is a complete answer.

Lag behavior differs by channel type in a direction everyone agrees on: direct-response channels (paid search, retargeting) concentrate their effect within days; broad-reach brand channels (TV, video, audio) stretch effects over weeks. The specific window for your brand is an empirical question the model should estimate, not a constant to be copied from someone else's deck, and this site will not pretend there is a universal number.

What mis-specification does

Suppose TV's true effect spreads over eight weeks but the model caps carryover at two. The revenue arriving in weeks 3 through 8 still exists; the regression must hand it to something. It goes to base ("brand is strong, cut TV") or to whichever channel spends steadily ("search is carrying us"). TV's measured ROI drops below its true value, next quarter's budget follows the measurement, and the error compounds. A too-long window fails in mirror image: noise gets attributed as lingering effect and a weak channel looks durable. Ask any vendor how their lag windows are chosen and validated; the good ones estimate them from data and test them out of sample.