The halo you build today pays back for 4-6 weeks.
Ad-stock physics, both directions. Each day's TV deposit takes ~12 days to reach half-strength on the rest of the funnel, ~6 weeks to fully decay. Build today, the halo compounds for a month. Cut today, the lag damage lands 12-14 days later. Same curve. The CFOs who cut TV at week 8 because MER reads flat are killing it the moment before it pays back.
"Cut TV today" simulator · how acquisition CAC reacts day-by-day
Geometric decay model: stock_t = α · stock_(t-1) + spend_t. Drag the carryover slider to see how the half-life shifts. TV default α=0.944. Search default α=0.65 (almost no carryover). Source: standard ad-stock model.
TV / CTV is the driver of this curve, not a row in the comparison. The α=0.94 carryover above is TV's deposit decaying into these downstream channels.
Carryover = % of yesterday's halo still working today. Higher = longer tail. Direct-response channels turn over fast; brand-builders compound.
The flighting discipline. If a brand cuts TV in week 8 because "MER looks flat" - they're cutting at the t+1 to t+3 lagged-peak window. The dashboard makes the lag visible. CFO sees halo emerging at week 2-3, not waiting for a 90-day mix model refresh that arrives after the budget cut.
Show the mathThe geometric decay model · why α=0.94 for TV
The model. stock_t = α · stock_(t-1) + spend_t. Each day's effective brand impact is the previous day's impact multiplied by carryover α, plus today's new spend. Half-life formula: half-life = -ln(2) / ln(α).
Why α=0.94 for TV. Industry empirical. TV has the longest carryover of any paid channel - the brand impression sticks for ~12 days at half-strength. Search and direct-response social have α≈0.5-0.7 (carryover under 2 days). This is why TV cuts hit late + why search cuts hit immediately.
Per-brand re-fitting. The default α is the population mean. With 60+ weeks of brand-level holdout data, refit α specific to the brand using non-linear least squares on the lagged response curve. Brand-specific α typically lands within ±0.02 of the population default.
Sources. standard ad-stock model, (Ad Stock), Hanssens-Parsons-Schultz textbook. Open-source implementations: open-source ad-stock transforms, Meridian, Robyn (Meta open-source mix model).