Adstock: Why an Ad's Effect Doesn't End When It Stops

A stone dropped in a pond doesn't stop moving the moment it hits the bottom. The ripples keep spreading outward for a while afterwards, getting smaller, until they're indistinguishable from the water's ordinary murmur. Advertising works the same way: an ad's effect on a person doesn't end when it stops airing, or when they scroll past it. It keeps producing a smaller and smaller nudge toward purchase for days, sometimes weeks, after the exposure. Marketing mix modelling has a deliberately unglamorous name for that lingering effect: adstock.

Adstock is the decayed, carried-over effect of an ad exposure on demand in the periods after it ran. The term was coined by the researcher Simon Broadbent in a 1979 paper on how television advertising actually works (Advertising adstock, Wikipedia), and it captures something every marketer already half-knows from experience: a campaign's real return includes the sales it generates after it's finished running, not just the ones that land while it's live.

A click can only ever credit the memory that led to it

Last-click attribution gives 100% of the credit to whichever touchpoint happened right before someone converted, which is usually a paid search ad or a retargeting banner. It has no way to reach further back than that, so it has no way to see the TV spot someone half-watched three weeks earlier that made the brand name feel familiar enough to search for in the first place (What Last-Click Attribution Doesn't Capture, The Media Manager).

That's not a measurement bug that a better pixel would fix but much more structural: attribution can only see the touchpoints it can track, at the moment it can track them. Anything with a delayed or offline effect, TV, out-of-home, radio, a lot of brand-building social, reads as underperforming, while search looks artificially efficient because it's harvesting demand that something else already created.

A simple correlation between spend and sales has a related but different problem: it can only sum weighted variables together within the same period, so it has no mechanism at all for a pound spent this week producing a sale next month.

This is where adstock comes in. Without it, models cannot account for the variable delay between seeing an ad and making a purchase.

The right shape of the decay depends on the conversion journey

The simplest version of the decay curve is called geometric decay: each period's carried-over effect is a fraction of the effect from the period before, shrinking a little further each time it's carried forward. Software built for this, like Google's Meridian and Meta's Robyn, generally offers a choice of shapes for that shrinking curve. One version front-loads the effect and decays quickly, suited to media whose impact is felt almost immediately. Another lets the effect persist for longer before fading, suited to media whose impact builds more slowly (Set the adstock_decay_spec parameter, Meridian).

That standard curve assumes the effect is strongest the instant the ad runs and fades from there, which isn't true for every category. A car ad or a mortgage offer doesn't convert anyone on day one; there's a consideration window before the effect peaks. Robyn and Google's LightweightMMM both offer a shape that can build to a delayed peak before decaying, which Robyn's own documentation recommends specifically for products with a longer decision-making window, like cars or credit (Robyn Key Features; LightweightMMM models).

Judgement is needed to get the adstock shape just right

TV and out-of-home tend to carry over for weeks. They're doing brand-building work, sitting in memory and building familiarity, and that pays out gradually. Paid search and paid social tend to decay in days, because they're closer to the moment someone was already going to act.

The more important point sits underneath that ranking, though: adstock isn't a fixed property of a medium. It's what emerges from the interaction between the advertising, the audience and the category. Two brands running TV in the same slot can have genuinely different decay rates. This is also why Google's own Meridian documentation stops short of prescribing channel-specific decay values at all, leaving the choice to the modeller's judgement rather than a lookup table (Meridian: media saturation and lagging).

Getting the decay curve right matters for a second reason too: it interacts directly with how much of a channel's spend has already hit a ceiling, which is where diminishing returns picks up.

If nobody at your company could say with confidence how long your TV or paid social spend keeps working after it stops running, that's usually the first sign an MMM would earn its keep. Get in touch with Day1data to talk through what your own channels' decay curves might look like.

FAQ

What is adstock in marketing mix modelling? Adstock is the carried-over, decaying effect of an ad exposure on demand in the periods after it ran. It's the mechanism that lets an MMM credit a campaign for sales that land weeks after it stopped, not just while it was live.

Why can't attribution or simple correlation capture adstock? Attribution only credits the touchpoint closest to conversion, so it structurally can't see a delayed effect from an earlier exposure. A basic correlation model can only sum variables within the same period, with no mechanism for spend in one period to affect sales in a later one.

What's the difference between geometric and Weibull adstock? Geometric decay assumes the effect is strongest immediately and fades from there. Weibull-shaped decay can produce a delayed peak before fading, which suits categories with a longer consideration window, like cars or credit.

Does TV really decay more slowly than digital channels? Usually, but treat it as a starting assumption rather than a law. Decay rates vary by brand and category, not just by media type, so they should be estimated from the company's own data rather than assumed.

What is a half-life in adstock terms? The time it takes for an ad's carried-over effect to fall to half its original strength. A slower-decaying channel has a longer half-life; a faster-decaying one has a shorter one.

Can adstock be estimated without running a full MMM? Not reliably. Estimating a decay parameter requires enough historical variation in spend and timing to separate the lag from everything else moving sales at the same time, which is exactly the kind of statistical separation an MMM is built to do.

Does a longer adstock automatically mean a channel is more effective? No. Decay speed describes the shape of a channel's effect over time, not its size. A channel can decay fast and still be highly effective within that shorter window, and calibrating the model against real experiments (see calibrating MMM with incrementality testing) is what confirms the size is right, independent of the shape.