Brand in marketing mix modeling: sales that arrive later.

The shortest version we have: brand is sales that arrive later. Same sales, later.

The split per channel starts from published industry evidence and the customer's own view of each channel's job, and the data adjusts it inside stated bounds. We are open about what that means. The short term is measured. The long term is a reasoned belief, written down.

A marketing mix model links spend to sales through a response curve and a carryover effect. On weekly data the carryover is short. Most of a channel's measurable effect lands in the week of the spend and the few weeks after.

That does not mean brand effects are missing from the sales. They are in the data. But with nowhere else to go they sit in the baseline, where nothing specific can be said about them.

The optimizer then does what it was built to do. It sees TV and other brand-building channels earn little inside the window, and it moves the budget to search and performance. Everyone in the room suspects that is wrong, and the model cannot say by how much.

The usual workarounds each have a cost. Stretching the carryover tail lets it absorb slow movements in demand that belong in the baseline. Adding a brand tracker as a control variable hands part of marketing's own effect to a noisy monthly survey. Locking the TV budget by hand works, and many of our customers have done exactly that, but it takes the question away from the model instead of answering it.

Each channel's effect is split

For every channel the model carries a brand fraction: the share of the channel's response that builds brand instead of converting now. The immediate share keeps the channel's own short carryover. The brand share goes into the bucket. A performance channel sits near zero. TV, audio and sponsorship sit high.

The split conserves the total. It changes when the sales land, not how much the channel earns.

One shared stock

The brand shares of all channels flow into one stock of brand value, measured in sales. Each period a small part of the stock is paid out as sales and leaves the bucket. What goes in comes out, spread over time.

Here Σ_c sums over channels, q is the channel's brand fraction and d is the share of the stock paid out per period.

The bucket leaks

Without new investment the stock falls by roughly two percent a month. The prior range is one to four percent, set from the purchase cycle and from published evidence on how brands decay. It is not left for the model to fit freely against a short history. At two percent a month, half of what a campaign put into the bucket has been paid out after just under three years.

One consequence is easy to miss: a stable brand still costs money to hold. Stop refilling and the level falls.

The bucket is not empty on day one

An established business enters the data with brand value already built. The model seeds the stock at the level the customer's own spending would sustain, and where brand tracker readings exist we use them to sanity-check that level.

This is the part to be most careful about. The data cannot settle the split between short and long term. Two models with almost identical fit can carry very different splits, so forecast accuracy cannot tell them apart.

So the split is a prior with bounds, per channel, and the data moves it inside the bounds. The short-term effect is driven by the data. The long-term share leans on published evidence and on what the customer believes about each channel's job.

The direction is stable across the evidence. TV, audio and sponsorship lean long. Search and Performance Max lean short. Paid social and online video sit in between, and depend on how they are bought.

Three rules we apply when mapping the evidence to a customer:

• Split by campaign objective, not only by platform. A sales campaign and a reach campaign on the same platform sit at opposite ends of the range.
• Content modifies format. A discount flyer is print by format and activation by content, so it sits low.
• Published evidence is an average. Profit Ability 2 is UK and cross-category, so a promotional e-commerce business starts at the low end of each range.

Why not model awareness directly?

A common alternative is a two-stage model: media drives a brand metric such as awareness or branded search, and the brand metric drives sales. With Meridian 2.0, Google announced a full-funnel approach along these lines, using brand equity signals such as branded search volume.

We chose not to make a tracker the target. A tracker is typically one reading a month, and a noisy one. In the tracker data we have seen there is too much variation to treat it as the truth the model should be built against. We model brand's effect on sales directly, and use tracker readings as a cross-check on the level.

Once part of the effect arrives later, "what does this channel return?" needs a second question: by when? The model answers with a payback horizon. Sales that land inside the horizon count. Sales that land after it do not.

A pure performance channel is credited in full at any horizon. A brand-building channel is credited with more the further out you look.

The figure uses two invented channels with the same spend. Scored on three weeks, the performance channel wins clearly: 216 against 155. Scored on one year they are level. Scored on three years, brand TV is well ahead: 331 against 228. Nothing in the model changed between the three. Only the horizon did.

That is how we suggest using it. Run the same budget on a short horizon and a long one, and take both plans to finance. The gap between them is the trade-off the business has to take a position on. You do not plan next year as if it were your last year in business.

Two rules keep this honest. The horizon changes what a recommendation counts, never the fitted model. And scenarios scored on different horizons are not comparable with each other. Compare like with like.

• It does not prove that brand drives sales. It assumes so, and estimates how much inside stated bounds. The long horizon is more belief and less measurement than the short one, and its ranges are wider.
• One shared stock, not brand value per channel. The model learns how much each channel fills the bucket. It cannot trace a sale that arrives two years later back to one channel's spend.
• No awareness or consideration stages. The model goes from spend to sales. It does not produce a brand health score.
• Brand means later sales, nothing else. Pricing power, retention, cheaper performance media and competitors' share of voice are not in it.
• Flat brand spend stays invisible. Spend that never changed level cannot be told apart from a constant baseline by any model.
• It is optional. The brand component is switched on per model, where a sizeable share of spend is brand-building. A business that is mostly performance media does not need it.

• Where does the long-term share per channel come from: the data, a benchmark or a slider?
• Can forecast accuracy tell two different short and long splits apart in your model?
• What happens to the brand level if we stop spending, and how fast?
• Does the optimizer count sales that land after the planning period, and can we choose how far out?
• Is brand tracker data a target, an input or a cross-check?

We are happy to answer all five about our own model, on your data.

• Thinkbox (2024), Profit Ability 2, with Ebiquity, EssenceMediacom, Gain Theory, Mindshare and Wavemaker UK.
• Binet, L. and Field, P. (2018), Effectiveness in Context, IPA.
• Ehrenberg-Bass Institute for Marketing Science, research on what happens to brands when advertising stops.
• Google Meridian, Meridian 2.0 announcement, September 2026.