Why does MMM need variation in marketing spend to work?

If your thermostat were locked at 70 degrees all year, you'd have no idea how well your air conditioner works. The house would feel comfortable most days, but you couldn't say how much of that was the AC and how much was mild weather. You'd only learn what the system can do by watching the room respond when the setting changes, or when a 95-degree afternoon tests it.

Marketing mix models (MMMs) face the same problem. So, why does an MMM need variation in marketing spend to work? Because a model can only measure what moves. An MMM estimates a channel's impact by watching how sales respond when that channel's spend goes up, down, or pauses. If a brand spends the same amount on Meta every day for six months, sales still come in, but there's nothing in the data to show how much of that revenue Meta produced and how much would have arrived anyway.

That matters because the way you spend shapes what your measurement can tell you. Budgets that never change, or that all change at once, leave even a good model with less to work with, and the result is wider uncertainty right where you need the clearest answers.

Key takeaways

• An MMM needs spend variation because it measures marketing's impact by comparing how sales respond when spend changes, and flat spend gives it nothing to compare.
• Two kinds of variation matter most: changes in how much a channel spends over time, and changes in one channel that don't happen in every other channel at the same moment.
• Real-world variation is often correlated with demand, because brands spend more when shoppers are already buying, which can make ads look more efficient than they are.
• Traditional and open-source MMMs commonly read the highest spend level they've seen as a channel's ceiling, confusing a budget limit with true saturation.
• Prescient's model uses funnel roles, flexible response shapes, and benchmark-informed starting points to get more out of the variation a brand has, and it shows wider ranges when that variation is thin.
• Marketers can create useful variation without turning campaigns off by staggering changes, flighting spend, and holding some channels steady while others move.

Why an MMM can't measure what doesn't move

An MMM works by finding patterns between marketing activity and sales over time. When a channel's spend rises and sales rise more than you'd expect from seasonality and everything else, that's evidence the channel is contributing. When spend falls and sales dip, that's evidence too.

Take away the change and the evidence goes with it. With perfectly flat spend, the model sees a constant input and a sales line that moves for other reasons. It can't separate that channel's contribution from baseline demand, because the channel never did anything different that the model could line up against sales.

That's why variation is the raw material of measurement. Scaling up, cutting back, pausing, and shifting budget between channels all give a model information it needs.

The two kinds of spend variation that matter

Not every budget change tells a model the same thing. Two types do most of the work, and they answer different questions.

Volume variation maps the curve. Without seeing a channel at both low and high spend, a model can't tell whether returns are holding steady or leveling off.

Independent variation separates the channels. When every channel scales on the same day, as often happens around Black Friday, their effects arrive tangled together. When one moves on its own, the model gets a cleaner contrast.

When variation works against you

In practice, spend rarely varies in neat, random ways. Brands deliberately spend more when demand is already rising, such as during holidays, product drops, and big promotions. That kind of variation is real, but it's tangled up with the very thing a model is trying to separate it from.

Spend that rises with demand

When the ad budget and shopper demand both jump in early November, sales rise for two reasons at once. A model that can't tell those apart may hand holiday demand to the ads. That's called baseline leakage, and it makes peak-season ads look more efficient than they are.

Budget caps that look like saturation

Many MMMs describe each channel's returns with a curve that's built to level off. The flattening point gets placed at the top of whatever spend range the brand has actually run. If a brand never spends more than $15,000 a day on Meta because of an internal budget rule, the model may conclude Meta saturates at $15,000. The cap was a company policy, but the model reads it as a limit of the channel itself.

When Prescient studied 20 real Facebook campaigns, a plain straight line matched how revenue tracked spend more closely than the usual curved shapes for most of them. That suggests many campaigns hit a budget ceiling long before they hit a real one.

How traditional, open-source, and Prescient's models handle spend variation

The three main types of MMM differ in how much they can learn from the variation a brand has, and in how they behave when that variation is limited or tangled.

Traditional MMMs

Traditional MMMs, often built by consultancies or in-house analysts, typically use regression on weekly data and are refreshed every few months. Rolling daily activity into weekly totals smooths out a lot of the day-to-day movement in spend, which leaves fewer distinct changes for the model to learn from. These models also tend to group campaigns into channels, so variation between individual campaigns disappears into the total.

Open-source MMMs

Open-source MMM libraries make it much easier to build a model in-house, and many vendor tools are built on top of them. Out of the box, they usually model sales as one baseline number with each channel's contribution stacked on top, and they bend every channel's returns toward a ceiling. When variation is limited or correlated, they lean on statistical penalties and starting assumptions to settle on an answer. That steadies the output, but the answer then reflects those settings as much as the data, and a small change to them can move a lot of credit.

Prescient's model

Prescient's model is built to get more out of the variation a brand has without pretending it has more than it does. A few design choices make the difference.

• Funnel roles give variation more meaning. Prescient's model treats demand-creating channels, such as YouTube and CTV, differently from demand-capturing ones, such as branded search. When upper-funnel spend changes, the model expects the payoff to surface over the next several weeks in the channels that convert that demand—branded search, direct visits, and retail—so it isn't relying on same-day patterns alone.
• Response shapes come from the data. Prescient's model doesn't assume every channel levels off where spending stopped. A campaign whose returns hold steady across its range is shown that way, so a budget cap isn't mistaken for a ceiling.
• Holiday demand isn't handed to the ads. Prescient's model doesn't carve sales into "baseline" and "media" after the fact. Seasonal swings and marketing act on the same underlying demand, which lowers the odds that holiday shoppers get counted as ad results.
• Low-variation channels get realistic starting points. When a channel has little spend history or little movement, Prescient's model leans on benchmark data across many brands, plus patterns from channels that play a similar role. Estimates stay within realistic bounds instead of swinging wildly or dropping to zero.
• Daily data captures more movement. Prescient's models update daily for online channels, so day-to-day changes in spend stay visible instead of being averaged into weekly totals.

None of this replaces variation. When spend has barely moved, Prescient's model shows that through wider forecast ranges and lower confidence, especially at spend levels a brand has rarely or never run.

Side by side

The table below summarizes how each approach tends to behave in common spend situations.

How to create useful spend variation without hurting performance

You don't need to turn campaigns off to give an MMM what it needs. Small, deliberate changes in how you schedule budget moves can make a big difference in how clearly each channel shows up.

• Stagger budget changes. Instead of raising every channel on the same day, space increases a few days or a week apart, so each one shows up on its own.
• Hold some channels steady. When testing a new budget level on one channel, keep the others flat for that window, so the contrast is clean.
• Flight upper-funnel campaigns. Run brand and video campaigns in planned bursts, with quieter stretches in between, so their multi-week effect on search and direct traffic is easier to see.
• Use regional changes where it makes sense. Raising or lowering spend in some regions while others stay steady creates contrast without putting the whole business at risk. Keep in mind that regions differ and can be hit by local events, so treat what one region shows as one piece of evidence.
• Vary within a sensible range. Moving a campaign's budget up and down within ranges you're comfortable with helps map its returns, and it gives Prescient's model more history to forecast from at those levels.
• Keep a clear record. Note when and why budgets changed, including promotions and launches, so unusual patterns have context.

Prescient also needs a minimum amount of history to measure a campaign on its own: at least seven days with spend in the past year. Campaigns below that threshold are better read as part of their channel until they've built a track record.

Where Prescient comes in

Prescient's model is built for omnichannel brands whose spend moves with launches, promotions, and seasons, which is when telling ads apart from demand matters most. Your website, Amazon, and retail revenue all sit in one model. Each campaign gets its own Modeled ROAS, halo effects and retail halo effects are part of the picture, and daily refreshes mean a budget change you made this week starts informing results right away.

Our Media Forecaster feature then shows which channels and campaigns to grow or cut back, with ranges and confidence that reflect how much variation supports each recommendation. If you'd like to see how much your current spend patterns can tell you, book a demo.

FAQs

Can an MMM work if my spend never changes?

Only to a limited extent. If a channel's spend is completely flat, there's no change for the model to compare against sales, so its effect can't be separated from baseline demand using that data alone. Some models, including Prescient's, can draw on benchmarks and similar channels for a starting estimate, but the read will carry a wide range until spend varies.

How much spend variation does an MMM need?

There's no single threshold, but more distinct changes give a clearer read. A channel that has run at several different spend levels, and that has sometimes moved on its own rather than in step with every other channel, is much easier to measure than one held at a single level.

Will varying my spend hurt performance?

It doesn't have to. Variation can come from staggering planned changes, flighting campaigns, and moving budgets within ranges you're already comfortable with. These are often things teams do anyway. The difference is spacing them out so each change shows up clearly.

Why do MMMs confuse budget caps with saturation?

Many MMMs fit each channel to a curve that's designed to level off, and they place that point based on the highest spend they've seen. If a budget rule kept a channel from going higher, the model may treat that rule as the channel's limit. Testing higher spend in steps, or using a model with flexible response shapes, helps avoid that mistake.

Do geo tests give an MMM the variation it needs?

They can help. Changing spend in some regions while holding others steady creates contrast a model can learn from. Regions aren't identical, though, and local events can affect results, so geo tests work best as one source of variation alongside staggered changes and campaign flighting.