Marketing Measurement ·

How do MMM tools account for seasonality, promotions, and price changes without over-crediting marketing?

How do MMM tools account for seasonality, promotions, and price changes without over-crediting marketing? Here's how models separate them, and where they fail.

Linnea Zielinski · 13 min read

Listen
0:00 / 0:00
AI-generated audio
How do MMM tools account for seasonality, promotions, and price changes without over-crediting marketing?

A real estate agent who closes twice as many sales in April as in January may be getting better at her job. Or it may just be spring, when more families list homes, mortgage rates ease, and buyers come out in force. If her brokerage paid bonuses based only on her April numbers, it would reward the season as much as the agent. A fair review would compare her results with what the spring market was going to deliver anyway.

Marketing measurement has the same problem. So, how do marketing mix modeling (MMM) tools account for seasonality, promotions, and price changes without over-crediting marketing? They include those factors in the model alongside ad spend and try to estimate what sales would have been without the marketing, so ads get credit only for the lift above that level. How well that works depends on the model. Brands raise budgets during holidays, sales, and launches, so ad spend and natural demand climb together, and a model that can't separate the two will hand marketing credit for shoppers who were coming anyway, or occasionally the reverse. As you can imagine, that gets expensive.

Key takeaways

  • MMMs account for seasonality, promotions, and pricing by modeling them alongside ad spend, so marketing only gets credit for sales above what those factors would have produced on their own.
  • The hard part is that brands deliberately time ad spend to match holidays and promotions, so spend and natural demand rise together and are difficult to separate.
  • Seasonality affects both how many people buy and how well ads work, and a good model needs to tell those two effects apart.
  • Promotions interact with advertising and can pull future purchases forward, while price changes can raise revenue and ROAS even when ads aren't working any harder.
  • Many traditional and open-source MMMs treat these factors as a separate baseline added to independent media slices, which can credit holiday demand to ads, or bury real marketing gains in the baseline.
  • Prescient's model learns each holiday's impact on your business, treats seasonality as affecting both baseline demand and marketing effectiveness, and models promotional periods separately.
  • Brands get better results by giving their MMM a clean record of promotions, pricing, launches, and stockouts, and by staggering budget changes so they don't always coincide with sales events.

Why a sales spike can't be taken at face value

When sales jump in the same week ad spend jumps, it's tempting to credit the ads. But a sales spike usually has several causes at once, and an MMM's job is to estimate how much each one contributed.

Those causes generally fall into three groups.

Type of factorExamples
Marketing activityPaid social, paid search, CTV, podcasts, creators, direct mail
Business decisions outside mediaPromotions, price changes, product launches, retail distribution, inventory
Outside forcesSeasonality, holidays, weather, economic conditions, competitor activity, long-term trends

Take a home goods brand that raises Meta spend 50% for a Black Friday sale and sees revenue jump 90%. Holiday gift shopping, the discount, a bigger email push, and a retailer's own deals all contributed. If the model hands Meta most of that 90%, the next Meta scale-up won't come close to repeating it.

How MMMs account for seasonality

Seasonality is the recurring rise and fall in demand that would happen even without your current marketing: swimwear in spring, fitness gear in January, gifts in December. Marketers know these patterns and plan budgets around them, which is exactly why they cause trouble for measurement.

MMMs handle seasonality in several ways, and many use more than one.

ApproachWhat it does
Calendar markersFlags weeks, months, or days of the week with different typical demand
Holiday flagsMarks specific events, such as Black Friday or Mother's Day
Repeating curvesFits smooth yearly cycles to capture recurring patterns
A baseline that changes over timeLets underlying demand drift up or down instead of staying fixed
Outside demand signalsAdds category search interest, weather, or similar indicators

Seasonal demand isn't the same as seasonal effectiveness

This is the distinction that separates a decent model from a good one. Holidays change two things at once:

  • How many people buy. More shoppers are in the market in November, regardless of your ads.
  • How well your ads work. Shoppers who are already buying gifts may respond better to an ad than they would in February.

A model that only adjusts the baseline for the season treats every ad dollar as equally effective year-round. That can understate marketing's impact in exactly the weeks it works hardest. A model that adjusts nothing will credit the ads with holiday demand. The goal is to capture both: more baseline demand, and different ad effectiveness.

How MMMs account for promotions

Promotions are tricky because they almost never run alone. A 25%-off sitewide sale usually comes with a bigger paid social budget, promotional search campaigns, extra emails, more retargeting, and sometimes creator posts announcing the deal. When revenue climbs 80%, the model has to work out how much came from the discount and how much from the ads that told people about it.

Models usually represent promotions with details like these.

Promotion detailWhat it captures
Whether a promotion was runningThe simplest on/off marker
Discount depth10% off behaves differently from 40% off
LengthA 48-hour flash sale versus a two-week event
CoverageOne product, one category, or the whole store
TypePercentage off, buy-one-get-one, free shipping, gift with purchase
Where it ranYour site, Amazon, specific retailers, or certain regions

Ads and promotions amplify each other

A discount can make ads convert better, and ads can make more people aware of the discount. The combined effect is often larger than either one alone. A model that treats promotions and advertising as completely separate can't see that, so it tends to misstate both.

Promotions can borrow from future sales

Not every promotional sale is new demand. A customer who stocks up on six months of supplements during a Black Friday sale might otherwise have bought in December and February. That's called pull-forward: the promotion boosts November and dents the following months. A model that only measures the spike will overstate the promotion's value. The more useful question is how much extra revenue the promotion produced over a meaningful stretch, not just during the sale.

How MMMs account for price changes

Price changes affect both how many people buy and how much each purchase is worth, which makes them easy to mistake for marketing performance when revenue is the outcome being measured.

Here's a simple example. A brand raises the price of its best-selling product from $80 to $96 while keeping ad spend flat.

BeforeAfter
Price$80$96
Units sold12,00011,000
Revenue$960,000$1,056,000
Ad spend$150,000$150,000
Blended ROAS6.4x7.0x

Revenue rose 10% and ROAS improved, even though the brand sold fewer units. The ads didn't get better. Each sale just got bigger. A model that doesn't know about the price change may read that as stronger ad performance.

MMMs handle pricing with inputs like list prices, price indices, or prices relative to competitors. Tracking orders or units alongside revenue also helps, since a price change moves them in different directions.

One caution: average selling price can change without any price change. If a campaign pushes premium products, the average order gets bigger. A model that treats that as a pricing effect could strip credit away from the campaign that caused it. Actual price histories are cleaner inputs than realized averages.

Where traditional and open-source MMMs go wrong

Most traditional and open-source MMMs describe revenue as a baseline, covering trend, seasonality, holidays, and promotions, plus a separate slice for each marketing channel. Each piece is assumed to act on its own. That structure causes problems in two different directions.

When spend and demand rise together, the split becomes a guess

Because brands time spend to match holidays and promotions, ad spend and the baseline move up together. When that happens, many different splits of revenue between "baseline" and "ads" fit the past equally well. The data can't pick one, so the model's built-in settings do. Holiday or promotional demand that ends up booked as ad performance is known as baseline leakage, and it's common: in Prescient's peak-season simulations, two popular open-source approaches inflated holiday effects by a fifth to almost a third.

Penalties and more history don't settle it

Statistical penalties and default assumptions are the usual fix for unstable results. They make the output consistent, but consistency comes from the settings rather than the evidence, so a different setting would land on a different split. Extra years of history rarely rescue it, because each year repeats the same pairing of holiday demand and bigger budgets, and confidence grows while accuracy doesn't.

The baseline can also swallow real marketing gains

Under-crediting is a risk too: a baseline that's too flexible can absorb growth marketing caused. A brand that invests in podcasts, CTV, and creators for six months may see branded search and direct traffic climb steadily. A model that reads that as a natural upward trend will under-credit the very campaigns that built it.

The same happens when a model assumes ads are equally effective all year. Any extra holiday effectiveness gets pushed into the seasonal baseline, so marketing looks weaker in the weeks it was working hardest.

Some controls subtract marketing's own effect

Adding more variables doesn't automatically help. If a model treats branded search volume, site traffic, or coupon redemptions as independent drivers of demand, it may be controlling for the results of the marketing it's trying to measure. A podcast that drives branded searches will look less effective if branded search is treated as a separate cause of sales.

How Prescient's model handles seasonality, promotions, and context

Prescient's model is built to separate outside demand from ad performance without forcing them into independent slices. Spend, the calendar, and promotions all push on the same underlying shopper interest in Prescient's model—none of them gets its own isolated slice—so the split isn't left to default settings.

  • It learns each holiday's impact on your business. Holiday effects come from your own history, informed by how similar businesses perform in the same periods, so a brand with a slow Black Friday isn't assumed to follow the typical retail curve.
  • Holidays affect both demand and effectiveness. Prescient's model treats a holiday as changing both baseline revenue and how well marketing works, so peak-season ad performance isn't just credited to the calendar or just to the ads.
  • Events evolve year to year. Black Friday this year isn't assumed to match last year's. The model adapts to changes such as promotional intensity, market conditions, and brand growth.
  • Promotional periods are modeled separately. Conversion rates, ROAS, cross-channel effects, and what happens after a promotion ends are accounted for rather than averaged into normal weeks.
  • Long-term trend is kept apart from seasonality.
  • Funnel roles are built in. When a brand campaign lifts branded search weeks later, that lift is credited back to the campaign through halo effects instead of disappearing into the trend line.
  • Returns follow your data. No channel is assumed to hit a ceiling before your results show one.

Traditional, open-source, and Prescient's models side by side

Each type of MMM handles outside factors differently. Here's a quick comparison.

What to compareTraditional MMMOpen-source MMMPrescient's model
How outside factors are treatedFixed controls inside a separate baselineSmoothing curves and controls in a separate baselineOutside factors act on the same demand as marketing
Spend timed to holidays and promotionsProne to crediting holiday demand to adsSplit set largely by penalties and defaultsSeasonal demand and ad effects modeled together
Ad effectiveness during peaksUsually assumed constantUsually assumed constantAllowed to change with holidays and promotions
PromotionsOften a simple on/off markerCan be added as a separate inputPromotional periods modeled separately, including after-effects
Upper-funnel growthCan be absorbed into the trendCan be absorbed into the trendCredited through funnel roles and halo effects
Peak budget advice (Prescient simulation)Not testedAbout 81% too high or 11.5% too lowRoughly 1% off

What the peak-season research shows

Prescient put this to the test in a simulated holiday season with a known best budget, pitting its model against two open-source MMMs (one regression-based, one Bayesian). The work is internal and not peer-reviewed, and scoring happened per channel since not every MMM measures campaigns.

Holiday outcomeRegression-basedBayesianPrescient's model
Recommended spend vs. the known best~81% over~11.5% under~1% off
Share of the possible result forfeited~32%~45%~6%
Error predicting each channel's holiday lift~0.56~1.06~0.28

The overspender read holiday shoppers as proof its ads were working. The underspender assumed marketing would run out of room early. Both had fit past sales well.

Why daily data makes promotions easier to untangle

How often a model reads your data affects how well it can separate a promotion from the ads around it. Consider a week where Meta spend rises on Wednesday, a 30%-off sale runs Friday and Saturday, and spend stays high through Sunday.

A weekly model sees one total for spend and one for revenue. A daily model can see that sales rose somewhat on Wednesday and Thursday before the sale began, jumped on Friday and Saturday, and settled on Sunday even though spend was still high. That gives it real evidence about what the ads did on their own and what the discount added.

Daily data is noisier, but weekly totals throw away exactly the timing differences that help separate overlapping effects. Prescient's model reads online data every day and retail data every week, matching how most retailers share it.

What omnichannel brands need to watch

Brands that sell on their own site, Amazon, and in stores face extra complications, because promotions and prices often differ by channel. A brand might run 20% off on its site, a separate Prime Day deal on Amazon, and a Target promotion, all while a national CTV campaign drives demand everywhere.

  • If the model only knows about site promotions, it may credit CTV with lift the retailer's own deal produced.
  • A site-only discount can pull sales away from Amazon, so site revenue alone overstates it.
  • Expanding retail distribution while raising ad spend can make marketing look stronger than it is.

Measuring all sales channels in one model, with each channel's promotions and distribution changes recorded, keeps these effects from getting mixed up.

What to give your MMM provider

The quality of your non-marketing data matters as much as your ad data. These are the records that help most.

RecordHelpful details
Promotional calendarStart and end dates, discount type and depth, eligible products, channels, and regions
Pricing historyList prices, temporary discounts, and permanent price changes
Product launchesLaunch dates and major assortment changes
InventoryStockouts and restocks
DistributionNew retail partners and store counts
Other marketingEmail and SMS send volume, loyalty campaigns

One rule of thumb: record what you did, not what happened as a result. "25% off, Friday through Sunday" is a cleaner input than "coupon redemptions," because redemptions are partly an outcome of your ads.

Questions to ask about how an MMM handles these factors

These questions will quickly show whether a provider has thought carefully about outside factors.

  • How do you separate seasonal demand from changes in ad effectiveness during peak periods?
  • How do you keep holiday and promotional demand from being credited to ads, and keep marketing gains from being absorbed into the baseline?
  • Can the model handle promotional depth, length, and channel, not just whether a sale was running?
  • How do you account for sales pulled forward by a promotion?
  • What happens when promotions and spend increases always happen at the same time?
  • How do you avoid controlling for things your marketing caused, such as branded search volume?
  • Can you account for different promotions and prices across your site, Amazon, and retailers?

Why this matters for forecasting

Getting this right isn't only about explaining last quarter. It determines whether your forecasts hold up. If a model says Meta returned 5x in November, the real question is whether more Meta spend in February will do the same.

The answer depends on what drove November. Holiday demand won't be there in February. A deep discount may not be running. Ad costs may be lower. A model that separated those factors can forecast February realistically, and one that didn't will promise November results in a February market. Our Media Forecasting feature uses what Prescient's model has learned about your seasonality when it projects results, so plans for a slow month aren't built on holiday performance.

Where Prescient comes in

If your biggest marketing moments line up with holidays, launches, and sales, Prescient was designed for exactly that problem. Our model learns how each holiday and promotion behaves for your brand, keeps seasonal demand separate from ad performance, and reports campaign-level Modeled ROAS across your site, Amazon, and retailers, refreshed daily.

Our Media Forecasting feature carries that into planning, so next November's budget is sized to what your ads contribute beyond the season itself. To see how much of your holiday lift your marketing actually earned, book a demo.

FAQs

How does an MMM separate seasonality from marketing?

An MMM includes seasonal patterns, such as holidays and recurring yearly cycles, alongside ad spend and estimates how much of each period's sales would have happened without the marketing. Ads get credit only for the lift above that level. Better models also recognize that ads can work differently during peak seasons, rather than assuming the same effectiveness all year.

Why do MMMs sometimes overstate marketing's impact during holidays?

Ad budgets go up in the same weeks holiday shoppers arrive, so spend and natural sales climb together. Many models can't reliably separate the two, and when holiday shoppers are booked as ad results, the model overstates marketing's impact. This is called baseline leakage, and it often leads to recommendations to overspend at peak.

How should promotions be included in an MMM?

Give the model a detailed promotional calendar, including start and end dates, discount depth, type, eligible products, and where it ran. A simple on/off marker misses much of what matters. A good model should also account for how promotions and ads amplify each other and whether a promotion pulled sales forward from later months.

Can an MMM under-credit marketing?

Yes. If a model's baseline is too flexible, it can treat growth caused by marketing, such as rising branded search after months of CTV or podcast ads, as a natural trend. Assuming ads work equally well all year can also push holiday effectiveness into the baseline, making marketing look weaker during its strongest weeks.

Do price changes affect MMM results?

They can. A price increase can raise revenue and ROAS even if ads aren't driving more demand, because each sale is worth more. Giving the model your pricing history, and measuring outcomes like orders or units alongside revenue, helps keep price effects from being mistaken for marketing performance.

Does daily data help an MMM measure promotions?

Often, yes. Daily data shows timing differences, such as spend rising a few days before a sale starts, that weekly totals hide. Those differences give the model evidence for separating what the ads did from what the discount did. Daily data is noisier, though, so the model still needs enough history and variation.

\