What happens when your MMM is missing marketing data?
When your MMM is missing marketing data, credit for that revenue shifts to other channels or the baseline. Here's which gaps matter most and how to fix them.
Linnea Zielinski · 13 min read
A lot of restaurants split tips based on the hours each person logs. It works fine until a bartender forgets to clock in for a busy Saturday. The tips she earned don't vanish. They get divided among everyone who did clock in, so the servers on shift that night look like stars and she looks like she barely showed up. The system just couldn't pay someone it couldn't see.
That's what happens when your marketing mix model (MMM) is missing marketing data. The revenue the untracked activity drove is still sitting in your sales numbers, and the model still has to explain it. So it hands that credit to something it can see: your baseline demand, a channel that happened to be running at the same time, seasonality, or the leftover noise. The tricky part is that the model can still predict total revenue well while giving the wrong channels credit for it.
That matters because your budget follows the credit. If a model hands an influencer push's results to branded search, you'll likely scale search and cut the creators who were doing the work, and you won't find out until performance slips. Knowing where your gaps are, and how your model treats them, is what lets you trust the recommendations for the channels you do track.
Key takeaways
- Missing marketing data doesn't remove revenue from your results. The model reassigns it to the baseline, to other channels, or to noise.
- A model can fit total revenue closely while misjudging which channels earned it, so a strong fit isn't proof of good attribution.
- The gaps that do the most damage come from activities that drive real revenue and tend to move at the same time as channels the model can see.
- Promotions, inventory changes, distribution changes, and missing sales destinations like Amazon or retail can distort marketing results just as much as a missing ad channel.
- Most traditional and open-source MMMs treat revenue as a sum of separate pieces, which makes them especially prone to pushing unmeasured lift into the baseline or onto tracked channels.
- When a gap can't be filled, proxies, sensitivity checks, and clear notes about uncertainty keep you from drawing conclusions the data can't support.
- Connecting all of your major media and revenue sources does more for MMM accuracy than almost any other single step.
Why an MMM needs to see the whole marketing picture
An MMM looks at how your revenue changes over time and works out how much of that change came from each marketing activity, how much came from demand you'd have had anyway, and how much came from things like holidays or promotions. In simple terms, most models think about it like this:
Revenue = baseline demand + each channel's contribution + other business factors + unexplained variation
Many traditional models treat that equation like a pie chart made of independent slices. Each channel gets its own slice, and the slices have to add up to the whole pie. If one slice is missing from the inputs, the pie doesn't get smaller. The other slices get bigger to fill the space.
Channels also don't work alone. A podcast ad might send someone to Google to search your brand name, and a creator video might lead to an Amazon purchase a week later. When the channel that started the chain is missing, the channel at the end looks like it did everything.
A quick example of credit moving
Here's a simplified look at how that plays out. Say a brand runs a podcast flight while it also raises its branded search budget, and revenue climbs by $120,000 over the period. The model has search data but no podcast data.
| Source of the lift | What actually drove it | What the model reports without podcast data |
|---|---|---|
| Branded search | $70,000 | $108,000 |
| Podcast | $50,000 | Not measured |
| Baseline or other factors | $0 | $12,000 |
| Total | $120,000 | $120,000 |
The total is right, but the story is wrong. Branded search looks roughly 50% more productive than it is, and the podcast that created much of that demand looks worthless. A budget plan built on these numbers would move money the wrong way on both channels.
The types of missing data MMMs run into
Not every gap is the same kind of problem. Missing an entire channel creates different issues than missing a few days of spend or not having impression counts. The table below covers the most common gaps and what each one tends to throw off.
| Type of gap | Common examples | What it tends to distort |
|---|---|---|
| Entire channels | Podcasts, connected TV, direct mail, out-of-home, sponsorships | Credit for that channel flows to the baseline or to channels running at the same time |
| Incomplete campaign histories | A campaign that launched mid-year, renamed campaigns, a new platform with only a few months of logs | Campaign-level results within a channel and how quickly the model learns new tactics |
| Missing spend or impressions | Gaps of days or weeks, flat monthly invoices with no daily breakdown | The timing of results, carryover estimates, and how spend relates to revenue |
| Untracked creator and affiliate activity | Gifted product, flat-fee creator deals, affiliate payouts tracked outside your ad stack | Paid social and search often absorb the credit |
| Missing revenue destinations | Amazon, Walmart, Target, or wholesale sales left out of the outcome data | The total value of marketing, especially upper-funnel channels that drive sales elsewhere |
| Missing non-marketing demand drivers | Promotions, price changes, stockouts, new store distribution, product launches | Marketing gets credit for demand it didn't create, or blame for drops it didn't cause |
Losing user-level tracking, like cookies or device IDs, isn't the same kind of gap. An MMM doesn't need to know which person saw which ad, only when marketing ran, how much ran, and how outcomes changed.
It also matters that missing doesn't mean zero, which the FAQs below cover in more detail.
What omitted variables do to your results
Statisticians call this problem omitted variable bias. In plain English, it's what happens when a model gives the wrong numbers to the channels it tracks because it missed something that also affected sales. The missing activity's revenue has to land somewhere, and it usually goes to one of these places:
- The baseline. Organic demand looks stronger than it is, which makes marketing look less necessary.
- A channel running at the same time. Whatever was active alongside the missing activity picks up its results.
- Seasonality, trend, or promotions. The model reads the lift as a calendar effect.
- Unexplained noise. The revenue gets treated as random, and the model's confidence drops.
Depending on how the missing activity lines up with everything else, a tracked channel can also end up with too little credit.
When missing data changes your decisions and when it just adds noise
How much a gap hurts depends heavily on whether the missing activity moves alongside something the model can see. Think of three versions of the same untracked creator program running next to Meta ads.
| Scenario | What usually happens | Risk to your decisions |
|---|---|---|
| Creator activity rises and falls with Meta spend | Meta absorbs the creators' results and looks more efficient | High: the model points you toward the wrong channel |
| Creator activity varies on its own schedule | Results show up as unexplained spikes or noise | Moderate: estimates get less precise, but the direction usually holds |
| Creator activity stays steady all year | Its ongoing contribution blends into the baseline | Moderate: the program looks unnecessary, and cutting it causes a surprise decline |
The first scenario is the one to worry about most, because it changes which way your budget moves. The other two mostly make results fuzzier or hide a channel's value inside the baseline. A program that's 5% of your budget can distort results badly if it always launches alongside your biggest channel.
The effects reach past attribution
A gap doesn't just warp the credit report. It can throw off most of the outputs marketers lean on to plan budgets.
- Return on ad spend and acquisition cost. A channel that absorbs extra credit shows a higher ROAS and a lower cost per customer than it really earns.
- Response curves. If an untracked activity tends to ramp up when a tracked channel's spend is highest, that channel will look like it keeps paying off at high spend levels. The model may miss where diminishing returns really start.
- Forecasts. A forecast assumes the missing activity keeps doing what it did before, so changing or stopping it can throw the forecast off.
Which missing variables matter most
Three questions tell you how much a missing input is likely to hurt: how much revenue it drives, how much it changes over time, and how often it moves at the same time as other inputs. The inputs below tend to score high on at least two of those.
| Missing input | Why it matters | Priority |
|---|---|---|
| Revenue from all sales destinations | If Amazon or retail sales are missing, the model can't see the full value of channels that drive purchases there | Essential |
| High-spend channels | Large investments leave large amounts of revenue to be reassigned | Very high |
| Channels that run alongside others | Overlapping timing is what lets credit slide onto tracked channels | Very high |
| Major promotions and price changes | Brands often raise ad spend during sales, so ads can take credit for the discount's lift | Very high |
| Inventory changes and stockouts | A sales drop from empty shelves can look like a channel stopped working | High |
| Distribution changes | New retail doors or marketplace launches raise demand in ways marketing didn't create | High |
| Campaign-level detail within a channel | Without it, the channel total may be fine but you can't tell which campaigns to grow | High for budget planning |
How to find out what your model is missing
Gaps rarely announce themselves, and standard fit statistics won't catch them. A high R² only tells you the model explained total revenue, not that it split the credit correctly. These steps tend to surface problems faster.
- Audit your integrations against your finance records. Pull every marketing line item from your ledger or invoices for the past year and check that each one maps to an input in the model.
- Compare your media calendar with the model's inputs. Look for launches, flights, creator drops, events, and PR moments that don't show up anywhere in the data.
- Review unexplained performance changes. If revenue jumped and the model credits the baseline, ask what else was happening that week.
- Look closely at periods where the model's predictions were weakest. Repeated misses around the same time each year, or around specific launches, usually point to a missing driver.
- Watch what happens when you add a new channel. If adding podcast data makes branded search ROAS drop sharply, the old model was likely overcrediting search.
- Check results against holdout or geo tests. A big gap between a credible test and the model's estimate is a reason to dig in.
Signs like a ROAS that seems too good to be true or a baseline that seems far too high don't prove there's a gap, but each one is worth a closer look.
What to do when the data can't be recovered
Sometimes the data simply doesn't exist, like when a creator agency never kept daily records. You can still get useful results if you handle the gap honestly.
- Use proxies for timing and volume. Post dates, email send counts, podcast download estimates, direct mail drop dates, or impression counts can show when an activity reached people, even without daily spend.
- Roll up incomplete campaigns. If campaign-level history is patchy, measuring at a grouped or channel level may give you a more reliable answer, even if it's less detailed.
- Run sensitivity checks. Compare results with and without a proxy, or across different assumptions about the missing activity. If a channel's results swing widely, treat its numbers with care.
- Write down the uncertainty and limit conclusions. Note which channels are likely absorbing credit from a known gap, and avoid big budget moves based on credit that could be borrowed.
- Use incrementality tests as a check. A test can show whether a channel drives real lift, but it covers one window and one spend level. It can't replace data about when and how much that channel ran across the rest of your history.
One caution: uncertainty ranges don't automatically account for missing variables. A model can report a tight, confident range around an answer that's off because of a gap it doesn't know about.
Comparing traditional, open-source, and Prescient's approaches
The structure of a model has a lot to do with how gracefully it deals with missing data. Here's how the three main approaches compare.
Traditional vendor MMMs
Most traditional MMMs use some form of regression that treats revenue as the sum of a baseline plus separate channel pieces. When something is missing, the math has to make the remaining pieces add up to the total, so unmeasured lift gets pushed into the baseline or the error term, or spread across whichever channels happened to be active. Many also want two or more years of weekly data, which newer channels rarely have.
Open-source frameworks
Open-source MMMs add techniques like ridge regression or Bayesian priors to keep their estimates stable. That steadies noisy results, but it doesn't bring back a missing signal. When several splits of credit explain revenue equally well, these techniques pick the tidiest one under the model's built-in assumptions, and adding more data on tracked channels only makes the model more confident in that skewed answer. Many vendors build on these frameworks, so the same limits often show up in paid tools.
Prescient's model
Prescient's model is built differently. Instead of treating revenue as independent slices, it tracks an underlying level of customer demand that rises and falls over time, with marketing pushing that demand up and other channels helping turn it into sales. When the model sees a surge it can't tie to a tracked input, it's designed to treat that as a shift in the underlying demand, rather than forcing it onto whichever tracked channel happened to be active.
A few other design choices help when data is incomplete:
- Funnel roles. The model separates channels that create demand from channels that capture it. That makes it harder for lower-funnel channels like branded search to claim credit for demand that something further up the funnel created.
- Continuity across gaps. Because the model follows demand as it builds and fades over time, a short window of missing spend doesn't wipe out what it has learned about how a channel's effect carries forward.
- Shared learning for short histories. A Bayesian hierarchical structure with priors informed by benchmarks from similar businesses helps the model produce reasonable estimates for newer campaigns. Campaigns can be modeled once they've had spend on seven or more days within the last 365.
- Separate handling for holidays and promotions. Holiday and promotional effects are modeled on their own, so a sale is less likely to be mistaken for ad performance.
None of this means the model can measure a channel it never sees. A missing channel's own results still can't be reported. What the design aims to do is keep that gap from scrambling the results for everything else.
| Traditional vendor MMMs | Open-source MMMs | Prescient's model | |
| Basic structure | Baseline plus separate channel pieces | Baseline plus separate channel pieces, with ridge or Bayesian stabilizing | Underlying demand that marketing builds and channels convert |
| Where unmeasured lift goes | Baseline, error term, or overlapping channels | Baseline, trend terms, or whatever the priors favor | Treated as a shift in underlying demand |
| Risk to tracked channels | High when timing overlaps | High, and more data on tracked channels doesn't fix it | Lower, since funnel roles limit how much credit lower-funnel channels can take |
| Short history | Often wants two or more years | Needs a lot of history and tuning | Shared learning across campaigns and similar businesses |
| Refresh speed | Often quarterly | Depends on the team running it | Daily online, weekly retail |
What Prescient's peak-season research found
Prescient's team tested how these approaches hold up during peak season, when large demand surges are easy to misassign to the baseline or to the wrong channel. The research used simulated data where the true answer was known, so each model's budget plan could be scored against the best possible plan. Because many MMMs don't measure at the campaign level, we compared the models channel by channel. The paper has not been peer-reviewed.
Here's how each model's peak-season recommendations stacked up:
- A regression-based model, similar to many traditional and open-source tools, recommended overspending by about 81% and gave up about 32% of the achievable revenue.
- A Bayesian model, similar to newer open-source tools, recommended underspending by about 11.5% and gave up about 45% of the achievable revenue.
- Prescient's model came within about 1% of the ideal spend level and gave up about 6% of the achievable revenue.
The study wasn't built to test missing data, but it measured the same weakness: what happens when a model pins a demand surge on the wrong source.
Where Prescient comes in
Prescient's view is simple: the more of the factors that influence how your business sells and earns revenue the model can see, the better it can forecast what will happen next. Missing channels and untracked influencer activity are some of the most common reasons we see unexplained swings in results. That's why onboarding with Prescient is about more than connecting an ad account. It means connecting all of your major media and all of your revenue sources, including 112 integrations that cover platforms like Amazon Ads, Walmart Connect, Target's Roundel, creator platforms, affiliate networks, and podcasts.
From there, the model updates daily for online data and weekly for retail, picks up halo effects when a campaign drives sales on another channel or in stores, and feeds our Media Forecasting feature so you can see where to add or pull budget across channels and campaigns. Curious how your current data gaps might be shaping your results? Book a demo to chat through it with our team of experts.
FAQs
Can an MMM still be useful if some marketing data is missing?
Yes. An MMM doesn't need perfect data to be helpful, but you need to know what it can and can't see. If the gaps are small, steady, or unrelated to your big channels, the results for your tracked channels may hold up well. If a missing activity is large and tends to run alongside channels you do track, treat those channels' results with more caution until you close the gap or test it.
How much historical data does an MMM need?
It varies, because it depends on how often your data is recorded, how many channels you run, and how much your spend actually changes. Traditional weekly models often ask for two or more years. Daily models can learn from a shorter window, though a year of daily data isn't the same as 365 fully separate data points. Prescient can start modeling a campaign after just seven days with spend inside a 365-day window, using what it learns from similar campaigns and businesses to fill in the early picture.
Should missing spend be entered as zero?
No. A zero tells the model nothing ran, while a blank means you don't know what ran. Entering zeros for unknown days teaches the model that sales held up without that channel, which can make it look less valuable than it is. It's better to leave those days flagged as missing, estimate them from known totals, or use a proxy like post dates or impressions.
Can incrementality tests make up for a missing channel?
They help, but they can't fully replace the data. A holdout or geo test shows whether a channel drove lift during the test window and at that spend level. It doesn't tell the model when and how much that channel ran during the rest of your history, which is what the model needs to keep it from crediting other channels. Tests work best as a check on your model, not a stand-in for missing inputs.
Can Prescient measure individual influencers?
Prescient measures marketing at the campaign level, so it can measure an individual creator only when that creator's work runs in its own campaign. If several creators share one campaign, the model sees their combined results. Connecting your creator platforms or tracking each partnership as its own campaign gives the model the clearest view.