- MTA needs identity resolution to work. If more than 30-40% of your traffic is on Safari or iOS (ITP and ATT both block persistent identifiers), your touchpoint data is already incomplete before you build a single model.
- MMM doesn't need cookies or pixels. It regresses spend against outcomes at the aggregate level, so it survives privacy changes that broke MTA in 2021, but it needs 18-24 months of weekly data and can't tell you which ad creative worked.
- Under $500K/month in media spend, skip both and run holdout tests. MMM needs statistical power you don't have yet, and MTA will just measure your retargeting claiming credit for organic sales.
- The real fix is running MMM for budget allocation across channels and incrementality tests to validate what MMM says, then using MTA (if you still need it) only for in-platform bid optimization.
The CMO wants to know if TikTok is working. The performance team pulls up the MTA dashboard: TikTok drove 340 conversions last month, third-touch attribution, cost per acquisition of $38. Looks great. Budget gets approved for another $50K.
Finance pulls up their model. It says TikTok's incremental contribution to revenue is close to zero. Same month, same channel, same company.
Nobody in the room can explain the gap without pointing fingers. Marketing says finance's model is a black box. Finance says marketing's dashboard is counting touches, not causation. Both are partly right, and the argument keeps happening because nobody stopped to ask what each tool is actually built to answer.
Two Different Questions, Not Two Versions of the Same Answer
Multi-touch attribution and marketing mix modeling aren't competing methods for the same measurement problem. They answer different questions, at different levels of granularity, using different data.
MTA answers: "Of the users who converted, which specific touchpoints did they interact with, and in what order?" It's a user-level, event-based approach. It needs to see individual sessions, tie them to a person (or a device, or a cookie), and stitch a path together.
MMM answers: "Given total spend by channel over time, how much of the outcome (revenue, signups, whatever) can be explained by each channel's spend level?" It's an aggregate, statistical approach. It doesn't know who converted. It knows that in weeks when paid search spend went up 20%, revenue went up 4%, and it isolates that relationship using regression across dozens of variables (seasonality, price changes, competitor activity, weather, whatever you feed it).
One is built from the bottom up, one user at a time. The other is built from the top down, one week at a time. That difference explains almost every disagreement you'll see between a marketing dashboard and a finance model.
MTA and MMM are not measuring the same unit of analysis. MTA counts touchpoints for converters. MMM estimates counterfactual lift for a whole channel. Even with perfect data, you should expect different numbers. Treating a gap as an error to reconcile misses the point.
Where MTA Breaks
MTA depends on being able to see the user across sessions and devices. That requirement got harder to meet every year since 2017, and it broke almost completely in 2021.
Safari's Intelligent Tracking Prevention caps first-party cookie lifespan at 7 days for cookies set via document.cookie from script (and blocks third-party cookies entirely). Apple's App Tracking Transparency, live since iOS 14.5, means any app asking to track you across other apps and websites needs an explicit opt-in, and something like 20-25% of US iOS users grant it, depending on the app category.
Run the math on your own traffic. Pull your GA4 or ad platform breakdown by browser and OS. If Safari and iOS Chrome (which uses WebKit and inherits ITP) make up 35% of your sessions, you're building an attribution model on incomplete data for over a third of your users before you've written a single line of model logic. The missing third isn't random either. It skews toward users who value privacy, which correlates with income and platform preference in ways that bias your channel mix conclusions.
The other failure mode is simpler and shows up even with perfect tracking: MTA measures touches, not causation. A user who was already going to buy sees a retargeting ad on day 6 of an 8-day research process. MTA gives that ad credit. The ad didn't cause the sale. It just happened to be there. I've written about this specific failure in retargeting before, an 8x attributed ROAS campaign that delivered a real incremental ROAS of 0.08x when we ran a holdout test. MTA had no way to catch that. It's not built to.
What MTA Is Actually Good For
Inside a single platform, with a single identifier, over a short window, MTA still earns its keep. Google Ads' data-driven attribution model, working off logged-in Google account signals within Search and YouTube, gives you a reasonable read on which keywords and ad groups within that platform deserve more budget. That's a narrower and more honest use case than "which of our 9 channels drove this quarter's revenue," and it's the one MTA is still suited for.
Where MMM Breaks
MMM doesn't care about cookies, doesn't touch a pixel, and doesn't know if a user is on Safari or Chrome. It works off aggregated spend and outcome data, usually at a weekly grain. That's the whole appeal: it survives privacy changes that gutted MTA.
But MMM has its own requirements, and they're statistical, not technical.
You need 18 to 24 months of historical weekly data at minimum to get a model with usable confidence intervals. Fewer data points and the regression can't distinguish real channel effects from noise, especially when channels are correlated (you tend to increase paid search and paid social budgets around the same promotional pushes, which makes it hard for the model to separate their individual contributions, a problem called multicollinearity).
You also need enough spend variation to identify an effect. If your Google Ads budget has sat within 8% of the same weekly number for two years, the model has almost nothing to learn from. MMM needs you to have turned dials up and down over time, on purpose or by accident, so it can observe what happened to the outcome when you did.
If every channel's weekly spend has been flat for 18 months, don't expect a useful MMM. The model identifies channel effects from variation in spend over time. Flat spend gives it nothing to regress against, and you'll get wide confidence intervals or a model that just reflects your priors back at you.
MMM also can't tell you which ad, which audience, or which creative worked. It operates at the channel or campaign-group level, not the individual ad level. If the question is "should we kill this Meta ad set," MMM has no opinion. If the question is "should we shift 15% of budget from Meta to YouTube next quarter," that's exactly what it's for.
The Spend Threshold Nobody Tells You About
Under roughly $500K a month in total media spend, both approaches struggle, and for different reasons.
MTA struggles because the identity resolution problem doesn't shrink with your budget. A $50K/month advertiser has the same Safari and iOS traffic mix as a $5M/month one, so the same fraction of paths are broken, but you have far less volume to compensate with, and the noise in any per-touch analysis swamps the signal.
MMM struggles because statistical power scales with data volume and spend variation, and a smaller advertiser usually has fewer channels, less budget to move around, and a shorter operating history. Running a formal MMM on $30K/month spread across 3 channels will mostly produce confidence intervals wide enough to be useless.
What actually works at that spend level: holdout tests. Turn a channel off for a geo or a segment for 2-4 weeks, compare conversion rate (not raw revenue) against a control group that kept seeing the ads, and you get a direct read on incremental lift without needing 18 months of history or a data science team. It's the same logic as the retargeting holdout test, applied to whichever channel is in question. It's cheap, it's fast, and it answers the actual question ("did this cause anything") instead of a proxy for it.
What to Actually Do
Stop treating this as a binary choice. For most companies spending over $500K/month across 4+ channels, the answer is both, used for different jobs.
Run MMM quarterly (or monthly if you have the spend and data volume) to set channel-level budget allocation. That's the tool built to answer "how much should we spend on paid social vs. paid search vs. TV next quarter," because it's not fooled by identity loss and it's designed for exactly this aggregate question.
Run incrementality tests (holdouts, geo-lift experiments) continuously, at least once a quarter per major channel, to validate what the MMM is telling you and catch cases where a channel's measured contribution is actually selection bias in disguise, the way the retargeting campaign was.
Use MTA, if you use it at all, only inside a single platform for bid and creative optimization, not for cross-channel budget decisions. Google's data-driven attribution helping you decide which search terms to bid up is a legitimate use. Google's attribution data helping you decide TikTok's total quarterly budget is not, because MTA was never built to compare across platforms with fundamentally different identity resolution and different degrees of missing data.
If your CMO and your CFO are getting different numbers from different tools, that's not a bug you fix by picking a winner. It's two instruments measuring two different things, and the fix is making sure everyone in the room knows which question each number is actually answering before the budget conversation starts.