Marketing Measurement ·

User-level attribution vs. MMM: Evaluating MTA, traditional, open-source, and Prescient's approach

Compare user-level attribution (MTA), traditional MMM, open-source MMM, and Prescient's approach, including how each handles privacy, saturation, and more.

Linnea Zielinski · 13 min read

Listen
0:00 / 0:00
AI-generated audio
User-level attribution vs. MMM: Evaluating MTA, traditional, open-source, and Prescient's approach

Wildlife biologists have two broad ways to estimate how many deer live in a forest. They can put GPS collars on individual animals and follow each one, which gives incredibly detailed paths until collars fail, deer wander out of range, or permits limit how many animals they can tag. Or they can count from the air and from tracks, then estimate the total population from those patterns. The aerial approach doesn't need to follow anyone, but it's only as good as the assumptions behind it. A biologist who forgets that deer move to lower ground every winter will think the herd collapsed in December and boomed in April, when it was the same herd.

Marketing measurement has the same split, and the same catch. User-level attribution follows individual shoppers, and privacy changes have taken away many of its collars. Marketing mix models (MMMs) measure from aggregate patterns instead, but traditional, open-source, and newer approaches make very different assumptions about how marketing works, and those assumptions decide whether the results hold up. For brands choosing a measurement partner, that choice shapes every budget decision that follows, so it's worth understanding what each approach can and can't see before committing to one.

Key takeaways

  • Multi-touch attribution (MTA) follows individual shoppers through clicks, cookies, and device IDs, and privacy changes have left it with growing gaps that tend to over-credit last-touch channels like branded search and retargeting.
  • Marketing mix models avoid user tracking by measuring from aggregated spend and revenue, which makes them more durable as privacy rules shift, but their accuracy depends heavily on how they're built.
  • Traditional MMMs often split revenue into independent slices for each channel, which can confuse holiday demand with marketing impact or bury real marketing impact in baseline demand.
  • Open-source MMMs give teams control and code visibility, but they usually share the same independent-slice structure, rely on statistical penalties to choose between equally good explanations, and use saturation curves that flatten by design.
  • Prescient's marketing mix model treats demand as something marketing builds over time, so it can represent funnel direction, seasonal changes in efficiency, campaign-specific decay, and response shapes that follow the data.
  • Each approach is best for a different job, so many brands use MTA or platform reporting for in-platform decisions and an MMM for cross-channel budget decisions.
  • Evaluate any approach on whether it can separate what marketing added from what would've happened anyway, not on how detailed or precise its dashboard looks.

Four ways to measure marketing, at a glance

It helps to see the four approaches side by side before digging into each one. Open-source MMM isn't a separate method so much as a different way of building an MMM, but it's worth evaluating on its own because its tradeoffs are so different.

User-level MTATraditional MMMOpen-source MMMPrescient's MMM
What it usesClicks, cookies, device IDs, and conversionsAggregated spend and revenueAggregated spend and revenueAggregated spend, impressions, and revenue
How it assigns creditRules or algorithms that split credit across tracked touchpointsIndependent slices of revenue for each channelIndependent slices, with statistical penalties to pick one answerDemand modeled over time, with funnel direction built in
Who runs itVendor or in-house teamVendorYour own data science team or a vendor that uses it as the base of their platformPrescient
Typical detailUser and touchpoint levelUsually channel levelDepends on your team's setupCampaign level, updated daily
Best suited forTracked digital journeys and in-platform tweaksPeriodic cross-channel planningTeams that want full control of the modelCross-channel budgets across website, Amazon, and other destinations

User-level attribution: detailed, but losing sight

Multi-touch attribution set out to show every step of a customer's path. By tracking individual users across devices with cookies, mobile ad IDs, and pixels, MTA tools split credit for each sale across the ads a shopper saw or clicked along the way.

Why MTA's view keeps shrinking

A series of privacy and platform changes has broken many of the links MTA depends on. Here's what changed and how it affects user-level tracking.

ChangeEffect on user-level tracking
Safari Intelligent Tracking Prevention and Firefox tracking protectionsThird-party cookies blocked by default for a large share of browsing
Apple's App Tracking Transparency (iOS 14.5 and later)Apps need permission to track users across other companies' apps and sites
Privacy laws like GDPR and CCPA/CPRAConsent and opt-out requirements reduce trackable data
Chrome's cookie plansGoogle kept third-party cookies and retired most Privacy Sandbox technologies, so tracking is patchy rather than gone

MTA didn't stop working overnight. Instead, more and more journeys are only partly visible, and the parts that stay visible are usually the last few steps before a purchase.

The pull toward last-touch credit

When MTA can only see the end of a journey, it gives most of the credit to the channels sitting there, like branded search, retargeting, and email. Upper-funnel channels such as YouTube, CTV, and podcasts often reach people who never click, so they get little credit even when they created the interest that led to the sale.

That creates a predictable distortion—one that gets worse as tracking gets patchier—where campaigns that capture demand look like top performers, and campaigns that create demand look like they're wasting money.

Where MTA can't see at all

Some sales are hard for user-level tracking to reach no matter how good the setup is:

  • Amazon purchases driven by ads running somewhere else.
  • Retail store sales influenced by digital ads.
  • CTV, podcast, and audio exposure, where viewers and listeners rarely click.
  • Halo effects across channels, like a YouTube campaign lifting branded search a week later.

Where MTA still helps

None of this makes MTA useless. It can still show how tracked digital journeys unfold and help compare campaigns or creative within channels it can see well. It's just an incomplete foundation for deciding how budget should be split across channels.

Traditional MMM: aggregate data with a pie-chart assumption

Marketing mix modeling avoids the tracking problem by looking at aggregated data, like daily spend and revenue totals, instead of individual journeys. That makes it far more durable as privacy rules change. But the way traditional MMMs split up revenue creates its own set of problems.

Revenue gets sliced into independent pieces

Most traditional MMMs treat revenue like a pie. They carve each day's sales into a baseline slice plus a separate slice for each channel, then add them back together. That assumes channels work independently, when in reality a CTV campaign can lift branded search and a Meta campaign can drive Amazon sales.

Spend rises and falls with demand

Brands naturally spend more during holidays, launches, and promotions, which is exactly when shoppers are already more interested. When spend and demand rise together, there are many ways to divide the resulting sales between baseline and marketing that all fit your history equally well, and the data alone can't tell the model which one is right.

Baseline leakage cuts both ways

That ambiguity leads to baseline leakage, where credit slides between baseline demand and marketing. Sometimes holiday demand gets credited to ads, making campaigns look more efficient than they are. Other times, real marketing impact gets absorbed into seasonality or trend, making campaigns look weaker than they are. Either way, budget decisions built on those numbers are off.

Granularity and timing can lag

Many traditional MMMs work at the channel level and refresh results a few times a year. That's useful for annual planning, but it won't tell you whether to move money from a Meta retargeting campaign to a Meta prospecting campaign this month.

Open-source MMM: control and visibility, with familiar limits

Open-source MMMs are publicly available modeling frameworks that a brand's own data science team can set up and run. They've become popular with teams that want to own their measurement and see exactly how the model works. They’re also used as the foundation for the model that some vendors offer through their platforms.

What open-source frameworks offer

There are real benefits to building in-house:

  • Code visibility, so your team can inspect the method rather than relying on a vendor's description.
  • Customization, including the ability to add variables and test different assumptions.
  • Control over data, since the model can run inside your own infrastructure.
  • No software license fee for the framework itself.

Statistical penalties pick one answer, not the right one

Most open-source MMMs share the same independent-slice structure as traditional ones, so they face the same problem when spend rises with demand. To get a stable answer, they apply statistical penalties or starting assumptions that nudge the model toward one of the many explanations that fit the data. That makes the output consistent, but it doesn't make it correct. The model picks one plausible split and presents it with confidence, and more data usually just makes it more confident in the same split.

Saturation curves are built in

Most open-source frameworks use response curves that flatten by design, so every campaign eventually looks maxed out. Those curves also place the flattening near the highest spend you've tried, which often reflects a budget cap rather than a market limit. Prescient's research on saturation assumptions found that a simple straight-line model matched real campaign behavior more closely than the most common saturation curves, which suggests many brands are underspending on campaigns that could scale.

The real cost of ownership

An open-source framework is free to download, but running it well isn't. Teams need data engineering, model setup, validation, ongoing maintenance, and someone who can explain results to leadership. Code being public also doesn't mean results are transparent in practice, since the choices your team makes when configuring the model matter as much as the code itself.

Prescient's MMM: modeling how demand actually builds

Prescient's marketing mix model uses the same privacy-safe aggregated data as other MMMs, but it doesn't split revenue into independent slices. Instead, our model treats demand as something marketing builds, sustains, and lets fade over time, so baseline demand and marketing impact are estimated together rather than carved apart.

That structure lets our model reflect how marketing works in practice:

  • Funnel direction: Upper-funnel campaigns build demand that lower-funnel campaigns, like branded search and retargeting, often capture. Our model credits the campaign that created the demand, not just the one that captured it.
  • Efficiency that changes over time: The same spend can perform very differently in July and December, so our model treats a campaign's efficiency as something that shifts with seasons and promotions rather than fixed.
  • Decay that varies by campaign: A CTV campaign's effects can build and fade over weeks, while a retargeting campaign's effects show up quickly, and our model lets each follow its own pattern.
  • Response shapes that follow the data: If a campaign is still returning at a steady rate, our model can show that instead of forcing it into a saturation curve.

Results are campaign level, updated daily, and include halo effects on branded search, direct traffic, and Amazon. Each result comes with a model confidence score, and our Media Forecasting feature shows what happens when you scale a campaign or shift budget into it from other channels.

What these differences can cost at peak season

The gap between approaches gets widest during peak periods like Black Friday and Cyber Monday, when budgets are biggest and spend rises alongside natural demand. That's when assumptions about baseline demand and saturation do the most damage.

In a simulation benchmark Prescient ran at the channel level, with correct answers built into the data, two widely used open-source MMMs gave very different peak-season advice. The benchmark hasn't been peer-reviewed, but the pattern is worth knowing:

ModelPeak-season budget recommendation vs. the correct amountMain reason
Open-source model AOverspend by about 81%Gave ads credit for holiday demand that would've happened anyway
Open-source model BUnderspend by about 11.5%Saturation curves underestimated how far campaigns could scale
Prescient's modelWithin about 1%Estimated baseline demand and response shapes together

The 81% overspending finding has been reported publicly. Errors that look moderate in a model's averages can turn into big budget mistakes once someone acts on them during the most expensive weeks of the year.

The full comparison

Here's how the four approaches compare on the issues that matter most to an evaluation committee.

User-level MTATraditional MMMOpen-source MMMPrescient's MMM
Exposure to privacy changesHighLowLowLow
Credit for lower-funnel channelsOften over-credits last touchProne to baseline leakageDepends on statistical penaltiesCredits the campaign that created the demand
Spend that rises with demandNot addressedOften splits credit arbitrarilyPicks one plausible answerEstimates baseline and marketing together
Response curvesNot applicableFixed curvesCurves that flatten by designShapes that follow the data
Halo effects and other destinationsLimited to tracked pathsOften not measuredOften not measuredMeasured across channels and Amazon
Who maintains itVendor or in-houseVendorVendor or in-housePrescient

How the approaches compare on privacy

Privacy is one of the clearest differences, though it's worth being precise about what "privacy-safe" means. Aggregate models don't need to track individuals, but they still handle your first-party business data, so data handling still matters.

User-level MTAAny MMM
Requires tracking individuals?UsuallyNo
Relies on cross-device matching?OftenNo
Affected by cookie and ID restrictions?SignificantlyOnly indirectly, if platform data inputs get noisier
Privacy law obligations?SignificantStill apply to how business data is handled

How to choose the right approach

The right choice depends on the decision you're making, the channels you run, and the resources your team has. This table matches common decisions to the approach best suited for each.

DecisionBest fit
Which tracked keywords or creative to optimizeMTA or platform reporting
How a tracked digital journey unfoldsMTA
How much revenue each campaign added across channels and destinationsA well-built MMM
Whether to shift budget from search into CTV or podcastsA well-built MMM with forecasting
Whether to build measurement in-houseOpen-source MMM, if you have the team to run it well

Whatever you're considering, a few questions help reveal how an approach really works:

  • For MTA: What share of our customer journeys can you actually see, and how representative are those journeys of everyone who buys?
  • For a traditional vendor: How does your model separate holiday demand from marketing impact, and how often do results update?
  • For open source: Who will configure, validate, and maintain the model, and who's accountable when results look implausible?
  • For any MMM: How do you handle funnel direction, delayed effects, and saturation, and how do you show uncertainty?

Using more than one approach

Many brands don't pick just one. A common setup gives each approach a clear job: platform reporting and MTA for in-platform optimization, and an MMM for cross-channel budgets and forecasting. Incrementality tests can add another data point, but Prescient's view is that these tests are locally accurate and globally inaccurate, since they describe one campaign in one specific window. That's why Prescient shows brands their model's accuracy with and without a test included, and the brand decides whether to build it in.

The important part is deciding in advance which source governs which decision, so teams aren't switching to whichever number supports the plan they already wanted.

Common mistakes when comparing measurement approaches

A few assumptions trip up evaluation committees again and again. Watching for them makes the comparison much more useful.

  • Treating MTA's detail as accuracy. A precise record of a click doesn't tell you whether that click changed anything.
  • Assuming any MMM measures what marketing added. That depends on how the model is built, not on the label.
  • Assuming open source means transparent results. Public code doesn't explain the choices your team made configuring it.
  • Judging a model by how well it predicts revenue. A model can predict totals well while splitting credit between channels incorrectly.
  • Ignoring total cost of ownership. Free software still requires people, infrastructure, and maintenance.
  • Comparing numbers with different definitions. Attributed revenue, modeled revenue, and website-only revenue aren't the same metric.

A plain-English guide to the technical terms

Measurement evaluations come with plenty of jargon. This quick reference can help translate for an evaluation committee.

Technical termWhat it meansWhy it matters
User-level tracking vs. aggregate modelingFollowing individual click trails vs. measuring from overall spend and revenue patternsExplains why pixel tools lose accuracy as tracking gets restricted, while aggregate models don't depend on it
Separable or additive modelTreating sales like a pie chart of independent slicesExplains why many MMMs struggle when ad spend and holiday shopping rise together
Statistical penaltiesRules that push a model toward one answer when several fit the dataShows why stable-looking results don't guarantee the model found the right split
Optimization regretRevenue lost by following a model's flawed budget advicePuts a dollar figure on the cost of a weak measurement approach

Where Prescient comes in

Prescient's marketing mix model measures from aggregated data, with no user tracking, and it was built to avoid the structural problems that trip up traditional and open-source MMMs. Our model estimates baseline demand and marketing impact together, credits the campaigns that create demand, lets each campaign's decay and response shape follow the data, and measures halo effects on branded search, direct traffic, and Amazon, all at the campaign level with daily updates.

That gives omnichannel brands one consistent view for cross-channel budgets, a model confidence score for every result, and our Media Forecasting feature to plan what happens when they scale or shift spend. If you want to see how your campaigns look under a model built for how marketing actually works, book a demo.

FAQs

Is multi-touch attribution still worth using?

Yes, for the right jobs. MTA can still show how tracked digital journeys unfold and help compare campaigns and creative within channels it sees well. But privacy changes have left it with large gaps, especially for CTV, podcasts, Amazon, and retail, and it tends to over-credit the last steps before a purchase. That makes it a weak foundation for deciding how budget should be split across channels.

What's the main difference between MTA and a marketing mix model?

MTA follows individual shoppers and splits credit for each sale across the touchpoints it can track. A marketing mix model looks at aggregated spend and revenue over time and estimates how much each campaign contributed, without tracking anyone. MTA answers which tracked ads came before a sale, while an MMM estimates how much revenue each campaign actually added.

Are open-source MMMs as accurate as commercial MMMs?

It depends far more on how a model is built and run than on whether it's open source or commercial. Many open-source frameworks share the same independent-slice structure as traditional MMMs, rely on statistical penalties to pick between equally good explanations, and use saturation curves that flatten by design. A well-resourced team can get real value from open-source tools, but accuracy still comes down to the model's assumptions and how carefully it's set up and maintained.

Can I use MTA and MMM together?

Many brands do. A common approach uses platform reporting or MTA for in-platform decisions, like which creative or audience to scale within Meta, and an MMM for cross-channel budget decisions and forecasting. The key is deciding ahead of time which source governs which decision, so conflicting numbers don't turn into debates about whose dashboard is right.

How can I tell whether an MMM is accurate?

Look beyond how well it predicts total revenue, since a model can match totals while splitting credit incorrectly. Ask how it handles funnel direction, delayed effects, and spend that rises with demand. Ask whether it's been tested on simulated data where the right answers are known, how it shows uncertainty, and how its results hold up across reasonable assumptions and known business events.