The flat case, where nothing interesting happens

Let’s start with a toy model where Sales is our response and everything else is a driver that points straight at it. I’ve written the contribution of each driver on the edge, as a percentage of total Sales volume.

flowchart LR

I2("Base (Sales)")
S2("Season (Sales)")
M(Media)
B(Brand)
S([Sales])

I2 --> |50%| S
S2 --> |0%| S
M --> |20%| S
B --> |30%| S

Read the graph right to left: Sales is explained by Base, Season, Brand, and Media, contributing 50%, 0%, 30%, and 20% respectively. Season sits at zero here only because we’re pretending we have exactly enough weeks for the positive and negative swings to cancel out. In the real world it won’t, but that’s beside the point.

This is easy. There’s one path from each driver into Sales, the numbers add to 100%, and there’s nothing to argue about. This post is dedicated to the case where things stop being that nice.

Adding a level: Direct attribution

Now let’s make it interesting. In the next graph we lift Brand up to being a response variable in its own right, with its own drivers. Media now affects Brand and Sales.

flowchart LR

I1("Base (Brand)")
I2("Base (Sales)")
S1("Season (Brand)")
S2("Season (Sales)")
M(Media)
B([Brand])
S([Sales])

I1 --> |60%| B
I2 --> |50%| S
S1 --> |30%| B
S2 --> |0%| S
M --> |10%| B
M --> |20%| S
B --> |30%| S

Same question: how much of Sales volume is attributed to each variable? Well, Base (Sales) is 50%, Season (Sales) is 0%, Brand is 30%, and Media is 20%. The numbers still add up.

But hang on. That’s the same Media contribution as before, even though Media now also helps build Brand, which in turn drives Sales. Something’s off. Or rather, we’ve silently picked a particular definition of attribution without saying so out loud. Let’s name it.

In Direct attribution we only count drivers one edge away from the response we’re decomposing. So when we attribute Sales, Base (Brand) and Season (Brand) never enter the picture because they don’t touch Sales directly. Media keeps its 20% edge to Sales and that’s all it gets credited with.

This is a perfectly legitimate way to report results. It’s just one specific question: given the thing immediately upstream of Sales, how much did each immediate driver contribute?

Following the paths: Total attribution

But maybe that’s not the question you actually want answered. Maybe you want to know the total effect Media had on Sales, including the bit that came through Brand. Then we need to follow every path Media can take to reach Sales.

Media has two paths in Figure 2. The direct one (20%) and the detour through Brand. The Brand detour contributes 10% of Brand’s 30% share of Sales, so , or 3%. Total Media effect: 23%.

But if Media is now 23% of Sales, Brand can’t also be 30%. Attribution is a zero-sum game. Brand’s 30% has to be split among whatever feeds into Brand. Base (Brand) gets credited with of Sales, Season (Brand) with , and Media picks up the remaining 3%. Brand itself? Brand gets nothing. It’s an intermediate, and its share has been fully passed through to the things upstream of it.

A leaf node is a node with no incoming edges. Only outgoing. The little graph below shows the convention: a square is a leaf (a true driver), a rounded box is a response (something being modeled).

flowchart TB

Leaf(Leaf node)
Response([Response node])

With that definition, Total attribution for Sales in Figure 2 looks like Table 1. Brand doesn’t appear because it’s not a leaf. And a small warning: don’t conflate Direct connection (an edge in the graph) with Direct attribution (the definition above). I picked the names; I’ll take the blame.

Variable	Direct connection	Indirect connection	Total
Base (Sales)	50%	0%	50%
Season (Sales)	0%	0%	0%
Base (Brand)	0%	= 18%	18%
Season (Brand)	0%	= 9%	9%
Media	20%	= 3%	23%
Total	70%	30%	100%

Five leaf nodes, five rows, summing to 100%. And the punchline worth staring at for a second:

The moment you decide to model Brand, you’ve committed to explaining it. And once it’s explained, its share of downstream Sales belongs, mechanically, to the things that explain it. You don’t get to count it twice.

Three levels deep

Since you’re still here, let’s go one level further. This time we have a first-level response (Brand), a second-level response (NPS), and a third-level response (Sales). Brand feeds both NPS and Sales; NPS feeds Sales.

flowchart LR

I1("Base (Brand)")
I2("Base (Sales)")
I3("Base (NPS)")
S1("Season (Brand)")
S2("Season (Sales)")
M(Media)
B([Brand])
S([Sales])
N([NPS])

I1 --> |60%| B
I2 --> |30%| S
S1 --> |30%| B
S2 --> |10%| S
M --> |10%| B
M --> |20%| S
B --> |30%| S
B --> |30%| N
I3 --> |70%| N
N --> |10%| S

Direct attribution for Sales is the easy bit. We take everything with an edge into Sales: Base (Sales) at 30%, Season (Sales) at 10%, Media at 20%, Brand at 30%, and NPS at 10%. Sums to 100%. Done.

Total attribution is where the bookkeeping starts earning its keep. The method is the same as before: trace paths either backwards from the response or forwards from each leaf. The answer’s in Table 2.

Variable	Direct connection	Indirect connection	Total
Base (Sales)	30%	0%	30%
Season (Sales)	10%	0%	10%
Base (Brand)	0%	= 19.8%	19.8%
Season (Brand)	0%	= 9.9%	9.9%
Base (NPS)	0%	= 7%	7%
Media	20%	= 3.3%	23.3%
Total	70%	30%	100%

Look at the Media row. Media has a direct path (20%), a path through Brand (Media -> Brand -> Sales = ), and a path through Brand and then NPS (Media -> Brand -> NPS -> Sales = ). That’s 23.3% total. The same logic applies to every other leaf. You do the algebra once and then it’s just a walk along the graph.

Visually, I find the Sankey diagram in Figure 4 the easiest way to internalize what’s happening. You can see the volume of Sales flowing backwards, through the intermediate responses, and fanning out into the leaves.

Why this matters for how you report

The technical argument is neat. The practical consequence is where teams get burned.

If you report Total attribution for your leaf drivers and also report that “Brand contributes 30% to Sales”, you have double-counted. Brand’s contribution has already been decomposed and handed out to Base (Brand), Season (Brand), and Media. Reporting it again, on top of those, makes your numbers sum to something greater than 100%, and makes it look like your marketing is doing more than it is. I’ve seen this in real decks. It’s not a rounding error. It’s a category mistake.

The principled way to report results in a hierarchical MMM is to pick your lens deliberately:

• For optimization decisions (where should the next euro go?), you almost always want Total attribution on leaf drivers. You’re asking about actionable levers, and Media or Base (Brand) are levers. Brand is not a lever you can buy directly.
• For diagnostic storytelling (what’s driving what?), Direct attribution gives you the immediate explanation. It tells you that Brand carries 30% of Sales in the most recent period, which is a useful thing to know when you’re trying to understand the shape of the business.

Just don’t mix them in the same table. Pick a lens. Label it clearly. Stick to it.

Where Alviss AI sits on this

We build hierarchical Bayesian MMMs at Alviss AI, and this is exactly the kind of thing we’ve had to get opinionated about. When you model a business holistically, with Brand, NPS, Customer Experience, or any other intermediate KPI as first-class responses, you inherit the attribution question in full. The platform enforces the distinction between Direct and Total attribution explicitly, and it won’t let you sum them into the same column. Not because we enjoy being strict, but because the arithmetic quietly breaks if you do.

The upside of modeling this way, of course, is that you get to ask richer questions. How much of my Media spend is earning its keep by building Brand rather than driving Sales this week? is a question that a flat MMM can’t answer at all. A hierarchical one can, cleanly, as long as you’re disciplined about which attribution lens you’re using when.

Conclusion

Attribution in a neural network based MMM isn’t one thing. It’s at least two things: Direct, which only counts edges one step from the response, and Total, which traces every path back to the leaves. Both are legitimate. Neither is wrong. But they answer different questions, and the moment you have intermediate responses in your model (Brand, NPS, anything you’ve chosen to explain rather than just consume), you have to commit to one.

The rule I’d leave you with: any response node in your graph contributes zero to the Total attribution of something downstream of it. If that feels counter-intuitive, go back to Table 1 and stare at it until it doesn’t. And if your current reporting has Brand and Media both contributing to Sales under the same definition, you have some bookkeeping to do.

If you want to talk about how this shows up in your own modeling setup, or you think I’ve missed something (which happens), reach out. I’d rather be corrected than comfortable.

The core idea

Marketing Mix Modeling is, at its heart, a statistical method for answering a deceptively simple question: which of the things I’m spending money on are actually working, and by how much?

You take historical data on your business outcomes (revenue, conversions, signups, whatever matters to you) and your marketing activities (media spend, impressions, GRPs, campaign flights) along with external factors that also affect your business (seasonality, weather, competitor activity, economic indicators). Then you build a model that separates the contribution of each driver.

That’s it. That’s the core idea.

Ok fine, but why is this hard? Because these signals are tangled together in the real world. You run TV and digital simultaneously. Sales go up in December regardless of what you do. A competitor launches a price war in Q3. A new product hits the shelves in the same week you increase your social spend. The model’s job is to disentangle all of this and give you a credible estimate of what each factor actually contributed.

How it actually works

Let’s walk through the mechanics without getting too deep into the math (we’ll save that for a future post on response curves and adstock).

At the most basic level, an MMM is a regression model. You’re predicting your KPI (let’s say weekly revenue) as a function of your marketing inputs and control variables. Something like:

Where captures the revenue you’d get even with zero marketing (brand strength, distribution, organic demand), each is a transformation of a media channel that accounts for how advertising actually behaves, captures external factors, and is the stuff the model can’t explain.

Now, those transformations are where the real modeling happens. Two concepts matter here:

Adstock (or carryover): When you see a TV ad today, it doesn’t just affect your behavior today. It lingers. You might buy something next week because of an ad you saw three weeks ago. Adstock transformations model this decay, the idea that advertising has a half-life. The shape of that decay varies by channel. TV tends to have longer carryover than paid search, for example.

Saturation (or diminishing returns): Doubling your spend on a channel does not double your return. At some point, you’re hitting the same people too many times, or you’ve already captured the easy conversions. Response curves model this saturation, and they’re critical for optimization because they tell you where each additional dollar starts to lose its punch.

Get these two things right, and you have a model that can decompose your revenue into contributions from each channel and external factor. Get them wrong, and you’re just doing expensive curve fitting.

Why MMM is back (and why it never really left)

MMM has been around since the 1960s. It was the standard approach for decades, then fell somewhat out of fashion when digital marketing arrived with its promise of tracking every click and attributing every conversion. Multi-touch attribution (MTA) became the shiny new thing.

So what happened? Reality happened.

Third-party cookies are effectively dead. Apple’s App Tracking Transparency gutted mobile attribution. Privacy regulations (GDPR, CCPA, and their descendants) have made user-level tracking increasingly difficult and legally risky. The “track everything” model of MTA has been eroding for years, and in 2026 there’s no pretending it’s coming back.

But the need to understand what drives your business hasn’t gone anywhere. If anything, it’s more urgent. Marketing budgets are under more scrutiny than ever, and CFOs want answers that go beyond “the algorithm says so.”

MMM doesn’t need cookies. It doesn’t need user-level tracking. It works with aggregated data (spend by week, impressions by channel) that you already have. It works for offline channels that MTA never could touch. And modern MMM, built on Bayesian inference rather than the frequentist regressions of the 1990s, can quantify uncertainty in a way that actually helps decision-makers understand what they know and what they don’t.

Classical MMM vs. modern Bayesian MMM

This is where it gets interesting, and where a lot of the confusion in the market lives.

Classical MMM, the kind that consulting firms have been running for decades, typically uses ordinary least squares (OLS) regression or some variant. You throw your data in, you get point estimates out, and you hope the confidence intervals are narrow enough to be useful. The model runs once, maybe twice a year. It takes months. It costs a lot.

Modern MMM is different in a few important ways:

Bayesian inference: Instead of just getting point estimates, you get full probability distributions over your parameters. This means you can say “we’re 90% confident that TV contributes between 8% and 14% of revenue” rather than “TV contributes 11%.” That range matters. It’s the difference between a number someone prints on a slide and a number someone actually uses to make a decision.

Prior knowledge: In a Bayesian framework, you can encode what you already know. If you know that TV doesn’t have a negative effect on sales (that would be pretty weird), you can build that into the model. If you have results from previous quarters, you can use them to inform the current model. This isn’t cheating. It’s being honest about the information you have. The prior should reflect your knowledge, not be adapted to fit your data.

Frequent updates: There’s no reason to wait six months between model runs. Modern platforms (including ours at Alviss AI) refit models daily or weekly as new data arrives. This turns MMM from a periodic strategic exercise into a continuous decision-support system.

Uncertainty-aware optimization: When you optimize budget allocation, the uncertainty matters. A channel with high average ROI but massive uncertainty is a very different bet than one with moderate ROI and tight confidence. Bayesian MMM gives you the information to make that distinction. Classical MMM doesn’t.

What MMM can (and can’t) tell you

Let’s be honest about both sides.

MMM is good at:

• Measuring the incremental contribution of each marketing channel to your business outcomes
• Estimating ROI and marginal ROI by channel
• Optimizing budget allocation across channels
• Capturing offline media effects (TV, radio, print, OOH) that digital attribution misses entirely
• Providing a privacy-compliant measurement framework
• Separating marketing effects from external factors like seasonality and economic trends

MMM is less good at:

• Real-time, campaign-level optimization (it works at a higher level of granularity, typically weekly or biweekly)
• Measuring creative effectiveness within a channel (it tells you “digital display drove X” but not which specific banner did the work)
• Working with very short data histories (you generally need 2-3 years of data for reliable estimates)
• Establishing true causality without experimental validation (it’s still an observational model, even if a good one)

Anyone who tells you MMM solves everything is selling you something. But for the question “how should I allocate my marketing budget across channels to maximize business outcomes,” it’s the best tool we have. Especially now that the alternatives have had their legs cut out from under them.

Where MMM fits in your measurement stack

MMM doesn’t replace everything else. It works best as part of a broader measurement approach:

MMM gives you the top-down, strategic view: which channels are working, how much to spend where, and what the diminishing returns look like.

Experiments (geo-lift tests, A/B tests, incrementality tests) give you causal validation. They’re the ground truth that calibrates your MMM. If your model says paid social drives 15% of incremental revenue, a geo-lift test can confirm or challenge that estimate. The best MMM implementations feed experimental results back into the model as calibration priors.

Attribution tools (whatever survives the privacy transition) give you the granular, tactical view within digital channels: which campaigns, audiences, and creatives are performing.

Think of MMM as the map, experiments as the ground surveys, and attribution as the GPS. You need all three, but the map is what you plan with.

Getting started with MMM

If you’re considering MMM for your organization, here’s what I’d think about:

Data readiness: You need historical data on your KPI and marketing activities, ideally 2-3 years at weekly granularity. You also need data on external factors: holidays, promotions, competitor activity, macroeconomic indicators. The quality of your MMM is directly proportional to the quality and granularity of your data.

Model transparency: Insist on understanding what’s inside the model. Can you see the response curves? The adstock parameters? The prior distributions? If a vendor hands you a black box with numbers coming out, walk away. I myself despise black boxes. The whole point of MMM is to build understanding, not to replace one opaque system with another.

Bayesian or bust: In 2026, there’s no good reason to use frequentist MMM. The Bayesian approach gives you uncertainty quantification, the ability to encode prior knowledge, and more stable estimates with limited data. If your vendor is still running OLS regressions, they’re a decade behind.

Continuous, not annual: The old model of running MMM once a year is dead. Markets move too fast. Your model should update as new data comes in, giving you a living view of performance rather than a stale snapshot.

Validation through experiments: Budget for geo-lift tests or other incrementality experiments to calibrate your model. An MMM without experimental validation is an opinion with math around it.

Conclusion

Marketing Mix Modeling is the most practical, privacy-compliant, and methodologically sound way to understand what drives your business outcomes and how to allocate your marketing budget. It’s been around for decades, but the modern Bayesian incarnation is a fundamentally different (and better) animal than what agencies were running in the 2000s.

The question isn’t whether you need MMM. If you’re spending meaningful money on marketing across multiple channels, you do. The question is whether you’re going to build it yourself, use an open-source framework, or work with a platform that handles the infrastructure, the modeling, and the continuous updates for you.

At Alviss AI, we’ve built our platform around Bayesian MMM with daily model refitting, full model transparency, and uncertainty-aware optimization. Not because it’s trendy, but because after a decade of building these models, we know it’s the right way to do it. If you want to see what that looks like in practice, get in touch.

The three approaches (and what they actually do)

There are three distinct measurement methodologies that matter. Not two. Three. People tend to frame this as “MMM vs. MTA” but that leaves out the one that arguably matters most for getting to truth: experiments.

Marketing Mix Modeling (MMM)

MMM is a top-down statistical approach that uses aggregate data (typically weekly spend, impressions, and revenue) to estimate how much each marketing channel contributes to your business outcomes (Hanssens et al. 2001). You don’t need to track individual users. You don’t need cookies. You need time-series data and a good model.

The model decomposes your KPI into contributions from media channels, baseline demand, and external factors (seasonality, promotions, economic conditions). It accounts for the fact that advertising has carryover effects (Broadbent 1979; Tull 1965) and diminishing returns (Hill 1910). Modern implementations use Bayesian inference (Jin et al. 2017) to quantify uncertainty around every estimate, which means you don’t just get “TV drove 12% of revenue” but “TV drove 9-15% of revenue with 90% probability.”

What it’s good at: Strategic budget allocation across all channels (including offline), understanding long-term effects, privacy-compliant measurement, capturing the full marketing mix.

What it’s not good at: Real-time optimization, creative-level insights, working with less than 1-2 years of historical data.

Why this isn’t a fair fight anymore

Let me be direct. Five years ago, the MMM vs. MTA debate was a genuine one. You could reasonably argue that MTA’s granularity and speed made it the better primary measurement system for digital-heavy businesses.

That argument doesn’t hold up in 2026.

The infrastructure that MTA depends on has been systematically dismantled. It’s not a temporary setback waiting for a new tracking solution. The direction of travel is clear: less user-level tracking, not more. Every major platform, browser, and regulator is moving in the same direction.

This doesn’t mean MTA is useless. It means MTA can no longer be your primary measurement framework. It can still be valuable for tactical optimization within the digital channels where you have good signal. But using it as the basis for budget allocation decisions is risky because it can’t see the full picture and the picture it can see has growing holes in it.

MMM, by contrast, was built for a world without user-level tracking. It never needed cookies. It works with the aggregated data that privacy regulations encourage. And modern Bayesian MMM (Jin et al. 2017; Chan and Perry 2017) has closed most of the historical gaps: it updates frequently, it quantifies uncertainty, and it can incorporate experimental results as calibration priors.

A real comparison

Ok, let’s get concrete. Here’s how these three approaches actually stack up across the dimensions that matter.

The table tells a clear story. MMM and experiments are complementary. MMM gives you the strategic allocation and experiments give you the causal validation. MTA fills a tactical niche for digital optimization but can’t anchor your measurement strategy.

How they should work together

The right answer isn’t picking one. It’s understanding how they fit together.

MMM sets the strategy. It tells you how much to spend on each channel and what the expected returns are. It sees the whole picture: offline, online, competitive effects, seasonality. This is your planning layer.

Experiments calibrate the model. You run geo-lift tests on your highest-spend channels to validate (or challenge) what the MMM says. If the model says paid social drives 15% of incremental revenue, a geo-lift test can tell you whether that’s in the right ballpark. The results feed back into the model as Bayesian calibration priors, tightening the estimates over time.

Attribution optimizes within channels. Once MMM tells you to spend €500k on digital display this quarter, attribution helps you figure out which campaigns, audiences, and creatives are working best within that envelope. It’s the tactical layer, not the strategic one.

This isn’t a theoretical framework. It’s how the best-performing marketing organizations actually operate. Google’s own measurement guidance recommends exactly this structure (Chan and Perry 2017), and it’s the approach we’ve built Alviss AI around.

The decision guide

Still not sure where to start? Here’s how I’d think about it.

If you spend across both online and offline channels: You need MMM. There’s nothing else that can measure TV, radio, OOH, and digital in the same model. MTA literally cannot do this.

If you’re a pure-play digital business: You still benefit from MMM (it captures cross-channel effects and diminishing returns that attribution misses), but you can get more tactical value from attribution in the short term. Start with MMM for budget allocation, use attribution for campaign optimization.

If you need to prove to your CFO that marketing works: You need experiments. Nothing else establishes causation. Run a geo-lift test on your biggest channel, show the incremental lift, and suddenly the measurement conversation gets a lot easier.

If you want to do this properly: Combine all three. MMM as the strategic backbone, experiments for calibration, attribution for tactical optimization. This is the unified measurement approach, and it’s where the industry is heading.

The common mistakes

I’ve seen a lot of measurement programs go sideways. Here are the patterns:

Treating last-click attribution as truth. Last-click gives all credit to the final touchpoint before conversion. This systematically overvalues bottom-of-funnel channels (paid search, retargeting) and undervalues top-of-funnel (TV, display, social). Research by Gordon et al. showed that observational attribution methods frequently disagree with experimental ground truth (Gordon et al. 2019), sometimes by a wide margin. If you’re making budget decisions based on last-click, you’re almost certainly underinvesting in awareness and overinvesting in channels that are just catching demand that already exists.

Building MTA as your primary measurement system in 2026. The fundamental challenge with MTA isn’t just missing data. It’s that even with complete data, observational attribution struggles to distinguish correlation from causation (Dalessandro et al. 2012). If your attribution data has holes (and it does), using it for budget allocation is like navigating with a map that’s missing half the roads. You’ll stay on the roads you can see, even if a better route exists on the ones you can’t.

Running MMM once a year. The old consulting model of annual MMM is dead. Markets move too fast. If you’re making Q4 decisions based on a model that was last updated in Q1, you’re working with stale information. Modern MMM platforms update weekly or daily.

Skipping experiments entirely. MMM gives you estimates. Experiments give you proof. Without experiments, your MMM is an educated guess. A good educated guess, but still a guess. Budget for at least 2-3 geo-lift tests per year on your highest-spend channels.

Confusing measurement and optimization. Measurement tells you what happened. Optimization tells you what to do next. MMM does both. Attribution mostly just measures (though some platforms attempt optimization). Experiments just measure. Knowing which tool answers which question is half the battle.

Conclusion

The MMM vs. attribution debate made sense ten years ago when both approaches had clear, distinct strengths and you had to pick a primary framework. In 2026, the answer is clearer: MMM is your strategic measurement backbone, experiments validate it, and attribution handles tactical optimization within digital channels.

The organizations getting this right aren’t the ones with the most sophisticated attribution models. They’re the ones with a clear measurement hierarchy: MMM for planning, experiments for proof, attribution for execution.

At Alviss AI (Alviss AI 2026), we’ve built our platform around this philosophy. Bayesian MMM with continuous model updates, experimental calibration, and the transparency to see exactly how the model reaches its conclusions. If you’re ready to move beyond the MMM vs. attribution debate and build a measurement stack that actually works, let’s talk.

What is Marketing Mix Modeling (MMM)?

Marketing Mix Modeling is a top-down, macro-level statistical analysis used to estimate the impact of various marketing tactics on sales or other Key Performance Indicators (KPIs). It relies on historical aggregate data, typically looking at weekly or monthly time series across several years.

At its core, MMM uses multivariate regression, often employing Bayesian methods or frequentist econometrics, to decompose a KPI into its constituent drivers. These drivers include:

• Media Variables: TV, radio, out-of-home, and digital spend.
• Non-Media Variables: Price changes, promotions, and distribution.
• External Factors: Seasonality, economic indicators, and competitor activity.
• The Baseline: Sales that would occur without any marketing activity.

For the marketing scientist, the beauty of MMM lies in its ability to account for diminishing returns and carryover effects, often referred to as adstock. Because it uses aggregate data, it is inherently privacy-friendly and unaffected by the “cookie apocalypse” or changes in mobile tracking IDs.

What is Multi-Touch Attribution (MTA)?

Multi-Touch Attribution is a bottom-up, granular approach that seeks to assign credit to every individual touchpoint a user interacts with before converting. Unlike MMM, which looks at the forest, MTA looks at the individual trees.

MTA tracks the user journey across digital channels, assigning fractional credit to clicks, impressions, or engagements based on specific rules or algorithmic models. Common models include:

• Linear: Evenly distributes credit across all touchpoints.
• Time-Decay: Gives more credit to touchpoints closer to the conversion.
• Data-Driven: Uses machine learning, such as Shapley Value or Markov Chains, to determine which touchpoints truly move the needle.

While MTA provides tactical insights into keyword performance or creative efficacy, it has grown increasingly difficult to execute. The rise of iOS 14.5 restrictions and the deprecation of third-party cookies have created significant “dark patches” in the user journey, making it harder for analysts to maintain a clean chain of causality.

Key Differences: A Comparative Analysis

To help you decide which model fits your specific analytical needs, here is a breakdown of the primary differences.

The Shift Toward Unified Measurement

The most sophisticated marketing organizations no longer view this as an “either-or” scenario. Instead, they are moving toward Unified Marketing Measurement (UMM), a framework where MMM provides the strategic guardrails and MTA (or click-based attribution) provides tactical direction.

By using MMM to set the “North Star” for ROI, analysts can calibrate their granular attribution models to ensure they aren’t overvaluing digital channels that happen to be at the bottom of the funnel.

For teams looking to bridge the gap between complex statistical modeling and actionable business insights, tools like Alviss.io offer advanced solutions. By automating the heavy lifting of data ingestion and model training, analysts can focus more on interpretation and less on manual data cleaning.

Why MMM is Reclaiming the Spotlight

As privacy regulations like GDPR and CCPA tighten, the “bottom-up” data required for MTA is becoming increasingly fragmented. Marketing scientists are returning to MMM because it provides a “ground truth” that does not rely on invasive tracking.

Furthermore, modern MMM has evolved. We are no longer limited to static, once-a-year reports. With modern computing power, models can be updated frequently, incorporating Bayesian priors to account for new market realities.

Conclusion: Building Your Measurement Stack

Choosing between MMM and MTA depends on your business model and your data maturity. If you are a direct-to-consumer brand with a heavy offline presence, MMM is non-negotiable. If you are a pure-play SaaS company with a short sales cycle, MTA might still provide significant tactical value.

However, for most, the future lies in a hybrid approach. Use MMM to understand the “why” and the “how much” at a high level, and use granular attribution to optimize the “where” and “when” of your digital execution.

To explore how you can leverage state-of-the-art modeling to optimize your marketing spend and drive higher ROI, visit Alviss.io to see our suite of analytical tools in action. By empowering your team with the right technology, you can turn complex data into a competitive advantage.

What to Look For in MMM Software in 2026

Before evaluating specific providers, it is worth establishing the criteria that distinguish capable MMM platforms from limited ones. The requirements have shifted materially over the past three years.

The 15 Best MMM Software Providers in 2026

Tier 1: Modern SaaS MMM Platforms

These platforms are purpose-built for in-house marketing teams that want to own their measurement infrastructure. They offer software interfaces, managed data pipelines, and optimization tools without requiring teams to write and maintain code from scratch. The Tier 1 landscape spans a wide range of methodological rigor, flexibility, and price points.

Tier 2: Open-Source MMM Frameworks

Open-source MMM frameworks offer rigorous methodology at zero software cost. The trade-off is substantial: implementation, maintenance, infrastructure, and interpretation all require in-house data science expertise. These frameworks are not products, they are libraries. The total cost of ownership, when talent and infrastructure are included, is often comparable to or higher than mid-market SaaS solutions.

Tier 3: Enterprise and Consultancy-Led MMM

These providers serve large enterprises that prefer to outsource measurement to specialist vendors or consultancies. They offer deep industry expertise, broad media coverage, and managed delivery, at the cost of speed, transparency, and in-housing capability. For organizations with the budget and the preference for managed measurement, they remain viable options. For teams building internal analytical capability, they are generally unsuitable.

MMM Software Comparison: Full Table

As stated this guide is published by Alviss AI. Please see disclosure in the introduction.

How to Choose the Right MMM Software

The right MMM platform depends on three variables: the technical capability of your team, the complexity of your measurement requirements, and your organizational appetite for in-housing vs. managed delivery.

For small and DTC e-commerce brands with limited analytics resources, Cassandra offers a low-cost starting point for basic media effectiveness measurement with optional geo-experiment calibration. Northbeam is a stronger option for brands running high-volume paid social and search campaigns that need near-real-time feedback. LiftLab is a strong option for US-based DTC and consumer brands that want geo-experimentation tightly integrated with their MMM rather than operated as a separate workstream. Its weekly cadence and full-funnel measurement make it particularly relevant for teams running active paid media across multiple channels simultaneously. The trade-off relative to Recast or Alviss AI is methodology: LiftLab’s proprietary model does not expose the computational graph or produce Bayesian posteriors, which limits auditability for advanced marketing science teams. If the team has data science capability, Meta Robyn provides rigorous methodology at no software cost.

For mid-market brands building in-house capability, Alviss AI and Recast are the strongest options. Both are Bayesian, both support in-housing, and both offer budget optimization. The key differentiators are model flexibility and multi-KPI support. Teams that need to model brand and performance KPIs in a unified structure, or that want to inspect and edit the computational graph, should evaluate Alviss AI specifically for those capabilities. On pricing, it is worth noting that methodology-rich platforms in the 2,000–3,000 EUR/month range may deliver more modeling capability than managed-service alternatives at comparable or higher price points, a comparison worth running for any shortlist. Odins.ai and Sellforte are relevant for European brands that prefer a managed model, but both involve trade-offs on transparency, capability, and value for money. LiftROI (Blackwood Seven / Kantar) is worth evaluating for European enterprise brands that need sub-channel publisher-level granularity within the model and want to extend the same Bayesian framework to churn and brand objectives beyond a pure sales MMM. The Kantar Group backing makes it a credible enterprise procurement choice. The trade-offs are significant for teams wanting model ownership: in-housing is not supported, the computational graph is proprietary, and pricing requires a direct sales engagement.

For enterprise teams with complex media mixes and multi-KPI requirements, Alviss AI is, to our knowledge, currently the only commercial SaaS platform that combines Bayesian methodology, multi-KPI modeling, an editable computational graph, full in-housing capability, and an integrated budget optimizer in a single product. Keen Decision Systems is a strong option for enterprises prioritizing continuous measurement and financial planning integration. For enterprises that want fully managed, outsourced measurement, Analytic Partners and Nielsen remain viable, with the understanding that transparency, iteration speed, and in-housing are explicitly not part of the offering.

For data science teams with engineering resources and a preference for open-source tooling, Google Meridian represents the current state of the art in open-source Bayesian MMM. Meta Robyn remains the most community-supported framework and is the practical default for teams already working in R or Python.

Conclusion

Marketing Mix Modeling in 2026 is defined by three shifts: from annual to continuous, from black-box to transparent, and from vendor-managed to in-housed. The tools that are winning are those built around Bayesian methodology, fast iteration, and genuine model ownership for internal teams.

The open-source frameworks, Meridian and Robyn, have raised the methodological floor across the industry. The legacy enterprise providers, Nielsen and Analytic Partners, remain relevant for specific large-enterprise contexts but are structurally misaligned with the direction of travel. The most interesting competitive space is in modern SaaS, where a generation of Bayesian platforms are competing on methodology, flexibility, and in-housing capability.

Within that space, Alviss AI (alviss.io) occupies a distinct position. Based on publicly available product documentation, Alviss AI is currently the only commercial SaaS platform that supports multi-KPI modeling, exposes and enables editing of the full computational graph, runs on Bayesian hierarchical methodology, and is explicitly designed for full in-housing, in a single integrated product. Customers including G-Star, Allianz, Saxo Bank, and Airalo use Alviss AI to run owned, transparent, and continuously iterated MMM across brand and performance KPIs.

The Fragmentation Problem

Most MMMs are built as dedicated models, each offering an isolated view of one side of the business. This creates a fundamental limitation, as marketing and media optimizations can be designed either to maximize sales or to minimize churn, but not both at once. As a result, marketing managers are left juggling multiple what-if scenarios, stitching together KPIs, and calculating trade-offs manually. The process is time-consuming, inaccurate, and ultimately incapable of reflecting the true complexity of the business.

Why Holistic Modeling Matters

Business does not happen in silos. Acquisition, churn, brand equity, and customer experience are interconnected forces that shape overall performance. Optimizing one without considering the others risks undermining long-term growth.

Holistic modeling changes the game. By connecting these dependencies into a unified framework, leaders can:

• Optimize for net positive outcomes rather than isolated KPIs
• Forecast with confidence, knowing the model reflects the full business reality
• Run meaningful scenarios that balance acquisition, retention, and brand-building simultaneously

This is not just a technical upgrade, it is a strategic evolution.

Alviss AI: Bridging the Gap

Alviss AI was built to overcome the limitations of traditional MMMs. The platform supports modeling the business as it truly operates, with all dependencies and hierarchical structures intact. The result is true holistic business optimization and forecasting.

And this is not just theory.

A Real-World Example: One Model for Net Growth

We recently implemented a holistic modeling approach for one of our insurance clients. In this case, the model did not just include acquisition and churn as KPIs, it also incorporated intermediaries such as brand equity and customer experience within the hierarchical structure.

The following graph illustrates the structure of the holistic model. Blue nodes are actually modeled while grey nodes are inputs to the models, such as media, price or other factors, for example including competitor activities.

This setup allowed the client to use each component individually, for example optimizing marketing to maximize acquisition or strengthen brand equity, while also combining them to drive net positive growth. This is also a great example how business units can use the same model for their individual purposes:

• The growth department can optimize their activities to maximise new policies,
• The churn department can use the model to derive actions to mitigate churn,
• And the brand team can quantify and act upon maximising brand awareness and consideration.

The outcome was a model that reflects the business as it truly is, interconnected, dynamic, and capable of guiding smarter decisions, across the organization.

Conclusion

The future of MMM lies in holistic modeling. Fragmented models may explain parts of the past, but only holistic models can shape the future. By embracing frameworks that integrate acquisition, churn, brand, and customer experience, businesses can move beyond reactive trade-offs and toward proactive, sustainable growth.

With Alviss AI, that future is already here. If you want to move beyond a siloed approach in your predictive business modeling, let’s get in touch!

Introduction

Marketing Mix Modeling (MMM) has long been the go-to framework for quantifying business drivers. Yet, despite its widespread adoption, the industry often falls into the trap of celebrating vanity metrics rather than true predictive power. Whether MMMs are built in-house or delivered by consultancies, the critical question remains: are we validating models in a way that actually drives better decisions?

The Illusion of High R²

For years, R² has been paraded as the gold standard of model quality. Consultants proudly present models with R² values north of 90%, as if that alone signals success. But let’s be honest: anyone who has built an MMM knows how easy it is to game this metric. Add enough dummy variables, and you can make the past look perfect. The problem? Business leaders don’t need a model that explains yesterday. They need a model that guides tomorrow. And this is where most MMMs fall short.

Step 1: Demand Predictive Proof

Validation must go beyond statistical fit. A true test of quality is whether the model can predict unseen data. That’s why holdout testing should be non-negotiable. Feed the model data it hasn’t seen before and measure how well it performs. Shockingly, holdout tests are often neglected in MMM projects. Without them, predictive quality is little more than guesswork. If your provider isn’t showing you holdout results, you should be asking why.

Step 2: Sanity Check Against Business Reality

Even the most statistically sound model can fail if it doesn’t align with business logic. That’s why sanity checks are essential. However, too often analysts assume it is enough to validate the logic of static results that merely explain the past. That is a dangerous illusion, especially when the model functions as a black box with hidden structures and obscured data transformations. True validation requires more. You need to pressure-test the model with what-if scenarios that mirror real-world decisions, exposing whether it can guide the future rather than just retell the past.

Run what-if scenarios that reflect real-world decisions:

• If you want to grow brand awareness by 1 percentage point, how much should you invest in media, assuming the same media mix as last year?
• If the model suggests spending less to achieve more, you’ve uncovered a fundamental flaw.

Models must not only predict - they must make sense in the context of how businesses actually operate.

The Thought Leadership Takeaway

The industry needs to move past the obsession with high R² and embrace a more holistic view of validation. True model validation is a two-step process: statistical rigor through predictive testing, and business relevance through sanity checks.

Anything less is just analytics theater.

Marketing leaders should demand more from their MMMs. Don’t settle for models that look good in PowerPoint but collapse in practice. The future of MMM lies in models that not only explain the past but actively shape smarter, more confident decisions for the future.

If you want to know more about how we build and validate our models with full transparency, do not hesitate to get in touch. We believe that models should not only be statistically sound but also business-relevant, open, and trustworthy.

With what frequency should the models and results be updated?

Any person in a marketing department involved in campaigns and media buying will typically be extremely impatient in relation to getting fresh insights into the current performance. Who wouldn’t like to measure the effect of the campaign that started yesterday? An MMM setup should definitely be able to deliver insights with a high frequency, but there are a few things to consider in relation to this topic:

Even though it sounds old-school, it is not recommended to start your MMM journey by setting up automated data flows from day one. It is much more valuable to start out by building robust models, understanding your business drivers fully and identifying the right and relevant data sources for your setup. That’s where you should start.

Once the initial models have been developed and are fully aligned, you should establish a system based on automated dataflows and instant model updates. This will save a lot of time in the modelling processes and ensure fast delivery of fresh results.

Automating the data flows will give you the possibility for very frequent (even daily) updates of results, which to some extent would be a dream scenario. Even though this is tempting, you have to think about whether your organization is able to react to results that are updated on a daily basis. Do you believe that the high frequency in the outputs will give information that is valuable, or would you be better off with automated reporting by week or by month? In many cases the latter actually turns out to work best when it comes to implementation of an MMM.

What KPIs should we use to measure campaign success?

The amount of potential KPIs to measure is enormous. Choosing the right ones for your organization is a fundamental part of the design.

An important rule of thumb is that you want to choose KPIs that are as closely linked as possible to the overall business performance. It can be tempting to track your campaign performance on metrics like clicks, engagement scores and view-through rates, but they don’t really tell anything about the real business outcome of your marketing efforts. A strong measurement setup should be based on real business KPIs such as revenue, transactions, in-take of new customers or similar.

It is better to start out with high-level KPIs for your business and plan for increased granularity as next step. By focusing on solid model building for high-level KPIs in the initial phase, you will reduce complexity and secure a quick and successful initial implementation of your MMM. Following that, you should go for a setup with an ambitious granularity level in the KPIs, which will enable you to deliver detailed insights into your marketing effects across sales channels, regions, by product, by segment etc. Once you reach this point in the implementation, the MMM setup will successfully obtain even broader relevance within your organization and among your main stakeholders.

Conclusion?

There are a number of things to look out for and consider when going into a design phase of an MMM. In this blogpost we have touched a couple of the most essential ones. The main point is to start “simple and focused” and then go from there to “automated and granular”. By following this approach, you are well underway to ensure a successful and future-proof MMM development for your organization!

If you want to learn more about how we help businesses and organizations obtain new levels of growth through data driven decision-making, don’t hesitate to reach out.

How to find the growth pockets within your organization and ensure successful implementation: The scoping phase, potential data quality challenges and the risk of developing overcomplicated systems

A key principle in successful implementation is to make a structured approach to define exactly where in the business operation it will deliver the highest expected value and how it can potentially work out. This can be approached by addressing a few basic questions:

1. What exact business challenge or operational task do we want to solve or optimize?

It should not be underestimated how important it is to scope and formulate the actual business need correctly, as a very first step. Be sure that you have internal alignment of the goal, as this is very important for success in the implementation!

2. Do we have a solid data foundation in that area of the business?

An understanding of data quality and whether it allows to solve the desired scope, is crucial in order to obtain success. If data is not 100 % in place yet, your focus should be to ensure better data quality going forward, before implementing any decision making based on it!

3. Are we able to react on it and make it operational in our day-to-day work?

Don’t get tempted into building decision making systems that you are not able to react on. The possibilities for creating data driven decision making are endless today and you can end up with overwhelming outcomes. Don’t forget to be realistic about the organizational or operational constraints, if you want it to succeed in the long run!

This may sound as an oversimplification of the task 😊 And it can of course be more complex than this. But it is crucial to address at least these above questions, before investing in the creation of any data driven decision making setup within your organization.

If you want to learn more about how we help businesses and organizations obtain new levels of growth through data driven decision-making, don’t hesitate to reach out.

What are Time-Varying Parameters?

Traditional marketing mix models often assume that the effects of different advertising channels remain consistent over time. However, in reality, the effectiveness of media spend can vary based on numerous factors such as seasonality, economic conditions, competitive dynamics, and consumer behavior patterns. These variations necessitate the use of time-varying parameters to better reflect how these factors impact campaign performance.

Time-varying parameters allow MMM models to dynamically adjust for changes in the marketing environment. This means that the coefficients assigned to each media channel are not fixed but can change based on specific market conditions or other external influences, providing a more accurate representation of real-world scenarios and enhancing predictive power.

Importance of Time-Varying Parameters in Media Attribution

Media attribution is critical because it helps marketers understand which channels contribute most to sales and brand lift. Without considering time-varying parameters, models might overestimate the effectiveness of certain media types or underestimate others, leading to suboptimal allocation of marketing budgets. By incorporating time-varying parameters, MMMs gain many benefits.

Enhancing Planning Through Time-Varying Parameters

For effective future planning, time-varying parameters enable more informed decisions about where to allocate resources:

1. Budget Optimization: By understanding how each channel responds to different conditions, marketers can make better decisions about budget allocation across channels. For instance, during economic expansions, a higher spend on high-cost media might be justified if the model predicts positive returns based on historical time-varying effects.

2. Strategy Adjustment: MMM with time-varying parameters allows for more flexible strategic planning. If a new product launch is expected to be highly dependent on digital marketing due to consumer behavior shifts online, then greater investment can be planned in this area rather than traditional media which might not yield the same ROI during that specific period.

3. Performance Benchmarking: Establish realistic benchmarks and expectations for future performance by considering historical time-varying effects. This helps in setting achievable targets and making strategic decisions based on factual predictions rather than guesswork.

Conclusion

In conclusion, incorporating time-varying parameters into marketing mix modeling is crucial for achieving accurate media attribution and enhancing the effectiveness of future planning in a rapidly changing market environment. By dynamically adjusting for changes in consumer behavior, economic conditions, and competitive landscapes, these models offer more reliable insights that drive better decision-making across various aspects of marketing operations. As such, businesses would do well to embrace MMM with time-varying parameters as an essential tool in their strategic marketing toolkit.

How to Use AI to Improve Your Marketing

The best marketers don’t just follow trends, they stay ahead of them. And right now, the most powerful tool for doing that is AI.

If you’re not already using AI for marketing, you’re missing out. Not because AI is some magical solution that will instantly fix everything, but because it’s the most efficient way to optimize what you’re already doing. AI marketing optimization isn’t about replacing human judgment, it’s about sharpening it.

Why AI Works

Marketing has always been about finding patterns. What messages resonate? Which channels perform best? When is the ideal time to reach your audience? The difference now is that AI can analyze these patterns at a scale and speed humans can’t match.

For example:

• Personalization at scale. AI can tailor messages to individual users based on their behavior, something that would be impossibly time-consuming manually.
• Predictive analytics. Instead of guessing what might work, AI can forecast trends based on historical data.
• Automated A/B testing. AI can run hundreds of variations simultaneously, learning and adjusting in real time.

The key is that AI doesn’t just give you answers, it helps you ask better questions.

How to Start

You don’t need a massive budget or a team of engineers to use AI for marketing. The simplest way to begin is by integrating existing tools:

• Chatbots for instant customer engagement.
• Email optimization tools that adjust subject lines and send times for maximum open rates.
• Ad targeting algorithms that refine your audience segments automatically.

The best approach is to pick one area where optimization would make the biggest difference and experiment there first.

The Pitfall to Avoid

The biggest mistake people make with AI marketing optimization is treating it as a set-it-and-forget-it solution. AI works best when it’s guided by clear goals and human oversight. If you automate without strategy, you’ll just scale mediocrity.

The most successful marketers use AI to enhance their creativity, not replace it. They let the AI handle the repetitive tasks so they can focus on the big picture crafting messages that actually matter.

What’s Next

AI in marketing is still evolving. The tools will get smarter, and the best practices will shift. But the core principle won’t change: the more you use AI to refine your approach, the better your results will be.

The question isn’t whether you should use AI for marketing, it’s how quickly you can start.

The Quiet Power of AI Marketing Optimization

If you’ve worked in marketing long enough, you’ve seen the same cycle play out repeatedly. A new tool or technique emerges, promising to revolutionize the field. Most of the time, it fizzles. But occasionally, something comes along that actually changes how things are done. Not because it’s flashy, but because it quietly solves a real problem.

AI marketing optimization is one of those things.

At its core, marketing is about allocating resources efficiently. You have a finite budget, and you need to spend it in a way that maximizes returns. The problem is that the world is messy. Channels interact in unpredictable ways. External factors, like seasonality or economic shifts distort the picture. And human intuition, though valuable, is often wrong when dealing with complex systems.

This is where media mix modeling (MMM) comes in. MMM isn’t new. It’s been around for decades, used by big companies to estimate the impact of different marketing channels. But traditional MMM had limitations. It was slow, expensive, and often relied on oversimplified assumptions.

AI changes that.

Modern AI-powered MMM doesn’t just run faster, it learns. Instead of forcing rigid models onto messy data, it adapts. It detects subtle interactions between channels that a human analyst might miss. It adjusts for external noise more accurately. And it does this continuously, refining its understanding as new data comes in.

The result? You spend less time guessing and more time acting on insights that actually work.

How It Works

The best AI marketing optimization tools don’t just dump a pile of data on you. They answer specific questions:

• Where should I allocate my next dollar?
• How much of my sales lift is actually due to marketing?
• What’s the real ROI of that influencer campaign?

Traditional MMM might give you broad strokes: “TV ads work better in Q4.” AI-driven MMM can tell you, “Shift 12% of your spend from paid search to connected TV in the first two weeks of December, but only if competitor X hasn’t launched a promotion.”

This precision matters. Most marketing budgets are still allocated based on habit or hierarchy, the channels that “feel” important, or the ones that have always gotten funding. AI optimization cuts through that. It doesn’t care about politics. It just shows you what works.

The Catch

Like any powerful tool, AI marketing optimization has prerequisites. Good data, patience and a willingness to act is key.

The Future

Right now, AI marketing optimization is still mostly used by larger companies. But like most technology, it’s getting cheaper and more accessible. Soon, even small teams will be able to run sophisticated MMM without needing a PhD in statistics.

The winners won’t be the ones who adopt it first, but the ones who use it best and who integrate it into their decision-making rather than treating it as a black box.

If you’re still guessing where to spend your marketing budget, you’re leaving money on the table. The tools to fix that are here. The question is whether you’ll use them.

Introduction

The funny thing about technology is that it has a way of creeping up on us. For years, people have been saying AI would change the world, but the hype seemed to fade out as soon as it came up. Yet, while we were all talking about robots and self-driving cars, it was in areas like marketing that AI found real, practical applications. Marketing Mix Modeling (MMM) is one of those places where AI is making big changes not just minor tweaks, but fundamental shifts in how brands make decisions. If you’re looking to get ahead, or even just stay in the game, now’s the time to understand how this works.

Marketing Mix Modeling used to be something only the big players could afford. It was complex, expensive, and largely outsourced. If a brand wanted to know how much TV ads moved the needle compared to digital, they would hire a specialized firm to analyze historical data, run a regression, and tell them what worked. These models were built to work in a specific context usually big marketing budgets spread across a few channels. But now, with a more complex media landscape and the rise of data-driven marketing, that model is breaking down. Today’s market needs something faster, more adaptable, and far more detailed. That’s where AI, and specifically Alviss AI, comes in.

AI brings a new kind of adaptability to MMM. Traditional models run on past data, looking backward to predict the future. But the problem with static models is they’re slow to adapt. This is why AI is a game-changer for MMM it learns as you go. If customer behavior shifts suddenly (think economic downturns or viral trends), an AI-driven model like Alviss AI can re-calculate and update itself, giving you an up-to-date answer instead of one that’s months old. This level of adaptability can make a big difference when you’re trying to move faster than competitors.

But AI in MMM is about more than speed. The key to better decision-making is clarity about what works and what doesn’t. For that, you need precision. That’s where Bayesian approaches come in, allowing models to account for uncertainty. In marketing, there’s a lot of noise seasonal effects, brand perception, competitor actions. A Bayesian approach handles that noise by not just estimating effects but quantifying how sure you can be about them. This is what makes AI-driven MMM fundamentally more useful: you’re not only seeing what worked, but you’re also getting a measure of confidence for each insight.

Another change AI brings is that it’s finally making MMM accessible. With platforms like Alviss AI, brands no longer need specialized teams to run the numbers or manage the modeling process. Anyone with basic analytics knowledge can set up, adjust, and run an AI-driven MMM model, and the insights are clear enough to apply directly. This shift means that even smaller brands can in-house MMM without sacrificing accuracy. The result? More control, more customization, and, often, lower costs.

If you’re running a brand or an agency today, you’re looking at a landscape that’s changing faster than it has in a long time. Channels are multiplying, customer behavior is fragmenting, and marketing itself is becoming harder to measure. The traditional approach to MMM, with long, expensive projects and static models, won’t keep up. But with AI-driven MMM, you get a model that’s always learning, always adapting, and always working at the speed of your business.

So, the future of MMM isn’t about more data or bigger models; it’s about having a tool that learns as fast as you do and doesn’t need a team of specialists to work. For most marketers, that means moving from outsourced projects to in-house, AI-driven platforms. This shift might be gradual for some, but the trend is clear. The brands that get ahead will be the ones that harness AI to make better decisions, faster. And that’s the promise of AI-driven MMM.

This post is part of a 6 part series called “Mastering Marketing Effectiveness with In-Housed MMM”. The posts are outlined below.

• Post 1: The New Era of Marketing Mix Modeling
• Post 2: Top 5 Challenges in Marketing Mix Modeling
• Post 3: How Alviss AI’s Bayesian Approach Enhances Marketing Mix Modeling Accuracy
• Post 4: DIY Marketing Mix Modeling: A Step-by-Step Guide for Brands and Agencies
• Post 5: In-House Marketing Mix Modeling Success Stories with Alviss AI
• Post 6: Future-Proofing Marketing with AI-Driven Marketing Mix Modeling

Introduction

One of the strangest things about Marketing Mix Modeling is that most companies know they need it but struggle to make it work. The theory makes sense: analyze past performance, measure the impact of different channels, adjust the marketing budget accordingly. But most MMM solutions have a kind of secret flaw. They work, but only if you’ve got endless time and a high tolerance for outsourced complexity.

Alviss AI was built on a different idea: that companies could actually do this themselves. That they wouldn’t just outsource MMM as an obligatory checkbox, but actually in-house it as a practical tool for driving their business. What’s surprising is how fast this approach pays off when companies make the switch. Here are a few stories of brands and agencies who in-housed MMM with Alviss AI and saw real results not months down the line, but practically out of the gate.

The Retailer Who Cut Waste and Improved ROI

This first example is a retail company with a lot of moving parts regional campaigns, seasonal variations, and a wide array of product lines. They had always outsourced MMM, but it was a slow and unwieldy process. By the time they’d get results from their vendor, the market had often changed, or new campaigns were already running. So, they’d constantly feel like they were reacting rather than adapting.

When they switched to Alviss AI, things immediately changed. Suddenly, they had real-time insights into their marketing effectiveness. They could run simulations to see what would happen if they shifted budget from one channel to another. And, maybe most importantly, they had the data themselves. They weren’t dependent on waiting for reports from someone else, and they didn’t have to deal with complex data pipelines.

The impact? They cut wasted spend by nearly 15%, simply by reducing investments in lower-impact channels. And because they were running campaigns in real time, they also caught seasonal changes that would’ve otherwise been lost in an annual report.

The Agency That Replaced Guesswork with Precision

Then there was an agency that had always struggled with MMM. Their clients loved the idea of optimization but were frustrated by the practical limitations. Most agencies don’t have the in-house data science resources to set up custom MMM for each client. For this agency, each MMM project took weeks and often relied on dated models and approximations, which meant a lot of guesswork.

Alviss AI solved that problem by giving them a platform they could use with little setup or customization. They plugged in their clients’ data and ran highly targeted models that could be updated at will. Instead of re-running an entire analysis every time a client wanted a change, they could simulate new budget scenarios or see what would happen if they adjusted certain campaign factors.

Now, their clients were no longer making decisions based on guesswork but on clear, concrete data. They saw a measurable increase in customer acquisition rates across nearly all of their accounts. Even more importantly, the agency noticed something surprising: their clients actually trusted them more. Clients could see what was happening in real time and could adjust budgets with far more precision. This allowed the agency to move focus from selling hours to providing value, all because they’d found a way to deliver more than just “ideas.”

The Consumer Brand That Finally Got Transparency

Finally, there’s the consumer brand that had long been at the mercy of opaque marketing reports. They used a traditional MMM provider, and while the results were often helpful, the process was always a bit mysterious. They’d hand over data, wait for weeks, and get back a black-box report with suggestions but no real visibility into how the results were calculated.

Alviss AI didn’t just give them data. It gave them transparency. Now they could see how every dollar in their budget was performing, and they could break down results in ways they never could before. Not only did they gain insight into high-performing channels, but they could also understand why those channels were performing better.

With Alviss AI’s Bayesian modeling, they saw more than just numbers; they saw probability distributions and confidence intervals. They could make decisions with a level of precision they hadn’t had before. Over two quarters, they doubled the accuracy of their predictions, which meant they could push more budget into higher-performing areas without increasing their overall spend.

The Hidden Advantage of Owning the Model

One thing these stories have in common is that all of these brands and agencies were able to achieve better results because they owned the model. It was theirs, it was immediate, and it was transparent. They could test new ideas without needing a new report. They could adapt to seasonal changes without waiting for an outsourced partner to catch up. And they could see precisely how changes in one area would impact results in another.

This is the real value of in-housing MMM: it doesn’t just improve results. It also changes the way you think about marketing. It lets you move from a rigid, slow model to something that feels almost alive a system that adapts and learns as you do.

With Alviss AI, brands and agencies don’t just run their marketing; they understand it. And as these stories show, that understanding has a direct impact on everything from waste reduction to ROI to client trust. It’s not a checkbox anymore. It’s a tool that works exactly when you need it.

This post is part of a 6 part series called “Mastering Marketing Effectiveness with In-Housed MMM”. The posts are outlined below.

• Post 1: The New Era of Marketing Mix Modeling
• Post 2: Top 5 Challenges in Marketing Mix Modeling
• Post 3: How Alviss AI’s Bayesian Approach Enhances Marketing Mix Modeling Accuracy
• Post 4: DIY Marketing Mix Modeling: A Step-by-Step Guide for Brands and Agencies
• Post 5: In-House Marketing Mix Modeling Success Stories with Alviss AI
• Post 6: Future-Proofing Marketing with AI-Driven Marketing Mix Modeling

Introduction

In-housing your marketing mix modeling (MMM) doesn’t have to be complicated. Until recently, most brands and agencies had to rely on external consultants or third-party platforms to understand how their marketing efforts drive results. But as MMM technology has advanced, it’s become more accessible. Today, a lot of the data crunching, modeling, and testing that used to take weeks can be done in-house and faster. That’s the real advantage of tools like Alviss AI. They don’t just let you analyze your marketing mix; they give you the flexibility to do it yourself.

The steps to set up your own MMM model are simpler than you might think, especially if you’ve got a handle on your data. Here’s a quick guide to get started using Alviss AI.

1. Start with Good Data

The first step to setting up a marketing mix model is having the right data. For MMM to work well, you need historical data on your marketing efforts, ideally from different channels. This includes media spend, impressions, clicks, sales numbers, or any other key metric you track to understand performance.

Your data doesn’t have to be flawless, but it does need to be consistent. Make sure you have similar metrics over similar time frames across different channels. If TV ads are tracked monthly and digital ads are tracked weekly, adjust so that you’re working with matching intervals. If the data isn’t aligned, your model will get messy.

2. Upload Your Data to Alviss AI

Alviss AI was built to simplify the data upload process. Once you’ve collected and organized your data, the platform guides you through the upload step. You can either drag and drop files or import data from integrated sources.

During this step, Alviss AI will help you format the data so it fits with the model structure. That’s especially useful if you’re working with mixed datasets that cover different channels. And once your data is uploaded, Alviss AI immediately starts crunching numbers, cleaning outliers, and aligning variables.

3. Choose Your Model Parameters

This is where things get interesting. Traditionally, marketers have had to rely on complex statistical expertise to tune their models. With Alviss AI, the platform helps set your model parameters based on the data you’ve provided.

Alviss AI uses Bayesian modeling, which is great for marketing data because it adapts as new data comes in. You start by choosing the main parameters: your primary KPIs, the weight of each channel, and the budget you’ve allocated to them. Bayesian modeling lets you easily run scenarios to see how different variables like seasonal changes or special promotions affect your results.

4. Test and Simulate

After setting your parameters, Alviss AI can show you simulations of different marketing mixes. This is where the platform’s Bayesian approach really shines. You can adjust budget allocations and immediately see the effect of these changes on your overall goals. Instead of waiting for a consultant’s report to come back, you can test ideas on the spot.

Want to see what happens if you put more budget into social media? Or what about cutting back on TV to see if it impacts sales? Running these simulations lets you test assumptions before you commit any real spend.

5. Analyze and Refine

Once you’ve simulated different mixes, you’ll have a clearer picture of which channels are contributing the most to your KPIs. Alviss AI provides a breakdown of results, showing how each channel affects your outcome and how changes in budget influence performance.

At this point, you might see things you didn’t expect. Maybe a channel you’ve been relying on is actually underperforming. Or maybe a smaller channel, like influencer marketing, is giving a much higher return on investment than you’d realized. The insights from this analysis let you refine your strategy with data you understand because you generated it.

6. Implement, Track, Repeat

Marketing mix modeling isn’t a “set it and forget it” tool. Markets change, competitors adapt, and your strategies should too. One of the strengths of Alviss AI is that it allows you to continually track and update your model as new data comes in. You can keep refining your parameters, simulating new scenarios, and adjusting your marketing mix as needed.

The more you use it, the more intuitive your modeling becomes. Instead of being locked into quarterly reports or waiting for insights from outside, you can be agile with your budget. You can act on trends as they happen.

Getting Started with Your Own MMM

Running your own marketing mix model in-house gives you more control over your data, budget, and insights. It also means you don’t have to rely on generic models or wait weeks for results. With tools like Alviss AI, brands and agencies can get more from their MMM by bringing it in-house.

So, start with good data, set up your parameters, and experiment. Once you’ve run a few scenarios and seen what’s possible, you’ll wonder how you ever trusted your budget to outsourced MMM in the first place.

This post is part of a 6 part series called “Mastering Marketing Effectiveness with In-Housed MMM”. The posts are outlined below.

• Post 1: The New Era of Marketing Mix Modeling
• Post 2: Top 5 Challenges in Marketing Mix Modeling
• Post 3: How Alviss AI’s Bayesian Approach Enhances Marketing Mix Modeling Accuracy
• Post 4: DIY Marketing Mix Modeling: A Step-by-Step Guide for Brands and Agencies
• Post 5: In-House Marketing Mix Modeling Success Stories with Alviss AI
• Post 6: Future-Proofing Marketing with AI-Driven Marketing Mix Modeling

Introduction

If you’ve worked in marketing long enough, you’ve probably heard that Marketing Mix Modeling (MMM) is half science, half art. Traditional MMM can help companies make smarter choices about where to spend their budgets, but it’s not exactly foolproof. The more complex the data, the harder it gets to understand what’s actually driving results. Most models try to fit a lot of moving parts into one framework, and that leads to all kinds of uncertainty. And if you’re working with a lot of uncertainty, you’re bound to get hit with surprises some good, but often, expensive.

At Alviss AI, we take a different approach. Instead of just trying to squeeze every factor into a single predictive model, we use something called Bayesian modeling. And this isn’t just a fancy statistical trick. Bayesian modeling lets us embrace uncertainty rather than ignore it. By doing so, we get more accurate insights, which are exactly what marketing teams need to avoid guesswork and start making decisions based on what’s most likely to be true.

Embracing Uncertainty

The way most traditional MMM models work is that they take a bunch of data from different marketing channels and fit them into a single equation. Then, the model says something like, “Based on this data, we estimate that TV drives 20% of your sales, digital drives 40%, and so on.” But let’s be honest no one is ever 100% sure of those numbers. And what if a new data source or a competitor’s unexpected move shakes up the whole market? Traditional models don’t adapt well to that. They get locked into their initial assumptions, and you’re left guessing how reliable the numbers are.

Bayesian models are different. Instead of just making a single guess about how much impact each marketing channel has, Bayesian models give us a range of possible answers and the probability of each. This gives marketing teams a clearer picture of what’s likely and what’s not, so they can make decisions with confidence, even when the landscape changes.

Why Probabilities Matter

Here’s where Bayesian modeling shines. Let’s say you’re launching a big campaign and want to know how much to invest in different channels. With a Bayesian model, you don’t just get a hard number saying, “spend $500K on digital.” Instead, you get a range that says something like, “if you spend between $400K and $600K on digital, there’s a 90% chance of hitting your target.” Now you have something you can work with, something that reflects the real-world uncertainty of marketing.

In traditional models, there’s usually just one set of answers: “Digital has a 40% impact, TV has a 20% impact,” and so on. But in the real world, things are messier. Channels influence each other, customer behavior shifts, and data isn’t always precise. A Bayesian approach accounts for all this. It doesn’t just assume that digital will always have the same impact or that TV spending will produce predictable results. Instead, it shows you the most likely outcomes based on the data you have and keeps those probabilities updated as new data comes in.

Better Decisions, Better Outcomes

So why does any of this matter? Because when you’re working with better insights, you make better decisions. Marketing budgets are rarely flexible, and every dollar counts. With a Bayesian model, you’re not just placing your bets on a single number or relying on past data to predict the future. You’re looking at the probabilities, weighing the risks, and making decisions that are based on what’s most likely to work now, not last year.

At Alviss AI, we’ve built Bayesian models that are flexible and transparent. That means our clients can see exactly where the model is less certain and where it’s rock solid. It’s like having a compass that doesn’t just point north but also tells you if there’s a storm ahead. Instead of following an outdated map, you’re getting directions in real time, adapting to changes as they happen.

The Value of Transparency

There’s another benefit to Bayesian modeling that often goes unnoticed: transparency. Because Bayesian models offer a range of possible outcomes, they’re inherently more open about where the data is strong and where it’s shaky. Traditional models tend to hide these uncertainties, presenting a single answer as if it were gospel. But in marketing, there’s rarely a single “right” answer. Every market is a mix of shifting trends, customer behaviors, and unpredictable competitors.

With Alviss AI, we wanted to create a model that shows its work. If there’s high uncertainty around a specific channel’s impact, you’ll know about it. If the data points to a likely outcome, you’ll see that too. This transparency gives teams confidence in their decisions, knowing that they’re not just guessing. They’re working with a model that’s honest about its limitations and clear about its strengths.

Building a Smarter MMM

Using Bayesian modeling doesn’t just make Marketing Mix Modeling more accurate; it makes it smarter. The world of marketing is fast, and the smartest decisions come from understanding probabilities, not just hard predictions. With Alviss AI, we’re giving marketing teams the tools to see what’s most likely, what’s possible, and what’s risky.

When you have a model that accounts for uncertainty, you’re no longer working in the dark. You’re working with something that respects the reality of marketing that every decision has some risk, but with the right insights, you can make that risk work in your favor. That’s the kind of clarity that lets you invest confidently and get results that matter.

This post is part of a 6 part series called “Mastering Marketing Effectiveness with In-Housed MMM”. The posts are outlined below.

• Post 1: The New Era of Marketing Mix Modeling
• Post 2: Top 5 Challenges in Marketing Mix Modeling
• Post 3: How Alviss AI’s Bayesian Approach Enhances Marketing Mix Modeling Accuracy
• Post 4: DIY Marketing Mix Modeling: A Step-by-Step Guide for Brands and Agencies
• Post 5: In-House Marketing Mix Modeling Success Stories with Alviss AI
• Post 6: Future-Proofing Marketing with AI-Driven Marketing Mix Modeling

Introduction

Marketing Mix Modeling (MMM) has been a cornerstone of marketing strategy for decades, guiding brands on where to put their ad dollars for maximum impact. But for many teams, working with MMM has also been a bit of a headache. Between long turnaround times, limited flexibility, and high costs, traditional MMM models often feel more like a burden than a tool. So why do brands keep coming back to it? Because, at its core, MMM has value. It’s just that until recently, brands haven’t had a practical way to harness that value on their own terms.

Here are the top five challenges brands face with traditional MMM and how in-housing their MMM with a tool like Alviss AI can solve these issues.

Time-Intensive Processes

The traditional process of setting up, running, and updating an MMM model can be painfully slow. Many outsourced models are built on fixed cycles, often quarterly or annually, which means brands have to wait weeks or even months to see results. And that’s in a best-case scenario. If there’s a data error or a change in campaign strategy, that timeline can stretch even longer.

In-house MMM changes this dynamic. With a platform like Alviss AI, brands don’t have to wait on third-party schedules or deal with long data processing cycles. They can pull in their own data, run the models, and see results in real-time. This immediate feedback allows marketing teams to adjust campaigns on the fly, capitalizing on trends as they happen rather than analyzing them months down the line.

Lack of Transparency

Most traditional MMM models are built and operated as black boxes. Brands receive reports and metrics, but they rarely see the inner workings of the model. This makes it hard to understand which factors are really driving results and leaves brands at the mercy of their agency’s priorities or methodologies.

When brands bring MMM in-house, they get full transparency. They have access to every part of the model data, algorithms, and calculations. They’re able to see exactly how their investments are being measured and can dig into the details if something seems off. Alviss AI, for example, offers a customizable interface where teams can adjust parameters and see how changes affect results. Instead of getting pre-packaged answers, brands get insights they can tailor and trust.

High Costs of Outsourced Solutions

Marketing Mix Modeling can be expensive. When brands outsource MMM, they’re often paying not just for the model but for all the overhead that comes with an agency management fees, consulting hours, and complex software licenses. For many brands, this cost is a barrier to using MMM consistently.

In-housing MMM significantly reduces these costs. By moving MMM to an internal team and using a dedicated platform like Alviss AI, brands can cut out the middleman. They pay only for the tool itself and don’t have to worry about surprise fees or expensive consulting hours. And once they’re up and running, they have the flexibility to run the model as often as they want without incurring additional charges.

Lack of Customization

Every brand is different, but outsourced MMM solutions are often one-size-fits-all. Agencies tend to standardize their models because it’s more efficient, but this approach leaves brands with models that may not fully reflect their unique markets or marketing mixes. Brands with complex data needs, seasonality, or a specific target audience can find that a generic model doesn’t deliver the insights they need.

With in-house MMM, brands have the freedom to customize. They can adjust their model to reflect exactly what matters to them, whether it’s specific marketing channels, regional differences, or seasonal spikes in demand. Alviss AI makes it easy for brands to tailor models to their own strategies, allowing them to adjust the weight of different media types, add unique datasets, or create specialized models for different regions. This customization gives brands a level of insight that no off-the-shelf model can provide.

Challenges with Data Privacy and Security

When brands outsource MMM, they’re also sharing sensitive information about their campaigns, budgets, and customer behavior with a third party. This can raise privacy concerns, especially as data regulations become more stringent. Handing over this data to an external team isn’t always the most secure option, and it can limit how brands use and store their data.

In-house MMM addresses this issue by keeping all data in the brand’s control. Instead of sending data outside the organization, brands can store and process it internally, reducing the risk of data exposure. With Alviss AI, brands retain ownership of their data and can decide exactly how it’s used and shared. This added control makes it easier to comply with regulations and gives brands peace of mind about their data security.

Wrapping up

By bringing MMM in-house, brands are solving old problems with a new approach. They’re no longer waiting on long cycles or accepting a model they can’t see or customize. Instead, they get a tool that’s fast, transparent, cost-effective, and customizable. And with platforms like Alviss AI making in-house MMM easier than ever, brands have the chance to take full control over their marketing insights.

In-house MMM isn’t just about saving money or keeping data secure. It’s about having the power to react, adjust, and experiment at a pace that matches modern marketing. For brands that want to stay competitive, that’s the kind of control that makes all the difference.

This post is part of a 6 part series called “Mastering Marketing Effectiveness with In-Housed MMM”. The posts are outlined below.

• Post 1: The New Era of Marketing Mix Modeling
• Post 2: Top 5 Challenges in Marketing Mix Modeling
• Post 3: How Alviss AI’s Bayesian Approach Enhances Marketing Mix Modeling Accuracy
• Post 4: DIY Marketing Mix Modeling: A Step-by-Step Guide for Brands and Agencies
• Post 5: In-House Marketing Mix Modeling Success Stories with Alviss AI
• Post 6: Future-Proofing Marketing with AI-Driven Marketing Mix Modeling

Introduction

For years, Marketing Mix Modeling (MMM) has been a mysterious art handed off to specialists. Agencies or third-party providers managed the models, and brands waited for periodic reports to tell them how their marketing dollars were performing. But now, things are shifting. More brands are bringing MMM in-house, and it’s changing everything about how they approach marketing.

Drivers of Change

Why are brands choosing to take control of something they once happily outsourced? In one word: speed. Well, maybe three words: speed, control, and flexibility.

When MMM first became popular, data was scarce. You might have weekly or monthly numbers, so it made sense to work with a partner who would run the models on their own schedule. But now, data comes in constantly. With digital channels, brands can get data every minute. When you’re moving that fast, waiting weeks for an updated report is like navigating with last month’s map.

In-housing MMM lets brands update their models in real-time. They can see exactly how their campaigns are performing, make adjustments on the fly, and test new ideas without waiting for approval from someone else. It turns MMM from a lagging indicator into a tool that can drive immediate decisions.

Control is another big factor. Traditionally, MMM has been a bit of a black box. Most brands only saw the output, not the inner workings. They didn’t know how the model was set up or if it was giving them the full picture. In-housing gives brands access to the raw data and the model itself. They can tweak it to fit their specific needs, try out new data sources, or look at results from a different angle. They’re no longer dependent on the agency’s priorities or methodologies.

Flexibility is the third big driver here. Most agencies have a standard approach to MMM because it’s efficient to run models the same way for each client. But brands are not all the same. Some rely heavily on digital ads, while others might see better returns from TV or in-store promotions. An in-house team can customize the model to fit the unique needs of their business. They can even adjust the model for seasonality, regional differences, or any other factor that’s important to them. They’re no longer limited to a one-size-fits-all approach.

Enabling the Transition

The rise of platforms like Alviss AI is what’s making this shift possible. Alviss AI offers brands and agencies a way to do MMM in-house without needing a team of data scientists. It’s a platform that combines the latest machine learning techniques with an intuitive interface that anyone can use. Alviss AI handles the hard stuff data integration, model setup, algorithm selection so that marketing teams can focus on what the data is telling them.

There’s another side benefit to in-housing MMM: transparency. When brands outsource their MMM, they’re trusting a third party with sensitive information about their marketing spend, performance, and customer behavior. In-house MMM means they keep control over that data. They’re not handing over the keys to their marketing strategy to someone else. And in a world where data privacy and security are top concerns, that control matters more than ever.

We’re seeing a new era in marketing, one where brands have the tools and the data to make their own decisions. In-house MMM is just one part of this shift, but it’s an important one. It gives brands control over their data, lets them respond faster, and frees them from a one-size-fits-all approach.

Brands that bring MMM in-house are discovering it’s not just a way to analyze past campaigns. It’s a tool for shaping their strategy in real-time. And in a competitive market, that can make all the difference.

This post is part of a 6 part series called “Mastering Marketing Effectiveness with In-Housed MMM”. The posts are outlined below.

• Post 1: The New Era of Marketing Mix Modeling
• Post 2: Top 5 Challenges in Marketing Mix Modeling
• Post 3: How Alviss AI’s Bayesian Approach Enhances Marketing Mix Modeling Accuracy
• Post 4: DIY Marketing Mix Modeling: A Step-by-Step Guide for Brands and Agencies
• Post 5: In-House Marketing Mix Modeling Success Stories with Alviss AI
• Post 6: Future-Proofing Marketing with AI-Driven Marketing Mix Modeling

High-level domain knowledge

We think the correct thing is to express high-level dynamics that come from knowledge of the specific business, which ensures sane causality.

For this example, we will do a small model that consists of online media (Online), offline media (Offline), brand awareness (Brand), net promoter score (NPS), product price (Price) and number of items sold (Sales).

Just from pure causality we know that the business buys advertisements in online and offline media and that drives Brand, NPS, and Sales. So by knowing this we can eliminate relationships such as illustrated in the figure on the right.

Since we know this does not adhere to the correct causality of the real world. What one instead wants to do is to have a conversation with a person that actually understands the dynamics of the specific business to construct what is considered to be the correct causal relationship.

In this case, we assume that Brand and NPS are unrelated to each other since the brand measures a high-level awareness in the population and NPS is direct customer interaction. There is a possibility of NPS affecting the brand, but we will ignore that for the sake of this example. So for this case, we assume a causal relationship that looks like the right hand figure.

This is a simplified version for illustrative purposes, in reality, there are most likely more factors and relationships that will come into play in a production model such as macro economical factors, seasonality, etc.

What one also should be able to encode on this high level is if there are any strict relationships such as when Brand goes up Sales needs to go up i.e. if more people know about the business and like the business sales should not be negatively impacted by this.

Low-level domain knowledge

For high-level domain knowledge, one thinks about the causal relationship between the different variables. With low-level domain knowledge, one focuses on how to model the specific ways to model a variable such as offline media, and how to model the interaction between variables such as how Offline Media and price interact to impact sales.

Model combinations at scale

As the amount of possible ways of modeling each variable grows one quickly get hundreds to thousands of possible combinations, so in order to actually be able to try all of the combinations and still have the correct uncertainty estimation from Bayesian Inference we rely on a technique called Variational Inference.

Variational inference turns the normal sampling problem of Bayesian inference into an optimization problem that is a lot faster to solve. So instead of a model taking hours to train it takes minutes to train. Combining this with elastic compute resources, one can fit the models in parallel further reducing the time to actually try all the combinations.