Your last-click report shows brand search running at a great cost per acquisition (CPA), week after week. It looks solid, but it may say very little about what brand search actually does for your business.
MMM and attribution both promise to clear that up, yet they answer different questions. The useful comparison is which one you can trust for which decision. Below, we use accuracy and precision to compare them, and then look at why experiments make a good tiebreaker.
Attribution assigns credit for a conversion to the touchpoints (the ads and other interactions) a customer had along the way. Last-click credits the final one, first-click credits the first, and multi-touch attribution (MTA) splits credit across several.
Teams like it because it gives near real-time feedback. Itβs a useful daily signal for managing spend down to the ad level.
The catch is that it runs on correlational data, so you can't know whether a touchpoint caused a sale or just happened near one. Platforms also report on their own ads, which tends to skew their numbers upward. And MTA leans on user-level identity, so privacy rules and cookie loss weaken it.
MMM answers a different question: what moved the business. MMM is a statistical analysis of how marketing and outside factors together influence business outcomes. It tries to separate marketing effects from seasonal, economic, and competitive noise.
It runs on aggregated inputs, like weekly spend, impressions, and outcomes, plus things like holidays, pricing, and promotions. That means it needs no personal data or identifiers. It also covers offline and brand channels like TV, radio, and out-of-home, which MTA mostly can't see.
In the guide to multi-touch attribution vs. MMM, we describe MTA as the microscope and MMM as the telescope. One shows granular campaign detail. The other shows how the whole constellation of marketing works together.
Before you pick a method, ask how much you can trust its numbers. An earlier Haus post on accuracy vs. precision in measurementΒ offers a framework for doing that.
Accuracy is how close measurements land to the truth on average. Precision is how close measurement results land to each other. Picture darts: A tight cluster far from the bullseye is precise, but not accurate.
Marketers mix these up a lot, and the difference matters. Precision without accuracy is a costly miss that can send you down the wrong path.
Here's how common methods stack up:
Click attribution is consistent, which is why a last-click brand search CPA can look great week after week. It just can't say whether those buyers would have converted anyway.
Traditional MMM's accuracy problem comes from spend patterns, which we cover below. Its precision problem is that it often works from aggregate daily data across many channels, so the model has little to run on.
Geo experiments get their accuracy from random variation in how marketing is delivered across the country, which acts as the control. In our approach, we use synthetic control methods to build a weighted average of control regions, then clean outliers after the test to sharpen precision.
MMM has its own blind spots.
You might hear that MMM is the answer because it works without user-level data. That's true, and it protects the model from tracking loss. It doesn't protect the model from confounding, which happens when other factors muddy the result.
IncrementalityΒ asks how many conversions wouldn't have happened without a specific marketing action. Teams usually measure it with randomized tests that compare a test group against a control. The core question is what that group was going to do anyway.
Here's a hypothetical: A region that sees an Instagram campaign has 12% higher sales than a control region. If everything else is consistent, you can reasonably credit that lift to the campaign.
Tests need random assignment and accurate outcome measurement, not complete user journeys, so tracking limits hurt them less.
Calibrate instead of triangulate. "Triangulating" often turns into picking the model you like. Calibrating means using experiment results to filter out MMM variants that don't match reality, or to supply priors (starting assumptions) for Bayesian models. Organizations that do this well often run a few robust experiments per year and tune their ongoing models to match.
That's the thinking behind Haus' products. Causal MMM treats experiment results as ground truth (the trusted baseline), and the model is continually anchored to them to avoid drift. We think of GeoLift as answering whether a channel worked, and Causal MMM as answering what to spend on each channel.
Causal Attribution does the same for daily numbers. It starts with raw pixel data, sets a baseline with experiments, and adjusts for seasonality and spend-level changes. Where no experiment exists, the Incrementality Index (a collection of thousands of incrementality experiments) fills the gap. The output is calibrated return on ad spend (ROAS), CPA, average order value (AOV), orders, and revenue down to the ad level.
In the Causal Attribution announcement, Greg Robinson, head of data at GLD Shop, says: "We've seen an immediate improvement in top-line business performance that neither traditional MTA nor platform reporting alone was able to predict or capture."
Be fair about the costs. Experiments cost more and are more complex. They may require turning off spend in parts of the market in order to get a snapshot, and results may not translate to other periods, regions, or campaigns. That's why they work best as anchors for your other methods.
The marketing attribution fundamentalsΒ guide shows what a simple validation might look like: Your MTA shows paid social rarely gets last-touch credit but appears in 60% of first-touch journeys. A regional holdout test then shows whether conversions drop without it.
Complexity of your mix is typically represented by spend. Experts that Haus has talked to suggest that if you spend under about $100,000 a year, platform tools are usually enough. At that size, consistency matters more than sophistication.
As spend grows, channel mixes get more complex, and thereβs more measurement required to continue to make good decisions. Attribution gives fast feedback, MMM optimizes across channels and budgets, and experiments validate both.
Here's our rule of thumb, based on those roles: Use attribution for daily bid and creative calls, MMM for quarterly budget planning, and an experiment when you need to prove a channel or a big bet.
Whatever you choose, apply the same measurement standards across channels. A patchwork of conflicting methods is harder to act on. That same Open Haus episode goes deeper on calibrating rather than triangulating.
You rarely have to pick just one method. Instead, decide which one anchors the others. We think experiments should. That's the idea behind the Haus platform, which pairs experiments with Causal MMM and Causal Attribution.
Attribution assigns credit for a conversion to the touchpoints a customer interacted with. MMM is a statistical analysis of how marketing and outside factors together influence business outcomes, using aggregated inputs like spend and outcomes.
Not on its own. Attribution runs on correlational data, and traditional MMM can't tell you how much bias its spend patterns add. Randomized tests that compare a test group against a control get closer to the question of what would have happened anyway.
It holds up without personal data or identifiers, so tracking loss doesn't hurt it the way it hurts MTA. But that doesn't protect it from confounding, and two models can hit the same forecast with opposite coefficients.
Triangulating can turn into picking the model you like. Calibrating means using experiment results to filter out MMM variants that don't match reality, or to supply priors for Bayesian models.
Probably not yet. According to the Open Haus episode on the three methods, if you spend under about $100,000 a year, platform tools are usually enough, and consistency matters more than sophistication.
