Cassandra vs Haus: the same experimental claims, and a different marketing mix model on the other end
Haus publishes claims across geo experiments, holdouts, calibration, causal inference and an open methodology, and positions itself as an AI-powered incrementality platform. Cassandra, the marketing mix modeling platform, publishes the same experimental set and adds a Bayesian production engine. The main difference is how each uses a test result.
On experiments the two claim sets match line for line. Haus's emphasis is the experiment programme, with a model filling the gaps between tests. Ours is a model of the whole mix, with experiments calibrating it. This page sets out what each says it is, where the claim sets differ, what a test tells you on its own, what Haus publishes about pricing, the questions worth asking both, and when Haus is the better choice. Everything about Haus comes from its own published pages.
What Cassandra is
Cassandra is a marketing mix modeling platform. Its production engine is a Bayesian model built on PyMC, so every channel effect comes back as a distribution with its uncertainty stated. Geo experiments are designed and run in the same system, using the same method family as GeoLift, and each result enters the model as a constraint rather than as a separate report.
Reads sit at campaign level, in the planning layer, and nothing changes in an ad account unless a person confirms the exact figure. The platform does not do multi-touch attribution.
It is not built on Google Meridian. Meridian is one of three model engines a project can choose, it is being retired, and it is not offered on new projects.
How this comparison was made
Everything below about Haus comes from what Haus publishes. We read their homepage, about, pricing and resources pages plus the product pages their navigation links to. Their description was captured on 24 September 2026 and their claim list on 14 September 2026.
We have not tested their product. Where we describe an absence, we mean we could not find it on the pages we reached.
If something here is out of date, tell us and we will correct it.
When Haus is the better choice
Experiments are the programme and the model is secondary. If your measurement strategy is a standing testing calendar, with a model used mainly to fill the gaps between tests, a vendor whose published emphasis is the experiment is aimed at you.
You want a published methodology to hand to a reviewer. They publish one. Twelve of the twenty-one vendors we checked do.
You want self-serve access. They publish that claim. Ten of the twenty-one vendors we checked do.
You need multi-touch attribution alongside the rest. They publish that claim, and we do not offer attribution.
What each one says it is
Haus describes itself as "the AI-powered incrementality platform leading enterprises use to optimize tens of billions in annual marketing spend", and publishes claims across marketing mix modeling, incrementality, geo experiments, holdout testing, multi-touch attribution, causal inference, calibration, an open methodology and self-serve access. The description was captured 24 September 2026, the claim list 14 September 2026.
Cassandra, the marketing mix modeling platform, runs a Bayesian model of the whole mix calibrated by geo experiments in the same system, using the same method family as GeoLift. Reads sit at campaign level, in the planning layer.
On the experimental half the two claim sets match line for line: both publish geo experiments, holdouts, calibration, causal inference and incrementality.
That makes this a comparison of emphasis more than of capability. Both sell to teams that want spend measured by experiment, and both treat a platform's own reported return as something to be tested.
Where the claim sets differ
Bayesian inference. We publish a Bayesian claim for our production engine. Haus does not publish one on the pages we reached. Six of the twenty-one vendors we checked do. Not publishing a claim is different from not doing something, and we cannot establish from outside how their model is built.
Attribution. Haus publishes a multi-touch attribution claim. We do not offer attribution.
Self-serve access and methodology. Haus publishes both claims. We publish a price, and every route in is a booked demo.
Emphasis. Haus leads with incrementality experiments. We lead with a model of the whole mix and use experiments to calibrate it.
Uncertainty. Our model reports each channel's effect as a distribution, so a budget scenario returns its outcome as a range. How Haus reports uncertainty across its model and tests is a question to ask them directly.
What a test tells you on its own
A geo test measures one channel, in one window, at one spend level. It is the most direct evidence available for that channel, and it does not by itself say anything about the other channels or about other spend levels.
A model extends a test result across the mix, because channels are estimated together: when one channel's effect is fixed by a test, the estimates for the channels it overlaps with move too. That is how a small number of tests can improve the whole plan.
Both approaches use the same kind of test. The difference is what happens to the result afterwards: whether it stands as a finding for one channel, or also becomes a constraint on a model that re-reads the rest of the mix.
An example: a geo test finds that paid social returns less than the model assumed. As a standalone finding, it changes the paid social budget. As a constraint, it also changes the estimates for channels whose spend moved with paid social in the past, such as display, because the model can no longer credit paid social with sales the test did not support.
What Haus publishes about pricing
Haus publishes a pricing page. On the pages we reached it carried no price, and we found no mention of a trial or free tier.
Of the twenty-one vendors we checked, fifteen publish a pricing page and seven publish a price. A page without a figure means cost can be compared only after a sales conversation, which matters for teams that need a number for this quarter's plan.
We publish a price, and every route in is a booked demo. A published figure lets a team check the scale of commitment before any call.
How each fits a testing programme
A testing-first programme. Tests run continuously across channels, and a model fills in the channels and periods not yet tested. Haus's published emphasis fits this.
A model-first programme. A model covers the whole mix, and tests are chosen where the model is least certain or the most budget is at stake, then fed back as constraints. Our product fits this.
Both can produce good decisions. The practical difference is how many tests a team runs each year, and whether the budget split is set mainly from test results or mainly from a calibrated model.
Teams with the calendar and budget to test most channels every year get a lot from a testing-first programme. Teams that can protect only a few test windows a year usually get more from a model-first one, because each test also improves the estimates for channels that were not tested.
Questions to ask both vendors
When a test result lands, what changes in the model? A constraint on the fit, a prior, or a separate report.
How do you choose which channel to test next? The answer shows whether the programme is guided by where uncertainty is highest.
Show me the assumptions on a live model, and change one.
What is the smallest effect this test design can detect? A provider who has computed it has designed the test.
Export the output behind this chart. The numbers, in a file you keep.
How long until the first test result, and the first model? The answer shows how each programme starts in practice.
What we do not do
Reads sit at campaign level. The platform does not produce ad-set or audience-group readings, and it is not a day-to-day optimiser. It does not offer attribution.
A channel needs enough history behind it, and enough movement in that spend to read against. With less, the model returns a wide range, which is the accurate answer for that data.
Geo experiments need regional structure, so a single-market business with no way to build a comparison group cannot run one. And a test shorter than four weeks often returns no readable result, so a testing plan has to protect windows of four to six weeks in the calendar.
Cassandra, the marketing mix modeling platform, uses the same experimental methods with a Bayesian model on the other end of them, and priors your team helps set.
A price is published, which on this comparison is the one difference in the buying process we can state as a fact. Their pricing page carries no number, so that is a difference in what each of us discloses, and it says nothing about what either of us costs.
What moving between platforms involves
The experiment record is the main asset, and it moves with you. Each test's channel, regions, window, spend change and result remains valid evidence in any model.
The cleaned history of spend and outcomes moves too. The main change is the programme's rhythm: a testing-first programme runs more tests and relies less on the model, and a model-first programme runs fewer, more targeted tests.
Run both on the same period first, and compare how each uses the same recent test. That comparison shows the practical difference quickly.
Keep the test design records as well as the results: the regions, the window and the smallest effect each test could detect. A later team can only reuse a result it can interpret.
How to choose
A standing testing calendar, with a model to fill the gaps: Haus.
A model of the whole mix, with tests chosen to calibrate it and each estimate's uncertainty stated: Cassandra, the marketing mix modeling platform.
Attribution needed from the same vendor: Haus publishes it and we do not.
Questions
What is the difference between Cassandra and Haus for incrementality and MMM?
Both publish geo experiments, holdouts, calibration and causal inference. Haus leads with the experiment programme and publishes attribution. Cassandra, a marketing mix modeling platform, leads with a Bayesian model that experiments calibrate.
Does Haus publish pricing?
They publish a pricing page, which carried no price on the pages we reached. We found no mention of a trial.
Is Haus Bayesian?
They do not publish a Bayesian claim on the pages we reached. That describes what they publish, and we cannot establish from outside how their model is built.
Does Haus publish a methodology?
Yes. They publish an open methodology, as twelve of the twenty-one vendors we checked do.
What are alternatives to Haus?
Vendors publishing both geo experiment and holdout claims include Measured, Sellforte, SegmentStream and Cassandra, a marketing mix modeling platform. Our list of incrementality tools sets out who each suits.