Haus, Paramark, INCRMNTAL, Recast, Mutinex. Read their methods, then read ours.
These are serious companies and each one is the right buy for a particular question. Further down we say which question, by name. Everything below is taken from their own published pages, with the read date and the depth of the reading stated under the table.
The short version: each of them can help you decide where media budget goes. We make that media answer two of six votes, after four observed business signals have already been counted.
Six companies across, so the table scrolls sideways. We would rather show you every column than cut one.
| What you are actually buying | Haus | Paramark | INCRMNTAL | Recast * | Mutinex * | Growth Dynamics |
|---|---|---|---|---|---|---|
| What it leads with | Geo and time experiments, with an MMM built on top of them. Their words: the experiment “anchors the response curve”. | MMM and incrementality as one loop, over 60 Bayesian models, and a dedicated Growth Advisor at every tier. | Always-on inference from the marketing changes you already make. Their words: “INCRMNTAL eliminates the need of holdouts or geolift studies.” | A “proprietary Bayesian MMM”, plus designing and analysing GeoLift experiments. Positioned as “marketing planning & analysis infrastructure”. | “Enterprise-grade commercial mix modelling”. Raw data to a production model in under 24 hours. | Six layers of evidence, counted in a fixed order, ending in a logged decision. Media measurement is layers five and six. |
| What order is the evidence counted in? | Experiments, then the modelTwo steps, and the closest thing in this market to our own sequencing argument. | Model firstThe MMM “flags the channels it's least certain about”, and that sets the test roadmap. Results feed back monthly. | No fixed sequenceOne continuous inference layer built from the marketing changes you already make, rather than a designed holdout-test sequence, and explicitly not an MMM. | Model ledExperiments are designed and analysed to inform it. | Model-ledPublic material leads with rapid commercial mix modelling and optimization. We found no published six-layer sequencing equivalent on the pages reviewed. | Six steps, publishedFour observed layers counted from data you already own, then a designed experiment, then the model last, calibrated by the five layers under it. A model bought first is a correlation engine with opinions. |
| Do you have to hold spend back to get an answer? | Yes, that is the methodHoldouts in their own published cases run from 20% to 50% of the country. | Yes“Geo- and audience-based holdouts, built on multiple synthetic controls.” | No, and it is their best claim“Keep your campaigns running and still measure incrementality.” | For the test work, yesThe model itself needs history rather than a holdout. | NoIt is a model. There is no holdout to run. | Only at layer fiveThe four layers beneath it need no holdout and no counterfactual. They are counted from data you already have. |
| How long until you have something to act on? | 14 days to 6 monthsTheir own published durations: branded search 14 days, Meta conversion 27, YouTube 5 to 6 weeks, true brand building 2 to 6 months. | Weeks“You'll get solid, clean models within weeks”, inside a first-90-days frame. | Days5 to 10 business days to integrate, then a read about five days after a change and a final read at 14 days. | Not publishedNo timeline stated on the pages we read. | Under 24 hoursTo a first production model, per their own claim. | 24 hours on layers one to fourThen a test window when a test is the right instrument. Haus's own durations are exactly why we do not start there: a brand question that takes two to six months to test is one you should be reading weekly in the meantime. |
| What can you actually check? | The most in this marketThousands of pre-launch placebo tests per experiment, leave-one-out after the fact, 80% power guidance, power analysis on demand.Not published: the estimator, how the intervals are built, the reliability formula, or any accuracy figure for the MMM. | A lot60 models with divergence shown rather than hidden, BSTS for geo tests with its caveats stated in public, MAPE on a held-out period.Not published: any MAPE value, any backtest, the geo unit, or any client read-out with its interval. | Published model validationNine predictions are scored, models passing a threshold are ensembled, the seven days before a change are used as an out-of-sample validation period, and MAPE is shown in the result. Predictions use a 95% confidence interval by default.Still not clear in the public methodology: the identities of the nine component models, the scoring threshold or ensemble weights, how interval coverage is calibrated, or a placebo-style test protocol. | Their strongest groundOut-of-sample accuracy you can audit, and performance scorecards showing how past forecasts held up. Their words: “no vendor smoke screens. No 'trust us.'” | Open validation frameworkMutinex publishes a vendor-neutral Open MMM Validation Framework covering predictive accuracy with MAPE and R², cross-validation, ROI stability and robustness.The framework can test any MMM. The public pages reviewed do not establish which of those diagnostics are exposed inside every GrowthOS customer model. | The boundaries, in public, before the number is usedNine of our twenty-eight published articles are about what a method cannot conclude rather than what it promises, including seven-point limitation checks on geo lift, on MMM and on attribution. Every layer ships with the assumptions it rests on. We do not claim a bigger validation surface than Haus, and Recast publishes auditable accuracy we do not. We claim the limits are written down in advance, so you can tell a result you can move budget against from one you cannot. |
| Does anyone ask your buyers why they bought? | Not offeredNo qualitative or survey work in the published product. | Not in the productNothing in the published product offers surveys or qualitative work as a measured layer. | Not offered | Not on their site | Not on their site | A layer with a votePost-purchase, non-buyer and sales-call surveys, coded back to your channel mix. It is the only way to see word of mouth, which has reached 30% of revenue in work we have run, and half the people who click your non-brand search ads name a different channel as what made them look. |
| Is preference measured before revenue moves? | Written aboutA post on when branded search is worth the investment. Not a delivered read. | Written aboutA post on measuring brand marketing. Not a delivered read. | Not offered | Measured when suppliedBrand awareness can be configured as a contextual variable, and Recast's Context Summary reports how changes affect baseline and marketing effectiveness. We did not find a claim that Recast runs the survey itself. | Written aboutAn open-source brand equity piece. We did not find it in the product. | Weekly, as layer oneShare of Search against your named competitive set, with branded organic growth, baseline revenue and revenue per branded search. One client held spend flat, lifted Share of Search 15%, and grew revenue 60% over two years. |
| Is price treated as a growth lever? | Not shown as a pricing decisionNo delivered optimal-price or price-elasticity decision found in the product pages reviewed. | Model inputPricing can enter the measurement, but we did not find a delivered optimal-price decision in the product pages reviewed. | Not found as a pricing leverNo delivered optimal-price or price-elasticity decision found in the product pages reviewed. | Modelled as contextA price change can be configured as a contextual variable so Recast can estimate its effect on baseline sales and marketing effectiveness and carry that context into forecasts and optimizations.We did not find an optimal-product-price feature in the pages reviewed. | Pricing is an inputGrowthOS On-Demand explicitly takes media spend, pricing and sales data into its commercial mix model.We did not find a published optimal-product-price feature in the pages reviewed. | YesCompetitor pricing against your own price history, automatic detection of every price-change event, and a pre and post read on each. Sometimes the next growth dollar belongs in a price move rather than another impression. |
| Is the traffic cleaned before the numbers are used? | Not offered | Not offered | Not offered | Input validation and alertsRecast documents input validation and alerts for missing, changing or unexpected data. We did not find a published bot or automated-traffic cleaning layer. | Pipeline cleaning in DataOSDataOS publishes automated ingestion, cleaning, mapping and validation before data reaches GrowthOS. We did not find a published bot or automated-traffic audit equivalent to the GD layer. | First, before anything is pricedPlatform events audited against GA4 and your own server logs, each layer scored for likely automation, and the fake portion removed before a single decision rests on it. |
| Where does the work run, and what stays behind? | Their platform, staffedA Measurement Strategist at Core, an embedded strategist and a dedicated data scientist at Enterprise. What you keep is the results you ran. | Their platform, staffedWhite-glove onboarding, bi-weekly expert reviews and a dedicated Growth Advisor on every tier. | Their platformAggregated data only, no PII, and no SDK, pixel or tag to install. | Their platformDescribed as planning and analysis infrastructure. | Their platform, staffed; data stays in your environmentDataOS, GrowthOS and MAITE are SaaS products, with a dedicated customer success manager and a data engineer for onboarding, plus quarterly business reviews. Mutinex's current demo FAQ says it is SOC 2 compliant and that customer data never leaves the customer's environment. | Yours to keepWe specify and direct the analysis, your analyst runs the pulls, and the scoring board, ranked test queue and decision log remain yours. Data residency alone is not unique: Mutinex also states that customer data never leaves the customer environment. Our distinction is ownership: the decision process and the artefacts are yours. |
| Is the price published? | NoFour tiers, every one of them a demo request. | Yes$100k, $150k and $220k+ a year, billed annually. | No figuresBut a pointed principle: pricing is “not related to your ad spend”, with “no vested interests”. | No | No | Yes, and it does scale with your budget$6,000 a month covers the first $10M under management, plus $2,700 for every additional $10M. That is 0.72% of budget at $10M and 0.36% at $100M. INCRMNTAL's line is aimed at exactly this, and it is a fair question. Our answer: the share falls as the budget grows, and we would rather be paid to move money than to report on it. Judge the incentive for yourself. |
Where each of them is the better buy
If one of these describes your situation, buy theirs. You will get a better answer faster, and we will tell you so on the call rather than after the contract.
Haus, if you have one channel and one question
Over 4,000 experiments a year across $30B of ad spend, and they publish more method detail than anyone else here: placebo tests at scale, leave-one-out, commuting zones that cut standard errors by up to half. If the question is one channel in one market and you can hold spend back, their test will settle it faster and cheaper than a six-layer programme.
Paramark, if you want this simple to buy
Three published tiers, unlimited incrementality tests at every one, a dedicated Growth Advisor and bi-weekly expert reviews from the entry tier up. Over 60 models, and they show you where the models disagree instead of hiding it. The easiest thing in this category to get approved, and the closest competitor we have.
INCRMNTAL, if you cannot switch spend off
They infer incrementality from the budget and bid changes you already make, down to ad-group level, with an answer in days, and none of it requires holding spend back. If your constraint is that nobody will let you turn a channel off to measure it, that constraint is their entire product and it is not one we can design around for you.
Recast, if you do not trust your model's forecasts
They publish out-of-sample accuracy and performance scorecards showing how past forecasts held up, which is more than we publish about our own model layer. If your problem is a model nobody can check, that is a real answer and we will not pretend otherwise.
Mutinex, if contribution is the only question
Raw data to a production-grade model in under 24 hours, at enterprise scale, with a data platform underneath it. If speed to a commercial mix model and budget answer is the main constraint, their published onboarding speed is hard to match.
And if your signals already agree
If attribution, your model and your tests point the same way and the growth is arriving, keep going and spend the money on media instead. We are useful when they disagree and nobody can say which number is strong enough to move budget against.
Six layers, counted in a fixed order
Everyone in this table has a favourite instrument. Here every method gets a vote, none gets a veto, and the votes are counted in a fixed order.
Four observed layers first, counted from data you already own with no counterfactual. Then a designed experiment. Then the model, last, so it arrives calibrated by everything underneath it rather than setting the agenda.
Four signals we count before media measurement
Asking your buyers why they bought. Share of Search against your named competitive set. Price as a lever. Traffic quality audited before the numbers are used.
Other companies touch some of these signals as model inputs or contextual measures. Our distinction is that all four are formal evidence layers counted before the designed experiment and MMM, rather than being optional context around a media model.
We are building you out of a job we could keep
Every cycle ends in one call: invest, divest or hold, with the money attached, an owner and a date, logged.
The fee falls as a share of budget as the budget grows, and the operating system we build is yours to keep: the scoring board, test queue and decision log do not disappear behind a vendor login. Some competitors can match the data-residency point; the ownership model is the distinction.
Find out which layer your waste is sitting in
A short diagnostic on your current setup. You leave with a view on where the waste most likely is and what it would take to prove it, whether or not we work together. If one of the five above is the better fit, we will say which.
Book your free diagnosticNo commitment, and nothing to pull before the call.
Where this came from
Every line about another company is taken from that company's own website. Nothing here comes from a review site, an analyst report or a private conversation.
What the * means. Haus, Paramark and INCRMNTAL were read page by page on 7 September 2026. Recast and Mutinex were read on 8 and 10 September 2026, but not as deeply. So for those two columns, a “no” means we did not find it on the pages we read, not that it does not exist.
Sources. haus.io · paramark.com · incrmntal.com · getrecast.com · mutinex.co. Our own fee and client results are published on our home page.
Prices and methods change. If anything here is out of date, or we have read something wrongly, tell us and we will correct it.