The first thing Haus wants a marketer to do is mildly terrifying: stop advertising somewhere. Pick a set of regions, withhold or change the media, and watch whether sales wobble. The exercise is an affront to the modern dashboard, where every click arrives bearing a claim for credit. It is also the cleanest way to answer the question that matters. Did the ad cause the sale, or did the sale merely happen near an ad?
Haus Analytics, Inc. has built an enterprise software business around that distinction. Founded in 2021 by Zach Epstein, a former Google data scientist and product manager who spent seven years on advanced analytics and marketing science, the San Francisco company packages causal inference into tools that growth teams can operate without assembling their own small faculty of econometricians. Haus says its customers now run more than 4,000 experiments a year, optimizing more than $30 billion of annual advertising spend across businesses with more than $1 trillion in combined revenue.
The customer list makes the abstraction tangible: FanDuel, StockX, Grubhub, Mejuri, Ritual, ThredUp, Jones Road Beauty, HexClad and Pernod Ricard's Jameson, among others. These are companies selling shoes, skincare, cookware, debt help, food delivery and bets. Their shared problem is not a shortage of numbers. It is a surplus of confident, incompatible numbers generated by ad platforms, attribution vendors and internal models.
A holdout is a tiny act of rebellion
The original product, often described through its GeoLift experiments, divides geography into treatment and control regions. A brand uploads its own outcome data - sales, subscriptions, orders or another business KPI. Haus searches for regions that behave alike, uses randomization or synthetic controls to estimate the counterfactual, and helps the advertiser vary spend in selected markets. The gap between what happened and what probably would have happened becomes incremental lift. Divide the additional profit or revenue by the media cost and the marketer gets an incremental return, not a click-based fable.
That privacy angle matters. Haus does not need to follow a particular person from impression to checkout to learn whether advertising moved a business. Regions become the unit of comparison. As cookies disappear and mobile identifiers become less dependable, geographic testing survives because it measures aggregated outcomes. The pitch is less “we know everything about this shopper” and more “Detroit behaved differently from its statistical control after the campaign changed.”
The company has widened that wedge into a suite. Fixed Geo Tests handle campaigns whose locations cannot be randomized - a billboard push in New York or an activation on the Las Vegas Sphere. GeoFences let a brand exclude states where it cannot sell, markets distorted by regulation or regions reserved for another test. Haus Commuting Zones use mobility patterns to reduce the awkward reality that people see an ad in one county and buy in another. Time Tests estimate the impact of national moments where no clean geographic control is possible. They are more directional, a concession to the limits of the design rather than an attempt to perfume them.
“We actually got a 6x return on the cost of Haus just by redeploying these dollars from one channel to another.”Noa Gutterman, SVP of Marketing at Mistplay
What failed first
The most revealing case study begins with a zero. Supplement brand Ritual tested TikTok and found no measurable lift. A less useful vendor might have softened the result. A less curious client might have killed the channel. Ritual instead changed its creative and ran the test again. The second experiment measured an 8 percent lift at an efficient incremental acquisition cost. The sequence is more instructive than a single triumphant chart: stop, learn, alter the thing under test, and earn the right to scale it.
Zero came first
TikTok showed no lift. New creative and a retest produced 8% lift.
Six tests, five months
The company expected to redirect millions and estimated 4% more borrowers.
The Sphere gets a receipt
A fixed-geo study measured 4.71% lift in warehouse depletions.
OOH meets orders
A New York campaign produced a measured 9% lift in new orders.
Tally provides the operating version. The debt-payoff app ran six experiments in five months. It estimated that the resulting decisions would redistribute millions of dollars annually from lower- to higher-incrementality tactics and increase total borrowers by 4 percent. FanDuel said better measurement saved millions, while an Insight Partners account put one portfolio company's avoided non-incremental investment in the tens of millions and its increase in activations at 11 percent.
The sharpest “changed our mind” story belongs to footwear brand OluKai. Its former marketing-mix-modeling provider recommended more TikTok, Snapchat and YouTube investment while reducing Meta. OluKai followed the advice. The forecast and reality separated: customer acquisition cost arrived 20 percent higher than predicted and performance continued to deteriorate. Haus's geo experiments had indicated that TikTok and Snapchat were less incremental than Meta. A Causal MMM grounded in those tests brought the forecasts close to actual orders; the case study says performance held for weeks at roughly 15 percent lower total spend. Controlled evidence won the argument because the correlated model had already sent the money the wrong way.
From the quarterly verdict to Tuesday morning
Experiments are powerful, but they are episodic. Marketers still need to plan an annual portfolio and make decisions on an ordinary Tuesday. Haus's next two products close that gap. Causal MMM, opened in beta in October 2025, treats experiment results as ground truth for a marketing mix model. The model forecasts channel response and lets teams play with allocations. The distinction from a traditional MMM is philosophical and mathematical: correlations in historical spend do not get to overrule an intervention designed to identify cause.
Causal Attribution tackles the faster rhythm. Released broadly in July 2026 and included with Haus experiments by default, it takes familiar click-and-view signals, then debiases them using experiment results. It can report down to the ad level without pretending that proximity equals causation. GeoLift says what genuinely worked, Causal MMM asks how the whole budget should move, and Causal Attribution steers the smaller daily choices. The products form a feedback loop rather than three unrelated dashboards.
Haus is careful, at least in its published research, about a tempting mistake: confusing more tests with better science. In a 2026 exercise, the company simulated 36 million years of budget decisions. Precise measurement beat a frozen-budget baseline 82 percent of the time. Noisy measurement did so 62 percent of the time, and running 15 noisy tests instead of six barely changed the odds. Under one representative setup, the precise high-cadence approach produced 13.2 percent average revenue lift versus 5.6 percent for the noisy approach. The lesson is unfashionable in software: velocity does not rescue a weak instrument.
What the customer is paying for
Haus sells enterprise subscriptions through order forms, with professional services available under separate statements of work. It does not publish a standard price. The customer-supplied economics are more interesting anyway. Mistplay says reallocating existing channel spend produced a return equal to six times the cost of Haus. In other words, the software does not need to invent a new revenue stream. It needs to find waste in a budget that is already large enough to hide it.
That makes the ideal customer fairly specific: a brand spending meaningfully across several channels, collecting reliable first-party outcome data, operating across enough distinct markets and willing to alter spend for the sake of learning. The buyer may sit in growth or analytics, but the product's political value extends to finance and the executive team. A well-designed holdout can turn “marketing believes” into “the experiment observed,” a sentence with a better chance of surviving a budget meeting.
Investors have financed the expansion. Haus raised a $17 million Series A led by Insight Partners in August 2023, taking publicly reported funding at the time to $24 million. In July 2024, 01 Advisors led another $20 million investment with Rahul Mehta, Gokul Rajaram and existing backers, bringing publicly disclosed capital to at least $44 million. No public valuation is attached. The company is organized around hubs in San Francisco, New York and Seattle, with values unusually on-theme: “Call it what it is,” “Done is better than perfect” and “Experiment boldly.”
The part worth stealing
A company does not need Haus to borrow the habit. Write down the decision before seeing the result. Choose the business KPI that matters, not the metric that flatters the channel. Preserve a control. Set the acceptable uncertainty in advance. If the first read says zero, inspect the tactic and retest rather than negotiating with the number. When an attribution dashboard and an experiment disagree, give more weight to the method that changed one thing and observed the consequence.
ColdStart, launched in August 2026, extends that discipline to products without a past. A normal geo experiment may need roughly nine months of historical KPI data to learn how regions behave. A new film, retail partner, market or product has none. ColdStart borrows the regional pattern and noise estimate from comparable established KPIs, runs placebo tests, designs the holdout, then switches to the real KPI once the campaign begins. Haus first built the method for film studios, where every title arrives as a new metric with a very short box-office window.
When the red pen runs out of ink
- The campaign cannot be varied across geography and no credible untreated comparison exists.
- Expected lift is too small relative to ordinary sales noise or the KPI has too little volume.
- Customers, commuters or media spill heavily between treatment and control regions.
- The brand changes prices, promotions, distribution or creative mid-test and muddies the intervention.
- A national one-off event lacks both a holdout and a stable “business as usual” baseline.
Those constraints explain where Haus fits in the market. It is not a universal replacement for attribution, a traditional MMM, a brand study or judgment. It is an anchor: a causal reading used to calibrate the noisier instruments around it. The company competes with measurement vendors, internal science teams and the seductive convenience of platform reporting. Its real alternative, though, is continuing to make large budget decisions with numbers nobody in marketing or finance entirely believes.
The amusing fact about modern advertising is that the industry invented exquisite machinery for finding people and surprisingly shaky machinery for proving it changed their behavior. Haus is betting that marketers will tolerate the discomfort of a holdout in exchange for an answer they can defend. Sometimes that answer will be 9 percent lift. Sometimes it will be zero. The zero may be the more valuable product.