We are obsessed with what we can track but when your customers lose trust in you they stop complaining and giving feedback. They leave or wait until they can. They stop giving signals and instead we see transactional minimalism.

Breaking News: A judge ruled that Google must make changes to address its advertising technology monopoly but would not need to break up that business. nyti.ms/4xv5OoH

"All models are wrong, but some are useful." My friend @josephfwyer wrote the best summation I've seen of how we view the world at Haus. I'm saving you a click and posting his full substack article below. Please tell us where we are right, but more importantly, tell us where we are wrong! The Hitchhiker’s Guide to Marketing Measurement and Decisions A map of the Haus Analytics Platform The first two posts in Decisions and Data have been about what not to do: don’t trust statistical significance to make budget calls for you. Fair enough, but now we’re going to go from “stop doing that” to “do this instead.” So this post is the map. It’s the whole path from the data everyone starts with in marketing measurement to a system that recommends decisions and checks its own work. I’m really proud to say this is what Haus Analytics has built. We’ve walked down this intellectual map the last few years and now I want to walk you down it. Each stop on the path deserves its own deep dive, and those are coming. Today is just the strategic view of the problems and solutions. Last-click attribution has big shortcomings Speed-running last-click attribution real quickly: You click on a digital ad on a advertising platform. Later you buy something on the advertiser’s site, where a snippet of code reports the conversion back to the platform. Roll that up and you get the reporting most marketers watch daily. Suppose for a campaign there was a thousand clicks and a hundred tracked purchases on just a hundred dollars of spend. You’ve acquired those conversions for one dollar each. Streams of this kind of attribution data are flowing constantly and reporting ultra granular data. The problem is when you ask if this advertising spend “caused” those conversions. Around 2011, the economist Steve Tadelis was consulting at eBay and looked hard at their paid search spend. The ad in question sat at the top of search results page when you searched “eBay,” directly above the free organic link to eBay. His hypothesis was that people typing “eBay” into a search engine were largely already on their way to buying something at eBay. Radical, I know. He got his chance to look deeper when Ebay shut down brand keyboard bidding while negotiating with a major search engine. After 3 months of data had populated, he found that organic clicks rose to absorb almost all of the traffic the ads had been getting credit for. eBay was paying approximately $20M/yr to the search engine for customers it was already getting. He eventually published the results in a peer-reviewed journal. The word for what attribution measures is correlation. The word for how much customer activity advertising causes is “incrementality”: purchases that happen because of the ad and would not have happened otherwise. Modern targeting makes the gap between attribution and incrementality on digital channels worse because platforms can model each user’s propensity to buy your product and show ads to the people most likely to purchase anyway. Hold on to this: attribution is biased, but it is also fast and absurdly detailed. I’ll come back to it later. Experiments are where incrementality comes from If you want to know what customer activity an ad caused, you need a comparison: people who saw it versus equivalent people who didn’t. That’s called an experiment. Keeping track of individuals in experiments can be messy (although some ad platforms can pull it off) so a workhorse in this industry is the geographic experiment. You can randomly assign market areas in a country into receiving advertising (treatment) or not (control). Then you simply look at the difference in conversions between the two. Randomization removes that attribution correlational bias that tricked Ebay into lighting money on fire (and many, many other companies even today). On average, the differences between the treatment and control regions wash out, so the estimate is centered on the true incrementality. That sounds really simple, so why does Haus have all these experimentation scientists? It’s a lot of work to improve precision without messing up accuracy. Precision is the measure of how far off the estimate can be when it is off. Think of it like darts. A tight cluster in the upper right of the dart board is precise but inaccurate. A loose scatter centered on the bullseye is accurate but imprecise. Randomization gets your cluster centered on the bullseye. What the science team does all day is making the cluster tighter without dragging it off the bullseye. If you have a great experiment design and analytical model then you will get a precise estimate on the real number. At eBay, the real number for brand search keywords was close to zero. But that’s one company at one moment. I’ve seen brand search come back near zero for one advertiser and strongly incremental for another. You don’t know until you test. So everyone needs to test to find how much money they are leaking. As much as I love experiments, they have limitations. An experiment tells you the causal effect at the spend level you tested. That’s only one data point! You can’t trace a curve through one data point and you need a curve to allocate budgets because every channel eventually hits diminishing marginal returns. Causal MMM, debiasing the model Marketers have been fitting models to trace diminishing return curves for a long time. Media mix models (MMMs) trace your sales on your spend across multiple ad channels and let statistics tell you how much each is driving. In theory this gives you the full diminishing curve for every channel at once. In practice, MMM has two mortal flaws. Multicollinearity. Businesses tend to turn all their channels up and down together, so when sales change, the model struggles to tell which channel did it. Seasonality. Businesses spend the most going into Black Friday and Christmas season, which is exactly when people’s propensity to buy surges without needing ads. So were the gains in sales caused by the ads or the season? All models are wrong, but some are useful. MMM is very wrong, but very useful because it can go all the way to a budget recommendation when an experiment can’t. So we want to fix MMMs. We do that by stopping treating experiments and MMM as rival methodologies and instead merge them together. You ran a geo experiment and learned that at last quarter’s spend, a particular ad channel truly drove some specific number of purchases. We require the model’s curve for that channel to pass through the experiment data point. We call this experiment calibration. Pinning one channel’s curve to ground truth also disciplines the others, because the remaining sales have to be explained by the remaining channels plus organic demand. The seasonal bias gets squeezed out the same way. That’s causal MMM (cMMM): the curve-tracing power of an MMM, anchored to the causal truth of the experiments. And Haus refreshes it weekly instead of the traditional once-a-quarter-two-quarters-later read. Causal attribution, debiasing the daily feed A weekly model of whole channels is still too slow and too coarse for the person making changes every day. Remember what attribution had going for it: daily, granular, ad-level. Attribution is wrong, but what if how wrong is predictable? So apply a similar calibration move we did with cMMM but to Attribution. If your experiments show that only five percent of the purchases the channel claims are truly incremental, then discount its daily feed by 95%. Do that per channel. You can’t have experiments everywhere so let a model reason about how the causal correction shifts over time and across spend levels. Now the daily numbers marketers already watch become numbers they can trust. We also show the raw platform-reported figures next to the corrected ones, so you can always see what the feed said and what we did to it. Architect, where measurement becomes decisions Everything up to this point is measurement, and measurement adds no value unless it impacts decisions. We’re now going to get into the most exciting part of the Haus stack. We call it Architect. It takes the experiments, the causal MMM, and the causal attribution feed, and turns them into specific recommendations: move this much budget from here to there. Then, it measures what happened after you made the move and reports back within a couple of weeks. Did revenue improve? By how much? Continually updating its expectations and recommending again. Across the first companies acting on these recommendations so far, the average adopted Architect recommendation has improved conversions by 10%. Some companies have stacked changes and watched the gains compound. Architect earned the trust to make those changes by tracing the logic through the experiment, cMMM, and cAttribution data. That is why we’ve been working so hard to mak every layer underneath an automated and scalable system. An experiment alone is an isolated data point. A calibrated model alone is a forecast. The loop of recommend, act, verify, and update is the thing that turns marketing measurement from a reporting function into the situation room where strategic decisions are made. One more note, since the obvious question in 2026 is “why not just have an AI do all this?” Some companies will sell you exactly that right now. Personally, an AI reading raw attribution data inherits every bias in it and without each layer built, tested, debiased, and made explainable, you can’t trust if it’s going to work or know why it told you to move the money. The entire stack is what makes an automated recommendation trustworthy. Wrapping it up That’s the map! Attribution gives you speed and detail but with debilitating bias. Experiments remove the bias but are solitary data points. cMMM extends the causal truth to actionable curves. cAttribution pushes causal truth back into the daily feed. Architect reads the outputs of the whole stack and closes the loop by recommending and verifying decisions. In the coming weeks I’ll talk through some stops on the map with some illustrative math and examples of where things can go wrong. If you only take one thing from the altitude view, take this: never trust a marketing number that hasn’t been anchored to an experiment somewhere.

Zuma, The Trade Desk’s Kokai Upgrade, Brings Intuitive UI, Agentic AI and Simplified Measurement TTD just dropped Zuma, the latest update to its Kokai media buying platform. The update pushes hard into AI, while aiming to greatly improve usability and simplify measurement. The Agents are Coming TTD is going full-bore on agentic AI. Koa Assistant becomes the conversational front door to specialized TTD agents handling campaign creation, audience building, troubleshooting, performance analysis and frequency optimization. The agents make for more efficient campaigns: TTD said Zuma has already produced an average 32% improvement in CPA performance. What’s more, its natural-language interface also eliminates repetitive grunt work that consumes traders’ days. Strategically, the most exciting part is that TTD is opening Kokai to agentic advertising beyond its own platform. TTD’s Open Agentic Kit lets buyers connect outside agents to its own - potentially making Kokai a platform operated by agents rather than software operated by humans. For example, buyers could access TTD’s agents through Claude. Measurement gets an Easy Button The second big improvement is less sexy, but just as important commercially. Zuma makes lift measurement much easier. A study that previously involved vendor coordination and roughly a week of setup can now be launched from Kokai with a click. That helps move programmatic measurement away from CTRs, completion rates and other proxies toward the question CEOs actually care about: Did my advertising generate incremental business? Goodbye, Periodic Table When Kokai launched in 2023, its most obvious novelty was an interface inspired by the periodic table. It was meant to simplify the UX, surface key campaign information - and perhaps lock in users as well. The only trouble: traders hated it, judging by the grumbling on social media. The periodic table is losing its throne as Kokai’s default interface. In comes a more conventional dashboard, prominent performance metrics, consolidated settings, bulk editing and previews showing what proposed changes could do before traders unleash them across hundreds of ad groups. Great technology that frustrates users eventually becomes somebody else’s revenue. TTD listened to its clients: Zuma feels less like management telling traders how they ought to work and more like a product designed around how traders actually like to work. TTD’s jump into agentic advertising is good to see. Competitively, agentic capabilities are rapidly becoming table stakes - Amazon DSP, Yahoo DSP, Google DV360 and others are moving in the same direction. TTD’s opportunity is to combine agents with its independent positioning, open-internet reach and increasingly integrated measurement stack. Zuma feels like The Trade Desk is becoming a little less stubborn about its product - and more focused on making buyers faster, happier and harder to steal. The Trade Desk is a client of W Media Research, LLC.

This is the best articulation of why physical media works that I've seen in a while. Infinite scroll expands the consideration set; the real world collapses it. Out of home isn't competing for attention in the loop, it's the exit from it.