The Attribution Models Are Getting Smarter. Most Brands Are Still Too Broke to Use Them.

Every few months, a new paper lands claiming to have solved marketing attribution. This time it's causal structure learning bolted onto neural marketing mix models. The methodology is real progress. It also isn't the problem most brands actually have.

That's not a minor bug. It's the whole argument for causal inference in one sentence. Correlation-optimized systems, including most of what Meta and Google hand you inside their own ad dashboards, get rewarded for prediction accuracy, not truth about cause and effect. A newer framework called DeepCausalMMM forces a directed acyclic graph onto the model, so it has to explain how channels interact before it earns credit for how much revenue they drove. A companion line of Bayesian causal MMM research goes further, attaching confidence intervals to every attribution weight instead of a single deceptively clean number.

Here's the part the papers don't say out loud: none of this works if your data is a mess. Causal structure learning needs six-plus months of clean, granular spend and outcome data across every active channel. Most brands don't have that. They have a Meta pixel that stopped firing correctly eighteen months ago, three UTM naming conventions used interchangeably, and a Shopify export that doesn't match what the ad platform reports. You cannot model your way out of broken plumbing. Tagging hygiene and cross-device stitching will move your attribution accuracy further than any model architecture, and almost nobody wants to fund that work because it doesn't produce a paper or a keynote.

Top-quartile testing programs run 24 or more experiments a year. The median brand runs 14. Compound a modest 6% median lift across eight to ten winning tests on a checkout funnel, and you can produce a 47 to 61% total lift over a year. That number gets repeated at every growth conference as proof that testing volume is the answer.

It's survivorship bias with a statistic attached. It counts winners and says nothing about the losers, the null results, or the traffic you needed to reach power in the first place. A brand doing $3M in annual revenue with 40,000 monthly sessions cannot run a properly powered test on most of its funnel within a calendar year. Telling that brand to test more isn't strategy. It's a recipe for false positives dressed up as rigor, because underpowered tests called early produce exactly the noisy, unreliable wins that erode trust in testing altogether.

The more consequential blind spot sits one level deeper. New benchmark data pulled from SEC 10-K filings shows CAC has roughly tripled since 2015, and that paid CAC now runs 2.4x to 3.1x higher than blended CAC. Most internal dashboards report blended CAC because it's flattering and easy to calculate. But blended CAC hides the exact decision a CMO needs to make: which channels are losing money on every single new customer, right now. A board that only sees the blended number will approve budget increases into channels that are quietly bleeding cash, because the average looks fine even when half the portfolio isn't.

If your tracking is broken, fix the tracking before you fund a causal model. If your traffic can't reach statistical power, stop pretending your A/B test results mean anything and go do the qualitative work instead. If your board has only ever seen blended CAC, that conversation is overdue, and it's cheaper than any dashboard you'll ever build.

The growth leaders who win the next two years won't be the ones with the fanciest model. They'll be the ones honest enough to know which rung of the measurement ladder they're actually standing on, and disciplined enough to fix the rung below before reaching for the one above it.