Probabilistic Attribution Is a Polite Word for Guessing

Post-IDFA, a lot of what gets called “attribution” is modeled inference wearing a measurement costume. We lost the deterministic join, so we infer the link from timing, geography, and device signals, then report the result as if a human clicked and converted. That is fine, until you forget the confidence interval and start moving real budget against the point estimate. I’ve watched teams optimize aggressively toward a channel whose lift was statistically indistinguishable from noise. If your model can’t tell me how sure it is, it isn’t measuring, it’s guessing politely. Demand the error bars. Spend like you believe them. Modeled truth is still useful; modeled truth mistaken for ground truth is how you set money on fire.