Attribution Algorithm Evaluation

We tested Meta’s new incremental attribution model. The results broke a few assumptions. Out of 4 campaigns last week, 1 was optimized for incremental attribution. The other 3 were ASC and ABO campaigns: And yet: - 1-day click CPA was 48.2% lower than the ASC - Incremental CPA was 49.0% lower than the ASC - New audience ROAS was atleast 55% higher than the ASC Why does this matter? Because 1-day click is the most conservative attribution model, it only counts conversions that happen fast and post-click. It’s historically been the cleanest proxy for incrementality. So lower 1DC CPA is exactly what short consideration DTC products need - faster conversions, lower CAC, and better cash efficiency 🚨 How It Works: Meta now runs always-on holdout tests in the background. It splits users into treatment and control groups and continuously asks: “Did this person convert because they saw the ad, or would they have purchased anyway?” The difference between those groups is considered incremental lift and that becomes the basis for attribution, optimization, and reporting. It’s not about what happened within a time window anymore. It’s about what happened because of the ad. 🚨 Why This Matters - Eliminates the guesswork of choosing between attribution windows - Shifts focus from tweaking settings to scaling what actually works - Enables budget consolidation and simpler account structures - Reduces dependency on exclusions and segments, Meta already accounts for it - Moves us toward causal measurement inside the algorithm itself We’ll continue to validate this. But so far, incremental attribution has outperformed our default benchmarks. Including ASC.

We just wrapped a test where Meta's new Incremental Attribution (IA) optimization drove a +23.1% increase in iROAS for prospecting. Setup: We first ran a holdout test on Meta Prospecting using standard optimization. Then, we switched those same Meta Prospecting campaigns over to Incremental Attribution and ran a second holdout test. Results: iROAS +23.13 Spend +6.76 Contribution -4.82% Adjustment Factor* -8.26% What it means: The Adjustment Factor* decreased by -8.26% moving closer to 100% which would be a "perfect score" where every platform tracked conversion was incremental. This strongly suggests that Meta's IA optimization is more effective at identifying when it’s truly influencing a purchase. This is really good and very critical: Meta needs to avoid claiming credit for sales that would have occurred anyway. If it can accurately recognize organic conversions, it can avoid wasting impressions and optimize for genuinely incremental outcomes. It's also worth noting in both tests, Meta Prospecting platform attribution was under crediting (for every 100 platform purchases there were more than 100 incremental conversions). These tests were not run concurrently which introduces some noise due to seasonality or macro factors. Since the overall contribution declined slightly despite a small increase in spend, it's reasonable to believe the overall buying climate was slightly better during the second test period. *Think of the Adjustment Factor like a currency exchange rate between platform-reported conversions and incremental conversions. Example: If Meta reports 100 conversions, but only 50 are actually incremental, the adjustment factor would be 50%.

How 8-figure brands are increasing new customer ROAS on Meta (1 simple change) They switched to incremental attribution. It's Meta's newest attribution model. Their most advanced machine learning model for measuring ad performance. And almost nobody is using it. I've been running it across client accounts for close to 10 months now. Here's what I keep seeing: When I compare incremental attribution results against store revenue and first-click UTM data using SARIMA/ARIMA correlation analysis, it tracks tighter than 7DC, 1DC, or 7DC/1DV. Every time. Especially on new customer revenue. The fluctuations match. The direction holds. It just lines up with what's actually hitting the store. And it's not just better measurement. Incremental attribution tells Meta to find conversions that wouldn't have happened without the ad. That changes who Meta targets and how it spends your budget. The result: → Better new customer efficiency → More net new reach → Data you can actually trust against your backend It also exposes what isn't working. I've audited brands running tight bid caps and cost controls on 7DC/1DV. Campaigns looked great. Flipped to incremental and the numbers fell apart. Same with view-through or or overly restrictive cost controled accounts. Incremental strips that away and shows you what's actually going on. Who should test it: You need enough data for the ML model to work. At minimum: $1,000+/day in spend AND OR 50+ purchases/week. Higher AOV, longer sales cycle brands get the most out of it. But I've seen it perform well across brand types. It's just a better model. If you're below those thresholds stick with 7DC. And if you're running a brand new account use 1DV + engaged at the start to collect data. How to test incremental attribution: 1. Duplicate your top ads in a new campaign set to incremental attribution 2. Compare both sets apples-to-apples using the retroactive attribution breakdown If performance holds, start shifting budget from original campaign, ~20% every few days. Then repeat across campaigns. No rip and replace. Controlled test, gradual migration. If you have the means an incrementality study or correlation analysis looking at actual new revenue/order data is ideal. I've audited so many accounts recently that are perfect candidates for this. Premium brands, long sales cycles, well over $1k/day. Still on 7DC or 1DC... Huge missed opportunity. This is Meta's newest attribution model since 7DC1DV became the default in 2021. It's clearly where the platform is headed. Will it work for every brand? No. But is it worth every brand within the criteria testing it? Yes.

"Natural Geometry of Robust Data Attribution: From Convex Models to Deep Networks"by Shihao Li, Jiachen li, Dongmei Chen "Data attribution methods identify which training examples are responsible for a model's predictions, but their sensitivity to distributional perturbations undermines practical reliability. We present a unified framework for certified robust attribution that extends from convex models to deep networks. For convex settings, we derive Wasserstein-Robust Influence Functions (W-RIF) with provable coverage guarantees. For deep networks, we demonstrate that Euclidean certification is rendered vacuous by spectral amplification -- a mechanism where the inherent ill-conditioning of deep representations inflates Lipschitz bounds by over 10,000×. This explains why standard TRAK scores, while accurate point estimates, are geometrically fragile: naive Euclidean robustness analysis yields 0% certification. Our key contribution is the Natural Wasserstein metric, which measures perturbations in the geometry induced by the model's own feature covariance. This eliminates spectral amplification, reducing worst-case sensitivity by 76× and stabilizing attribution estimates. On CIFAR-10 with ResNet-18, Natural W-TRAK certifies 68.7% of ranking pairs compared to 0% for Euclidean baselines -- to our knowledge, the first non-vacuous certified bounds for neural network attribution. Furthermore, we prove that the Self-Influence term arising from our analysis equals the Lipschitz constant governing attribution stability, providing theoretical grounding for leverage-based anomaly detection. Empirically, Self-Influence achieves 0.970 AUROC for label noise detection, identifying 94.1% of corrupted labels by examining just the top 20% of training data." Paper: https://lnkd.in/dmNJhsWM #machinelearning