Lowe's Companies (2026) — Integrated Marketing Attribution (IMA) for campaign-level measurement

Summary:

This paper is about a new way to measure marketing at the campaign level without tracking individual people. The authors call it Integrated Marketing Attribution, or IMA.

The idea is to keep the privacy-safe, big-picture view of marketing mix modeling and use it to guide more detailed campaign-level estimates.

The paper was written by researchers at Lowe’s Companies, Inc., the retailer. Their main goal is to show that they can break channel-level marketing results down into campaign-level results using only aggregated data, while staying consistent with the original marketing mix model.

They say they tested the approach on three years of real-world historical marketing data. The main model worked in three steps.

First, they ran a weekly marketing mix model on aggregated sales and media data.

Second, they split those weekly channel results into daily values using adstock rules about delayed media effects.

Third, they fit separate Bayesian regression models for each channel, using the channel-level marketing mix model results as priors to guide campaign-level estimates.

They evaluated the method by rolling the daily campaign estimates back up to weekly channel totals and comparing those with the original marketing mix model outputs.

The main findings are that the method stayed very close to the weekly channel-level benchmark while giving more detailed campaign-level results.

In a representative channel, the weekly fit to the benchmark was very high, with R² of 0.98. Reported weekly MAPE was 4.1% for Search, 5.97% for Social, and 3.25% for Display.

The authors say this is only a small loss in accuracy compared with the original marketing mix model, while giving much more campaign detail.

Note that the campaign estimates are model-based and are anchored to another model, the MMM, rather than validated against outside ground truth. The paper also excludes channel names, dates, geography, sample counts, and revenue levels for privacy, which makes independent checking hard. Evaluation is mostly against the model it starts from, so strong agreement mainly shows the second-stage model can reconstruct the first-stage outputs. Because the data come from one company’s internal history, it is also unclear how well the method would generalize elsewhere.