Cassandra (2025) — Google Ads incremental ROI benchmarks

Summary:

This report from Cassandra, a company that sells marketing mix modeling (MMM) software, looks at how much different Google Ads campaign types actually return when you measure incremental lift instead of Google’s own attribution. It compares Search Brand, Search Non-Brand, Performance Max, Shopping, Display, Demand Gen and YouTube, and breaks some results out by industry.

Cassandra analyzed 253 MMM models from 59 advertisers, covering about $383M in media spend from 2023 to 2025. Each record is one channel from one model run, sorted into seven campaign types and industry groups; the sporting sector was excluded. The report gives medians, P10–P75 ranges and 95% bootstrap confidence intervals, and says any cut with fewer than 5 records is directional only. These are modelled results, not experiments.

Key findings:
• Search Non-Brand has the highest median incremental ROI at 5.21x, followed by Performance Max at 4.64x and Search Brand at 4.14x.
• Incremental ROI is 2 to 5 times lower than platform-reported ROAS, depending on the channel, because MMM tries to remove revenue that would have happened anyway.
• Performance Max is the most predictable format – its confidence interval [3.11–6.41] is tighter than Search Non-Brand’s or Display’s, but advertisers usually can’t see where its budget goes.
• YouTube returns a median 2.70x on just $3.6M of spend, and Cassandra calls it an underpriced upper-funnel channel. Demand Gen is lowest at 2.22x, and the report itself says its numbers need closer checking.
• Industry changes the picture more than format – Search Brand ranges from 3.19x in Retail & Ecommerce to 9.32x in Healthcare and over 10x in Media & Entertainment; Fashion & Luxury reaches 8.07x on Search Non-Brand.

Margin notes:
This is a vendor report from a company that sells MMM, and the article doubles as an argument for its own product, so the founder’s interpretation (especially on YouTube) is mixed with the data. The method is described only at a high level, with no detail on model quality, how advertisers were picked, how revenue was defined, or where they are based. Several channel and industry cuts rest on small samples – YouTube has 14 models, and vertical medians come from few clients – and some numbers don’t match between the report’s tables, charts and text.