The skill
Install once in Claude. It designs the test before you spend and reads the lift result after, running the market matching, power analysis and iROAS math with code.
What's inside
Three files, one job: measure causal lift
Built for performance marketers, not data scientists.
Install guide
Step by step setup, turn on code execution, upload the skill, toggle it on. Includes a no-install option for anyone who cannot add skills.
Plan B prompt
The same logic as a plain prompt. Paste it into any Claude chat with your data and run the design or the readout without installing anything.
How it works
Three steps, from inputs to a budget verdict.
Step 1Bring your inputs
For a design, attach the channel, its planned spend, and pre-period outcome data, meaning geo-level conversions or revenue. For a readout, give the lift result and its interval, the incremental conversions or revenue, and the channel spend during the test.
Step 2Claude does the work
In design mode it picks a method, matches control markets, sizes the holdout and duration by power analysis, and checks feasibility. In readout mode it computes iROAS and iCAC and applies the reallocation rules.
Step 3Get a spec or a verdict
You receive a Test Design Spec or an iROAS and iCAC readout, a reallocation verdict naming which rule fires, and a short list of what is uncertain or missing.
There is no live connection to your ad accounts. You bring the data, the skill does the design and the math.
How to run an incrementality test for a paid channel
An incrementality test measures the conversions your advertising actually caused, not the ones it merely touched. It compares an exposed group against a holdout that was eligible but not served, or test regions against matched control regions, and reads the difference as causal lift. Platform ROAS does the opposite, it credits any conversion it can attach an ad to, which is why it runs high on the channels that capture demand you already had.
What an incrementality test actually measures
Causal lift is the gap between what happened with the ads on and what would have happened with them off. A clean test holds out part of the audience or part of the map, so the comparison is like for like. This is the part platform attribution cannot do for you, because the optimization engine deliberately serves ads to people already likely to convert, so exposure and conversion are correlated no matter what the ad did.
Which channels to test first
Test the highest-spend, most-suspect channel before anything else. Branded search and retargeting are the usual offenders because they harvest existing intent. Branded search cannibalizes organic clicks the brand would have won for free, and retargeting claims last-touch credit for purchases that were going to happen. When eBay turned off branded search in a geo experiment, the measured return was about minus 63 percent because almost all the traffic substituted to organic. The point is not that branded search never works, a replication at Edmunds found far less substitution, but that the size is firm-specific, so you measure your own rather than copying a number.
How the skill designs the test
The skill picks a method and explains why. A geo holdout with synthetic control is the default, randomizing exposure by region so the platform does not choose who sees the ad, fitting a counterfactual from matched control regions in the CausalImpact style, and capturing offline outcomes. User-level conversion lift with ghost ads is the cleanest split when only in-platform outcomes are needed. Switchback testing fits marketplaces where geo splits are not feasible. The skill then matches control markets on their pre-period trends, excludes adjacent regions to avoid spillover, and sizes the holdout and duration by power analysis so the minimum detectable effect is smaller than the lift you expect. It also checks spend against the current floor, about $5,000 on the tested channel since Bayesian methods cut the minimum from around $50,000 a month.
Reading the result into iROAS and iCAC
After the test, the skill computes incremental ROAS as incremental revenue over incremental spend, and incremental CAC as incremental spend over incremental customers, then applies four reallocation rules. If measured iROAS comes back 30 to 50 percent or more below the platform number, reallocate away from the channel. If the holdout shows near-zero lift, cut the spend and redeploy it. If branded-search substitution stays under about 50 percent in your own test, keep defensive brand bidding. If a marketing mix model disagrees with the experiment by more than its credible interval, trust the experiment. Platform ROAS gets relabeled as an in-platform optimization signal, and a blended revenue-over-spend check catches the double-counting that walled gardens create.
Why this matters now
Measurement got harder and the cost of guessing got higher. Apple's App Tracking Transparency cut user-level iOS signal to a minority, Google reversed its cookie deprecation but Safari and Firefox still block cookies by default, and ecommerce customer acquisition cost rose roughly 40 to 60 percent from 2023 to 2025. Attribution is less accurate exactly as each misallocated dollar gets more expensive. For the full evidence base, including the Uber, P&G and Chase budget cuts that changed nothing, read Your ROAS Measures What Ads Touched, Not What They Caused.