Lab

A public log of marketing AI experiments.

1 experiment live · more in flight · Last updated May 2026

Hypothesis. Method. Result. What changed in the thinking. The frameworks on the rest of the site are the conclusions; this is where the work happens before it becomes a conclusion. Sometimes it works. Sometimes it does not. Either way, it gets published. The current experiment tests an idea developed in Performance Max as a signal feedback loop.

Rule 01
Hypothesis before run.
Every experiment starts with a stated hypothesis. No retro-fitting the conclusion to the data.
Rule 02
Failures count.
An experiment that disproves the hypothesis is a successful experiment. It still gets a write-up.
Rule 03
What changed.
Every entry ends with one line: what changed in the thinking. If nothing changed, that gets said too.
Experiments
001
First entry
In progress

Can a single Performance Max signal change the audience mix in 14 days?

Hypothesis: feeding a value-based conversion signal weighted by predicted retention will shift PMax audience mix toward higher-AOV cohorts within two weeks — without changing creative, budget, or targeting. The constraint: we have to see the shift in audience reporting, not just in conversion value reporting.

Performance Max Value-based bidding 14-day window May 2026
002
In flight
Week 4 of 8

Can an agentic loop compress a Performance Max signal-tuning cycle from 6 weeks to 48 hours?

Hypothesis: an agent reading weekly PMax diagnostics, proposing 3–5 signal changes, and staging them for human approval can collapse the cycle by an order of magnitude. Mid-run finding: ~5× faster on the first three iterations, and the bottleneck moved — the new constraint is the org's approval cadence, not the model.

Performance Max Agentic AI 8-week window May 2026
003
Queued
Partner wanted

Re-running last quarter's MMM with vision-model creative features as covariates.

Hypothesis: vision-extracted creative features (hook archetype, pacing, claim density, brand-asset prominence) explain a measurable share of MMM residual variance — enough to redistribute channel coefficients and give the CMO a defensible creative-value number. Looking for a partner with 12+ months of clean media and creative metadata.

MMM Vision model Creative measurement Q3 2026 target
More experiments coming. Subscribe to Stay Sharp to get them as they're published. Or browse the Files library — templates, audit checklists, and operating tools from the experiments that have already shipped.
Stay Sharp

Get the experiments as they happen.

New Lab entries delivered before they're indexed. Plus the frameworks, teardowns, and counter-intuitive POVs.