Treating exploratory spend as a fixed R&D line item is a step in the right direction, but that just moves the argument from "is this click wasted?" to "why is this R&D budget 5% and not 2%?" The percentage-of-main-spend model is a trap, because it ties a fixed discovery cost to a variable performance budget. If main campaign spend doubles next quarter, should your discovery budget automatically double too? Probably not.
You pitch it as insurance. A flat, non-negotiable monthly amount, like your observability tool bill. You decide the size by asking what it would cost you to miss a new, high-intent query cluster for a quarter. Run a small, ugly broad match campaign for a month, tally the total waste, and that's your baseline. That's the premium you pay to avoid flying blind.
The real trick is proving the negative - what you *didn't* miss. Good luck getting that into a spreadsheet.
Your k8s cluster is 40% idle.
The sandbox analogy is spot on, especially the part about mirrored data. For these exploratory campaigns, we use a mirrored conversion tracking setup but with a dedicated, lower-fidelity attribution window. That way, the noise doesn't contaminate our core attribution model.
On the trigger question, we started with a consistent weekly interval but found it too rigid. We now use a signal-based approach, with the primary trigger being a quantile drop in new, converting query variants from our exact/phrase campaigns over a rolling 14-day period. This acts as our "synthetic probe." We've defined thresholds that essentially flag when the high-fidelity feed's discovery rate falls outside its expected statistical boundaries, which prompts a manual review and often triggers a probe campaign.
However, this requires establishing a baseline for what "normal" discovery velocity looks like, which can take a few months of clean data after the broad-to-exact switch. Without that, you're just reacting to noise.
—chris