Two failure modes cost the same money: killing a channel that was working, and funding a channel that never will. Both come from acting on a number that is not yet distinguishable from noise. Put in what you actually spent and what actually came back — this returns the interval, not a feeling.
| Measure | Value |
|---|---|
| Conversion rate (point estimate) | — |
| 95% confidence interval (Wilson) | — |
| Cost per conversion | — |
| Cost per unit | — |
| Zero-conversion floor (rule of three) | — |
Kill when the upper end of the interval is below the rate you need to hit your acceptable cost per conversion. Not when the point estimate is low — with 200 sends, a rate of 0% and a rate of 1.5% are the same evidence, and the interval is what says so.
Kill a zero-conversion channel only after 3 × the units the target rate requires.
The rule of three: if nothing converted in n attempts, the 95% upper bound on the rate is about
3/n. At n=100 that bound is 3% — higher than most target rates, so the result proves nothing yet.
Scale only when the lower end of the interval already clears your target. Everything else is "keep running and do not write a conclusion", which is the least popular answer and the only honest one.
This is the sample-size and kill-rule arithmetic from the open marketing-mindset skill (landing page, npm) — the same numbers exposed as CLI tools and as an MCP server, so an agent can call them instead of guessing.
Related free tools: test-size calculator · agent-readiness checklist.
Want the channel built and instrumented instead of measured after the fact? Fixed packages are published: Sprint $900, Engine $1,900/month, Full build $2,900. What a marketing engineer does and what it costs · cost breakdown · Telegram.
Related free tool: llms.txt generator — the file that tells models what the site answers and which pages are canonical.
Also: Service / Offer schema generator — copy-ready Service + OfferCatalog JSON-LD with the price, url and description checks.
Also: UTM + GA4 event naming builder — the tagged url and the event name that ends a test, with the checks that decide whether the number is readable.
Also: AI crawler robots.txt generator — who may cite you and who may train on you, written as robots.txt groups you can paste.
Also: A/B verdict calculator — was the difference separated from zero, and was the volume ever above the floor the observed lift needs?
Also: honest test queue planner — turn volume into a queue: the floor per arm, the days each test needs and the day each verdict lands.
Also: what one honest test costs to finish — the contacts a verdict needs, the days it takes and how many funded tests a budget carries.
Also: which step of your funnel can fund a verdict — the floor, the events each arm sees in a month, and the months a readable unit takes at your volume.
Also: which of your channels can actually fund a verdict — floor per arm, months to verdict, the cost of one full two-arm test and the channel to run first at your volume.