Custom Performance Benchmark Builder

Build performance benchmarks from your own account history instead of borrowing an industry average — so you stop investigating noise and start catching real change.

flowchart LR
  A["<b>1 · Your own history</b><br/>12+ periods of one metric<br/>per channel"]
  B["<b>2 · The builder</b><br/>clean → baseline → control<br/>limits → run rules"]
  C["<b>3 · A baseline register</b><br/>normal / watch / signal zones<br/>and when to recalibrate"]
  A --> B --> C
  style A fill:#f6f8fa,stroke:#57606a,color:#1f2328
  style B fill:#ddf4ff,stroke:#0969da,color:#1f2328
  style C fill:#dafbe1,stroke:#1a7f37,color:#1f2328

Loading

A month at 0.71% CTR is normal against your own baseline. Against a borrowed industry benchmark of 0.45% it reads as a 58% outperformance, and someone spends a week working out what they did right.

Built on the Historical Performance Benchmarking Framework.

---

Run it

1. Put SKILL.md in your AI tool's instructions field — a Claude Project, a Custom GPT, a Gemini Gem, or an API system prompt.
2. Put reference/volatility-multipliers.md into its knowledge base or file uploads.
3. Give it 12 or more consecutive periods of one metric, for one channel.

---

What it needs, what you get back

You give it	12+ consecutive periods of history for one metric, one channel, with volume per period
You get back	A baseline register — mean, control limits, three zones, run rules, and when to recalibrate

Missing a required input, it asks instead of guessing.

---

How it works

1. Picks a window of genuinely stable history, starting after any structural change — a baseline spanning a replatform is measuring two different businesses
2. Excludes one-time events, and tells you what each exclusion did to the mean rather than quietly applying it
3. Sets control limits using a multiplier matched to how volatile that metric actually is
4. Adds run rules, which catch the slow deterioration that control limits miss — and slow deterioration is what actually kills accounts
5. Runs a leakage test to confirm no ad's own performance is feeding the benchmark it gets scored against

---

Worth knowing

Three zones, and the important one is Normal. Most of what marketing teams investigate is noise, and the cost isn't the investigation — it's the changes made in response to it.

There's also a protocol for a brand with no history at all: use an industry anchor for four weeks, clearly labeled provisional, build your own baseline over the next eight, and only then turn control limits on. A provisional benchmark presented like an established one gets acted on as though it were.

---

Files

File	What it is
SKILL.md	The skill. This is the thing you paste in
reference/volatility-multipliers.md	How wide the limits should be per metric type, and why

---

Related

marketing-measurement-validator → run first · ad-performance-scorecard-builder → uses this

---

MIT © Raneq Barber