Daily operator loop

A small remnant fleet does not give you enough signal to change two things in one day.

We run tests on the order of a few thousand impressions and a couple dozen clicks per day. If you change targeting and creative on the same morning, tomorrow's number cannot tell you which move mattered. The loop below is the public version of how we operate: health first, one change, revert written before apply, then a later judgment.

This is operator-judged procedure. It does not describe internal logins or exchange credentials.

1. Check status before you touch anything.

Are campaigns running, paused on a daily cap, or stopped for a reason you did not choose? A daily-capped campaign that spent its budget and is waiting for the next UTC day is a healthy enrolled state, not an outage. A young campaign with no delivery is too early to tell, not broken. Fix real failures before you "improve" a working flight.

2. Read the right meters.

Use the supply-quality readout checklist. Judge traffic on the first-party traffic grade: real visits, bot traps, and engagement on your own pages. Treat vendor IVT as a hint, and only once a vendor tag is attached. Do not judge on click-through rate. Delivery clicks use the clicks plural field. Do not invent a number you cannot measure.

3. Verify the product path you use, not only the ads.

If you launch from Cursor, MCP, or the dashboard, confirm that path still creates targeting and creative through the public API. A broken launch path plus a live campaign is two problems. Do not debug them as one.

4. Make exactly one change to how the campaign buys.

Write the experiment row first, including the revert command. If you cannot write the revert, you do not understand the change well enough to ship it. State the metric and the window that will decide it. A visit- or click-based metric needs many days at this volume. A delivery metric can move sooner.

Apply the change to the whole fleet only when it is clearly reversible. Otherwise split cells.

5. Record KEEP, REVERT, or EXTEND.

Reverting is a normal outcome. The state to avoid is "applied, never judged," which silently becomes permanent policy. Do not raise budgets to make a metric look better.

Daily operator loop

Date (UTC): _______________
Fleet status (running / daily-cap pause / stopped / too early): _______________

Meters
[ ] Window written
[ ] Clicks field is clicks plural
[ ] Concentration noted, or "not pulled"
[ ] First-party traffic grade noted, or "not enough visits yet"
[ ] IVT blank unless vendor tag is on and fields are non-zero

Change (one only)
[ ] Revert command written BEFORE apply
[ ] Metric and decision window written
[ ] Applied to: whole fleet / split cells
[ ] Expected proof (one sentence)

Later judgment
[ ] KEEP / REVERT / EXTEND
[ ] What we will not claim

If something urgent is broken, fix it and skip the experiment that day. Two moves on a thin remnant book are how you learn nothing and keep the second move forever.