Diagnosing CPA / ROAS problems
"My results got worse - what do I do?" is the question you'll face most. Today: a systematic diagnostic tree so you respond with a precise cause instead of random changes. Where calm, first-principles thinking becomes a visible professional edge.
1First, is it even a real problem?
Before touching anything, rule out false alarms - the most common "problem" is no problem at all:
- Learning-phase volatility (Day 5) - new/edited ad set swinging? Wait.
- Too-short a window - judging on one day? Daily noise is huge. Look at 3-7 days.
- Attribution timing (Day 18) - conversions report on a delay; today looks bad because credit hasn't landed.
- Normal variance - outside the account's usual range, or just a slightly-down day?
Half of "emergencies" dissolve here. Reacting to noise is itself a cause of bad performance (you reset learning).
2The diagnostic tree
A CPA rise traces to one of three. Localize before you act.
| Symptom | Likely cause → fix |
|---|---|
| 🔺 CPM up | More competition/seasonality (Day 2), audience too narrow/saturated, quality ranking dropped. Often external → better creative or broader reach, not panic. |
| 🔻 CTR down | Creative fatigue (Day 14) or weak creative → fresh creative/iterations. Most common controllable cause. |
| 🔻 Conversion rate down | Landing page, offer, price, checkout, or a tracking/signal break (Day 10). Often NOT the ad. |
| 📉 Volume collapsed | Budget too low, cost cap too tight (Day 7), audience exhausted, or learning limited (Day 5). |
3The checklist, in order
A bad doctor prescribes everything at once. A good one localizes: vitals, history, "did anything change?", rule out false alarms, isolate the system, THEN treat one thing. Changing five settings at once because CPA rose is malpractice - you'll never know what worked, and you've likely reset learning. Diagnose one cause, change one thing, observe. Precision over panic is the entire professional difference.
(1) Reacting to noise/learning as a real trend. (2) Changing many things at once - resets learning, unscientific. (3) Never checking signal, so they "fix" a creative problem that was a broken Pixel. (4) Blaming the ad for a landing-page issue. (5) No change-history check. For clients, your value is the calm diagnosis: "Here's exactly what broke and the one thing we're changing." That builds the trust your startup runs on.
Recap - 30 seconds
- First rule out false alarms: learning volatility, short windows, attribution delay, normal variance.
- CPA = CPM ÷ (CTR × conversion rate) - localize the rise before acting.
- Check signal (Events Manager) early - a broken Pixel mimics a performance crash.
- CTR down = creative fatigue; conversion rate down = landing page/offer; CPM up = auction/saturation.
- Change ONE variable at a time, check change history, judge over days not hours.