Your Meta agency should inspect your account every day. It should not change it every day.
You are paying for management, so a busy change log feels like proof you got it. I understand the appeal. I used to mistake a busy week for a productive one when I managed accounts by hand. But daily tweaks can restart learning, disrupt tests, and turn one clear result into three explanations nobody can defend.
A crowded change log is not evidence of control. Often, it is evidence of indecision.
Daily observation and daily intervention are different products. I want the first without paying for unnecessary doses of the second. That means intervention rules, patience with conversion lag, and experiments that survive long enough to answer a question.
Inspect daily. Save ordinary decisions for the evidence window.
What I want from daily management starts with a five-minute emergency scan:
- Check spend against expectations.
- Check that purchases are reporting normally.
- Check for rejected ads.
- Check for spend with zero results.
If all four look normal, close the tab. That routine comes from ROAZ’s weekly optimization playbook, which puts the main budget and creative decisions in a weekly review using seven-day data.
That distinction matters because real emergencies do not wait for Monday. Tracking failures, delivery anomalies, and guardrail breaches need attention the same day. A weak day of ROAS does not automatically qualify.
Free tool: the ROAS Calculator works out your return on ad spend, your break-even ROAS and the budget a revenue goal takes.
The same routine prescribes a Monday main review, a Wednesday pacing check, and a Friday read on early signals without changes. There is work happening between decisions. It just does not all produce an edit.
Restraint is part of the job. The scan proves someone is watching. Leaving a healthy campaign alone proves they understand what they are watching.
Define an emergency before someone calls Tuesday one
I count three kinds of emergency:
- Tracking breaks. The pixel stops firing, Conversions API stops sending, or purchase values go blank. An overnight gap between Ads Manager and Shopify deserves investigation, but it is not automatically broken tracking. Facebook attribution can take up to 72 hours: Events Manager shows receipt, while Ads Manager reports after attribution. That is why the guidance says to wait 72 hours before comparing Ads Manager with third-party data. A confirmed collection failure needs attention now; a reporting delay needs time.
- Delivery breaks. An ad gets rejected, a campaign stops spending despite available budget, or spend piles up with zero results. Normal pacing can look alarming here. Meta may spend up to 75% above a daily budget on a particular day, subject to its weekly limit of seven times that budget. So a $300/day budget can spend $525 in one busy day. That spike alone is not an emergency. An unexplained delivery stop needs investigation.
- Guardrails break. Spend crosses a limit agreed before launch: maximum weekly CPA, minimum blended ROAS before budget cuts, or a daily loss limit for a new test. When that threshold trips, intervention follows a decision already made—not a fresh panic.
Changing a headline because Tuesday looked soft belongs in none of those categories. Define the trigger before the bad day arrives. Otherwise, every uncomfortable number becomes permission to meddle.
A weak day of ROAS is not a verdict
Say you spend $20k/month and Tuesday prints a 1.1 ROAS against your 2.5 target. Your instinct says cut it. Mine says check the emergency rules, then let the evidence window finish.
Daily CPA can swing 40% from one weekday to the next without a structural problem. A single day tells me little about where an ad set will settle. That is why I grade ordinary optimization decisions on seven-day windows rather than morning-after panic.
Edits also carry a mechanical price. Meta ad sets usually exit learning after about 50 optimization events in the week following the last significant edit. Changes to targeting, creative, optimization events, or bid strategy can restart that process; budget changes matter when the move is large enough.
Make significant changes on five separate days and you can pay for repeated restarts. Where several necessary operational changes belong together, batching avoids dragging that disruption across the week. That does not make changing five variables at once a clean experiment.
The same guidance warns against judging an ad set while it is still learning, when performance is less stable. An edit has a cost even when it takes ten seconds to make. A change log that records only the action hides that cost.
A busy change log can destroy the explanation
Change the budget, swap the audience, and launch new creative on the same day. ROAS goes up. Which change did it?
Nobody knows.
I used to tell clients that kind of busy week showed hustle. I was wrong. Pulling every lever and claiming credit for whichever result looks good in hindsight is not a hypothesis. If I want to explain a test result, I need to isolate what changed.

Budget moves deserve the same discipline. The rule I use is to move in 10–20% steps every few days and grade against the account blend: scale above 1.2 times blended ROAS, hold between 0.8 and 1.2 times, and cut below 0.8 times for two weeks or more.
Daily reallocation based on yesterday’s ROAS breaks that logic. It replaces a decision window with whatever happened most recently.
The same playbook warns against raiding the testing budget to feed established campaigns. It suggests protecting roughly 15–25% of spend for testing below $300/day and notes that the highest-ROAS campaign can already be saturated. The best-looking number is not automatically the best place for another dollar.
Starve the test to feed the winner and you eat next month to pay for this week.
Conversion lag makes this worse. Someone clicks Tuesday, buys Friday, and Meta credits the click date once attribution settles. Cut the ad set on Wednesday morning and you have acted before that recovery appears. The later conversion can still arrive; your decision was simply premature.
I read blended ROAS alongside platform ROAS and give conversions time to arrive before touching budgets. Yesterday’s apparent loser is not a settled result.
Make every discretionary change earn approval
Every change I approve starts with one sentence:
We believe X will lower CPA because Y, and we will judge it after Z conversions or Z days.
That sentence forces an operator to name the mechanism, the intended outcome, and the evidence window. “Performance looked soft” does none of those things.
Effective creative testing starts with a single-variable hypothesis and reads ROAS or CPA, not just CTR. The practical guidance here is three to five creatives running for at least three to four days, ideally seven to smooth weekday effects.
If an agency cannot state the hypothesis and the decision window in plain language, it is not testing. It is fiddling.
Separate emergency controls from optimization
Emergency responses follow pre-agreed rules. A tracking failure, rejected ad, unexplained delivery stop, or breached loss limit triggers the appropriate action: investigate, pause where necessary, fix, and log it.
Everything else is discretionary. Discretionary changes wait for their evidence window.
That split lets daily management coexist with learning. The account can be watched constantly without being rearranged constantly. Continuous execution includes enforcing a decision not to intervene. Otherwise, “always on” becomes a polite way of saying “never left alone.”

Write the thresholds on a calm Monday
For a $20k/month account, the policy might look like this:
- Scale: only after seven days above 1.2 times blended ROAS.
- Hold: between 0.8 and 1.2 times blended ROAS.
- Cut: after two full weeks below 0.8 times blended ROAS.
- Adjust gradually: move budgets in 10–20% steps.
- Protect testing: reserve an agreed share of spend so established winners cannot raid it.
Those rules do not replace emergency controls. They stop ordinary optimization from becoming an emotional response to the latest number.
They also impose a cost: I have to tolerate an uncomfortable day without buying relief through an edit. I have to leave test spend alone when another campaign looks more attractive. And I have to tell a client why “no change” is the correct recommendation.
A threshold set in advance turns restraint from laziness into policy. Set it on Monday. Then Tuesday cannot bully you.
Protect creative tests from account-management theater
Busy accounts kill tests first. Someone launches three new angles on Monday, shifts budget on Tuesday, and declares a winner by Friday. The test never held still long enough to support the explanation attached to it.
The question determines the setup:
- If I want to understand why a creative wins, I need a controlled comparison that isolates the creative difference.
- If I want Meta to find the cheapest winner quickly, an auction or dynamic test can concentrate spend, but it will not reliably explain why that winner won.
Finding a winner and explaining a winner are different jobs. Most agencies pick the fast method and then pretend they got the careful answer.
Uneven delivery is not proof that the creative caused the result
A large-scale study of 3,204 Lift tests and 181,890 A/B tests found almost no imbalance in the Lift tests but clear divergent delivery in the A/B tests. In the latter, 22% of standardized mean differences exceeded 0.20.
In plain English: Meta showed variants to different mixes of people. That matters because a performance difference can reflect who saw the ad as well as what it said. The problem was worst for conversion objectives.
The study also examined tightly matched configurations: identical objectives, targeting, budgets, bids, and placements. Only three tests met the final matching criteria. That is not a basis for pretending a simple account checklist makes audience imbalance disappear.
My operational requirement is narrower: keep the test configuration stable, protect its agreed window, and do not turn a delivery preference into a causal explanation.
Nor do I treat roughly 50 optimization events as a magic test-validity threshold. That number belongs to the learning-phase discussion. It does not, by itself, prove one creative beats another.
Before hiring anyone, I would run the process through a creative-testing acceptance checklist. If the budget changed mid-flight, I want that disruption disclosed—not buried beneath a winner label. Protect the test, or stop claiming it answered the original question.
Hire for judgment, not edit counts
I want a report that explains changes and deliberate non-changes. The first list should show:
- What changed.
- Why it changed.
- Which hypothesis it tested or emergency rule it followed.
- When the result will be judged.
Then I want the second list: what looked bad, what stayed untouched, and why.
That second list separates management from motion. I would rather hire the team that writes “no change, only 31 conversions since the last edit, judging Monday” than the team that ships twelve tweaks to fill a slide. That is what auditing the change log instead of ROAS screenshots forces into the open.
I would hold groas to that standard, too: continuous execution must mean enforcing guardrails and protecting evidence windows, not manufacturing activity. Automation does not excuse a bad decision rule. It just makes that rule run faster.
The accountable strategist should be able to explain both the intervention and the decision to wait. The action log should preserve that reasoning. The test should remain protected until its agreed read. Hire the team that can defend doing nothing on Tuesday.
Hire for edit counts and you will get edits. Your learning will keep restarting, your tests will stop reading cleanly, and every monthly report will celebrate activity while ROAS drifts sideways. Then the agency will sell you more activity to fix it. Demand restraint with receipts instead, or keep paying 2015 prices for someone to restart your learning five times a week.

