"checked ads manager before brushing my teeth. and at the school play. and at 2am. dread, not learning."
That is how a Shopify founder described the job to us. The check is compulsive because the stakes feel binary: either something is wrong and every hour you wait costs money, or nothing is wrong and every edit you make costs learning. Both fears are correct. The problem is that panic-checking cannot tell you which one you are in. Signal or noise is the judgment question underneath every other question you ask your ad account, and almost nobody writes it down.
This is for founder-operators and growth leads at $500K–$30M running Meta and Google who either over-correct (and keep resetting learning) or freeze (and let a real problem run for a fortnight). It gives you the three questions to ask before any edit, the six intervention states an experienced media buyer actually works in, and what a recommendation has to contain before it earns an "act now".
Key Takeaways
- → Before touching anything, answer three questions: is this a real signal, normal variance, or the effect of a change you made yourself?
- → Over-intervention is the expensive failure. Every material edit on Meta can restart learning, and stacked edits mean you can't read any of them.
- → There are six intervention states, not two. "Hold and observe" and "do not act" are decisions with a reason, a window and a trigger, not neglect.
- → An "act now" has to be earned: alternative explanations ruled out, guardrails set, a rollback rule written before the change goes in.
- → One read a day, on yesterday's complete numbers, beats twelve reads on this morning's partial ones.
Signal or Noise? The Three Questions Before Any Edit
Ads Manager shows you a number and a colour. It never shows you the reason. So the operator supplies one, usually the scariest available, and acts on it by lunch.
Stop at the number and ask three things, in order.
- → Is this a real signal? A move that survives a full conversion cycle, on enough conversions to mean something, with the business underneath it unchanged.
- → Is this normal variance? A daily swing inside the range this account produces every week when nothing has happened. Most "drops" are this.
- → Is this the effect of my own recent change? A budget step, a new creative, an audience edit, a landing-page change, a price move. Recent edits make short-window comparisons untrustworthy, and the platform's own learning period is the most common culprit of all.
The stakes at the top of the market are instructive. In April 2023 Meta's delivery system malfunctioned and accounts spent roughly 90% of their daily budgets by 9am, with CPAs roughly tripling, according to CNBC's reporting at the time. A CMO putting $75–100 million a year into Facebook described trying to raise a human at Meta afterwards with one word, repeated: "Buehler? Buehler? Buehler?" (AdExchanger). That is the service ceiling at $100 million of spend. At $2 million, the intervene-or-wait call is yours, every morning, and nobody is picking up.
Here is the practical test for each question.
| Check | Reads as noise when… | Reads as signal when… |
|---|---|---|
| Window | The move is one or two days old. Yesterday's ROAS is always lower than it will be once conversions finish reporting | It persists across a full conversion cycle (seven days for most Shopify brands) against the same days of the previous cycle |
| Volume | The "drop" is built on a handful of conversions. On 20 orders, one big basket moves ROAS 15% on its own | The window holds enough orders that the swing sits outside the account's usual week-to-week range |
| Your own edits | You changed budget, creative, audience, bid or the landing page in the last three to seven days | The account has been untouched long enough that learning has settled and the comparison is clean |
| Calendar | It is Tuesday after a strong Sunday, the day after payday, or your own email send moved the mix | Same weekday, same position in the month, and the move is still there |
| Business reality | Something outside the ad account moved: a hero SKU sold out, a discount code broke in the cart, a price changed, a promotion ended | Stock, site conversion rate, pricing and promotions are all steady, so the ads are the remaining suspect |
| Measurement | Shopify orders are flat while platform-reported conversions fell: the pixel, CAPI or an attribution setting changed | Shopify orders and platform conversions fell together |
Notice that four of the six checks live outside the ad account. A Shopify merchant once watched conversion rate fall from about 1.8% to 0.3% and sit there for weeks; the cause was a discount code that had stopped applying in the cart drawer, and the ads spent against the broken store the whole time (Shopify Community, a verified post in our research). No edit in Ads Manager would have fixed that, and every edit would have destroyed the evidence.
Why Over-Intervention Is the Expensive Failure Mode
Freezing costs you money at the rate of one leak. Over-editing costs you the ability to see anything at all.
Three reasons, and the first is mechanical. Meta treats most material edits (budget steps beyond a modest band, new creative, audience changes, bid changes) as a reason to re-enter learning, and learning needs a run of conversions before delivery stabilises. Edit on Monday, panic on Wednesday, edit again on Thursday, and the campaign has spent the week learning instead of selling. Google's Smart Bidding has the same rhythm with a longer fuse.
The second is evidential. If you change two things at once, you cannot attribute the result to either. Operators who "just try a few things" on a bad Tuesday end up on Friday with a different account and no idea which change helped. Do not make multiple causal edits together unless you are containing an emergency, and even then, write down what you changed.
The third is reversibility. A budget cut is easy to undo. Deleting an ad set is not; the history goes with it. When your confidence is incomplete, prefer the action you can reverse, and prefer no action to an irreversible one.
So the honest tally looks like this: a real problem left for three days costs three days of one leak. A reflex edit on noise costs a learning reset, a muddied comparison and, quite often, the winner you just killed for having a bad Tuesday. Plenty of accounts are not underperforming. They are over-managed.
The Six Intervention States
Experienced buyers do not work in "act" or "wait". They work in six states, and the state is chosen from the evidence, not from the mood.
| State | When it applies | What you do | What you don't do |
|---|---|---|---|
| Act now | Evidence and urgency are both high: a tracking break, spend to a sold-out product, a broken checkout, a guardrail breached | The one change that addresses the cause, with a rollback rule written first | Fix a measurement failure with a media-buying edit |
| Act cautiously | The evidence points one way but volume is thin or the window is short | A small, reversible step, then observe a full cycle | Make the full-size move you would make on strong evidence |
| Run a controlled experiment | The uncertainty itself is worth money to resolve (does brand exclusion help, does the new offer convert) | Isolate the variable in its own campaign or period with a fixed budget and a fixed window | Force equal spend just to make the test look tidy; that has an opportunity cost |
| Investigate first | The move is real but the cause is not established | Work the diagnostic order: measurement, business reality, account state, then the funnel | Touch the account before you know which layer moved |
| Hold and observe | A recent change, reporting lag or plain variance makes action premature | Name the window and the trigger that would change the call, then wait it out | Re-check hourly. The number will not have finished reporting |
| Do not act | Performance is acceptable, or the proposed change adds risk without expected value | Record why waiting is the better decision, and close the tab | Make a change to feel like you did something |
Most operators only ever use the first row. The account they are managing needs the last three far more often than the first three.
"No Action" Is a First-Class Decision
There is a difference between a deliberate hold and neglect, and it is written down.
A deliberate hold has a reason ("budget stepped up Monday, learning will settle by Thursday"), an observation window ("re-read Friday on the full seven days"), and a trigger ("if CPA is still 30% above the prior cycle on Friday with Shopify orders confirming it, cut back to Monday's budget"). Neglect has none of those. It is the same non-action, but one of them is management and the other is avoidance. If you cannot say what would make you act, you have not decided to wait; you have decided not to look.
This is also the answer to the 2am check. The check is trying to resolve uncertainty that a partial day's data cannot resolve. Write the window and the trigger down at 9am, and the 2am version of you has nothing to do.
How to Earn an "Act Now"
An "act now" is expensive if it is wrong, so it has to be earned. Before the change goes in, a real recommendation contains the following, whether it comes from a media buyer, from you, or from an AI agent.
- → The observation: what moved, at what level (account, campaign, ad, product), over which windows compared with which.
- → Materiality: how much money this is actually about. A 40% CPA rise on a $30-a-day ad set is not an emergency.
- → The diagnosis, and the alternative explanations: the most likely cause and the two or three others that could produce the same numbers. If you cannot name an alternative, you have not diagnosed anything.
- → Evidence and confidence: which numbers and which business facts support the diagnosis, and an honest low, medium or high on how sure you are.
- → Platform state and business context: is the campaign learning, recently edited, in a promotion? What do margin, stock and new-customer value say about the cost of being wrong?
- → The action, with magnitude and timing: not "reduce budget" but "reduce this campaign from $800 to $600 tomorrow morning, before the daily reset".
- → Guardrails: the limits the change must stay inside, and the entities it must not touch.
- → The observation window and the rollback rule: when you will judge it, and what evidence would reverse it.
Do this once by hand for a real change and you will notice something. Half the "act now" impulses fail at alternative explanations, and most of the rest fail at materiality. That is the discipline working.
Where Each State Sends You
The states are a router. Each one has a next page.
- → Investigate first sends you to the diagnostic order: measurement, variance, your own changes, creative, auction, conversion path, product mix, learning state. That order is written out on Why did my ROAS drop?, and the fifteen-minute daily version is the daily root-cause framework.
- → Act now or act cautiously on a scaling move sends you to the scaling method: read what the last budget step actually earned before deciding the next one, and size cuts and raises with the cut-or-feed method.
- → A suspected creative problem sends you to the fatigue read first, because a winner that "stopped working" has five impostors and only one of them is fatigue.
- → The rules behind the states, the ten principles anyone or anything touching the account should follow, get their own piece in this series.
Making It a Daily Discipline Without Living in Ads Manager
You know the drill: the tab is open all day, and the day's numbers are wrong all day, because conversions have not finished landing. Here is the cadence that replaces it.
One read a day, at the same time, on yesterday's complete numbers, with Shopify orders next to the platform figures. Ask the three questions. Assign a state. If the state is anything other than "act now", write the window and the trigger and close the tab. Once a week, a longer read for the allocation questions: marginal efficiency, creative rotation, what to feed and what to cut. Once a month, the things that only show up slowly: new-customer share, product mix, payback.
Daily is also the right cadence for the signal-versus-noise call itself. Intraday numbers are incomplete by construction; a weekly read misses a leak for six days. Yesterday, read once, on complete data, is where the signal lives.
How Drew fits. Ask Drew "what changed yesterday, and is it real?" Its diagnosis separates real signal from normal variance, working through the same order as above (measurement, the Shopify business behind the ads, account state, then the funnel), and it includes the case where the honest answer is to leave the account alone. Drew's checks run daily: it reads yesterday, not the last hour, and a daily read is the right cadence for telling signal from noise. The six states above are published method, the vocabulary a good buyer works in; Drew gives you the diagnosis and the reasoning, and the call on whether to touch the account stays yours.
If you want to see it on your own account, pricing is published and flat and the free plan doesn't need a card.
Key Takeaway
Every anomaly in your ad account is one of three things: a real signal, normal variance, or the echo of something you changed yourself. Answer that before you touch anything, because over-intervention is the more expensive failure: each reflex edit can reset learning, muddy the comparison and kill a winner for having a bad Tuesday. Work in six intervention states rather than two, treat "hold and observe" and "do not act" as decisions with a written reason, window and trigger, and make an "act now" earn its place with alternative explanations, guardrails and a rollback rule. Then read the account once a day on yesterday's complete numbers, and close the tab.
Frequently Asked Questions
Should I change my Facebook ads now or wait?
Wait if any of these is true: the move is under a full conversion cycle old, it sits on a small number of conversions, you edited the campaign in the last three to seven days, or the business underneath the ads changed (stock, site conversion, pricing, a promotion). Act now only when the evidence and the urgency are both high, such as a tracking break, spend going to a sold-out product, or a broken checkout, and write the rollback rule before you make the change.
How long should I run Facebook ads before deciding?
At least one full conversion cycle after the last material edit, which for most Shopify brands means seven days, and long enough to collect a meaningful number of conversions in the window. Judging a campaign on two days of data is judging it on incomplete reporting and unsettled learning. If conversion volume is low, extend the window rather than lowering the bar.
When should I kill a Facebook ad?
When it has had a fair run (a full cycle, enough impressions and conversions to read), its efficiency sits below your break-even on Shopify revenue rather than only on platform-reported numbers, and the cause is the ad itself rather than a site, stock or measurement problem that would sink any ad. Kill it as a reversible move where you can: pause rather than delete, so the history and the option to restart both survive.
How often should I make changes to my ad account?
Less often than you want to, and one causal change at a time. A daily read is right; a daily edit usually is not. Most accounts do well on one deliberate change per campaign per conversion cycle, with creative refreshes on their own schedule, and emergency changes reserved for genuine emergencies. If you cannot say which of last week's edits caused this week's result, you made too many.
Is a one-day ROAS drop a real signal?
Almost never on its own. Yesterday's ROAS is always understated because conversions keep reporting for days after the click, one large or small basket moves a low-volume day by 15% or more, and day-of-week effects are bigger than most operators assume. Treat a one-day drop as a prompt to check business reality (stock, site, pricing) and measurement, then hold and re-read on the full seven days.
What counts as normal variance in Meta ads?
The range your account produces week to week when nothing has changed. Measure it: take the last eight clean weeks and note the spread of weekly ROAS and CPA. Daily swings of 20–30% around the weekly figure are common at modest conversion volumes. A move that stays inside that spread is noise by definition, whatever it feels like at 2am.
Written by Sumit Bansal, co-founder of Datadrew. Published 3 September 2026. Quotes and figures are linked to their sources inline; the merchant lines come from our operator research and are quoted as written.