How to check your Facebook ad account health, weekly

Search for this and you mostly get checklists. Look at spend, look at frequency, look at your relevance diagnostics. All reasonable advice, but none of it tells you what a bad number actually looks like, which is the part you need.

You open the account, CPM is 14 dollars, and nothing on the screen tells you whether 14 is normal for this account. So the call becomes a judgement you make on the day. Two people can look at the same week and reach opposite conclusions, and the same person can reach different conclusions on a Monday and a Thursday.

Industry benchmarks do not solve it either, because they describe other people's accounts. Your seasonality, your offer and your audience are all different, so a fixed CPA target will mark a perfectly decent week as a bad one. The comparison that works is the same account, the week before.

The date maths matters more than it looks. Meta's last_7d preset ends yesterday, not today, so the prior window has to be the 7 days immediately before that. Get this wrong and your two windows overlap by a day, which quietly flattens every change you are trying to detect:

const thisEnd    = yesterday;
const thisStart  = new Date(yesterday.getTime() - 6 * day);
const priorEnd   = new Date(thisStart.getTime() - day);
const priorStart = new Date(priorEnd.getTime()  - 6 * day);

With the windows fixed, it pulls account totals for both, plus ad set and ad level insights for the current week so the report can name what is dragging the score down rather than just reporting a number.

There is a bigger problem than date maths, and most week-over-week reporting gets it wrong. Meta attributes a conversion to the date of the click, not the date the conversion happened. On the default 7-day click window, a purchase today can be written back to a click six days ago. Your most recent days therefore keep filling in for about a week after the fact.

The consequence is that a naive comparison is biased against the current week every single time. You are measuring a window that is still filling against one that has already settled, so the recent week looks worse than it will turn out to be, and the score flags problems that resolve themselves by Thursday.

Two ways to deal with it. Lag both windows by 7 days so you compare two settled periods, which costs you a week of freshness and is the honest option for accounts with long consideration cycles. Or keep the windows where they are and read the trend rather than the absolute, accepting that the current week is a floor and not a reading. What you cannot do is compare a fresh week to a settled one and treat the difference as real, which is what most weekly reports do.

One node runs the check battery and scores it. A Claude step then ranks the flagged checks into the top three fixes, with a hard instruction to use only figures that appear verbatim in the input, so it cannot invent a number. Delete that node and the score still works.

1. Config node. Every threshold in one place, plus the list of conversion events that count as a result for your account. That list matters. A lead gen account and an ecommerce account do not share a definition of success:

KPI_FLAG_PCT: 15,        // primary KPI worse by this much = flag
KPI_CRIT_PCT: 30,        // ...= critical
SPEND_UP_PCT: 10,        // spend up this much while KPI worsens = flag
CPM_FLAG_PCT: 25,
VOLUME_DROP_PCT: 25,     // impressions down

2. Build the two windows. Pure date maths, no API calls, producing a this-window and prior-window range.

3. Fetch. Account totals for each window, then ad set and ad insights plus the ad set entities, because insights give you performance but not names or budgets.

4. Run checks and score. One node, pure transforms. Each check returns pass, flag or critical, and the score falls out of the results.

5. Triage and print. The AI step ranks fixes, the final node assembles the scorecard. Swap it for Slack or a sheet without touching anything upstream.

Tune the thresholds first. A high-volume account is more stable, so 15 percent is a real signal. A small account swings 15 percent on noise. Raise it until the flags mean something.

Set your result actions. The single most important edit. Get this wrong and every other number is meaningless.

Compare month over month. Widen both windows for low-volume accounts where a week is too noisy to read.

Log the score.Append each run to a sheet and the score becomes a trend line, which is far more useful than any single week's number.

Put it on a schedule. Swap the manual trigger for a Monday Schedule Trigger. That is the change that turns a check into monitoring.

Want a score in your inbox every Monday?

30 minutes, free. Bring your ad account. You leave with a plan for a weekly health score and the top three fixes, whether you build it or we do.