Meta ads anomaly detection: two ways of returning nothing

You ask Claude why last week went sideways. It reaches for the anomaly tool, gets one sentence back, and tells you nothing is being flagged. That sounds like reassurance, so you stop looking.

Sometimes it is reassurance. Sometimes the examination never happened and the tool is telling you so in wording almost nobody reads closely. The reply gives you very little to tell those apart, and the difference is the entire value of having asked.

What the connector actually hands back

Point it at an ad account and the reply is a single key holding a sentence of English. No list, no severity, no timestamp, no count of what was examined. Across every ad account I can reach, on 29 August 2026, that sentence arrived in exactly two forms:

No anomaly signal data available for the given criteria.

No anomaly signal issues found in the analyzed ad sets.
Checked for auction overlap, creative fatigue, and narrow audience.

Most of the accounts gave me the first one. One gave me the second. The one that answered properly is the one with a real history of delivery behind it; the ones that shrugged include a dormant account and one with no payment method attached. That pattern is worth more than the sentences themselves, because it tells you what the tool needs before it will say anything: enough recent ad-set activity to have a baseline to deviate from.

The second sentence is also the only place the tool tells you what it is actually testing. Auction overlap, creative fatigue, narrow audience. That is the list. Not cost per result, not pacing, not conversion rate — three structural conditions, all of them properties of how an ad set is configured and how much room it has to spend.

So I pushed on the scope. The tool takes an optional list of entity IDs, and its own description says those can be campaigns, ad sets or ads. I passed it two ad IDs. The answer came back about the analyzed ad sets. I passed a campaign ID instead, and got the same wording. Whatever level you hand it, the thing it reasons about is the ad set, and the sentence quietly says so. If your question is "which of my ads is behaving strangely", this is not the tool that answers it.

The two replies, side by side

A model summarising either of these for you will fairly reasonably say "no anomalies found". That summary is correct about the first sentence and useless, because a check that did not run cannot find anything. This is the ordinary failure of an assistant reading tool output: the wording is faithful and the meaning has gone.

Most of the skill here is knowing which question a tool answered. See how the weekly review is structured →

How to get something usable out of it

1. Ask for the reply, not the summary START HERE

   Tell Claude to quote the tool's answer back verbatim before it interprets anything. The whole distinction lives in a few words, and it is exactly the sort of detail that gets smoothed away in a paraphrase. The trade: your chat fills up with raw output, and on an account with several tools in play that gets tedious fast.

2. Run it when there is something to see

   The shrug reply tracks with quiet accounts. Point it at a period and a set of ad sets that have actually been spending, and it has a baseline to work against. Asking a dormant account what is unusual about it is a question with no answer. The trade: that is precisely the moment you are least likely to bother asking, because everything looks fine.

3. Treat it as a pre-read, never a verdict

   The tool's own description is unusually careful here. It tells the model to distinguish observations from causes, says anomalies indicate areas for deeper analysis rather than definitive conclusions, and rules itself out for delivery-blocking setup problems, policy and publish errors — pointing at the errors tool territory instead. It also defers to the opportunity score for anything claiming a quantified lift.

A check that never ran and a check that passed do not deserve the same sentence in your head.

One vendor write-up I found says the tool returns a learning-phase flag on any ad set. I cannot corroborate that. I never got a populated response out of it, so I do not know what a hit looks like, and neither the tool's description nor any Meta page mentions learning phase. Take that claim as unverified until someone publishes an actual flagged response — including me.

• Reading "no anomalies" as "the account is healthy." The three checks it names are narrow, and everything outside them was never in scope.
• Asking it about a single ad. It rolls whatever you give it up to ad sets and answers there, which you only notice if you read the wording.
• Using it to explain why spend stopped. Its description excludes delivery-blocking issues by name. You will get a confident nothing.
• Letting it close the review. It is the cheapest first look available, and it is a first look.

Try this tonight Ask for the anomaly check on each ad account you hold, and insist on the raw sentence each time. Sort the accounts into the ones that answered and the ones that shrugged. The shruggers are the accounts where this tool will keep telling you nothing no matter how often you ask, so stop counting its silence as information there — and go through those by hand with the audit questions instead.

Every diagnostic on this connector deserves the same treatment: call it, read the exact words, work out what it did not test. The counted manifest is where to find the rest of them, and the connection walkthrough is the place to begin if none of this is wired up on your machine yet.