Every ads tool now says it uses AI. For an agency, that claim is hard to check from a demo. The demo account is clean and the recommendations sound sharp.
The real question is simpler: is the tool reading this client’s account, or is it running the same fixed rules on every account and writing the output in nicer sentences? Rules have their place. But if you are paying for judgment, you should be able to see judgment.
This guide gives you a five-step test you can run on any AI ads tool, a one-week side-by-side trial against your own team, and a scorecard you can paste into a client doc or a vendor review.
Key takeaways
- Test the tool on a client account you know well, so you can tell a good read from a lucky one.
- A real read names the ad set, the creative, what changed and why now. A rule names a metric and a threshold.
- Run it on a second account. The calls should differ where the accounts differ.
- Ask what it chose not to do, and why. Restraint with a reason is a strong sign of a real read.
- Generic “raise budget” or “pause” advice, with nothing tied to the account, is a rule with a new label.
Why rules give every account the same advice
Start with what a rule is, because Meta already gives you one for free. Automated rules in Ads Manager check your campaigns, ad sets and ads, then take an action or notify you. You choose the criteria that trigger the rule, the action it takes and what it applies to. Meta’s own examples include pausing a low-performing ad when it hits a specific metric, or raising budget based on ROAS.
Meta calls those criteria conditions, and an ad set must meet all the conditions on a rule to trigger it. That is the whole mechanism: if this number crosses that line, do this.
Say you set a rule: if cost per purchase is above $60 over the last 3 days and spend is above $200, pause the ad set. That rule will say the same thing to a skincare brand in its best month and a furniture brand in a slow week. It does not know that one ad set just had its budget doubled, or that the top creative has run for six weeks. It only knows the threshold.
Plenty of “AI” tools are that same logic with a chat window on top.
Steps 1 and 2: Use an account you know, and check for specifics
Step 1: Run the tool on a client account you know well
Pick the account your team understands best: its history, its seasonal swings, the creative that carried last quarter. On an account you don’t know, a confident but wrong read looks the same as a correct one.
Before you open the tool, have the account lead write down the three things they would change this week and why. This is your answer key.
Step 2: Check that it names the ad set, the creative and what changed
Read each recommendation and ask four questions. Which ad set or ad? Which creative? What changed? Why now, and not last week?
Here is the difference, with hypothetical numbers. Say the client’s target CPA is $40.
- Rule output: “CPA is above target on 2 ad sets. Consider pausing underperformers.”
- A real read: “The broad prospecting ad set went from about $38 to about $61 CPA after its daily budget doubled last Tuesday. Its Delivery column reads Learning again. Hold the budget and judge it after it has gathered more results, rather than pausing it now.”
The second one is tied to a date, a change and a reason you can check in Ads Manager. It also lines up with how Meta describes the learning phase: an ad set usually exits learning after about 50 results in the week after its last significant edit, and Meta recommends avoiding frequent budget changes that can send an ad set back into learning. A rule watching CPA alone would have paused it.
Step 3: Run it on a second account and compare the calls
Now run the same tool on a second client with a similar target but a different story. Say client B also has a $40 target CPA and also saw CPA climb to about $60. But here, nothing was edited. The top video has run for seven weeks, and its cost per result has risen week after week while the newer ads hold steady.
A real read should land somewhere different on each account. On client A, the call is “hold budget, wait for learning.” On client B, it is “the top video is fatiguing, brief new hooks on the same footage.” Same symptom, different cause, different action.
If both accounts get “pause ad sets above target CPA,” you are looking at a threshold. Lay the two lists side by side. Where the accounts differ, the advice should differ.
Steps 4 and 5: Ask what it chose not to do, and spot the relabeled rule
Step 4: Ask what it chose not to do, and why
This is the most revealing question you can ask. A rule cannot answer it, because a rule has no options it weighed. It either fired or it didn’t.
A real read can say something like: “I did not raise budget on the retargeting ad set, even though ROAS is above target, because it has only about 20 results since its last significant edit and a budget change could restart learning.” That is restraint with a reason, and it should match what your account lead would have said.
Step 5: Treat generic advice as a rule with a new label
Some phrases should make you look closer: “raise budget on winners,” “pause underperformers,” “refresh creative,” “test new audiences.” None of these are wrong. They are just true of every account, which is why they tell you nothing about this one.
If you can remove the client’s name from a recommendation and it still makes sense for any other client, it is a rule. If the tool only ever produces “raise” or “pause,” with nothing tied to an ad set, a creative or a date, you have your answer.
A one-week side-by-side trial: your team vs the tool
Keep it to one week and one account, and judge the quality of the read, not the results. One week is too short to credit a tool with a better ROAS or CPA, and some changes need time in learning before they show anything.
- Day 1: Pick the account from step 1. The account lead writes their own read first: top three changes, the reason for each, and one thing they would not touch. Seal it. Note the date range and the attribution setting you used.
- Day 1, later: Connect the tool to the same account. Give it read access, or make sure nothing runs without your approval. Collect its recommendations using the same date range and attribution setting.
- Day 2: Compare the two lists line by line. Mark each tool item as: matched the team, found something the team missed, or wrong. Check every “missed” and “wrong” item in Ads Manager.
- Days 3 to 5: Run the account as you normally would. Each day, log what the tool recommends and whether your team would act on it. Ask it step 4’s question at least twice.
- Day 4: Run the tool on a second account for step 3. Compare its calls across the two clients.
- Day 7: Fill in the scorecard below with the account lead and one other person on the team. Decide: extend the trial, use it as a helper, or drop it.
Copy-ready scorecard
Paste this into your vendor review or client doc. Score each line 0 (no), 1 (sometimes) or 2 (yes, consistently).
- Specific: names the exact ad set, ad or creative. Score: __
- What changed: says what moved, and when. Score: __
- Why now: explains why this week, not last week. Score: __
- Different accounts, different calls: advice differs where the two accounts differ. Score: __
- Restraint: says what it chose not to do, and why. Score: __
- Checks out: each reason can be verified in Ads Manager, and it matched or beat the team’s sealed read. Score: __
- Total (out of 12): __. 10 or more: worth a longer trial. 6 to 9: a helper, not a decision maker. Under 6: rules with a new label.
Test account: ____. Second account: ____. Date range and attribution setting: ____. Scored by: ____.
Where Vibelets fits
Vibelets audits each client’s Meta ad account once it is connected and the 14-day trial starts, so you can run this test on an account you know. Each week it writes a plan of what to pause, test and launch, with the reason for each proposed change, which gives you something concrete to score. Several client ad accounts sit in one workspace, each with its own audit and plan, and the All accounts view shows them side by side for the second-account check. It runs only changes your team approves, and the activity log records each change and who approved it. See Vibelets for agencies.
FAQ
Are Meta’s automated rules a bad idea?
No. Rules are good at simple, known limits, like turning off ads on a schedule or alerting you when spend passes a line. Meta itself says you should still monitor overall performance when using them. The problem is only when a tool sells rules as judgment.
Should the tool be allowed to make changes during the trial?
Not without your approval. The point of the trial is to grade its read against your team’s. If it changes the account mid-week, you can no longer tell whether results came from the tool, your team or normal swings.
What if the tool disagrees with my account lead?
That is useful. Ask for its reason and check it in Ads Manager. If the reason holds up, it found something your team missed. If it doesn’t, mark it wrong on the scorecard.
Is one week enough to judge an AI ads tool?
It is enough to judge whether the tool reads an account like a person who knows it. It is not enough to judge performance. If it passes the scorecard, extend the trial before you credit it with any change in CPA or ROAS.