# When to pause an ad: tired ad, used-up audience or weak from the start

> Pause an ad when nothing you changed explains the drop, when its share of the results is smaller than its share of the spend by more than luck explains, and when it is worse than its peers in both halves of the window. With 20 results expected at the peers' cost, an ad needs 11 or fewer to count as clearly worse; with 5 expected, only zero results counts. Then pool the ads with the same angle: if the pool is weak too, the angle failed, and a new ad with the same argument will fail as well.

Source: https://nimblet.io/guides/when-to-pause-an-ad. Updated 2026-10-23.

Someone asks in the team channel: is this fatigue? A pause decision often starts with one bad week and a feeling. Two mistakes cost the most. You pause an ad that was only unlucky, or you keep a weak ad because it won once. The decision tree below removes both, and its arithmetic holds at the small result counts of business-to-business (B2B) accounts.

## How do you know the drop is real and larger than noise?

Compare the ad's share of the spend with its share of the results. An ad as good as its peers gets a share of results that matches its share of spend. The drop is real when the ad's share of results is smaller than luck alone would give. How big that gap must be depends on how many results stand behind it.

Results are counts, such as leads or registrations. A count moves by luck. For a count of N, luck moves it by about the square root of N (one standard deviation), because the variance of a Poisson count equals its mean ([Wikipedia: Poisson distribution](https://en.wikipedia.org/wiki/Poisson_distribution)). We call this the swing. As a share of the count, the swing is 1 divided by the square root of N.

| Results | Swing from luck, as a share of the count |
| --- | --- |
| 5 | 45 percent |
| 10 | 32 percent |
| 25 | 20 percent |
| 100 | 10 percent |

An ad that would average 10 results lands between 7 and 13 in about three periods out of four, with no change in the ad. So a comparison on 10 results says little.

To test an ad, work out the results it would have at its peers' cost. Call this number E. Add up the peers' spend and the peers' results, divide spend by results to get the peer cost per result, then divide the ad's spend by that cost. Now ask how often an ad that is as good as its peers lands this low. The table below answers it. We built it from the exact Poisson probability. It flags an ad when that chance is 2.5 percent (1 in 40) or lower. It looks at one side only, because only a worse ad matters for a pause.

| Results you expect at the peers' cost (E) | Flag the ad at this many results or fewer | How far below E that is |
| --- | --- | --- |
| 5 | 0 | 100 percent |
| 10 | 3 | 70 percent |
| 15 | 7 | 53 percent |
| 20 | 11 | 45 percent |
| 25 | 15 | 40 percent |
| 50 | 36 | 28 percent |
| 100 | 80 | 20 percent |
| 200 | 172 | 14 percent |

Read it like this. If you expect 20 results at the peers' cost, an ad with 11 or fewer is clearly worse. An ad with 12 may still be unlucky. If you expect 5, only an ad with zero results counts. Below 4 expected results we make no call at all.

In a spreadsheet one formula gives the exact chance. `BINOMDIST(results, results + peer results, spend share, TRUE)` works in Google Sheets ([Google help](https://support.google.com/docs/answer/3093987?hl=en)), and Excel has `BINOM.DIST` with the same four arguments ([Microsoft help](https://support.microsoft.com/en-us/office/binom-dist-function-c5ae37b6-f39c-4be2-94c2-509a1480770c)). It works because a test of two Poisson rates reduces to a binomial test once you fix the total count ([R manual: poisson.test](https://stat.ethz.ch/R-manual/R-devel/library/stats/html/poisson.test.html)). The cut-off table treats the peers' cost per result as known. That is fair when the peers hold most of the spend. When the ad holds 30 percent of the spend, the real chance behind a flag from the table is about 4 to 5 percent for 20 or more expected results, not 2.5 percent (our computation). The formula has no such flaw, and the template below has a by-hand method that follows it.

Two rules keep the arithmetic honest. First, add up the peers; never average their costs. Take two peers: one spent 100 euros for 10 results (10 euros each), the other spent 100 euros for 1 result (100 euros each). The average of the two costs is 55 euros. The true cost of the pair is 200 euros for 11 results, which is 18.18 euros. Second, use the ad's own peers, not the account. Peers share the goal result, the audience and the format family.

Comparing two single ads is harder, because both counts are noisy. This table assumes equal spend. The better ad has the results in the first column. The weaker ad is flagged at the number in the second column or lower.

| Results of the better ad | Weaker ad is flagged at | Its cost per result is then at least |
| --- | --- | --- |
| 5 | no count flags it | not possible |
| 10 | 2 or fewer | 5.00 times |
| 25 | 12 or fewer | 2.08 times |
| 50 | 31 or fewer | 1.61 times |
| 100 | 73 or fewer | 1.37 times |
| 200 | 161 or fewer | 1.24 times |

So two ads with about 25 results each can only be told apart when one costs about twice as much per result. Small budgets cannot see small differences.

The tables are a minimum. Four things make real decisions harder than the tables show.

- Results come in clumps. Several leads can come from one company or one day. In our reading this makes counts swing more than the Poisson model expects, an effect called [overdispersion](https://en.wikipedia.org/wiki/Overdispersion).
- Looking again and again inflates false alarms. Evan Miller shows that peeking at a running test ten times turns what you think is 1 percent significance into about 5 percent ([Evan Miller](https://www.evanmiller.org/how-not-to-run-an-ab-test.html)). Pick one review day per window and look only then.
- Many ads make a false alarm likely. With six ads checked at 1 in 40 each, the chance that luck flags at least one ad is about 14 percent (1 minus 0.975 to the sixth power). The second half-window check and the angle check below exist for this reason.
- Ads in one ad set compete for the same people, so their results are not fully independent. Use the chances as a guide.

A wrong pause costs you one ad, so you may set a looser bar. LinkedIn tells advertisers that a 10 percent level of statistical significance is typically acceptable for its A/B tests ([LinkedIn help](https://www.linkedin.com/help/lms/answer/a525922)). The page does not say whether that level counts one side or both sides. Read as 1 chance in 10 on the worse side, the flag at 25 expected results moves from 15 results to 18.

## What is the difference between creative fatigue and audience saturation?

Creative fatigue hits one ad. Audience saturation hits every ad in one audience. Meta defines the first as an audience seeing the same creative too many times ([Meta help](https://www.facebook.com/business/help/1346816142327858)). It defines the second as a high share of your audience seeing your ads without acting ([Meta help](https://www.facebook.com/business/help/388669258876309)). A tired ad needs a new idea. A used-up audience needs new people.

Think of the person, not the chart. Someone who saw your ad three times and did not act has given an answer: not now, or not you. A fourth view adds no news. If everyone on a small account list has that answer, a new picture of the same idea changes nothing. A new reason to act does, and a new reason is a different concept. A new color on the same idea is not one. This is our reading.

Frequency alone proves neither case. Meta says to watch frequency together with results ([Meta help](https://en-gb.facebook.com/business/help/1546570362238584)). Jon Loomer writes that "a high frequency, by itself, is not problematic" ([Jon Loomer](https://www.jonloomer.com/creative-fatigue-meta-ads/)).

Both can happen at once. To tell them apart, put a new concept into the same audience. If it does well, the audience still has people to give. Do not test the old ad on a fresh audience for this. Fresh people have not seen it, so it recovers in both cases. This is our reading, and one new ad is a hint, not proof.

## What does each one look like in the numbers?

Read several signals together. One falling number has many causes.

| Signal | Tired ad | Used-up audience | Where you read it |
| --- | --- | --- | --- |
| Cost per result of this ad against the other ads in the ad set | Rises for this ad, the others hold | Rises for all ads together | Both platforms, per ad |
| Unique reach, read over a date range that grows week by week | Grows more slowly for this ad; other ads still add people | Grows more slowly for all ads together | Meta per ad; LinkedIn per ad set |
| Frequency | Rises faster on this ad than on the others | Rises across the ad set | Meta per ad; LinkedIn average per ad set |
| Share of the audience reached | Not readable for one ad, because penetration covers whole ad sets | High, and it stops rising | LinkedIn audience penetration, per ad set |
| Video: people who stay past the first seconds | Falls on this ad | Falls on every ad | Both platforms |
| A new, different ad in the same audience, judged with the same tables | Does well | Is weak too | Both platforms |

Reach counts unique people, so do not add up weekly reach. Read it over a date range that grows. LinkedIn shows reach and average frequency per ad set. It models these numbers, asks you to use them directionally, and needs 24 to 36 hours to show them ([LinkedIn help](https://www.linkedin.com/help/lms/answer/a426154)). Audience penetration is the share of your target audience that your ad sets reached. It is approximate and needs an audience of 300 or more (same page). We found no fatigue label in LinkedIn's help pages on delivery metrics, rotation and creatives, and no per-ad reach on the delivery metrics page. On LinkedIn you diagnose with the cost per result of the other ads in the same ad set.

Meta shows reach and frequency per ad, and it estimates frequency from sampled data ([Meta help](https://en-gb.facebook.com/business/help/1546570362238584)). Meta also writes Creative limited or Creative fatigue in the Delivery column when it judges that people saw the same ad too often and the cost per result is higher than for past ads. The label reads limited below twice the past cost and fatigue at twice or more. It exists only for ad sets with one creative, and not for dynamic creative or Advantage+ catalog ads ([Meta help](https://www.facebook.com/business/help/1346816142327858)). An ad set with several ads may never show it.

For video, hook rate is 3-second plays divided by impressions, and hold rate is ThruPlays divided by 3-second plays. People build both from Meta's counts ([Meta help](https://en-gb.facebook.com/business/help/743427195703387), [Meta help](https://www.facebook.com/business/help/2051461368219124)). They are custom metrics, not default columns. LinkedIn reports view rate and views at 25, 50 and 75 percent ([LinkedIn help](https://www.linkedin.com/help/linkedin/answer/85305)). A falling start rate says the first seconds stopped working. A falling hold ratio with a steady start rate says the middle of the video lost people. That is our reading.

## What looks like fatigue but is not?

A change you made, a change the platform made, or a change in the market. Rule these out before you touch the ad.

- You changed something. On Meta, a new ad or a significant edit starts the learning phase. Results are less stable then, and cost per action is usually higher ([Meta help](https://www.facebook.com/business/help/112167992830700)). Significant edits include pausing the ad set and changes to the optimization event, audience or creative ([Meta help](https://www.facebook.com/business/help/942374239243867)). That page does not say whether pausing one ad counts, so treat the days after any pause as noisy.
- Delivery moved to other ads. LinkedIn's recommended rotation setting, Optimize for performance, serves ads evenly at first and then gives more impressions to the ads that perform best ([LinkedIn help](https://www.linkedin.com/help/lms/answer/72149)). Meta says some ads are delivered more than others ([Meta help](https://www.facebook.com/business/help/112141358874214)). A starved ad has a small count. It looks tired in a chart, but it never had enough delivery to tire. A small count gives a small E and no flag, so the tables above handle it.
- The budget moved. With Advantage+ campaign budget, Meta may spend up to 75 percent over or below the daily budget on one day. A budget change late in the day can shift the over-spend or under-spend to the next day ([Meta help](https://www.facebook.com/business/help/388669258876309)). Judge weeks, not days.
- The audience changed. On LinkedIn, audience expansion and the LinkedIn Audience Network are ad set settings you can switch on or off ([LinkedIn help](https://www.linkedin.com/help/lms/answer/a425248)). Either one changes who sees the ad.
- The market moved. Meta says that when other advertisers bid more aggressively, your cost per result goal or bid cap may no longer be high enough to compete ([Meta help](https://www.facebook.com/business/help/388669258876309)). Then a rising cost can come from the auction and not from the ad. Holiday and summer weeks also change who reads and acts. No platform page we read states this. It is practice, not a measured fact.
- The page, the form or the tracking changed. A change to a conversion window affects only conversions counted after the change, and in April 2025 LinkedIn lengthened the windows for event registrations, which can lift counts ([LinkedIn help](https://www.linkedin.com/help/lms/answer/a426359)). PPC Land reported in October 2025 that from 12 January 2026 the Meta Ads Insights application programming interface (API) would stop returning the 7-day view and 28-day view windows ([PPC Land](https://ppc.land/meta-restricts-attribution-windows-and-data-retention-in-ads-insights-api/)). A report built on those windows can show a drop that no ad caused.
- The last days are not final. A conversion can be credited to an ad at any time inside the conversion window, which defaults to 90 days for most conversions on LinkedIn ([LinkedIn help](https://www.linkedin.com/help/lms/answer/a426359)). Leave out the last two days, and more for slow conversions.

Write each change on a dated line. Step 1 of the tree uses that list.

## How do you tell fatigue from an ad that was weak from the start?

A tired ad worked first. A weak ad never did. For each week, divide the ad's cost per result by its peers' cost per result. Call this the weekly index. An index of 1.0 means equal to the peers, and above 1.0 means worse. These three cases are constructed, with costs in euros.

| Case | Ad cost per result, weeks 1 to 4 | Peers' cost per result, weeks 1 to 4 | Weekly index | Reading |
| --- | --- | --- | --- | --- |
| A | 40, 42, 55, 78 | 41, 40, 42, 41 | 0.98, 1.05, 1.31, 1.90 | Tired ad: a good start, then a slide while the peers hold |
| B | 90, 85, 95, 88 | 45, 44, 46, 45 | 2.00, 1.93, 2.07, 1.96 | Weak from the start: worse in week one and flat after |
| C | 42, 50, 61, 74 | 41, 49, 60, 72 | 1.02, 1.02, 1.02, 1.03 | Peers slide too: used-up audience or market shift, not this ad |

Case C is the trap. The index stays near 1.0 because every ad slides together, so a test against peers finds nothing. Look at the absolute cost per result of the whole ad set, then at reach and frequency, then at the rule-outs above.

A new ad needs a full window of its own. Meta says a new ad enters the learning phase, where performance is less stable ([Meta help](https://www.facebook.com/business/help/112167992830700)). LinkedIn serves ads evenly at first ([LinkedIn help](https://www.linkedin.com/help/lms/answer/72149)). The first days of a new ad prove little in either direction.

In our reading, weak from the start means the problem sits in the concept: the hook, the offer, the proof or the match with the buyer. A refresh does not fix that. Retire the ad and learn from it. A slide means the ad had something and used it up. Keep what it proved and carry it into the next concept.

## When should you pause an ad?

Pause when four things are true:

1. Nothing you changed explains the window.
2. The share of results of the ad, or of the pool of ads with its angle, is smaller than its share of spend by more than luck explains.
3. The ad is worse than its peers in both halves of the window.
4. A new ad can fix the cause. The cause is the ad or its angle, not the audience, the page or the market.

The block below is the whole rule. Paste it into your team notes and fill it in once per review. It works on LinkedIn and Meta.

```template
AD DECLINE DECISION TREE

Ad: [ad name]
Platform: [LinkedIn or Meta]
Ad set: [ad set name]
Goal result: [what the ad set was bought for, for example lead form completions]
Window: [first date] to [last date]. Leave out the last 2 days.
Review day: [weekday]. Look once per window, on this day.

STEP 1. Was the window clean?
List every change inside the window, with its date: budget or bid, audience or placement, an ad added or edited, landing page or form, tracking or conversion window, a holiday week, an ad set paused and restarted.
- A change happened: restart the window on the change date. Stop. Come back when the new window has data.
- No change: go to step 2.

STEP 2. Who are the peers?
Peers are the other ads with the same goal result, the same audience and the same format family (for example single image, video or document), in the same window. Use the other ads of the same ad set when you can.
- Fewer than 2 peers: stop. Keep the ad running.
- 2 or more peers: go to step 3.

STEP 3. Is the gap larger than luck?
Ad spend A = [amount]. Ad results a = [count].
Peer spend B = [amount]. Peer results b = [count]. Add the peers together. Never average their costs.
Peer cost per result = B divided by b.
Expected results E = A divided by (B divided by b). This is what the ad would have if it were as good as its peers.
- E below 4: we make no call. Go to step 6 and pool by angle.
Chance of luck = the chance that an ad as good as its peers shows [a] results or fewer. Use one method, not both.
  Spreadsheet (exact). Google Sheets: =BINOMDIST([a], [a]+[b], [A]/([A]+[B]), TRUE)  Excel: the same with BINOM.DIST
  By hand (close to the spreadsheet):
    p = A divided by (A + B). This is the share of the spend that the ad holds.
    n = a + b. This is the number of results of the ad and its peers together.
    M = n times p. This is what the ad would get if results followed spend.
    Swing = the square root of (M times (1 minus p)). This is the typical size of luck.
    Gap = (a minus M) divided by the swing. A negative gap means the ad is below M.
- Chance at or below 0.025 (1 in 40), or a gap of -2 or lower: the ad is a suspect. Go to step 4.
- Chance above 0.025, or a gap above -2: the gap can be luck. Before you keep running, do two checks.
  First, compare the cost per result of the peers in the first and the last week. If it clearly rose, go to step 7, case F.
  Second, if the gap is negative (the ad is below M), go to step 6 and pool by angle.

STEP 4. Is the gap in both halves?
Split the window into two halves of equal length. For each half, work out the ad's cost per result and the peers' cost per result.
First half: ad [cost], peers [cost]. Second half: ad [cost], peers [cost].
A half with no results for the ad has no cost per result. Count that half as worse.
- Worse than the peers in both halves: go to step 5.
- Worse in one half only: keep running. Look again on the next review day.
This check catches one bad stretch. It is not a second statistical test.

STEP 5. Did it ever work?
For each week, divide the ad's cost per result by the peers' cost per result. This is the weekly index. 1.0 is equal. Above 1.0 is worse. A week with no results for the ad counts as worse.
- Clearly above 1.0 in the first full week and still there: weak from the start.
- Near or below 1.0 at first, then rising for 2 or more weeks: a slide. A slide needs at least 3 weeks of data. With fewer, you can say "weak from the start" or nothing.
Also check the peers that are not suspects. If their own cost per result clearly rose over the same weeks, the audience or the market moved. Go back to step 1, then read reach and frequency (Meta, per ad) or average frequency and audience penetration (LinkedIn, per ad set).

STEP 6. Ad or angle?
An angle is the reason the ad gives the buyer to act. Two ads that make the same argument share an angle, even when they look different.
Pool all ads that share this ad's angle: add up their spend and results. Run step 3 on the pool. The peers of the pool are the ads of other angles.
- The pool is flagged: the angle failed for this buyer. Run steps 4 and 5 on every ad of the pool that is not a suspect yet. Every ad of the pool then gets the verdict of step 7, flagged alone or not.
- The pool is not flagged but the ad is: this execution failed.
- Neither is flagged: go to step 7, case A.
- E of the pool is below 4: say "too early". Do not retire the angle.
If step 5 found a slide, also work out the weekly index of the pool against the ads of other angles. A pool that slides too means the angle is tired.

STEP 7. Decide.
A. Neither the ad nor its pool is flagged, or E below 4: keep running. Look again on the review day.
B. Weak from the start, angle flagged: pause the ad. Retire the angle. Brief a new concept.
C. Weak from the start, angle not flagged: pause the ad. Keep the angle. Write a new execution (hook, visual or proof).
D. A slide, and the other ads of the angle slide too: the angle is tired. Rest the angle. Brief a new concept.
E. A slide, and the other ads of the angle hold: this ad is tired. Add a new execution of the angle. Pause the old ad when the new one has a window of data.
F. The peers slide too, and nothing changed in step 1: used-up audience or market shift. Do not pause ads for it. Widen or refresh the audience, cut overlap, then look again.
Write a pause note for every pause: the date, the numbers, the cause you believe, and what you would see that proves the cause wrong.
```

Here is the tree run on a constructed example. LedgerLoop is an invented month-end close tool. One LinkedIn ad set holds six single-image ads for two weeks. The goal result is lead form completions. Every number is invented.

| Ad | Angle | Spend (euros) | Results | Cost per result | Share of spend | Share of results | Chance of luck |
| --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | Close in 3 days | 1,000 | 22 | 45.45 | 18.2 percent | 22.4 percent | ad is better than peers |
| 2 | Close in 3 days | 900 | 20 | 45.00 | 16.4 percent | 20.4 percent | ad is better than peers |
| 3 | Audit trail | 1,100 | 24 | 45.83 | 20.0 percent | 24.5 percent | ad is better than peers |
| 4 | Audit trail | 800 | 17 | 47.06 | 14.5 percent | 17.3 percent | ad is better than peers |
| 5 | Peer story | 1,000 | 9 | 111.11 | 18.2 percent | 9.2 percent | 0.0103 (about 1 in 100) |
| 6 | Peer story | 700 | 6 | 116.67 | 12.7 percent | 6.1 percent | 0.0270 (about 1 in 37) |

Take ad 5. Its peers spent 4,500 euros for 89 results, so their cost is 4,500 / 89 = 50.56 euros per result. At that cost the ad's 1,000 euros would give E = 1,000 / 50.56 = 19.8 results. The ad has 9. Now the by-hand method. The ad holds p = 1,000 / 5,500 = 0.182 of the spend, and the ad and its peers have n = 9 + 89 = 98 results together. If results followed spend, the ad would get M = 98 × 0.182 = 17.8. That is a little less than E, because M counts the ad's own results too. The swing is the square root of (17.8 × 0.818), which is 3.8. The gap is (9 - 17.8) / 3.8 = -2.3, below the line of -2. So ad 5 is a suspect. The exact chance is `BINOMDIST(9, 98, 1000/5500, TRUE)`, which is 0.0103. Its peers include the other weak ad, which makes the check a little cautious.

Steps 4 and 5 need the weeks. Each ad spent the same in both weeks.

| Group | Week 1 spend (euros) | Week 1 results | Week 2 spend (euros) | Week 2 results |
| --- | --- | --- | --- | --- |
| Ads 1 to 4 | 1,900 | 42 | 1,900 | 41 |
| Ad 5 | 500 | 6 | 500 | 3 |
| Ad 6 | 350 | 3 | 350 | 3 |

Ad 5 cost 83.33 euros per result in week one and 166.67 in week two. Its peers (ads 1 to 4 and ad 6) cost 50.00 and 51.14. Ad 5 is worse in both halves. The weekly index is 1.67, then 3.26, so the ad was weak from the start and got worse. Ads 1 to 4, which are not suspects, cost 45.24 and 46.34 per result. They did not slide, so the audience did not move.

Ad 6 has E = 13.4 and 6 results. Its share of the spend is 700 / 5,500 = 0.1273, so M = 98 × 0.1273 = 12.47 and the swing is 3.30. The gap is (6 - 12.47) / 3.30 = -1.96, just short of -2. Alone, ad 6 is not flagged (exact chance 0.027). That is the small-count problem. The gap is negative, so step 3 sends the ad to step 6. The two Peer story ads together spent 1,700 euros, which is 30.9 percent of the spend, and got 15 of the 98 results, which is 15.3 percent. Their E is 37.1 and they have 15. For the pool, M = 98 × 0.309 = 30.3 and the swing is 4.6, so the gap is (15 - 30.3) / 4.6 = -3.3. The exact chance is 0.0003, about 1 in 3,000. The two other angles run better than their peers.

Ad 6 is also weak from the start. It cost 116.67 euros per result in both weeks. Its peers (ads 1 to 5) cost 50.00 and 54.55, so its weekly index is 2.33, then 2.14, and it is worse than its peers in both halves. The verdict is B for both Peer story ads: pause them and retire the angle. Pausing only ad 5 and writing another quote card would repeat the failure.

You can have an assistant run the same steps on an export. The prompt below holds a by-hand method, so it works without code. If your assistant can run code, it computes the exact chance. If you connect an assistant such as Claude to Nimblet through the [Nimblet Model Context Protocol (MCP) server](/mcp), the assistant can read the numbers itself.

```prompt
You are a careful paid media analyst. I give you a table of ads from one ad set, or from one group of peer ads, on [LinkedIn or Meta]. Decide which ads to pause and why. Use only the numbers I give you. Do not invent numbers. Show every calculation.

The table has one row per ad per week. Columns: ad name, angle, week start date, spend, goal result count. If a column is missing, ask me before you start. Peers share the goal result, the audience and the format family. If the table mixes ad sets, audiences, formats or goals, ask me which ads are peers before you start.
The goal result is [what the ads were bought for, for example lead form completions]. Money is in [currency].

[paste the table here]

Changes I made in this period, with dates: [list them, or write "none"]

Follow these steps in order.

1. Change check. If I list a change, say which weeks it touches. Use only the weeks after the latest change. If no week is left, stop and tell me to restart the window.

2. For each ad: A = its total spend, a = its total results. Peers = all other ads in the table. B = peer spend, b = peer results. Add up the peers. Never average ad costs. Peer cost per result = B / b. Expected results E = A / (B / b).

3. Chance of luck. This is the chance that an ad as good as its peers shows a results or fewer. If E is below 4, make no call and write "too small to test". Otherwise, if you can run code, compute the chance exactly: the binomial cumulative probability with n = a + b, k = a, p = A / (A + B). Flag the ad as a suspect if the chance is 0.025 or lower. If you cannot run code, say so and work by hand: p = A / (A + B). n = a + b. M = n * p. Swing = square root of (M * (1 - p)). Gap = (a - M) / swing. Flag the ad as a suspect if the gap is -2 or lower. Show all five numbers.

4. Halves. Do this for each suspect, and for each ad in a pool that step 6 flags. Split the weeks into a first half and a second half of equal length. If the number of weeks is odd, leave out the middle week. For each half, give the ad's cost per result and the peers' cost per result. A half with no results for the ad counts as worse than the peers. A suspect passes only if it is worse than its peers in both halves. If there is only one week, say so and skip this step.

5. Start or slide. Do this for the same ads as step 4. For each week, give the ad's cost per result divided by the peers' cost per result. Write "weak from the start" if the ratio is clearly above 1 in the first week and stays there. Write "slide" if it is near or below 1 at first and then climbs for two or more weeks. A slide needs at least 3 weeks of data. Also say whether the cost per result of the peers that are not suspects clearly rose over the same weeks. If it did, the verdict for the ad is "the audience or market moved, do not pause for it".

6. Angle. An angle is the reason the ad gives the buyer to act. Pool the ads that share an angle: add up their spend and results. Run steps 2 and 3 on each pool. The peers of a pool are the ads of other angles. If a pool is flagged, run steps 4 and 5 on every ad in it, and every ad in the pool gets the pool's verdict, flagged alone or not. If an ad shows a slide, also give the weekly ratio of its pool against the ads of other angles. A pool that slides too means the angle is tired.

7. Give one row per ad in this form, and write n/a where a step does not apply to an ad:
Ad | Spend | Results | Cost per result | Expected results | Chance of luck | Worse in both halves | Start or slide | Angle pool flagged | Verdict
The verdict is one of: keep running; too small to test; pause the ad and keep the angle; pause the ad and retire the angle; the angle is tired, rest it; the ad is tired, add a new execution and pause it when the new one has data; the audience or market moved, do not pause for it.

Rules. Never recommend a pause on cost per result alone. Never call a drop fatigue unless steps 4 and 5 show a slide. Write short plain sentences and use no em dashes or en dashes. End with three lines: what to pause, what to keep, what to test next.
```

On the LedgerLoop table the steps of the prompt give these verdicts: keep ads 1 to 4, and pause ads 5 and 6 with the angle Peer story retired.

### What if your numbers are too small to decide?

Then say so. Many B2B ad sets cannot buy a clean answer, and the tables show why. LinkedIn's own A/B tool asks for at least 14 days and recommends 21, with a minimum budget of 700 dollars in total and 20 dollars a day, and LinkedIn says conclusive results are not guaranteed ([LinkedIn help](https://www.linkedin.com/help/lms/answer/a525922), [LinkedIn help](https://www.linkedin.com/help/lms/answer/a529753)). One vendor page puts a full-scale LinkedIn test at 15,000 to 30,000 dollars per variant. The same page asks for at least 100 conversions per variant and quotes a cost per lead of 75 to 150 dollars. That multiplies to 7,500 to 15,000 dollars per variant, so the page disagrees with itself ([aimers.io](https://aimers.io/blog/linkedin-ads-ab-testing)). Treat the vendor figures as one vendor's rough numbers.

A small account decides on concept contrast and a stated cause, not on significance. Four moves help.

- Pool ads by angle. A pool of two or three ads has two or three times the counts of one ad, so it speaks sooner. That is step 6.
- Test ideas that differ in the idea. A second ad with the same argument and a new picture is a variation. Jon Loomer writes, about Meta ads, that unless variations are distinctly different, they are unlikely to make a meaningful impact ([Jon Loomer](https://www.jonloomer.com/andromeda-3/)).
- Write the cause before you decide. Use the pause note below. A cause you wrote down can be proven wrong, and an unwritten feeling cannot.
- Use an earlier action to find suspects, never to clear one. A landing page view or a video view at 75 percent happens more often than a lead. Use it to decide where to look first. Do not use it to keep an ad that the goal results flag.

## Did the ad fail, or did its angle?

Pool the ads that share the weak ad's angle and run the same test on the pool. If the pool is weak, the angle failed for this buyer, and new versions of it will fail the same way. If only the single ad is weak, its execution failed: a softer hook, a worse picture, a thinner proof.

An angle is the reason the ad gives the buyer to act ([angle guide](/guides/b2b-ad-angles)). Two ads that make the same argument share an angle, even when they look different. Nimblet gives each ad with recent spend an AI read that records its core message, hook, offer and proof, so you can line up the weak ads and see whether they make the same argument. If you tag by hand, the [reading card](/guides/ai-prompts-to-analyze-b2b-ads) gives the fields.

In the LedgerLoop example the pool decided what ad 6 alone could not. A pool needs enough spend and enough results to speak. If it has too few, write "too early" and keep the angle. Do not retire an angle on a pool that cannot be tested.

## What do you do next in each case?

| Situation | Do | Avoid |
| --- | --- | --- |
| Not flagged, or too few results | Keep running. Look again on the review day. Pool by angle. | Pausing on a bad day |
| Weak from the start, angle failed | Pause the ad, retire the angle, brief a new concept ([brief template](/guides/b2b-creative-brief-template)) | Writing another version of the same argument |
| Weak from the start, only this ad | Pause the ad, keep the angle, write a new hook, picture or proof ([hook guide](/guides/b2b-ad-hooks)) | Retiring a good angle |
| A slide, and other ads of the angle slide too | Rest the angle, brief a different concept | A new color on the same idea |
| A slide, other ads hold | Add a new execution, pause the old ad when the new one has data | Pausing before a replacement runs |
| Peers slide too, no change in the log | Widen or refresh the audience, cut overlap between ad sets, then look again | Pausing ads for an audience problem |
| A change you made explains the window | Restart the window on the change date | Judging the old numbers |

Meta backs the add-first, pause-later order in the table. For creative fatigue it recommends a new ad that is materially different from the original, and it says that keeping the original active instead of pausing it may maximize results. It also lists expanding the audience, to reach people who have not seen the ad too much yet ([Meta help](https://www.facebook.com/business/help/1346816142327858)).

For the Meta system around this decision, read [Facebook ads for B2B](/guides/facebook-ads-for-b2b). For the container choice on LinkedIn, read [LinkedIn ad formats](/guides/linkedin-ad-formats). When cheap results turn out to be the wrong people, read [low quality leads from LinkedIn and Meta ads](/guides/low-quality-leads-linkedin-meta-ads).

Write down why you paused. Without a note, the team tests the same idea again in three months. This card takes two minutes and works as one line in your [learnings log](/guides/share-creative-learnings).

```template
PAUSE NOTE
Date: [date]
Ad: [ad name]
Angle: [angle name]
Platform: [LinkedIn or Meta]
Window: [first date] to [last date]
Goal result: [what the ad set was bought for]
Numbers: the ad spent [amount] for [count] results. Its peers spent [amount] for [count] results. Chance of luck: [number].
Worse than the peers in both halves: [yes or no]
Weak from the start, or a slide: [weak or slide]
Changes inside the window: [none, or the list with dates]
Cause we believe: Because [what we saw across the ads], we believe [cause].
Angle verdict: [the angle failed, or only this execution failed, or too early to say]
Next: [the new concept or execution in one line], judged on [goal result] from [start date]
We were wrong if: [what we would see that proves the cause wrong]
```

Nimblet does not detect fatigue and it never pauses an ad. It reads your LinkedIn and Meta accounts, ranks each ad against ads bought for the same goal, and says when too little data stands behind a rank. See how it works on the [creative intelligence page](/features/creative-intelligence).

## Nimblet runs this check

Nimblet ranks each ad against ads bought for the same goal, and it says when too little data stands behind a rank. Ads with recent spend also get an AI read that records core message, hook, offer and proof, so you can see whether weak ads share an angle. Nimblet only reads your LinkedIn and Meta accounts and never changes a campaign, so you pause the ad yourself in the platform.

## Frequently asked questions

### How many days should an ad run before I pause it?

Count results, not days. Work out E: the results the ad would have at its peers' cost. If E is below 4, wait. If E is 20 or more, one window can already give a clear answer. For reference, Meta says an ad set usually leaves the learning phase after about 50 results in the week after its last significant edit ([Meta help](https://www.facebook.com/business/help/112167992830700)). LinkedIn's A/B tool asks for at least 14 days ([LinkedIn help](https://www.linkedin.com/help/lms/answer/a525922)).

### How far below average should an ad be before I pause it?

It depends on the result count. With 25 expected results, an ad must fall 40 percent short of its peers. With 100, 20 percent is enough. With 5, only zero results counts. The cut-off table above shows why no single percent fits every account.

### Should I pause an ad that has spent money and has no results?

Yes, once its spend reaches about 4 times the peers' cost per result, step 1 finds no change and the form and the tracking work. At that point an ad as good as its peers would show zero results by luck only 1.8 percent of the time (e to the power of minus 4, where e is about 2.718). Below that, zero results says little.

### Does pausing one ad hurt the others?

The other ads take over the delivery, so results can shift. For campaigns with Advantage+ campaign budget, Meta says frequent pauses throughout the day can interfere with delivery optimization ([Meta help](https://www.facebook.com/business/help/388669258876309)). Pause one or two ads at a time, so you can see what changed.

### What if all my ads are below target?

Then look first at what all the ads share, because one weak ad cannot explain it. In our reading that is the target itself, the page, the form, the tracking, the audience settings, the offer or the market. Run the rule-outs above before you touch an ad. If the numbers look cheap but the people are wrong, read the guide on [low quality leads](/guides/low-quality-leads-linkedin-meta-ads).

### Is a falling click-through rate enough to call an ad fatigued?

No. LinkedIn says the click-through rate often declines when the same ad runs week after week, so a decline can be normal ([LinkedIn help](https://www.linkedin.com/help/lms/answer/a420488)). Check the goal results of the ad against its peers, and run the weekly index.

### How often should I refresh ads?

LinkedIn advises at least once a month with new text or images ([LinkedIn help](https://www.linkedin.com/help/lms/answer/a420488)). Ad tool blogs give ranges from 2 to 8 weeks for LinkedIn. The ones we read call their numbers a starting point or a default, and none names a dataset behind them. We give no calendar. The decision tree reacts to your own numbers.

### What frequency is too high?

The help pages we read from LinkedIn and Meta give no safe limit. Watch the trend. Frequency that climbs while results per euro fall against the peers is the warning. Meta says to watch frequency together with results ([Meta help](https://en-gb.facebook.com/business/help/1546570362238584)).