When to pause an ad set

Admaxxer is a DTC analytics platform with built-in Meta + Google ad ops. Pausing ad sets too early wastes learning-phase spend; pausing too late burns budget on losers. The short answer: the math says you need ~40 conversions to detect a 10% CPA difference with statistical confidence, so pause rules should gate on spend multiples of target CPA, not arbitrary dollar amounts. ## TL;DR - Rule 1: Spend > 3× target CPA with zero conversions → investigate or pause - Rule 2: 40+ conversions at 1.5× target CPA → statistically confirmed loser - Rule 3: CPM spike >40% day-over-day with no creative refresh → platform reset signal - Audience overlap >20% with a proven winner → pause to preserve learning - Use Admaxxer's [Claude agent](/docs/ai-agent) to run the rules across all ad sets daily ## Three pause thresholds ### Threshold 1: Zero conversions at 3× target CPA If your target CPA is $40 and an ad set has spent $120 with zero purchases, something is broken. Either the creative doesn't resonate, the audience is wrong, or the landing page is broken. Investigate before pausing — spend that returns zero is diagnostic gold. Check [High spend zero purchases](/guides/high-spend-zero-purchases) for the nine most common causes. ### Threshold 2: 40+ conversions at 1.5× target CPA This is the statistically confirmed loser threshold. With 40 conversions, you have enough power to detect a 50% CPA difference at 95% confidence. If the observed CPA is 1.5× target with that sample size, you can be confident the ad set is genuinely underperforming, not just unlucky. Pause and reallocate budget. ### Threshold 3: CPM spike without creative refresh A sudden 40%+ day-over-day CPM increase without a creative change usually signals one of: - Meta algorithm reset (happens periodically, often after major platform updates) - Audience saturation (you've reached everyone; frequency is climbing) - Bid competition spike (a competitor launched; unfixable from your side) The right response depends on which cause. Audience saturation = refresh creative. Algorithm reset = wait 48 hours. Bid competition = accept the new floor or move budget. ## Diagnostic steps ### Step 1: Pull the ad set's learning phase status Meta's learning phase requires 50 optimisation events (usually purchases) in 7 days. An ad set in learning mode should not be paused based on short-term CPA — the algorithm is still training. Let it reach 50 conversions or 7 days before judging. ### Step 2: Check spend vs. target CPA multiple In Admaxxer's [campaign view](/docs/campaigns), each ad set shows `spend_vs_target_cpa_multiple`. Values: - <2×: keep running (not enough data) - 2–3× with 0 conversions: investigate - >3× with 0 conversions: pause or fix - >1.5× target CPA with 40+ conversions: pause ### Step 3: Compute statistical significance if you're second-guessing With n conversions at observed CPA c vs. target t: - Standard error SE ≈ c × sqrt(1/n + 1/reference_n) - z-score = (c - t) / SE - |z| > 1.96 → significant difference at 95% This is simpler than the two-proportion test because you're comparing means, not proportions. ### Step 4: Check audience overlap with winners Meta's Audience Overlap tool shows if your ad set is competing with a proven winner for the same impressions. Overlap >20% means you're self-cannibalising. Pause the weaker one. ### Step 5: Use the Admaxxer Claude agent for nightly sweeps Ask the [Claude agent](/docs/ai-agent) via `pause_all_low_roas` to surface ad sets breaching these thresholds. The tool requires `confirmed: true` before executing, so you review before pausing. ## Statistical significance math The core math is about sample size and variance. For a target CPA of $40 with typical DTC conversion volatility (CV ≈ 1), you need: - 10 conversions to detect a 50% difference (large effect) - 40 conversions to detect a 25% difference - 160 conversions to detect a 10% difference Most DTC teams pause too early because they don't realise how much sample they need. 10 conversions is noise; 40 is marginal; 100+ is solid. If you pause an ad set after 15 conversions because CPA looks high, you're 40% likely to be wrong (it's actually average, not underperforming). The implication: don't pause below 40 conversions unless Threshold 1 (zero conversions at 3× spend) applies. Let winners and losers separate naturally. ## Common mistakes Three patterns that kill performance: 1. **Pausing during learning phase.** Meta's optimiser needs 50 events to stabilise. Pausing at 20 events means you never know if the ad set was viable; you just bought 20 expensive data points. 2. **Pausing on day-of-week noise.** A Tuesday spike in CPA on an ad set that was fine Mon/Wed/Thu is random variance, not a signal. Always aggregate to rolling 3-day or 7-day. 3. **Pausing without checking audience overlap.** If you pause the wrong ad set, you'll see the remaining ad set's CPM spike as it picks up the freed impressions — because they were competing. Always check overlap first. See [Blended MER vs ROAS](/guides/blended-mer-vs-roas) for why ad-set-level CPA should be cross-checked against account-level MER before making dramatic changes. ## FAQs **Should I ever pause a winner?** Only if creative fatigue sets in (CTR drops >30% from peak while frequency climbs past 4). Winners rarely break spontaneously. **What about a "3-day CPA" rule?** 3-day rolling CPA is better than 1-day. It smooths day-of-week noise. Use it as the basis for threshold checks, not single-day spikes. **Does this apply to Google Ads?** Mostly yes — the math is the same. Google's thresholds are slightly different because Search traffic is higher-intent and converges faster (20–30 conversions can be enough for Search, vs 40 for Meta prospecting). **Can the Claude agent pause ad sets?** Yes — `pause_all_low_roas` with `confirmed: true` will pause any ad set matching the 3× spend / zero conversion rule or the 40+ conversions / 1.5× CPA rule. You review before confirming. **What about TikTok ad sets?** Same thresholds apply; TikTok's learning phase requires 50 conversions in 7 days. The Admaxxer TikTok connector is on the roadmap for v1.5.