Frequency hits 3.4 and CPA climbs 40% in two weeks. I still would not call that creative fatigue. Rising frequency is a reason to investigate, not a diagnosis, and a red square on a heatmap does not change that. Before paying for another subscription or replacing an ad, I want Claude to separate four things: weaker creative response, higher auction costs, a failing post-click path, and data too thin to judge. Run these nine checks on the real numbers before spending money on the presumed cure.

Set up Claude before running the checks

  • Choose a data route you can repeat: paste an Ads Manager export into Claude, or connect Claude to groas over MCP at https://mcp.groas.ai/mcp for live access.
  • Keep the price comparison straight: groas Unlimited is $19 a month, flat, for a whole team, with unlimited teammates, ad accounts, and websites, not a per-seat fatigue module.
  • Start with analysis, not execution: Today is [date]. Analyze only. Do not pause, duplicate, or edit anything. Account: [ad account ID]. Ad: [ad ID/name]. Conversion event: [purchase / lead]. Window A: [newer completed period]. Window B: [equally long preceding period]. Data source: [CSV export dated X / live groas connection]. Tell me if either window is too recent to judge.
  • Treat that instruction as an approval boundary, not proof of technical read-only access; require Claude to wait before making changes through a connected account.
  • Stop if the account is not ready: an account launched four days ago, a window interrupted by a budget, bid, or landing-page change, or roughly $2k a month producing two purchases a week does not give this comparison much to work with.
  • Save the prompts for repeat use; one screenshot is not a diagnosis, and repeating an immature comparison does not make it mature.

Chart showing rising frequency alongside a flat response line under a magnifying glass

1. Confirm exactly which ad and event Claude is judging

  • Ask Claude: List the ad account ID, campaign, ad set, ad IDs, and exact conversion event you are using. Confirm that no other ad or event is mixed into the comparison.
  • Check IDs, not just names: two ads sharing a name are still two ads, and a purchase comparison is not a lead comparison.
  • Require Claude to repeat the two reporting windows before interpreting any movement; Window A is the newer period throughout this checklist.
  • Keep the same conversion event in both windows so that a reporting change does not masquerade as a performance change.
  • Stop if Claude cannot identify the account, ad, event, and periods cleanly. Everything after that would be arithmetic applied to fog.

2. Compare equally mature windows

  • Ask Claude: Compare Window A with Window B using the same conversion event and equal-length periods. Flag days that have not completed the applicable attribution window as provisional.
  • Check the attribution setting before judging CPA or ROAS: with a 7-day click window, recent clicks can still produce purchases that appear later in the report.
  • Allow for delayed conversion reporting; a report pulled on the 3rd can miss purchases later attributed to the 1st.
  • For a conservative purchase comparison, use the last completed 7-day period ending at least 7 days ago against the immediately preceding 7-day period.
  • For lower volume, use 14 days against the prior 14 days, again leaving enough time for the newer window to mature.
  • Read current-week impressions and clicks as delivery signals, but do not treat current-week CPA and ROAS as settled results.
  • Keep this rule simple: compare finished windows with finished windows, not a complete past against an incomplete present.

3. Refuse a verdict when delivery is too thin

  • Ask Claude: Report spend, impressions, clicks, and conversions per window. Return insufficient evidence if either window has fewer than 1,000 impressions or 10 conversions.
  • Treat those numbers as this checklist’s screening floor, not a guarantee that everything above them supports a confident diagnosis.
  • Accept “insufficient evidence” as a useful answer: three bad days on a $30-a-day ad can move an average without establishing that the creative has worn out.
  • Keep impressions, clicks, and conversions visible beside the percentage changes; a dramatic percentage without its underlying counts is an invitation to overreact.
  • Do not buy a fatigue score to escape a volume problem; scoring the same thin data does not supply the missing evidence.
  • Wait for mature windows and enough delivery before running the full separation again. A longer checklist cannot manufacture conversions.

4. Read frequency beside response, never alone

  • Ask Claude: Show impressions, reach, frequency, outbound CTR, and cost per result side by side for both windows. Describe what changed without diagnosing fatigue from frequency alone.
  • Keep frequency in its proper job: it measures repetition, not whether the repetition has stopped working.
  • Use the ad’s own response history rather than a universal cutoff; frequency alone cannot establish fatigue.
  • Read rising frequency with stable outbound CTR and stable cost per result as repetition without demonstrated deterioration, not an automatic instruction to refresh.
  • Treat rising frequency plus falling response on the same ad as a reason to continue the checks, not permission to skip them.
  • Allow for context: a retargeting pool can remain profitable at a frequency of 6 while a cold audience weakens at 2.
  • Keep why frequency alone cannot prove fatigue beside the report if the urge to pause at 3.0 keeps returning.

5. Separate auction costs from creative response

  • Ask Claude: Show CPM, outbound CTR, and CPC for both windows. Explain whether the increase in click costs came from higher CPM, lower CTR, or both before interpreting CPA.
  • Read CPM before blaming the creative: if CPM rises 32% while outbound CTR holds at 1.1%, the auction has become more expensive without a corresponding decline in that response measure.
  • Do not turn that example into a complete CPA explanation; carry the result into the post-click check before deciding what caused the final cost increase.
  • If outbound CTR falls as well, keep both explanations in play rather than forcing every movement into one fatigue label.
  • Require Claude to show the numbers behind its explanation, not simply repeat that performance has deteriorated.
  • Keep the distinction plain: CPA is the scoreboard; CPM and CTR help explain how the cost changed. Neither substitutes for the rest of the diagnosis.

6. Check the post-click path before commissioning an ad

  • Ask Claude: Show landing-page conversion rate, add-to-cart or lead rate, and cost per landing-page view for both windows where available. Flag any known tracking or page changes, and identify missing data.
  • Require a written post-click check: a stable click-through rate does not explain a conversion rate falling from 3.2% to 1.8%.
  • Treat that pattern as a reason to inspect the page, offer, or tracking before calling the creative tired; it points beyond the click rather than proving a single cause.
  • Check for a landing-page change inside the comparison period, including a slower template, before accepting the windows as comparable.
  • Do not let Claude invent a page or tracking failure from a rate change; require it to distinguish an observed drop from a confirmed explanation.
  • Keep creative replacement off the automatic-fix list while a broken checkout remains a plausible explanation. A new hook does not repair checkout.

7. Inspect placement and audience shifts

  • Ask Claude: Break down spend share, CPM, and CTR by placement and available age or audience segments for both windows. Include reach trend and first-time impression ratio if available; identify anything unavailable.
  • Check the delivery mix before interpreting the blended CTR: Feed falling from 70% of spend to 35% while Audience Network takes the balance changes what the account average represents.
  • Treat a placement shift as an alternative explanation to investigate, not proof that the creative stayed healthy in every placement.
  • Compare response within placements as well as across the account so Claude does not mistake a change in venue for a change in the same audience’s reaction.
  • Look for reach flattening, first-time impressions falling, and multiple ads weakening together; that pattern points toward audience saturation rather than one tired ad.
  • Match the proposed action to the problem: tired creative calls for materially different creative; saturation points toward a broader or new audience, or less budget.

Cutaway diagram showing Meta delivery shifting between placements while the creative stays unchanged

8. Compare sibling ads, then price the next step

  • Ask Claude: Rank ads running to the same offer in the same ad set or campaign by change in outbound CTR and cost per result from Window B to Window A. Show spend beside each result and flag unequal delivery.
  • Look for an isolated decline: one weakening ad among healthy siblings makes creative fatigue more plausible; every ad weakening together points toward a shared problem worth investigating.
  • Keep offer, page, tracking, auction conditions, and saturation in that shared-problem list rather than assuming that every creative expired simultaneously.
  • Treat the sibling table as diagnostic evidence, not a controlled test; delivery rotation, budget shifts, and learning changes can muddy the comparison.
  • Do not describe unequal delivery as an equal-spend comparison just because Claude puts the ads in adjacent rows.

Use free scores as calculators, not judges

  • Run a checker after the preceding checks, not instead of them; a score does not rule out auction costs or a post-click problem.
  • AdRiseLab’s free checker takes five Ads Manager numbers and weights CTR decay at up to 60 points, with days active and frequency at 20 each; it also warns against pausing on the score alone.
  • The day-by-day CSV checker flags the first day the 3-day CTR moving average falls 20% below the first-3-day baseline while frequency exceeds baseline.
  • Note what that second rule excludes: a CTR drop with flat frequency is deliberately not flagged as fatigue, which is a scoring choice rather than a complete account diagnosis.
  • Use either output to organize a review, not to overrule stronger evidence from CPM, conversion rate, placement mix, or sibling ads.

Compare monitoring costs before adding another subscription

  • Price the diagnosis you need, not just the dashboard: GoodMorning lists $500 per month per ad account, with a fatigue heatmap in every plan.
  • Its frequency-based heatmap coloring does not, by itself, provide the separation this checklist requires; a red frequency cell still needs investigation.
  • For this workflow, Claude with groas at $19 a month flat is the stronger-value route than adding a per-account monitor: use the connected data to work through the competing explanations in plain English.
  • Keep using an Ads Manager export if another subscription is not justified; buying access and reaching a sound verdict are separate decisions.

Define a meaningful replacement before approving it

  • If fatigue remains plausible, propose materially different creative: the same offer with a new hook, new first three seconds, or a different proof point, not merely a new background color.
  • Run the proposed creative beside the suspect ad and compare CTR and cost per result while accounting for unequal spend; do not call ordinary delivery a controlled experiment.
  • Name the problem the replacement is supposed to solve: at a hypothetical $20k a month in spend, refreshing on frequency alone still leaves an auction-cost problem untouched.
  • Keep the proposed creative change to one next action so the final approval has something specific to approve or reject.

9. Write the verdict before spending or changing anything

  • Require the written verdict: Give one verdict: fatigue plausible, another explanation stronger, or insufficient evidence. Cite the numbers supporting it, name the strongest competing explanation, propose exactly one next action, and wait for approval.
  • Accept fatigue plausible only as the conclusion the evidence supports, not as a claim that Claude has proved the cause beyond doubt.
  • If another explanation is stronger, address that explanation rather than commissioning creative because a frequency number looks uncomfortable.
  • If evidence is insufficient, keep the gap explicit; a confident paragraph is not a substitute for missing delivery or immature conversions.
  • When the call is close, use a two-assistant check before the next ad change and keep the stricter verdict; disagreement is useful data, not permission to choose the more convenient answer.
  • Set the approval instruction: Propose one change, the spend it affects, and how long to wait before judging it. Make no change until I reply approved. Then log what changed and why so it can be undone.
  • Save the verdict with the report and prompts so the next rushed review starts with evidence rather than another red square.
  • Do not skip the written verdict: this is the item most often missed, and the cost is replacement creative, a possible learning reset, another monitoring invoice, and the original problem still sitting in CPM, placement mix, or the page. Rising frequency starts the review. The written verdict ends it.