# Losing Ad Diagnosis Ad Creative > Losing Ad Diagnosis - Skill: `mardab96/losing-ad-diagnosis-ad-creative` (Agent Skill) - Install (CLI): `npx skillmds@latest add mardab96/losing-ad-diagnosis-ad-creative` - Raw SKILL.md: https://api.skillmd.com/api/skills/mardab96/losing-ad-diagnosis-ad-creative/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: mardab96 (https://skillmd.com/u/mardab96) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/mardab96/losing-ad-diagnosis-ad-creative --- # Losing Ad Diagnosis Shared quality bar: `../references/output-standard.md`. All numbers cited here live in `../references/thresholds.md`. A weak ad is usually rewritten at the wrong layer. The stage where people stop tells you which layer to touch, and rewriting any other layer is expensive guessing. ## Use this skill when One named ad is underperforming its ad set and the proposed action is a rewrite, a new image or a new offer. It applies to an ad that never worked. An ad that worked and slid belongs to the fatigue read. "The clicks do not convert" is this skill's trigger, not `ad-to-landing-page-match-ad-creative`'s. That one compares one ad against one destination and reads no impressions, clicks, sessions or conversions, so it cannot establish that the loss sits on the page rather than at the hook or at the click to session step. This skill locates the step first; where the located step is session to conversion, it hands the destination to that skill, which is the point at which a page check is worth doing. ## Required input - For that ad: impressions, link clicks, landing page sessions or views, conversions, spend, over one stated window. - The same four figures for the ad set as a whole, or a median across its live ads, as the comparison. - The destination URL and what it promises above the fold. - Audience or targeting definition, and whether placements are automatic. - Whether the ad carries the same offer as the ads it is being compared with. - Which Advantage+ creative enhancements were on during the window and whether any changed inside it (`test.creative_enhancements_rule`). With image expansion or text variation running, the asset that served differs per impression, so the ad being diagnosed is not exactly the one that was uploaded. ## Analysis workflow 1. State the window as dates and confirm the ad has no creative edit inside it. 2. Check the ad against `delivery.never_delivered_floor`. An ad under it did not lose, it did not run. Mark it `never_delivered`, stop, and say the finding is allocation rather than creative. Then check it against `delivery.starved_flag`: an ad that cleared exposure and still holds a thin slice of its ad set's spend has been read by delivery and repeatedly passed over, which is a verdict on the creative rather than an absence of one. That ad may still be diagnosed here, but every funnel row on it carries the starved state, because its counts came from whatever narrow slice of the audience delivery kept giving it. Where its share is also under `delivery.ranking_floor`, say that the ad set median it is being compared against is not a like-for-like comparison. 3. Build the funnel as four counts in order: impressions, link clicks, landing sessions, conversions. Missing any one of these makes the stage below it unreadable, and that is reported rather than modelled. 4. Compute the rate at each step and put each against the ad set comparison. Confirm the click figure is link CTR as defined in the denominators table, not all clicks. 5. Find the first step where this ad falls materially behind the comparison. Judge a rate step with `test.material_gap`, and say in the output that the key is being used outside its home. Its home is the relative difference between variants inside one test; here it is being applied to one ad against an ad set median, which is not a test and was never randomised. The reuse is defensible on one ground and only one: that entry is a business relevance floor, a difference too small to act on whatever produced it, and that half travels. The statistical half does not travel, so no confidence is drawn from clearing it. Where the located step is a cost question rather than a rate one, cost per outbound click or cost per acquisition against the ad set median, use `money.relevance_floor` instead, which is written for exactly that comparison. 6. Read the located step. A shortfall from impression to click is a hook, a visual or a first-line problem. A shortfall from click to session is delivery: load time, redirect, or broken tracking, not creative. A shortfall from session to conversion is offer or message match, and routes to the landing page match check rather than to a new headline. 7. Check the volume behind the located step against `test.click_floor` and `test.volume_floor`. Under either, the diagnosis is a `hypothesis` and carries `low` confidence, whatever the gap looks like. 8. Where any part of the read compares this window's conversions against a prior window, either wait out the account's attribution window before reading it or state `reporting.conversion_lag_bias` in the same sentence as the number. The current window is still filling in while the prior one has finished, so the comparison is biased against the present by construction, and it will read as a decline that is not there. Click and impression comparisons do not carry this and may be read as they stand. 9. Check the destination against `match.above_fold_rule` before blaming the ad's copy. A promise the page does not repeat produces a session to conversion collapse that looks like a bad offer. 10. Count the distinct live angles in the ad set against `coverage.angles_per_ad_set`, applying `coverage.angle_distinctness` first. Below the floor a weak angle and a weak execution cannot be told apart, so the whole diagnosis is capped at `low` confidence and the output says which count capped it. 11. Name the one change to make, at the located layer only, and name what would confirm the diagnosis was right. ## Decision rules - Impression to click behind by more than `test.material_gap`, that key read as a relevance floor outside its home as the workflow states: `rewrite` the hook, one element at a time under `test.one_change_rule`. - The located step is a cost step and the gap is inside `money.relevance_floor`: no rewrite. A difference that does not move cost per acquisition by that much is not worth production time, whatever the rate table shows. - Ad under `delivery.never_delivered_floor`: `investigate` allocation, issue no creative verdict, and keep the ad out of any cut list. It never had the delivery to fail with. - Ad flagged by `delivery.starved_flag`: diagnose it, but say in every row that the counts come from a slice of delivery the platform kept shrinking, and cap the diagnosis at `low`. The starved state is itself a finding and belongs in the output beside the funnel step. - Distinct live angles in the ad set below `coverage.angles_per_ad_set`: every row in this diagnosis drops to `low` confidence, including rows that would otherwise qualify, because the angle and the execution are confounded. Say the coverage count is what capped it. - A conversion figure compared against a prior period without the attribution window having closed: do not report the change as a change. Either wait, or carry `reporting.conversion_lag_bias` in the same line as the number. - Click to session loss crossing `funnel.click_session_loss_flag`: `investigate` delivery and tracking. Do not touch the creative. - Session to conversion behind while the click rate is at or above the comparison: `investigate` the destination, and route to the ad to landing page match check. - Volume under `test.click_floor` at the located step: `monitor`, and state how much longer at current spend makes it readable. - An Advantage+ enhancement changed inside the window, or the change dates are unknown: cap the whole diagnosis at `low` under `test.creative_enhancements_rule` and say the funnel counts span two different served assets. The located step is still worth reporting; the rewrite instruction that follows from it is not, until the account can say which asset produced the counts. - Restricted category in `claims.restricted_topics`: `approval_needed`, and route to the platform's own policy page rather than issuing a verdict. ## Output format Open with one line naming the stage where this ad loses people. | Funnel step | This ad | Comparison | Relative gap | Material | Reads as | Evidence | Decision | Confidence | |---|---|---|---|---|---|---|---|---| Then the one change to make, then `What this could not see`, `Missing data`, `Approval gates`. ## Practical example Illustrative made-up account data, not a real advertiser. Input: one ad, 96,000 impressions, 118 link clicks, 96 landing sessions, 1 conversion. Ad set median link CTR 0.41 per cent. No landing page benchmark supplied for the other ads. This ad took 1,180 of the ad set's 4,300 spend in the window, which is 27 per cent, clear of `delivery.never_delivered_floor` and clear of `delivery.starved_flag`, so it genuinely ran and may be diagnosed. The ad set holds 4 distinct live angles once `coverage.angle_distinctness` is applied, above `coverage.angles_per_ad_set`, so the coverage cap does not fire and the confidence label below comes from volume alone. Reading: this ad's link CTR is 118 over 96,000, which is 0.12 per cent. Against 0.41 per cent, that is a relative shortfall of about 70 per cent, far past `test.material_gap`, and it is the first step where the ad falls behind. Stated in the output: that key is being read here as a relevance floor against an ad set median rather than inside a test, so it says the gap is worth acting on and says nothing about whether it is statistically there. Click to session loses 22 of 118, which is normal and not the finding. Declined: the session to conversion step is not diagnosed. One conversion on 96 sessions is under `test.volume_floor` by a wide margin, and no comparison rate was supplied for the other ads, so that row is `needs_data` and the offer is neither cleared nor blamed. Calling the offer weak here would be an assumption dressed as a funnel reading. Output: "This ad loses people at the first step, impression to click." The hook row is `export` with decision `rewrite`, but confidence is `low`, because 118 clicks sits under `test.click_floor` and the diagnosis is therefore a `hypothesis` about the hook rather than a measured fact. Confirmation: one hook variant against this ad as control, read once it clears the click floor. ## Guardrails - Never rewrite a layer the funnel did not implicate. - Never diagnose a step whose count was not supplied. - Never treat a click to session gap as a creative defect. - Never call an offer weak on volume under `test.volume_floor`. - Do not edit or pause the ad. Recommend, and let a human approve.