markifact/markifact-mcp/diagnose-underperformer

Goal

Pull the right report, walk a structured decision tree, and hand the user a ranked list of likely causes with the specific next op to call to fix it. No writes — diagnosis only.

Inputs to confirm (batch)

1. Account name (substring) — required.
2. Target — campaign / ad set / ad name or ID — required.
3. Date range — default last 14 days; if the entity is < 14 days old, use its lifetime.
4. Comparison window (optional) — previous 14 days, for trend.
5. Primary KPI — conversions / ROAS / CPA / CPL — required if the entity has a conversion goal.

Workflow

1. Discover ops. For Google Ads: gads_select_accounts, gads_list_report_fields, gads_get_report, gads_get_account_history. For Meta: meta_ads_select_accounts, meta_ads_list_report_fields, meta_ads_get_report, meta_ads_get_account_history, meta_ads_get_adset_settings, meta_ads_get_campaign_settings.
2. Inspect with get_operation_inputs. Always call `*_list_report_fields` before `*_get_report` — never guess metric/dimension names.
3. Resolve account with substring match.
4. Pull the report at the right granularity (campaign / ad set / ad). Required metrics:

• Spend, impressions, clicks, CTR, CPC, conversions, conv rate, CPA / ROAS, frequency (Meta), search impression share + lost-to-budget + lost-to-rank (Google Search).

1. Pull settings/history to check for recent edits that could have triggered the dip:

• Google: gads_get_account_history filtered to this campaign — look for budget changes, bid strategy switches, status flips, asset changes.
• Meta: meta_ads_get_campaign_settings / meta_ads_get_adset_settings for current state, meta_ads_get_account_history for recent changes.

1. Walk the decision tree below and rank the matching diagnoses by likelihood + impact.
2. Present: top 3 diagnoses, evidence from data, and the exact next op to call (or the matching slash command). Do not run any write op.

Decision tree

Branch A — Low impressions / low spend (entity not delivering)

Branch B — Spending normally, low CTR

Branch C — Good CTR, low conversion rate

Branch D — Good CVR, low overall conversions

Branch E — Sudden performance drop

Platform-specific rules the model MUST respect

• `list_report_fields` first. Field names differ per platform; never guess.
• Account-name substring match for select_accounts.
• Frequency is a Meta concept — not directly comparable on Google.
• Search impression share is a Google Search-specific metric; PMax has its own (search-categories report).
• Don't propose changes that need to wait. If the entity is < 7 days old or made a major change in the last 7 days, say "still in learning, don't touch yet."
• Read-only. This command never calls a write op. Recommend the right write command instead.

Output to user

Top diagnoses for <entity>:
 
1. <Diagnosis> — confidence: high/medium
   Evidence: <numbers>
   Next: run /markifact:<command> OR call <op_id>
 
2. ...
 
3. ...
 
Other observations:
- ...

End with: "Want me to do any of these? I won't make changes without your go-ahead."