Account health diagnostics

Account Health Diagnostics

Two modes for one job: figuring out what's wrong with a Google Ads account.

Mode selection

Mode	Question	Data needed	Use when
Spike Diagnosis	Why did my CPA jump?	Two comparison periods of campaign data	The user knows something broke and roughly when
Drift Scan	Is anything quietly going wrong?	60+ days of DAILY campaign data	No known problem; health check, monitoring, "what looks weird"

Selection heuristics:

• User names a specific symptom with a timeframe ("CPA went from $40 to $65 this month") → Spike Diagnosis
• User wants a general scan, alert-style review, or asks "what changed" with no specific symptom → Drift Scan
• Drift Scan finds a serious anomaly → offer to run Spike Diagnosis on it as the follow-up. This chain is the ideal workflow: Scan finds it, Diagnosis explains it.

State which mode you're running at the top of the response.

Shared cause library (both modes reference this)

CPA problems have six well-understood causes. Spike Diagnosis checks all six formally; Drift Scan uses them as hypotheses when cross-referencing anomalies.

1. Bid strategy changes: automated bidding shifted targets or learning reset. Signature: CPC rose sharply with stable conversion rate. Learning period spikes are normal for 7-14 days after a strategy switch.
2. Search term drift: new irrelevant queries entered the auction. Signature: new high-spend zero-conversion terms in the recent period, CTR often drops. (Fixing it is the negative-keyword-engine skill's job.)
3. Quality Score drops: lower QS → higher CPC → higher CPA. In play if campaign avg QS dropped 1+ points. NEVER claim this cause without QS data; it's the most common scapegoat and often wrong.
4. Competitor activity: "IS lost to rank" up, absolute top impression share down. Auction insights data nails this.
5. Landing page issues: conversion rate dropped while CTR held steady. People click but don't convert. Check by device; mobile conv. rate drops often mean a speed or layout regression.
6. Audience shifts: demographic, geographic, or device mix changed. New geos receiving spend, or mobile share growing in a desktop-heavy account.

Cross-reference signatures (used by both modes):

• CPA spike + CPC spike + IS lost to rank up → competitor entry (cause 4)
• CPA spike + conv. rate drop + CTR stable → landing page (cause 5)
• CPA spike + CTR drop + new search terms → search term drift (cause 2)
• Spend spike with no conversion spike → bid strategy or budget (cause 1)
• CPC up + conv. rate stable + no IS change → bidding (cause 1)

These are hypotheses to surface, not proof. Always use "possible cause" framing.

---

MODE 1: Spike Diagnosis

Required inputs (ask before proceeding if missing)

1. The spike quantified: "CPA went from $X to $Y over [period]." Don't accept "CPA is bad."
2. A baseline period: typically the 30 days before the spike. No baseline, no diagnosis.
3. Campaign data for both periods with the columns below.
4. Known recent changes: new campaigns, bid strategy switches, creative updates, budget shifts, landing page edits. Ask; even small changes can be the cause.

If the user just said "my CPA spiked," reply with one message:

Before I diagnose, I need:

1. The spike numbers: CPA before and after, over what dates?
2. Campaign performance data for both periods: CSV/Excel with spend, conversions, CPA, impression share, and Quality Score columns (or live via your Google Ads MCP)
3. Any recent changes: new campaigns, bidding, creative, budgets, landing pages?

Required data columns (or close equivalents)

Per campaign, per period: Campaign name; Spend; Impressions, Clicks, CTR; Conversions, CPA, Conv. rate. For cause 4: impression share (overall, lost to budget, lost to rank) and top/absolute-top %. For cause 3: average Quality Score. Search terms and keyword-level QS are bonus.

If a cause's data is missing, say so explicitly ("I can't evaluate competitor activity without impression share columns; re-export with those") rather than extrapolating.

Workflow

Step 1: Compare periods. For each campaign compute deltas: ΔCPA, ΔSpend, ΔConv, ΔConv. rate, ΔCTR, ΔCPC, ΔIS. Sort by absolute spend impact (ΔSpend × ΔCPA / baseline CPA) so the biggest offenders surface first.

Step 2: Run the six checks from the shared cause library against the deltas. Note: bid strategy and competitor causes often produce smaller, harder-to-spot signals than search term drift but are more often the real culprit. Don't blame the most dramatic-looking metric by default.

Step 3: Ranked diagnosis output. For each cause with evidence:

Field	Content
Severity	Critical / High / Medium / Low
Cause	Plain-language description
Evidence	Specific numbers from the data
Specific fix	Concrete steps
Expected CPA impact	Range, e.g. "5-15% reduction"
Timeline	How long until results show

Severity rubric: Critical = explains most of the spike and is actively bleeding budget; High = major contributor with clear evidence; Medium = contributing factor or strong hypothesis needing one more data point; Low = present but minor.

Sort by severity, then expected impact. End with a most likely cause verdict (one cause, or two if clearly multi-factor) and the first action to take this week.

Spike Diagnosis must NOT

• Diagnose without a baseline.
• Recommend pausing campaigns reflexively; most spikes are fixable without nuking volume.
• Claim QS is the cause without QS data.
• Extrapolate causes from missing data.

---

MODE 2: Drift Scan

Required input

Daily campaign metrics for the last 60 days minimum. Weekly data is too coarse for statistical baselines. If only weekly/monthly data is available:

"Anomaly detection needs daily-level data for at least 60 days. Weekly data is too coarse for statistical baselines. Please re-export at daily granularity."

Columns per campaign x day: Date, Campaign name, Spend, Impressions, Clicks, CTR, CPC, Conversions, CPA, Conv. rate, Search impression share.

Why this catches what monitoring misses

A 5% daily change feels small, but 5% drift sustained over 30 days compounds and is often invisible until it has bled budget for 7-14 days. Statistical baselines catch what eyeballing doesn't.

Workflow

Step 1: Build the baseline. For each campaign x metric, use days 1-30 (the OLDER half) as baseline: compute mean and standard deviation. Days 31-60 (recent) are what you're testing. The split prevents recent anomalies from contaminating the baseline. Filter out zero-spend and paused-day rows first; they distort SDs. Skip campaigns with <14 days of baseline data.

Step 2: Flag spike anomalies. For each recent day x campaign x metric: z = (value - baseline_mean) / baseline_std.

Z-score	Severity
> 3 SD	Critical
> 2 SD	Warning
> 1.5 SD AND 3+ consecutive days trending	Watch

Only flag the LAST 7 days for spike alerts; older spikes are historical, not actionable. Where data allows, use day-of-week-adjusted baselines: Sunday CPA legitimately differs from Tuesday in many B2B accounts. Don't flag known weekly cycles.

Step 3: Detect trend drift. Compare recent-14-day avg to prior-14-day avg within the recent 30: drift = (recent14 - prior14) / prior14 x 100. Flag if |drift| > 5%. Classify direction: drift UP is bad for CPA, CPC, IS lost to budget/rank, and spend (if budget-flat); drift DOWN is bad for conversions, conv. rate, CTR, revenue, ROAS, and IS.

Step 4: Cross-reference. For each anomaly, check co-moving metrics on the same days against the shared cause library signatures. Surface hypotheses, not verdicts.

Step 5: Alert dashboard output.

# Anomaly Detection Report: [Date Range]

## Critical (>3 SD)
- **[Campaign]**: [Metric] [direction] [magnitude] on [date(s)]
  - Current: [value], baseline avg: [value], deviation: [Z] SD
  - Possible cause: [hypothesis from the cause library]
  - Recommended action: [specific step]

## Warning (>2 SD)
[same structure]

## Watch (sustained drift)
[same structure]

## All Clear
- Campaigns within normal bounds: [list]

ALWAYS include the All Clear section; stakeholders want to see what was checked, not just what failed. If the user is setting up a recurring scan for email/Slack, format with bullets and short lines, no wide tables.

End Critical/Warning findings with an offer: "Want me to run a full spike diagnosis on [campaign]? I'd compare the anomaly window against the prior baseline and check all six causes."

Drift Scan must NOT

• Run on weekly data.
• Flag every movement; apply the SD thresholds. False positives erode trust.
• Claim certainty on causes.
• Include zero-spend days in baselines.

Implementation note (both modes)

Do the math with pandas in the code environment; don't estimate z-scores or deltas by eyeballing tables. Show the thresholds used at the top of every report so the user can challenge them.