Inside the PMax Diagnostics Engine

A composite score from 0–100 calculated from two penalty components: a waste penalty (based on your leakage rate) and a maturity penalty (based on how much of your 90-day cumulative spend sits in unstable campaign stages). Lower penalties = higher ACI. A score below 50 means the algorithm is operating without structural constraints.

ACI = 100 − waste_penalty − maturity_penalty

Waste penalty tiers:
  Leakage < 10%  → 0 pts
  Leakage 10–20% → 10 pts
  Leakage 20–30% → 25 pts
  Leakage > 30%  → 40 pts

Maturity penalty tiers:
  Unstable spend < 20% → 0 pts
  Unstable spend 20–40% → 10 pts
  Unstable spend > 40% → 20 pts

PMax campaigns move through distinct learning stages tracked over a 90-day cumulative window. The distribution of your spend across these stages directly feeds the ACI maturity penalty. Campaigns in EXPLORATION or DEGRADATION stages are considered unstable — the algorithm hasn't converged on an efficient strategy yet.

Maturity penalty = f(unstable_spend / total_tracked_spend)

Stages tracked: CALIBRATION, EXPLORATION, SCALING, DEGRADATION
Unstable stages: EXPLORATION + DEGRADATION

Note: 90-day window means maturity figures
will exceed your 30-day spend total.

PMax always allocates a portion of budget to exploration — testing audiences and inventory beyond its highest-confidence asset group. The EER measures whether that cost is justified. It divides the CPA of your top-spending asset group (exploitation) by the blended CPA of all remaining groups (exploration). Guardrails are applied: the metric is only calculated when campaigns have sufficient volume to be statistically meaningful.

EER = exploit_CPA ÷ explore_CPA

Where:
  exploit_CPA = spend_rank_1_spend ÷ spend_rank_1_conversions
  explore_CPA = all_other_spend ÷ all_other_conversions

Guardrails (metric shown as NOISE if not met):
  • Campaign total spend < 100 → NOISE_LOW_CAMPAIGN_VOLUME
  • Exploration < 25 spend or < 5% of total → NOISE_LOW_EXPLORATION_SHARE
  • Exploitation conversions < 3 → NOISE_LOW_BASELINE_CONVERSIONS
  • Exploration conversions < 3 → NOISE_LOW_EXPLORATION_CONVERSIONS

The Black Box Tax is the sum of all spend that produced zero conversions in the reporting window, sourced from the waste_management_intelligence_daily table. It is broken down by waste_category into the waterfall — each category represents a distinct structural failure type. Both Exploration and Creative waste are recoverable through different fix paths.

Black Box Tax = SUM(wasted_spend) from waste_management_intelligence_daily
Waste Leakage Rate = Black Box Tax ÷ total_spend × 100

Waste categories (from waste_category field):
  EXPLORATION — algorithm learning spend with no return
  CREATIVE — low-performing asset groups funded despite zero conversions

Sourced from signal_health_intelligence_daily. Compares Ads-reported conversions against GA4-reported conversions at the campaign level over the last 30 days. When the gap exceeds 20%, the algorithm is bidding on phantom signals — inflating CPCs account-wide and suppressing Quality Scores. The signal_status field is pre-calculated in the pipeline.

Drift % = |ads_conversions_30d − ga4_conversions_30d|
              ÷ MAX(ads_conversions_30d, ga4_conversions_30d) × 100

Status classification (from signal_status field):
  HEALTHY        → drift < 15%
  WARNING_DRIFT  → drift 15–40%
  CRITICAL_BLIND → drift > 40%
  LOW_VOLUME     → insufficient conversion volume for meaningful drift

Sourced from google_ads_campaign_performance, filtering for PERFORMANCE_MAX channel type. PMax reports multiple conversion types simultaneously: click-attributed (user clicked then converted), view-through (user saw ad, didn't click, converted later), cross-device (interaction on one device, conversion on another), and modeled (Google's statistical estimates). This breakdown reveals what percentage of your reported conversions are based on hard click data vs. softer attribution signals.

Per campaign (PERFORMANCE_MAX only):
  click_conversions = SUM(conversions)
  view_through = SUM(view_through_conversions)
  cross_device = SUM(cross_device_conversions)
  all_conversions = SUM(all_conversions)
  modeled = all_conversions − click_conversions

Percentages:
  view_through_pct = view_through ÷ all_conversions × 100
  cross_device_pct = cross_device ÷ all_conversions × 100
  click_attributed_pct = click_conversions ÷ all_conversions × 100

Key threshold:
  view_through_pct > 30% = algorithm optimizing on impressions, not clicks

Sourced from google_ads_shopping_performance. Analyzes the structural health of your product feed by checking for missing attributes (category, brand, condition), measuring conversion coverage (what percentage of active products have converted), and counting zero-conversion products with meaningful traffic. Products with missing attributes receive lower placement quality from Google's Shopping algorithm, degrading PMax performance across the entire account — not just for the individual products.

Feed health metrics:
  total_products = COUNT(DISTINCT product_item_id) WHERE clicks > 0
  missing_category = COUNT WHERE product_category_level1 IS NULL
  missing_brand = COUNT WHERE product_brand IS NULL
  converting_products = COUNT WHERE conversions > 0
  zero_conv_products = COUNT WHERE conversions = 0 AND clicks >= 20
  conversion_coverage_pct = converting_products ÷ total_products × 100

Key threshold:
  zero_conv_products > 30% of total = structural feed problem
  missing_category > 0 = immediate Merchant Center fix needed

Sourced from google_ads_shopping_performance (Google Ads API). A product is flagged as a zombie when it meets all three conditions simultaneously: zero units sold in the 30-day window, at least 20 clicks (ensuring spend is intentional, not accidental), and spend exceeding a currency-scaled guardrail. The guardrail adapts dynamically — it takes the greater of the fixed threshold or 1% of total shopping spend for that account.

Zombie conditions (ALL must be true):
  1. units_sold = 0 (or NULL) in 30-day window
  2. clicks ≥ 20
  3. wasted_cost > MAX(currency_threshold, total_shopping_spend × 0.01)

Currency thresholds:
  USD / EUR / GBP / CAD / AUD / CHF
  KWD / AED / SAR / SGD / NZD → $50
  All other currencies (INR etc.) → ₹2,500

Sourced from campaign_maturity_daily. Average ROAS tells you what happened historically. Marginal ROAS tells you what the next incremental dollar of spend will return. When marginal ROAS falls below average ROAS, you have hit a diminishing returns threshold — scaling budget at this point destroys efficiency. The breakeven ROAS threshold is set at 1.5x.

Marginal ROAS = change in revenue ÷ change in spend
             (calculated in campaign_maturity_daily pipeline)

Status classification:
  Marginal ROAS < 1.5         → CRITICAL: below breakeven
  Marginal ROAS < Average ROAS → WARNING: diminishing returns active
  Marginal ROAS ≥ Average ROAS → HEALTHY: scaling efficiently

The recovery figure is pre-calculated in the waste_management_intelligence_daily table as recovery_potential — it is not computed on the client side. The pipeline applies category-level recoverability rates against confirmed waste spend. The 70% baseline described in the audit methodology is a conservative floor, not a flat multiplier — actual recovery varies by waste category.

Recovery = SUM(recovery_potential) from waste_management_intelligence_daily
             (pre-calculated per waste category in the pipeline)

Capture rate rationale:
  • Some exploration spend is genuinely necessary
  • Budget remapping is never perfectly efficient
  • Market conditions introduce variance
  • Real-world accounts recover 55–85% of identified waste
  • 70% baseline is the conservative floor used in methodology descriptions

Sourced from google_ads_pmax_placement_performance via Google's new network segmentation reporting. Each placement row now includes an ad_network_type field that classifies whether the impression was served on DISPLAY, YOUTUBE_WATCH, CONTENT, or SEARCH networks. This breaks the "black box" of PMax placement reporting by revealing which network is responsible for each placement type.

Query groups by:
  ad_network_type × placement_type

Output per combination:
  total_impressions = SUM(impressions)
  unique_placements = COUNT(DISTINCT placement)

Key interpretation:
  DISPLAY + Mobile App = highest junk risk
  YOUTUBE_WATCH + YouTube Channel = brand spend without conversion signal
  CONTENT + Website = mixed quality, requires per-site review

Sourced from google_ads_budget_recommendations via Google's budget reporting endpoint. Each campaign budget includes the current daily amount, Google's recommended amount, and the estimated weekly cost change if the recommendation were applied. The audit divides the weekly figure by 7 to display daily variance — aligning the unit with the daily budget column for direct comparison.

Daily variance = estimated_change_weekly_cost ÷ 7

Status classification:
  recommended > current → INCREASE RECOMMENDED
  recommended < current → DECREASE RECOMMENDED
  recommended = current → OPTIMAL

Data joins:
  google_ads_budget_recommendations (costs & recommendations)
  × google_ads_budget_details (budget names & metadata)
  Joined on campaign_budget_id
  Deduplicated via ROW_NUMBER() OVER(PARTITION BY campaign_budget_id ORDER BY date DESC)

Sourced from google_ads_audience_performance, filtering for criterion_type = 'USER_LIST'. The query identifies remarketing segments and customer lists that are consuming budget (SUM(cost) > 50) while generating zero conversions (SUM(conversions) < 1). These audiences should be excluded from PMax campaigns so the algorithm focuses on new customer acquisition.

Exclusion candidates where:
  criterion_type = 'USER_LIST'
  AND SUM(conversions) < 1
  AND SUM(cost) > 50

Output:
  audience_name = COALESCE(user_list_name, display_name, 'Unknown')
  wasted_spend = SUM(cost)
  action = 'Recommend Exclusion'

Ordered by: wasted_spend DESC
Limit: Top 5 highest-waste audiences

Sourced from google_ads_age_demographics and google_ads_gender_demographics via Google's upgraded audience reporting for PMax. Each demographic segment (age range or gender) reports impressions, clicks, conversions, and cost. CPA is calculated per segment to identify which demographics are converting efficiently and which are consuming budget disproportionately.

Per demographic segment:
  CPA = SUM(cost) ÷ NULLIF(SUM(conversions), 0)

Age ranges: 18–24, 25–34, 35–44, 45–54, 55–64, 65+, Unknown
Gender values: Male, Female, Unknown

Key interpretation:
  High impressions + low conversions + high CPA
  = algorithm over-indexing on expensive segment

  "Unknown" segment typically has highest CPA
  = unclassifiable traffic from low-quality inventory