GA4 Measurement Intelligence

⬡ Vertex AI Data Cleaning Pipeline

Raw GA4 hits scored server-side → ML clustering filters bots before BigQuery export

Raw Sessions

—

Cleaned

—

Bots Removed

—

Raw Sessions

rawSessions[day]

Loaded from pipeline output via summary.json.
Production: BigQuery events_* table · GA4 public dataset ↗

Cleaned Sessions

raw × (1 − bot_rate)

Vertex AI clustering removes bot sessions before export · Vertex AI ↗

Bots Removed %

(raw − cleaned) / raw × 100

Computed from totals across the 7-day window.

$ Value-Based Bidding Engine

Algorithmic $0 bidding on detected bots — replacing manual audience exclusions

Total Clicks

count++ per event

Human

score ≥ 0.70

Bot / Fraud

score < 0.70

Ad Spend Saved

bot_n × $2.40 CPC

Edge Score

Math.random() → 0.0–1.0

Cloud Armor ↗ + reCAPTCHA Enterprise ↗

Classification

≥ 0.70 → HUMAN
0.40–0.69 → SUSPICIOUS
< 0.40 → BOT

Conv. Value

HUMAN → $150.00
else → $0.00

$0 for bots = Smart Bidding deprioritises that traffic automatically.

Ad Spend Saved

bot_clicks × $2.40

$2.40 = simulated avg CPC.
Production: pulled from Google Ads API.

Time	Client ID	Source	Edge Score	Type	Conv. Value

↑ Live feed — new event every 2s · all values simulated client-side

◈ Channel Attribution — Bot Inflation by Channel

Bot-contaminated signals overstate channel lift — calibration reveals true media performance

Raw Calibrated

Raw View: Channel lift is overstated — LLM scrapers and ad fraud bots are triggering conversion pixels, inflating measured ROAS. Toggle to Calibrated to see bot-adjusted attribution using human-only sessions from the cleaned pipeline.

Channel Attribution — How the numbers are calculated

Raw Lift

GA4-reported conversion rate
(includes bot-triggered pixels)

From pipeline output via summary.json — loaded on page open.

Calibrated Lift

Raw Lift × (1 − bot_contamination_rate)

Pipeline output (numpy attribution model) · Production: feed cleaned export into Meridian (Bayesian MMM) ↗

Bot Inflation

(Raw − Calibrated) / Calibrated × 100

% by which bot traffic overstated the channel's true lift.

Channel	Raw Lift	Calibrated	Bot Inflation	Interpretation
Google Ads	37.2%	35.7%	48.3%	Moderate bot inflation — intent-based search audience has ~48% bot share in this dataset
Meta Ads	28.5%	29.3%	44.5%	High bot contamination on broad paid social audiences — calibrate before budget planning
Bing Ads	20.3%	21.0%	44.2%	Similar bot rate to Meta — smaller volume means lower absolute spend impact
Display	14.0%	14.0%	45.9%	Lowest share channel — high bot rate typical for programmatic display inventory

Pipeline output — numpy attribution model (raw vs. bot-cleaned sessions) · Install google-meridian ↗ for Bayesian MCMC in production · Media spend synthesized from ROAS assumptions [SYNTHETIC]

GA4 Measurement Protocol — Server-Side Event Stream

EDGE = header edge detection cycle (every 4s) VBB = simulated paid click (every 2s) MERIDIAN = toggle event Score = Math.random() 0.0–1.0 Conv = score ≥ 0.70 ? $150 : $0 All payloads simulated — no GA4 MP endpoint called