---
title: Analytics Health Check
description: Ten one-query diagnostics for publisher analytics — tracking continuity, identity migrations, pageleave coverage, window sensitivity, denominator traps — each with how to run it and what failing looks like.
url: https://preview.artificialpoets.com/resources/analytics-health-check/
site: Artificial Poets
type: a13s_content
date: 2026-08-09T20:37:16+00:00
modified: 2026-08-15T20:38:03+00:00
---
# Analytics Health Check

Ten one-query diagnostics that catch broken metrics before they decide anything. Each exists because it caught a real failure.

Ten diagnostics, roughly one query each, that catch broken metrics before they decide anything. Every one exists because it caught — or would have caught — a real failure in a year of continuous measurement on a live publisher panel. Run them quarterly, and always before a number leaves the analytics team.

**Free guide**

## Get the health check

Enter your work email and the ten diagnostics unlock — readable in the browser, printable for the team.

## The ten diagnostics

- **1 · Per-site last-seen date.** The most recent event date for every property — individually, never aggregated. *Failing looks like:* any site whose data ends inside a window you are reporting on. Our comparison group went dark for nine weeks before anyone looked.
- **2 · Identity continuity.** Count events keyed by the tool's site ID and separately by domain; compare. *Failing looks like:* a site that "died" on one ID while a "new" site was born on another — a migration splitting history in two.
- **3 · Balanced-panel check.** For any multi-site comparison: every site present in every counted week. *Failing looks like:* weeks where the panel is partial silently averaging in.
- **4 · Volume-anomaly screen.** Week-over-week pageview change per site; flag multiples. *Failing looks like:* a site whose volume quadruples on bot traffic and poisons whatever group contains it.
- **5 · Pageleave / unload coverage.** Unload events ÷ pageviews, per site, per device. *Failing looks like:* coverage below ~40% while duration metrics are quoted as facts. Our panel ranged 21% to 62% on identical instrumentation.
- **6 · Bot-filter parity.** The same bot exclusion applied to every period being compared. *Failing looks like:* a "lift" that appears the week the filter changed.
- **7 · Metric-definition drift.** Diff the current event taxonomy and metric definitions against the baseline period's. *Failing looks like:* a pageview that started counting auto-loaded content mid-series.
- **8 · Window-sensitivity sweep.** Recompute your headline number with the baseline start shifted ±2 months and the endpoint moved month by month. *Failing looks like:* a lift that swings by multiples or flips sign — we watched one metric run +32pp to −4pp on identical data.
- **9 · Denominator audit.** For every percentage metric, ask what the denominator does when the product or site changes. *Failing looks like:* scroll depth "collapsing" under infinite scroll because the document grows with the reader.
- **10 · Intervention-date anchor.** For any before/after claim: a named day the change went live, instrumentation identical on both sides. *Failing looks like:* "we ramped through April" — which is how attribution dies.

## Scoring

**9–10 pass** — your numbers are decision-grade; disagreements are about strategy, not data. **6–8** — fix the failures before the next reporting cycle; treat current trends as directional. **Under 6** — stop quoting trends until repaired: the dashboard is currently a random-number generator with a brand palette.

## You just estimated. Want it measured?

Every tool here approximates. In twenty minutes on your analytics we compute the real distribution: your single-page rate, your depth by device and channel. No deck.

[Run my real numbers](/request-a-demo/)