SuperAPI AI Directory

measurement-debug

Marketing & growth

View repository

Updated 10-16 10:30

About this skill

Spotify Ads API — Measurement Debug

Find the failing layer before recommending a change. Read troubleshooting.md before diagnosing.

Setup

PLUGIN_ROOT="${CODEX_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-.}}"
api() { "$PLUGIN_ROOT/scripts/api-request.sh" measurement-debug "$@"; }

Before the first Ads API v3 call, read and follow $PLUGIN_ROOT/skills/api-reference/references/live-openapi.md.

Measurement paths use business_id. Discover it with GET businesses when absent.

Triage

Capture:

  • expected event, expected volume, first/last known good time, timezone, and affected environment
  • Pixel, CAPI, or both; direct code, web GTM, server GTM, or another owner
  • business, dataset, integration, connection, and ad-account IDs
  • whether the problem is receipt, deduplication, sharing/selection, or attributed reporting
  • a sanitized failing request/response and trace ID when available

Do not ask for CAPI tokens, raw email/phone, cookies, device IDs, IP addresses, or production payloads. Ask for redacted field names and error metadata instead.

Read-only audit

1. Resolve topology

api GET "businesses/<business_id>/datasets"
api GET "businesses/<business_id>/datasets/<dataset_id>"
api GET "businesses/<business_id>/capi/<capi_connection_id>"
api GET "businesses/<business_id>/ad_accounts/<ad_account_id>/datasets"

The GET datasets endpoint does not accept limit or offset parameters. To discover Pixels, inspect the pixel field on each dataset response rather than calling a separate pixels endpoint.

Only call GET businesses/<business_id>/capi/<capi_connection_id>/tokens when the incident specifically requires token inventory. The response includes active token values, not just IDs: treat the entire raw response as secret, never show it in a command result or transcript, and expose only redacted token IDs/counts in the final audit.

Verify that:

  • the expected integration is in the expected dataset
  • the dataset is shared to the intended ad account
  • each integration's dataset_id field points to the correct dataset (a null value means events are ingested but not routed — see isolation table)
  • Pixel domain and CAPI connection ID match the intended environment/business
  • when authentication is in scope, the token ID exists on the same CAPI connection; never display token values
  • dataset flags such as is_receiving_events and is_receiving_lead_events fit the symptom

2. Inspect event receipt

api GET "businesses/<business_id>/datasets/<dataset_id>/diagnostics?granularities=HOURLY"
api GET "businesses/<business_id>/datasets/<dataset_id>/diagnostics?granularities=DAILY"
api GET "businesses/<business_id>/datasets/<dataset_id>/diagnostics?granularities=HOURLY&datasource_ids=<datasource_id>"

Compare each datasource's event names, counts, timestamps, and last_activity_ms. Convert timestamps explicitly and state the timezone. Use hourly data for recent incidents and daily data for trend comparison.

Receipt diagnostics answer “did Spotify receive events?” They do not prove trigger correctness, identifiers, deduplication, campaign data-source selection, or attribution.

3. Isolate the layer

Observation Likely layer Next check
No source receives events implementation/network base code or server request, auth, CSP, container, endpoint
Pixel absent, CAPI present browser base code, trigger, duplicate install, CSP, redirects
CAPI absent, Pixel present server token/connection pair, payload, 4xx/5xx logs, server GTM container
Both present, total inflated deduplication same real event uses same event_id on both sources
Diagnostics present, campaign metrics absent selection/attribution dataset shared and selected; date/window/privacy/report fields
Wrong site/account receives data topology/config domain, business, dataset membership, connection ID
Events stopped suddenly deployment/credential release time, GTM publish, token revocation, CSP/infrastructure change
Integration in dataset but dataset_id null dataset routing integration was moved via POST datasets with multiple integration_ids; events are ingested but not routed — fix by removing the integration and re-adding it to a properly-linked dataset

4. Inspect implementation evidence

If the user supplies a public page, browser access, GTM export, or sanitized CAPI example, inspect it read-only:

  • Pixel: compare pixel key and base/event code with Ads Manager; search direct installs for spdt; inspect console/network/CSP; verify event fires only at the intended action.
  • Do not conclude Pixel is absent merely because source HTML lacks spdt; tag managers can inject it at runtime.
  • Confirm the site is not using both direct and GTM installs.
  • Verify redirects preserve Spotify's spclid query parameter.
  • Web GTM is for Pixel; CAPI GTM requires a server container.
  • CAPI: validate envelope, event enum, unique stable event ID, ISO 8601 time, at least one identifier, SHA-256 normalization, optional source URL/action source, and revenue fields.

Never expose secrets or personal data in findings.

Mutating tests and fixes

Default to a remediation plan. Ask for explicit confirmation immediately before:

  • submitting any CAPI event, including a test event
  • creating/revoking a token
  • changing Pixel/CAPI/dataset metadata
  • sharing/unsharing a dataset
  • moving an integration between datasets

A test event changes measurement data. Prefer replaying a legitimate known event in a non-production integration with a stable unique event_id, and label/document it. Never blindly replay a production POST after a timeout or 5xx.

For source-code or GTM changes, provide the smallest isolated change and an acceptance test; do not claim the Ads API can deploy website or GTM code.

Attribution boundary

If receipt is healthy but reporting is not:

  1. verify the dataset is shared to and selected by the relevant campaign/ad set
  2. verify the report date range, entity, and conversion field
  3. distinguish total received event activity from attributed conversions
  4. account for attribution windows and privacy thresholding
  5. route reporting extraction to the report skill

Do not “fix” healthy ingestion to force attributed numbers to match internal analytics.

Incident report

Return:

  • symptom and affected scope
  • topology diagram in one line: business → dataset → source(s) → ad account
  • evidence table with check, result, timestamp/timezone, and confidence
  • most likely failing layer and alternatives
  • exact remediation by owner
  • safe verification steps and wait period
  • redacted trace IDs/errors for escalation

Label facts, inferences, and unverified assumptions separately. If the issue remains unresolved, recommend Spotify Ads API support with client ID, endpoint, sanitized request/response, timestamp, and sp_trace_id.

Guardrails

  • Begin read-only.
  • Wait at least 20 minutes after new event activity before diagnosing empty Ads Manager diagnostics.
  • Never request, echo, or store CAPI secrets or personal identifiers.
  • Never automatically retry POST, PATCH, or DELETE.
  • On 4xx, preserve the sanitized error and stop; on ambiguous POST failure, check receipt/resource state before any retry.