Agent Skills: GA4 → BigQuery Export

GA4 → BigQuery Export

The Data API is good but bounded — sampled past a threshold, capped at ~150 dimensions, no cohort joins. BigQuery export gives you the raw event stream as SQL-queryable tables, free at the standard GA4 tier (up to 1M events/day), with no sampling and full event payloads.

This skill: one-time setup, then the SQL recipes for what the Data API can't do well.

Setup — one time

1. Link your GA4 property to a GCP project

In https://analytics.google.com/:

1. Admin → Property column → BigQuery Links
2. Create a link → pick your GCP project
3. Data location: pick the BQ region for the export tables (US multi-region is fine for most cases; EU if you need data residency)
4. Export type:
   • Daily — single events_YYYYMMDD table per day, written ~24h after midnight. Fine for most reporting.
   • Streaming — events_intraday_YYYYMMDD table written ~near-real-time. Costs more, useful for hot ops dashboards.
   • Most setups: enable both. Streaming for "today", daily for everything else.
5. Include advertising identifiers — uncheck unless you specifically need device-graph data (most don't)
6. Save

The first daily table lands within 24h. The first streaming table is near-immediate. After that you have a new events_YYYYMMDD every day forever, no maintenance needed.

2. Verify the export is working

PROJECT=your-gcp-project
DATASET=analytics_123456789       # auto-named after the property ID

bq ls "$PROJECT:$DATASET" 2>&1 | head -10
# Expect: events_YYYYMMDD tables + events_intraday_YYYYMMDD if streaming enabled

If you see no dataset, the link is configured but the first export hasn't fired yet — wait 24h.

3. Authorize a service account for querying

The SA from ga4-auth-setup only has Data API access. For BQ queries, grant the same SA:

SA_EMAIL=ga4-reader@your-project.iam.gserviceaccount.com
gcloud projects add-iam-policy-binding "$PROJECT" \
  --member="serviceAccount:$SA_EMAIL" \
  --role="roles/bigquery.dataViewer"
gcloud projects add-iam-policy-binding "$PROJECT" \
  --member="serviceAccount:$SA_EMAIL" \
  --role="roles/bigquery.jobUser"

dataViewer reads tables; jobUser lets the SA run queries (queries are jobs in BQ's model).

The events table schema (the important columns)

Every row in events_YYYYMMDD is one event. The schema is denormalized — user + session + event + page + device all flat in each row.

| Column | Type | Notes | |---|---|---| | event_date | STRING | 'YYYYMMDD' | | event_timestamp | INT64 | Microseconds since epoch | | event_name | STRING | page_view, session_start, purchase, custom event names | | event_params | ARRAY<STRUCT<key, value>> | All event parameters; value is itself a union STRUCT (string_value, int_value, float_value, double_value) | | user_pseudo_id | STRING | GA4's cookie-based user ID (anonymous unless user_id is set) | | user_id | STRING | If you set user_id via gtag('set', {user_id: '...'}) | | user_properties | ARRAY<STRUCT<key, value>> | Same shape as event_params | | device.* | STRUCT | category / os / browser / model | | geo.* | STRUCT | country / region / city | | traffic_source.* | STRUCT | source / medium / campaign of FIRST session (not current) | | session_traffic_source_last_click.* | STRUCT | source / medium of CURRENT session — what you usually want | | ga_session_id (param) | INT64 | Pulled via (SELECT value.int_value FROM UNNEST(event_params) WHERE key='ga_session_id') | | ga_session_number (param) | INT64 | Same idiom — 1 = first session, 2 = second, etc. |

The event_params and user_properties arrays are the gnarly bit. Pulling a parameter requires UNNEST + filter. The idiom:

-- Pull the page_location for every page_view
SELECT
  TIMESTAMP_MICROS(event_timestamp) AS ts,
  (SELECT value.string_value FROM UNNEST(event_params) WHERE key = 'page_location') AS page,
  user_pseudo_id
FROM `your-project.analytics_123456789.events_*`
WHERE _TABLE_SUFFIX BETWEEN '20260513' AND '20260520'
  AND event_name = 'page_view'
LIMIT 100;

_TABLE_SUFFIX BETWEEN '...' AND '...' is the canonical way to scan a date range across the wildcard table. Always set it — without a suffix filter, you query the entire history and pay for it.

Recipe 1 — True cohort retention

The thing the Data API can't do cleanly:

WITH first_seen AS (
  SELECT
    user_pseudo_id,
    DATE(MIN(TIMESTAMP_MICROS(event_timestamp))) AS first_date
  FROM `your-project.analytics_123456789.events_*`
  WHERE _TABLE_SUFFIX BETWEEN '20260401' AND '20260520'
  GROUP BY user_pseudo_id
),
activity AS (
  SELECT
    user_pseudo_id,
    DATE(TIMESTAMP_MICROS(event_timestamp)) AS active_date
  FROM `your-project.analytics_123456789.events_*`
  WHERE _TABLE_SUFFIX BETWEEN '20260401' AND '20260520'
  GROUP BY user_pseudo_id, active_date
)
SELECT
  DATE_TRUNC(f.first_date, WEEK) AS cohort_week,
  DATE_DIFF(a.active_date, f.first_date, WEEK) AS weeks_since,
  COUNT(DISTINCT a.user_pseudo_id) AS active_users
FROM first_seen f
JOIN activity a USING (user_pseudo_id)
GROUP BY cohort_week, weeks_since
ORDER BY cohort_week, weeks_since;

Output: rows of (cohort_week, weeks_since, active_users). Pivot in your tool of choice for the classic triangle chart.

Recipe 2 — Sessions table (denormalized from events)

GA4's BQ export is event-rows, not session-rows. To reason about sessions, build the session table yourself:

WITH sessions AS (
  SELECT
    user_pseudo_id,
    (SELECT value.int_value FROM UNNEST(event_params) WHERE key = 'ga_session_id') AS session_id,
    MIN(TIMESTAMP_MICROS(event_timestamp)) AS session_start,
    MAX(TIMESTAMP_MICROS(event_timestamp)) AS session_end,
    COUNT(*) AS event_count,
    COUNTIF(event_name = 'page_view') AS pageviews,
    ANY_VALUE(device.category) AS device,
    ANY_VALUE(geo.country) AS country,
    ANY_VALUE(session_traffic_source_last_click.manual_campaign.source) AS source,
    ANY_VALUE(session_traffic_source_last_click.manual_campaign.medium) AS medium,
  FROM `your-project.analytics_123456789.events_*`
  WHERE _TABLE_SUFFIX BETWEEN '20260513' AND '20260520'
  GROUP BY user_pseudo_id, session_id
  HAVING session_id IS NOT NULL
)
SELECT * FROM sessions
ORDER BY session_start DESC
LIMIT 100;

You'd usually CREATE TABLE or CREATE MATERIALIZED VIEW over this — querying the events table directly every time is slow + expensive.

Recipe 3 — Top pages by source

SELECT
  (SELECT value.string_value FROM UNNEST(event_params) WHERE key = 'page_location') AS page,
  session_traffic_source_last_click.manual_campaign.source AS source,
  COUNT(*) AS pageviews,
  COUNT(DISTINCT user_pseudo_id) AS users
FROM `your-project.analytics_123456789.events_*`
WHERE _TABLE_SUFFIX BETWEEN '20260513' AND '20260520'
  AND event_name = 'page_view'
GROUP BY page, source
HAVING pageviews > 10        -- filter the long tail
ORDER BY pageviews DESC
LIMIT 50;

Cost considerations

BigQuery costs $5/TB scanned (first 1 TB/month free). What that translates to in practice:

| Scenario | ~TB / query | |---|---| | Small site, 1k events/day, 30-day window | < 1 GB | | Medium site, 100k events/day, 30-day window | ~10 GB | | Large site, 1M events/day, 30-day window | ~100 GB | | Same scenarios but querying the full 14-month history | 12-15x the above |

Stay under the free tier with normal usage. Cost-saving patterns:

1. Always set _TABLE_SUFFIX BETWEEN — don't scan all history if you only need 7 days
2. Materialize hot queries — CREATE TABLE / CREATE MATERIALIZED VIEW for sessions, daily aggregates, etc.
3. SELECT only the columns you need — BQ is columnar; selecting event_params array always reads the whole array even if you only want one parameter
4. Use --dry-run before any new query to see TB scanned: bq query --dry_run --use_legacy_sql=false "SELECT ..."

Streaming vs daily — which to query

| Table prefix | When | |---|---| | events_YYYYMMDD | Stable historical data. Use for reports, analysis. | | events_intraday_YYYYMMDD | "Today" data, near-realtime. Schema is identical, but rows may not be deduped yet. | | events_* (wildcard) | When the date range crosses both. BQ will scan both transparently. |

Today's data lives in events_intraday_TODAY; tomorrow it gets rolled into events_TODAY and the intraday table for today is dropped. So if you have a query that needs both stable history + today, use events_* and filter on _TABLE_SUFFIX.

Common gotchas

| Issue | Why | |---|---| | Numbers don't match the GA4 UI exactly | UI uses different identity-stitching for cross-device users; raw events are pre-stitching. Off by a few % is expected. | | event_params UNNEST returns NULL | The key doesn't exist on that event. Always wrap in (SELECT ... LIMIT 1) so missing-key events return NULL instead of erroring. | | Query scans way more than expected | Missing _TABLE_SUFFIX filter, OR using events_* without a _TABLE_SUFFIX BETWEEN clause | | events_intraday_* has 2x the rows you'd expect | Intraday tables aren't deduped; the same event may appear twice if it was buffered + retried. The daily rollup dedupes. | | No user_id even though I set it on the front-end | user_id is the explicit identifier; user_pseudo_id is the auto-generated cookie ID. If user_id is consistently NULL, the gtag('set', {user_id: ...}) call is firing AFTER the event you're checking, or it's set on a property that the export doesn't pull. |

Related skills

• ga4-auth-setup — for the service account that queries BQ
• ga4-data-api-query — when you DON'T need event-level granularity (Data API is faster + cheaper for aggregates)
• ga4-common-reports — the Data API recipes that BQ supersedes