The GA4 BigQuery export schema is one table per day, events_YYYYMMDD, and every row in it is a single event. Simple values like event_name sit in plain columns. Parameters, user properties and products live in nested, repeated records that you have to UNNEST before you can read them. Below is every field Google documents, in Google’s own groups, with its type and what it actually tells you.

You’ll find the same structure in your own project at BigQuery › Explorer › your project › Datasets › analytics_<property_id> › events_YYYYMMDD › Schema. That tab gives you names and types. It won’t tell you which fields stay empty, which three of them all claim to be “the source”, or why your totals don’t match the GA4 interface. That’s the part this page covers.

How the Tables and Rows Are Organized

Every linked property gets a single dataset called analytics_<property_id>. Inside it, Google writes up to two kinds of daily tables, depending on which export options are switched on for the link.

TableWhen it existsWhat to know
events_YYYYMMDDDaily export is onThe full day. Late events keep arriving for a while, so the table for January 1 can still change on January 4.
events_intraday_YYYYMMDDStreaming export is onFilled continuously during the day and deleted once the daily table is complete. The BigQuery sandbox doesn’t get it at all.

Report from the daily tables. Google recommends events_YYYYMMDD over the intraday table because it’s a stable dataset for the day, while streaming is a best-effort service that can have gaps. Keep intraday for checking that today’s tags fire.

Inside a table, one event is one row, but the row isn’t flat. A single page_view carries page_location, page_title and page_referrer as separate entries in the repeated event_params record. In the preview it looks like one event spread over several lines, with the name and timestamp on the first line and each parameter on its own line underneath.

The Schema, Group by Group

Google’s schema reference splits the columns into 11 groups. I’ve kept that order, cut the descriptions down to what you need while writing SQL, and noted the fields that cause trouble. Types are exactly as Google lists them, including a couple that look wrong at first glance.

Event fields

FieldTypeWhat it tells you
event_dateSTRINGThe event date as YYYYMMDD, in your property’s time zone.
event_timestampINTEGERWhen Google Analytics received the event, in microseconds, UTC. Events sent in one request can share it.
event_previous_timestampINTEGERWhen the previous event happened, in microseconds, UTC.
event_nameSTRINGThe event name, such as page_view or purchase.
event_value_in_usdFLOATThe event’s value parameter, converted to USD.
event_bundle_sequence_idINTEGERSequential ID of the bundle the event was uploaded in.
event_server_timestamp_offsetINTEGERGap between collection time and upload time, in microseconds.
batch_page_idINTEGERGoes up with each new page within an engagement.
batch_ordering_idINTEGERGoes up every time a page sends a network request.
batch_event_indexINTEGERThe event’s position inside its batch, in the order it happened on the device.

event_params is a RECORD that repeats once per parameter. Which keys you get depends on your tagging, so the list is different for every property. The shape of the record never changes, though, and the typed value columns are where most first queries go wrong.

FieldTypeWhat it tells you
event_params.keySTRINGThe parameter name, such as page_location or ga_session_id.
event_params.valueRECORDHolds the value in one of the four typed fields below.
value.string_valueSTRINGText values like URLs and campaign names.
value.int_valueINTEGERWhole numbers. Google’s own session queries read ga_session_id from here.
value.double_valueFLOATDecimal numbers.
value.float_valueFLOATDocumented, but Google says it isn’t currently in use.

User fields

The privacy_info fields are filled from the visitor’s consent state when consent mode is on. If you haven’t set it up yet, start with the guide to Google Consent Mode v2 and its four signals, because the first two columns mirror the consent signals your banner sends.

FieldTypeWhat it tells you
user_idSTRINGThe unique ID assigned to a user, if you send one.
user_pseudo_idSTRINGPseudonymous ID given on the first visit or app open. Count this when you don’t send user_id.
user_first_touch_timestampINTEGERWhen the user first visited or opened the app, in microseconds.
is_active_userBOOLEANWhether the user was active at any point that calendar day. Daily tables only.
privacy_info.analytics_storageSTRINGAnalytics storage consent: Yes, No or Unset.
privacy_info.ads_storageSTRINGWhether ad targeting is enabled: Yes, No or Unset.
privacy_info.uses_transient_tokenSTRINGYes when a web user denied analytics storage and measurement without cookies, based on transient tokens in server data, is enabled.
user_properties.keySTRINGThe name of a user property you set. The record repeats once per property.
user_properties.value.string_value, .int_value, .double_valueSTRING, INTEGER, FLOATThe property’s value, in whichever column matches its type.
user_properties.value.float_valueFLOATDocumented as currently unused.
user_properties.value.set_timestamp_microsINTEGERWhen the property was last set, in microseconds.
user_ltv.revenueFLOATThe user’s lifetime revenue. Empty in intraday tables.
user_ltv.currencySTRINGThe currency of that lifetime value. Empty in intraday tables.

Device and geo

FieldTypeWhat it tells you
device.categorySTRINGMobile, tablet or desktop.
device.mobile_brand_name, .mobile_model_name, .mobile_marketing_nameSTRINGThe device’s brand, model and marketing name.
device.mobile_os_hardware_modelSTRINGThe model as reported by the operating system itself.
device.operating_system, .operating_system_versionSTRINGThe OS and its version.
device.vendor_idSTRINGIDFV, present only when IDFA isn’t collected.
device.advertising_idSTRINGAdvertising ID or IDFA.
device.languageSTRINGThe OS language.
device.time_zone_offset_secondsINTEGEROffset from GMT, in seconds.
device.is_limited_ad_trackingBOOLEANThe device’s Limit Ad Tracking setting.
device.web_info.browser, .browser_versionSTRINGThe browser and its version.
device.web_info.hostnameSTRINGThe hostname the event was logged on. Handy for keeping staging sites out of your numbers.
geo.continent, .sub_continent, .country, .region, .metro, .citySTRINGLocation, worked out from the IP address.

Traffic source records

There are three of them, and they don’t answer the same question. The section after the reference explains which one to use. Here’s what each record holds. Every field in all three is a STRING.

Field in collected_traffic_sourceTypeWhat it tells you
manual_source, manual_mediumSTRINGutm_source and utm_medium collected with the event, plus values parsed from referral parameters.
manual_campaign_name, manual_campaign_idSTRINGutm_campaign and utm_id.
manual_term, manual_contentSTRINGutm_term and utm_content.
manual_source_platform, manual_creative_format, manual_marketing_tacticSTRINGThe matching utm_source_platform, utm_creative_format and utm_marketing_tactic tags.
gclidSTRINGThe Google Ads click ID that came with the event.
dclidSTRINGThe click ID for Display & Video 360 and Campaign Manager 360.
srsltidSTRINGThe Google Merchant Center identifier.
Part of session_traffic_source_last_clickFieldsWhat it tells you
manual_campaign9source, medium, campaign_name, campaign_id, term, content, source_platform, creative_format, marketing_tactic of the last clicked manual campaign.
google_ads_campaign6customer_id, account_name, campaign_id, campaign_name, ad_group_id, ad_group_name.
cross_channel_campaign5campaign_id, campaign_name, source, medium, source_platform of the last clicked cross-channel campaign.
sa360_campaign10Search Ads 360 campaign, ad group and account details.
dv360_campaign18Display & Video 360 partner, advertiser, campaign, creative, exchange, insertion order and line item details.
cm360_campaign20Campaign Manager 360 account, advertiser, campaign, creative, placement and site details.
Field in traffic_sourceTypeWhat it tells you
traffic_source.sourceSTRINGThe network that first acquired the user. Same as the User source dimension in reports.
traffic_source.mediumSTRINGThe medium that first acquired the user. Same as User medium.
traffic_source.nameSTRINGThe campaign that first acquired the user. Same as User campaign.

Stream and platform

Two plain columns. stream_id is the numeric ID of the data stream, stored as a STRING, so compare it to '1234567890' in quotes. platform is a STRING that tells you whether the event came from Web, IOS or Android.

Ecommerce

The ecommerce record holds order-level values. If your purchase tracking isn’t in place yet, the walkthrough on e-commerce tracking without a developer covers the events that fill it.

FieldTypeWhat it tells you
ecommerce.transaction_idSTRINGThe transaction ID of the order.
ecommerce.purchase_revenue, _in_usdFLOATRevenue in local currency and in USD. Filled on purchase events only.
ecommerce.refund_value, _in_usdFLOATRefund amount. Filled on refund events only.
ecommerce.shipping_value, _in_usdFLOATShipping cost.
ecommerce.tax_value, _in_usdFLOATTax.
ecommerce.total_item_quantityINTEGERThe sum of items.quantity in the event.
ecommerce.unique_itemsINTEGERUnique items, judged by item_id, item_name and item_brand.

Items

items repeats once per product in the event. Predefined fields like item_id get columns of their own. Only the custom item parameters you define end up in items.item_params, which has the same key and typed value shape as event_params.

FieldTypeWhat it tells you
items.item_id, .item_name, .item_brand, .item_variantSTRINGWho the product is.
items.item_category to .item_category5STRINGThe category and four sub-category levels.
items.price, .price_in_usdFLOATPrice in local currency and in USD.
items.quantityINTEGERQuantity. Set to 1 if you don’t send it.
items.item_revenue, .item_revenue_in_usdFLOATPrice times quantity. Filled on purchase events only.
items.item_refund, .item_refund_in_usdFLOATPrice times quantity. Filled on refund events only.
items.coupon, .affiliation, .location_idSTRINGCoupon code, supplying company or store, and the item’s location.
items.item_list_id, .item_list_nameSTRINGThe list the product was shown in.
items.item_list_indexSTRINGThe product’s position in that list. Yes, a STRING, so cast it before sorting.
items.promotion_id, .promotion_name, .creative_name, .creative_slotSTRINGThe promotion and the creative slot the product appeared in.
items.item_params.key, .value.string_value, .int_value, .double_value, .float_valueSTRING, STRING, INTEGER, FLOAT, FLOATYour custom item parameters.

App-only groups

app_info holds four STRING fields: the app’s package or bundle ID, its Firebase App ID, the store that installed it, and its version. publisher holds AdMob ad data: estimated ad revenue in USD on ad impression events, plus ad format, ad source and ad unit. It’s never filled in intraday tables. On a website-only property you can skip both groups.

Same Visit, Different Answers: Picking the Right Source Record

This is where most BigQuery traffic reports go wrong. All three records look like “source and medium”, and people grab whichever one autocompletes first. Here’s one visit, read through each of them.

A reader first found you through your newsletter in March, from a link tagged utm_source=newsletter&utm_medium=email. Today she clicks a LinkedIn post tagged utm_source=linkedin&utm_medium=social, lands on a blog post, then reads two more pages.

RecordScopeWhat it shows for today’s three page views
traffic_sourceUser, first acquisitionnewsletter / email, today and on every later visit. It doesn’t change when she clicks other campaigns.
collected_traffic_sourceThe event itselflinkedin / social on the landing page view, because that hit carried the tags. The two later page views carried no campaign data of their own.
session_traffic_source_last_clickThe sessionlinkedin / social, the last-click source attributed to the session.

So match the record to the question. “What brought this person to us the first time” is traffic_source. “How many sessions came from LinkedIn” is session_traffic_source_last_click. “What exactly was on this hit” is collected_traffic_source. If the First user versus Session split still feels slippery in the interface, the guide to how GA4 channel groupings classify traffic walks through the same logic in reports.

One more trick from Google’s own help. When traffic_source shows a “Not Available” record for an ad click, pull the gclid from collected_traffic_source and look the click up through the Google Ads API, Google Ads Scripts or the BigQuery Data Transfer Service for Google Ads. wBRAID and gBRAID identifiers aren’t in the export, so those clicks can’t be recovered this way.

Three Queries You Can Copy

Each one filters dates with _TABLE_SUFFIX, the way Google’s own examples do, so you only scan the days you ask for. Swap in your project and property ID. A bonus of filtering on digits is that intraday tables drop out on their own, because their suffix starts with intraday_.

1. Sessions by source and medium

SELECT
  session_traffic_source_last_click.manual_campaign.source AS source,
  session_traffic_source_last_click.manual_campaign.medium AS medium,
  COUNT(DISTINCT CONCAT(user_pseudo_id, '-', CAST(
    (SELECT value.int_value FROM UNNEST(event_params)
     WHERE key = 'ga_session_id') AS STRING))) AS sessions
FROM `your-project.analytics_123456789.events_*`
WHERE _TABLE_SUFFIX BETWEEN '20261001' AND '20261031'
GROUP BY source, medium
ORDER BY sessions DESC

A session here is user_pseudo_id plus ga_session_id, the same pair Google’s own session queries group by. Clicks from Google’s ad products have their own parts of the record, like google_ads_campaign, so add those columns if you run ads.

2. Which parameters each event carries, and in which column

SELECT
  event_name,
  ep.key AS param,
  COUNTIF(ep.value.string_value IS NOT NULL) AS string_rows,
  COUNTIF(ep.value.int_value IS NOT NULL) AS int_rows,
  COUNTIF(ep.value.double_value IS NOT NULL) AS double_rows
FROM `your-project.analytics_123456789.events_*`,
  UNNEST(event_params) AS ep
WHERE _TABLE_SUFFIX BETWEEN '20261025' AND '20261031'
GROUP BY event_name, param
ORDER BY event_name, param

Run this before any real query. It tells you whether a parameter lives in string_value or int_value, which saves you from the classic “my column is all NULL” moment. If a numeric parameter like value shows up in more than one column, wrap the typed fields in COALESCE, as Google’s revenue examples do.

3. Orders, revenue and products

-- Order level: one row per purchase event
SELECT
  event_date,
  COUNT(DISTINCT ecommerce.transaction_id) AS orders,
  SUM(ecommerce.purchase_revenue) AS revenue,
  SUM(ecommerce.total_item_quantity) AS units
FROM `your-project.analytics_123456789.events_*`
WHERE _TABLE_SUFFIX BETWEEN '20261001' AND '20261031'
  AND event_name = 'purchase'
GROUP BY event_date
ORDER BY event_date
-- Product level: one row per item in each purchase
SELECT
  item.item_name,
  SUM(item.quantity) AS units,
  SUM(item.item_revenue) AS item_revenue
FROM `your-project.analytics_123456789.events_*`,
  UNNEST(items) AS item
WHERE _TABLE_SUFFIX BETWEEN '20261001' AND '20261031'
  AND event_name = 'purchase'
GROUP BY item.item_name
ORDER BY item_revenue DESC

Keep these two apart. Once you UNNEST(items), every order-level value repeats on each product row, so SUM(ecommerce.purchase_revenue) in the second query would count a three-product order three times.

Why BigQuery Won’t Match Your GA4 Reports

You’ll compare totals at some point, and they won’t line up. Most of the gap comes from a handful of documented differences in how the export works.

  • No modeling. Google says the export leaves out the “value additions” it makes in standard reports and explorations, and the daily export is last click observed, with no modeling.
  • Two clocks. event_date follows your property’s time zone, while event_timestamp is UTC. Group by one or the other, never a mix.
  • Late events. Daily tables keep taking late events for up to three days, so yesterday’s number isn’t final yet.
  • Intraday holes. traffic_source, user_ltv, is_active_user and publisher aren’t filled in intraday tables. Streaming also skips first-user source for new users, and for existing users it needs about 24 hours.
  • Cookieless pings. With consent mode on, cookieless pings land in the export too, so check privacy_info before you count users.
  • The dead column. float_value isn’t in use. Read int_value or double_value instead.

Want to practice before touching your own data? Google publishes an obfuscated sample, bigquery-public-data.ga4_obfuscated_sample_ecommerce, with Google Merchandise Store events from November 1, 2020 to January 31, 2021. Two catches: it has no batch_ fields, and some values are placeholders like <Other>, NULL or an empty string.

Once your queries return numbers you trust, the next job is shaping them into something people act on. The guide on turning analytics data into decisions picks up from there.

Frequently Asked Questions

What tables does the GA4 BigQuery export create?

One dataset per property, named analytics_<property_id>. With daily export on, you get an events_YYYYMMDD table per day. With streaming on, you also get events_intraday_YYYYMMDD, which is deleted once that day’s daily table is complete.

What’s the difference between traffic_source and collected_traffic_source?

traffic_source is the source, medium and campaign that first acquired the user, and it never changes after that. collected_traffic_source is whatever campaign data arrived with the individual event, such as UTM tags or a gclid. For session-level reporting, use session_traffic_source_last_click.

Why is traffic_source empty in my intraday table?

Because Google doesn’t fill it there. The traffic_source record isn’t populated in intraday tables, and streaming export skips first-user source for new users. Query the daily events_YYYYMMDD table for user acquisition.

What are batch_event_index, batch_ordering_id and batch_page_id for?

They put events in the order they happened, even when several were sent in one batch. Google’s own page-sequence query sorts by user_pseudo_id, user_id, batch_page_id, batch_ordering_id and then batch_event_index.

Tom Bradley

About the Author

Tom Bradley

Marketing analyst with 8+ years in web analytics. I’ve completed 150+ GA4 implementations and helped 50+ brands turn data into growth strategies. Every guide on Viewing comes from real projects and real problems I’ve solved.