BigQuery Marketing Attribution Without the GA4 Blind Spots

Below: the recipe every tutorial teaches, the SQL in full, the four things Google's own documentation says the export does not contain, a query that proves each one against your own dataset in seconds, and what a dataset built for attribution looks like instead.

Search for BigQuery marketing attribution and you get ten results. Nine of them are the same five steps, written between 2021 and this year, and the tenth is a product page. The five steps are correct as far as they go.

1. Link GA4 to BigQuery and wait a day for the first `events_*` table.
2. Flatten `event_params`, which is a repeated key-value record, into columns you can read.
3. Group events into sessions on `user_pseudo_id` plus the `ga_session_id` parameter.
4. Pull the source, medium, campaign and any click ID off each session's first event to get one touchpoint per session.
5. Order the touchpoints per user, cut them at the conversion, and divide the credit: all to the first, all to the last, or evenly across the path.

Nothing there is wrong. The problem is that step one decides everything after it, and every one of those articles takes step one for granted. The GA4 export is not a marketing dataset. It is a log of what the GA4 tag observed in a browser, which is a much narrower thing, and the difference only shows up once the model is finished and somebody tries to make a decision with it.

Runnable against the documented GA4 export schema. Swap `my-project` and `analytics_123456789` for yours, and the two dates for your window. The three blocks are one statement: query 1 ends with a `SELECT` so you can inspect the touchpoints on their own, and you drop those last two lines when you chain all three. Nothing here needs a paid tool.

Query 1: one row per session, with the source that started it

`collected_traffic_source` is the record that holds what was actually on the URL: the UTM parameters and the click IDs. It is per event, so it is empty on most events in a session, which is why the second step takes the maximum rather than the first value.

WITH events AS (
  SELECT
    user_pseudo_id,
    (SELECT value.int_value FROM UNNEST(event_params)
       WHERE key = 'ga_session_id')                 AS session_id,
    event_timestamp,
    collected_traffic_source.manual_source          AS source,
    collected_traffic_source.manual_medium          AS medium,
    collected_traffic_source.manual_campaign_name   AS campaign,
    collected_traffic_source.gclid                  AS gclid
  FROM `my-project.analytics_123456789.events_*`
  WHERE _TABLE_SUFFIX BETWEEN '20260601' AND '20260831'
),

touchpoints AS (
  SELECT
    user_pseudo_id,
    session_id,
    MIN(event_timestamp) AS started_at,
    MAX(source)          AS source,
    MAX(medium)          AS medium,
    MAX(campaign)        AS campaign,
    MAX(gclid)           AS gclid
  FROM events
  WHERE session_id IS NOT NULL
  GROUP BY user_pseudo_id, session_id
)

SELECT * FROM touchpoints
ORDER BY user_pseudo_id, started_at;

Query 2: the path to a conversion, with a lookback window

`event_timestamp` is an integer of microseconds, not a timestamp, so a 90 day lookback is arithmetic rather than `TIMESTAMP_SUB`. That one detail is where most hand-written models quietly go wrong.

conversions AS (
  SELECT
    user_pseudo_id,
    MIN(event_timestamp) AS converted_at
  FROM `my-project.analytics_123456789.events_*`
  WHERE _TABLE_SUFFIX BETWEEN '20260601' AND '20260831'
    AND event_name = 'generate_lead'
  GROUP BY user_pseudo_id
),

paths AS (
  SELECT
    t.user_pseudo_id,
    t.source,
    t.medium,
    t.campaign,
    t.started_at,
    ROW_NUMBER() OVER (
      PARTITION BY t.user_pseudo_id ORDER BY t.started_at
    ) AS position,
    COUNT(*) OVER (PARTITION BY t.user_pseudo_id) AS touches
  FROM touchpoints t
  JOIN conversions c USING (user_pseudo_id)
  WHERE t.started_at <= c.converted_at
    -- 90 days, in microseconds
    AND t.started_at >= c.converted_at - 90 * 24 * 60 * 60 * 1000000
)

Query 3: three models over the same paths

First click, last click and linear are one `CASE` each. Time decay and position based are two more lines on the same shape. This is the whole of what a rules-based model is, which is worth knowing before anyone sells you one.

SELECT
  source,
  medium,
  campaign,
  ROUND(SUM(IF(position = 1, 1, 0)), 1)            AS first_click,
  ROUND(SUM(IF(position = touches, 1, 0)), 1)      AS last_click,
  ROUND(SUM(1 / touches), 1)                       AS linear
FROM paths
GROUP BY source, medium, campaign
ORDER BY linear DESC;

That is the model. It runs, it is auditable, it costs a few cents, and it will show you that your branded search campaign takes most of the last-click credit and almost none of the first-click credit, which is usually the first genuinely useful thing a company learns from doing this. See attribution model for what each rule is claiming.

The output above counts credit. Nobody makes a budget decision on credit. The question is cost per lead by campaign, and the GA4 export has no field for what anything cost, because GA4 does not receive spend from anywhere except a linked Google Ads account, and even that stays in the reporting interface.

You do not have to take my word for it. Ask your own dataset:

SELECT column_name
FROM `my-project.analytics_123456789.INFORMATION_SCHEMA.COLUMNS`
WHERE table_name LIKE 'events_%'
  AND (LOWER(column_name) LIKE '%cost%'
    OR LOWER(column_name) LIKE '%spend%'
    OR LOWER(column_name) LIKE '%impression%');

-- 0 rows.

So every article promising marketing ROI in BigQuery needs a second pipeline nobody mentions in the introduction: an ETL tool pulling cost from Google Ads, Meta, LinkedIn and Microsoft on their own schedules, into their own schemas, with their own currency handling and their own definition of a campaign. That is a second project with a second budget, and it is the reason most of these models are abandoned about three weeks in. See the best marketing ETL tools if you are going down that road anyway.

There is a subtler version of the same problem. Once cost arrives from four platforms, it has to be joined to the touchpoints, and the only join key available is the campaign name as each platform spells it. A campaign renamed mid-quarter splits into two rows on one side of the join and stays one row on the other.

The model above attributes a `generate_lead` event, which is a form submission. For an e-commerce shop where the purchase happens on the site, that is close enough to revenue and the export's `ecommerce` record carries the amount. For anybody selling to businesses, it is not revenue at all. It is the start of a sales cycle, and between that form fill and the money there is a qualification call, a proposal, a stage the deal sat in for six weeks and a close date in another quarter.

None of that is in BigQuery, and none of it can be, because it happens in a CRM the GA4 tag cannot see. So the model divides credit across the path to a form fill, and a form fill is precisely the metric that made Meta send you 84 leads of which sales called 22 real. You have moved the wrong number into a warehouse and made it queryable.

• What the export can attributeA browser event: a page view, a form submission, an on-site purchase. Anything the tag fired.
• What the budget meeting asks aboutCost per qualified lead, cost per closed deal, and revenue by first and last touch.
• The gap between themEvery CRM stage after the form, which is where the difference between a good and a bad campaign actually shows.

Joining a CRM export to a GA4 export afterwards is the obvious idea, and it fails on the key. The CRM knows an email address. The export knows a `user_pseudo_id`. Nothing in either table connects them unless you were already writing an identifier into both at the moment of the form fill, which is not something you can start doing retroactively for last quarter. See why GA4, Meta and Google never agree with your CRM for the full version of that argument.

This one surprises people who assume BigQuery is simply GA4 with more detail. The data-driven attribution model that GA4 shows in its advertising reports, the one built on Google's own path analysis, does not appear in the export. There is no column carrying its output, no per-conversion credit split, nothing. Google's schema documentation lists no attribution model results at all.

What you get instead is `session_traffic_source_last_click`, which is last-click attributed session source, and `traffic_source`, which is the campaign that first acquired the user and, in Google's words, does not change if the user interacts with subsequent campaigns. Two fixed rules, both useful, neither of them a model.

So the GA4 interface and the GA4 export will never agree, and the exercise of reconciling them is wasted. If you want a data-driven model over your raw events you rebuild it yourself, which in practice means a Markov chain or Shapley values in Python over the paths from query 2. That is a real project with a real maintenance cost, and it is worth doing only once the paths it runs on contain cost and revenue, which brings us back to the first two blind spots.

Your GA4 dashboard shows more conversions than your warehouse, permanently, and the difference is not a bug in your SQL. Under Consent Mode, visitors who decline the banner send no analytics event, and GA4 estimates what those visitors would have done. That estimate is what fills the gap in the interface.

It is not exported. Google's own description of both the daily and the fresh daily export is "Last click observed, no modeling". The rows for a visitor who declined consent do not exist in BigQuery, so a warehouse figure is always the observed subset and the dashboard figure is the observed subset plus an estimate. Both are labelled conversions.

You can measure the size of your own gap directly, because the export does carry the consent state per event:

SELECT
  privacy_info.analytics_storage,
  privacy_info.ads_storage,
  COUNT(*) AS events
FROM `my-project.analytics_123456789.events_*`
WHERE _TABLE_SUFFIX BETWEEN '20260601' AND '20260831'
GROUP BY 1, 2
ORDER BY events DESC;

In the European Union the denied share is routinely a third of traffic and on some accounts more than half. Run it before you present a warehouse number to anybody. See consent mode and GDPR compliant conversion tracking for the ways to lose less of it in the first place.

Every query above partitions on `user_pseudo_id`, and Google defines it as a pseudonymous ID assigned when a user first visits the site. It is per browser, on a device, and it is stored in a cookie. A person researching on a phone at lunch and converting on a work laptop is two of them. A person who cleared cookies is two of them. On Safari, a person who came back after the cookie's lifetime expired is two of them.

That does not just lose data, it biases the model in a specific and predictable direction. Splitting one path into two makes each fragment shorter, and the shorter a path is, the more the first touch and the last touch are the same session. Fragmented data makes single-touch models look reasonable and multi-touch models look pointless, which is the opposite of what is really going on.

If you send a `user_id` for logged-in users, you can measure the fragmentation on the population where you know the truth:

SELECT
  user_id,
  COUNT(DISTINCT user_pseudo_id) AS browsers
FROM `my-project.analytics_123456789.events_*`
WHERE _TABLE_SUFFIX BETWEEN '20260601' AND '20260831'
  AND user_id IS NOT NULL
GROUP BY user_id
HAVING browsers > 1
ORDER BY browsers DESC;

Whatever ratio that returns is a floor, not an estimate, because it can only see people who logged in. Everyone anonymous is fragmenting at least as badly and invisibly. See cross-device tracking.

A standard, free GA4 property exports up to one million events per day. That sounds enormous until you remember that an event is a page view, a scroll, an outbound click and a form interaction, so a moderately busy site clears it without being a large site at all.

What happens next is the part worth knowing in advance. Google notifies property editors by email, and a property that consistently exceeds the limit has its daily export paused. Previous days are not reprocessed. So the failure mode is not a warning banner on a chart, it is a hole in the middle of a table that nobody notices until a quarterly query returns a suspiciously flat week, and there is no backfill to repair it.

The usual answer is to filter events out of the export before they count against the cap, which works and means deciding today which events you will never want to query. The other answer is Analytics 360, which raises the cap to 20 billion and costs what an enterprise contract costs.

Step back from the five blind spots and they have one shape. Each is something that happened outside the browser: money was spent in an ad account, a salesperson moved a deal to closed won, a visitor declined a banner, the same human picked up a different device. The GA4 export is a faithful record of a browser, and every question that matters is about something the browser never saw.

Which means the modelling has to happen where those facts exist, and the warehouse gets the result. That inverts the usual architecture, and it is the right way round. BigQuery is superb at storing an attributed dataset, serving it to Looker Studio, Power BI and Tableau, joining it to finance data and holding years of it cheaply. It is a bad place to reconstruct a customer journey from browser logs, because the inputs are not there and no query language adds a column that was never exported.

Nothing in the right-hand column is exotic. It is what an attribution product is for, and the only real decision is whether the finished data stays in that product's interface or lands in a dataset you own. It should land in a dataset you own, which is the whole argument for a marketing data warehouse in the first place.

Tracking runs server-side on your own domain at 95%+ accuracy, so the click IDs are captured at the first visit and kept: gclid with gbraid and wbraid, fbclid, li_fat_id and msclkid, plus the UTM parameters and the landing page. That anonymous first click is joined to a person at the form fill, the call or the booking, and from there the record follows your CRM stages to the closed deal. Native integrations cover HubSpot, Salesforce, Pipedrive, Close, Attio, GoHighLevel, ActiveCampaign and Odoo, and anything else connects by webhook.

The BigQuery export is then switched on from the Apps tab: paste a service account key with write access to your dataset, map your columns, and clicks, conversions and leads stream in continuously, each already carrying the campaign it belongs to and the CRM stage it reached. The dataset is in your own Google Cloud project. There is no pipeline of ours to maintain and no scheduled job to babysit, and because Looker Studio, Power BI and Tableau read BigQuery natively, they reach the attributed data without a connector from us. That is a fact about BigQuery rather than a feature of ours, and it is worth saying plainly.

What you write in the warehouse changes shape as a result. The eighty lines above become a `GROUP BY` over rows that already know their campaign and their revenue, and the interesting queries stop being about reconstructing journeys and start being about the business: cohorts by close month, payback by channel, pipeline by first touch against last touch, marketing joined to the finance tables that were already in there.

Two honest limits. The export is on the Scale and Enterprise plans, not the entry tier, so it is priced for teams that own a warehouse. And Snowflake is on the roadmap rather than shipped: today the warehouse destination is BigQuery, and saying otherwise to a data team would be found out in the first call. Setup takes about 21 minutes, and you can click through the whole product before talking to anybody.

Whichever route you take, these six checks take an afternoon and decide whether the project is worth starting.

1. Run the consent query. If a third of your events carry denied storage, a warehouse model is measuring two thirds of your marketing and no SQL recovers the rest.
2. Run the fragmentation query. The browsers-per-user ratio is the floor on how badly paths are being split, and it decides whether multi-touch is even meaningful on your data.
3. Count your daily events against the cap. One million a day on a standard property, and the export pauses without a backfill.
4. Name the cost pipeline. Which tool pulls spend from which platforms, on what schedule, at what price, and who fixes it when a campaign gets renamed.
5. Name the revenue join key. Not the plan to invent one later. The field that exists today in both the CRM and the web data, written at the moment of the form fill.
6. Decide what the warehouse is for. If the answer is reporting and joining to finance, it is the right tool. If the answer is reconstructing journeys from browser logs, the inputs are missing and the project will stall at the same place everyone else's does.

Related reading: the BigQuery integration for what actually lands in the dataset, the best marketing data warehouse for choosing one, multi-touch attribution for B2B lead generation for the models themselves, server-side versus browser tracking for why the collection layer decides the ceiling, gclid explained and UTM parameters explained for the join keys, GA4 revenue attribution for the commonest version of this problem, and LeadJourney vs Google Analytics 4 for the side by side.