BigQuery's 2025 Features I Found Interesting - 今年出たBigQueryの機能で個人的に面白かったもの - ブログ

This is the 21st article of the JADE Advent Calendar 2025.

Yesterday's article was Hasegawa-san's PoC on adding visualization features to an AI agent

— worth checking out if you're curious.

Happy holidays!

I'm Jeremiah, and I work as a data engineer here at JADE.

I usually work between the consultants and engineering teams, helping build and maintain data pipelines when needed.

For this year's Advent Calendar, I didn't want to write anything too heavy.

Instead, I'll walk through some BigQuery features that came out in 2025 and explain why they felt genuinely useful from a data engineering and analytics perspective—especially around AI.

At JADE we frequently use GA4 data, so to keep things concrete I'll use one public GA4 dataset throughout as examples.

AI.FUNCTIONS — AI directly inside BigQuery

The biggest "oh wow" feature for me this year was AI.FUNCTIONS.

In short, you can now call LLMs directly from BigQuery SQL.

That means things like:

Classifying text
Extracting structure from messy strings
Generating summaries

…all without exporting data or standing up another service.

Classifying GA4 landing pages (AI.CLASSIFY)

Let's say we want to automatically classify landing pages by intent (product, blog, checkout, etc.), without maintaining a big rules table using the AI.CLASSIFY function.

WITH
  landing_pages AS (
    SELECT
      ep.value.string_value AS page_location
    FROM
      `bigquery-public-data.ga4_obfuscated_sample_ecommerce.events_*`,
      UNNEST(event_params) AS ep
    WHERE
      event_name = 'session_start'
      AND ep.key = 'page_location'
    LIMIT 100
  )
SELECT
  page_location,
  AI.CLASSIFY(
    page_location,
    ['Product', 'Category', 'Blog', 'Checkout', 'Other']) AS page_type,
  COUNT(*) AS sessions
FROM landing_pages
GROUP BY page_location, page_type
ORDER BY sessions DESC

This gives you an AI-generated dimension you can immediately use downstream, for example like a distribution graph above.

What I like here is how normal and easy to use it feels:

It works inside Dataform
Incremental models still work
IAM, billing, and logging all stay in BigQuery

It doesn't feel like "AI glued on the side".

It feels like SQL just got a new function.

Scoring GA4 sessions by engagement quality (AI.SCORE)

Labels are nice, but in analytics you often want a numeric signal you can rank or threshold.

That's where AI.SCORE fits really well.

Instead of defining rigid rules for "engaged sessions", we can let the model give us a continuous engagement score.

WITH session_summary AS (
  SELECT
    user_pseudo_id,
    COUNT(*) AS event_count,
    COUNTIF(event_name = 'purchase') AS purchases
  FROM `bigquery-public-data.ga4_obfuscated_sample_ecommerce.events_*`
  GROUP BY user_pseudo_id
  LIMIT 100
)

SELECT
  user_pseudo_id,
  event_count,
  purchases,
  AI.SCORE(
    CONCAT(
      'Events:', event_count,
      ', Purchases:', purchases,
        'How engaged does this session look from 1 to 10?'
    )
  ) AS engagement_score
FROM session_summary
ORDER BY engagement_score DESC

Instead of a binary "engaged / not engaged" flag, you get a ranking signal:

Higher scores tend to correspond to sessions with more interaction or conversions
Lower scores tend to look like quick exits or low-intent sessions

This isn't meant to replace GA4's built-in engagement metrics —

it's more of a cheap, fast heuristic you can layer on top of existing logic to see a quick overview of sessions.

TimeFM — forecasting GA4 traffic without suffering

Another feature I ended up liking more than expected is TimeFM, Google's foundation model for time-series forecasting.

Using the same GA4 data, we can build a daily sessions table and forecast traffic with very little setup.

SELECT
  *
FROM
  AI.FORECAST(
    (
      SELECT
        PARSE_DATE('%Y%m%d', event_date) AS session_date,
        COUNT(DISTINCT user_pseudo_id) AS sessions
      FROM `bigquery-public-data.ga4_obfuscated_sample_ecommerce.events_*`
      WHERE event_name = 'session_start'
      GROUP BY session_date
      ORDER BY session_date
    ),
    data_col => 'sessions',
    timestamp_col => 'session_date',
    horizon => 30)

No feature engineering.

No separate training pipeline.

No extra infrastructure to maintain.

Is it perfect? No.

Is it often good enough for basic forecasting, baselines, or anomaly detection? Absolutely.

One GA4 dataset, a lot of capability

Stepping back a bit, what I like about these examples is that they all use the same single public GA4 dataset, with no special setup.

Using just:

bigquery-public-data.ga4_obfuscated_sample_ecommerce.events_*

we were able to:

Enrich GA4 data with AI-generated page classifications
Create a lightweight engagement quality signal using AI.SCORE
Forecast sessions using TimeFM

All of this stays inside BigQuery:

No exporting data
No extra services to maintain
No changes to existing IAM or billing models

It feels less like "adding AI to analytics" and more like analytics itself getting more expressive just within BigQuery.

If you're already working with GA4 data in BigQuery, these features slot into existing pipelines surprisingly naturally.

Final thoughts

BigQuery in 2025 feels like it's moving beyond "just a data warehouse".

AI inference, and forecasting are becoming native, SQL-accessible features, not bolt-ons.

From a data engineering perspective, I really like this direction:

Fewer moving parts
Easier governance
Faster iteration from raw GA4 data to insight

If you already live in BigQuery, these features let you do some pretty powerful things without changing how you work.

Happy holidays, and thanks for reading! 🎄