New BigQuery tools for understanding metric changes

From noticing change to investigating it

On September 14, 2026, Google introduced augmented analytics functions for BigQuery. They address drivers of metric changes, change points, trends, and seasonality. Structured SQL outputs can become part of analytical workflows used by tools and AI agents.

AI.KEY_DRIVERS examines contributions to metric changes, while ML.DETECT_CHANGE_POINTS identifies structural shifts. Multiple analysis stages can be connected. These functions simplify investigation, but management interpretation still requires understanding the data and business context.

Automated analysis needs a precise question

Our suggested starting point is a stable metric definition. Lower sales might reflect changed order recording, delayed ingestion, or a real event. A precise function cannot make a comparison valid if the underlying definition differs between periods.

Choose comparison windows carefully. Holidays, customer mix, and channel changes can affect findings. Write down known factors and assumptions first so the team can distinguish a new insight from an expected pattern.

Correlation is not proof of causation. Coinciding changes can provide a lead, while action needs additional evidence. Present numerical findings with data limitations, analytical assumptions, and human interpretation to avoid overstating an automated result.

SQL examples for investigating a metric in BigQuery

These two queries use the public Austin Bikeshare dataset: first locate changed periods, then compare selected dimensions before and after a date. They are teaching examples; check function access, query cost, and the comparison window before running them.

1. Find changed periods with ML.DETECT_CHANGE_POINTS

WITH daily_trips AS (
  SELECT TIMESTAMP_TRUNC(start_time, DAY) AS trip_day,
         COUNT(*) AS trip_count
  FROM `bigquery-public-data.austin_bikeshare.bikeshare_trips`
  GROUP BY 1
)
SELECT begin_timestamp, end_timestamp,
       metrics.avg AS avg_daily_trips
FROM ML.DETECT_CHANGE_POINTS(
  (SELECT * FROM daily_trips),
  data_col => 'trip_count',
  timestamp_col => 'trip_day'
);

The result gives each period’s start and end and its average daily trip count; it does not establish the cause.

2. Inspect contributors with AI.KEY_DRIVERS

WITH trips AS (
  SELECT bike_type, subscriber_type,
         1 AS trip_count,
         DATE(start_time) >= DATE '2018-02-11' AS after_change
  FROM `bigquery-public-data.austin_bikeshare.bikeshare_trips`
  WHERE DATE(start_time) BETWEEN DATE '2018-01-12'
                             AND DATE '2018-03-13'
)
SELECT drivers, contribution
FROM AI.KEY_DRIVERS(
  (SELECT * FROM trips),
  metric_col => 'trip_count',
  interest_label_col => 'after_change',
  dimension_cols => ['bike_type', 'subscriber_type'],
  top_k => 10
);

This compares membership and bike types across two periods. A statistical contribution alone is not proof of causation.

Test a familiar metric

Select a metric with reliable history and compare the new analysis with the team's manual investigation. Measure execution cost, time to insight, and explainability. This trial shows where the capability can help the organization's existing analytical process.

Source publication date: 2026-09-14. Practical explanations and recommendations are Liyan Knowledge editorial analysis.Sources: Google Cloud — BigQuery Augmented Analytics TVFs