What is data freshness? How to measure it and how fresh a decision needs

Freshness, latency and “last updated” are three different things

These three get used interchangeably, and the confusion is behind most stale-number incidents:

“Last updated today” does not mean “covers today”. A load can run at 06:00 and bring nothing past last Thursday, because the source stopped sending. A backfill can rewrite a table today with figures from two years ago. Freshness is about the dates inside the data, not the time on the table.

How to measure data freshness

Measure freshness on the dates in the data, per source, against the date you expected:

SELECT
  'ad spend' AS source,
  MAX(date) AS latest_day,
  DATE_DIFF(CURRENT_DATE(), MAX(date), DAY) AS days_behind
FROM `owox-demo.ecommerce.ad_performance`
UNION ALL
SELECT
  'orders',
  MAX(order_date),
  DATE_DIFF(CURRENT_DATE(), MAX(order_date), DAY)
FROM `owox-demo.ecommerce.orders`

Three rules make the number trustworthy:

• Use the event date, not the load time. MAX(loaded_at) tells you the pipe ran. MAX(event_date) tells you what the data covers.
• Measure every input, not the report. A report that reads three sources has three freshness values. The report’s freshness is the oldest of them.
• State the expected lag. “One day behind” is on time for a daily load and three days late for an hourly one. Write down what each source should be, then compare.

CURRENT_DATE() in BigQuery is UTC unless you pass a time zone. Use the zone your business day is counted in.

A ratio is only as fresh as its stalest input

On 14 October 2026 we ran the query above on the [Demo] E-commerce data (a demo shop whose orders are generated daily and whose ad spend comes from ad platform connectors). Orders ran to 14 October. Ad spend stopped on 10 October, four days behind.

Then we computed ROAS, revenue from completed orders divided by ad spend, two ways:

The “last 7 days” ROAS is three times too high, and nothing in it is wrong except the window. Seven days of revenue are divided by three days of spend. No query fails and no table is empty. A dashboard set to “last 7 days” shows 42.5 to anyone who opens it.

One window could be a fluke, so we repeated it. For every 7-day window ending on a Tuesday from 7 July to 29 September 2026 (13 windows), we computed ROAS with complete spend, then with the last 1 to 4 days of spend missing:

On this shop, every missing day of spend adds roughly a fifth to a third to a 7-day ROAS, and four missing days more than double it. The true ROAS over those windows averaged 17.6. A team that reads this dashboard on a Monday after a connector failed on Friday would see a number above 30 and might shift budget towards ads that did nothing different.

If spend were perfectly even, the inflation would be pure arithmetic: seven days of revenue over six, five, four or three days of spend, or +17%, +40%, +75% and +133%. The measured ranges are wider because real spend is not even across the week. The size of the error depends on how even spend is, so measure it on your own data with the same query shape. Which way it moves depends on which input is stale; the checks article covers both directions.

The fix is to compute ratios only over days every input covers. Cap the window at the oldest input’s latest day, and show that date next to the number:

WITH latest AS (
  SELECT LEAST(
    (SELECT MAX(date)
     FROM `owox-demo.ecommerce.ad_performance`),
    (SELECT MAX(order_date)
     FROM `owox-demo.ecommerce.orders`)
  ) AS covered_to
)
SELECT covered_to,
  DATE_SUB(covered_to, INTERVAL 6 DAY) AS covered_from
FROM latest

Use covered_from and covered_to as the report’s date range instead of “today minus 7”.

How fresh does a decision need its data?

“Real-time” is rarely the right target. The right freshness depends on how often the decision is made and what a stale day costs:

Two things follow from the table. Completeness matters more than speed for most decisions: a weekly review needs every source to cover the full week, not the last hour. And ad platform data is not always final on day one. Conversions can still be attributed to a past day after it is loaded, so a “fresh” yesterday can still change. Google documents its own processing times for Google Analytics data in its help page on data freshness.

Ask one question per report: what is the oldest data this decision can tolerate? Then check each input against it.

Making freshness visible to the person reading the number

Often nobody knows that ad spend is four days behind until a number looks wrong, and then the analyst goes looking for which load failed. The person reading the dashboard never finds out at all. Most stale-number incidents are not a pipeline failure; they are a number read without its date.

Three practices close that gap:

1. Show “data through” next to every number. Not “last refreshed”, but the latest day all inputs cover.
2. Check freshness before a report is delivered, not after someone notices. Data quality checks on every Data Mart shows a freshness check that fails on an empty table too, and how to get alerted.
3. Write an expected lag per source. For hourly and real-time targets, data freshness SLAs covers how to set and monitor them.

How freshness is shown in OWOX Data Marts

In OWOX Data Marts, every Data Mart shows a Data Last Updated value: when the source tables behind it last changed in your warehouse. It appears on the Data Mart page with a Check now button, as a sortable column in the Data Marts list, on the Models canvas, and in every MCP query response, so an AI assistant answering from a Data Mart gets the time with the data. The documentation is explicit about what it means: a storage timestamp, and “updated today” does not mean “covers today”.

We asked the Unified Ad Spend Data Mart in [Demo] E-commerce for its latest date through the MCP server on 14 October 2026. The response carried the answer, 10 October, and a data_last_updated block: source tables last written on 11 October at 20:39 UTC, coverage complete, with the time for each of the seven ad platform tables behind it. The same response instructs the assistant to say “the source tables were last updated” and not to describe the data as fresh up to that time. That is the distinction this article is about, enforced on the AI side: written on the 11th, covering to the 10th.

Each scheduled report run measures it and records it in the run’s history, so the time is attached to the delivery a business user received in Google Sheets or elsewhere. For the “covers today” half, a data quality freshness check on a timestamp field, or the threshold query above, tests the dates in the data.

Data Last Updated works on Google BigQuery, Snowflake, AWS Redshift, AWS Athena and Databricks, with storage-specific limits listed in the documentation. Snowflake, for instance, reports the hour of the last change and can trail by several hours.

You can try OWOX Data Marts on your own warehouse.

Data QualityCommon Data Quality Issues and How to Fix ThemOctober 14, 2026→Data QualityWho is Responsible for Data Quality? A RACI for Analytics TeamsOctober 14, 2026→Data QualityData Lineage: How to Trace a Number Back to Its SourceOctober 14, 2026→

See all articles →