Why numbers need a history
Measurement changes constantly, and most changes are invisible in the data until someone compares two periods. A consent banner redesign changes how many visitors are measured. A container release fixes a purchase tag that was firing twice, and recorded revenue drops overnight. A checkout moves to a new domain and sessions split. A platform renames a metric or changes a default. A KPIA metric tied directly to an objective. is redefined at a planning meeting. Each is sensible on the day. A year later, the year-on-yearComparing a period with the same period a year earlier, aligned by weekday and by holidays that move, so that the seasonal pattern drops out of the change. report shows a step on that date and nobody remembers why.
Several parts of a measurement practice already point at the same need. The tracking plan guide versions the specification and records what changed in it. The tracking outage SOP ends each incident by marking the outage window. The BigQuery export basics guide tells readers to note the date the export was linked in the measurement log, because it is the first date any query can return. The dashboard design guide asks for notes on charts where a line dropped on a known date. Each assumes a shared log exists. This tutorial creates it. Once it exists, the measurement governance playbook gives it a monthly read against the version list and a place at every release, outage and audit, with a named owner for each. In general terms it is a change logA dated record of changes made to tracking, accounts, reports, lists or rules, with who made each and why, so later movements in the figures can be traced.: a dated record of each change, who made it and why.
Before you start
- A tracking plan. The tracking plan lists the events and parameters the log will refer to; the log records when they changed, and the plan records what they became.
- A list of the systems in the measurement stack. The GA4Google's event-based analytics platform. property, the Tag Manager container (GTMTag management system.), the consent platform, the site and its checkout, the ad accounts, the warehouse and the dashboards, each with an owner.
- Access to history. Read access to the container's version list, the release or ticket system, the ad accounts and any project notes from the last year.
- An owner for the log. One person, usually the analytics lead, who checks it monthly and chases gaps.
- Agreement from the people who make changes. Developers, agencies and campaign managers need to know that a change is not finished until it is logged.
Choose one home for the log
Pick a single place that everyone who reads reports can open and everyone who changes measurement can write to. A shared spreadsheet is the usual choice because it is easy to edit and dashboards can read it; a table in the warehouse works for teams that already live there. What matters is that there is one log for the whole stack, not a note in each tool, and that it is structured as a table with one row per change, so that a dashboard or a query can read it. Name it plainly, put the link in the measurement plan and the dashboard footers, and give the owner edit rights over its structure while everyone else adds rows.
Define the entry types
List the kinds of change the log records, so that people know what counts. Eight types cover the usual sources: container and tag releases; consent and banner changes; analytics settings and filters, including key events, custom definitions and the internal traffic rules the GA4 setup tutorial creates; site releases and migrations that touch tracked pages, the checkout or domains; campaign launches and pauses large enough to move totals; outages; changes to metric definitions, attribution rules and channel groupings; and changes announced by the platforms themselves, found in the Analytics help centre and the Tag Manager help centre or in their notices. Add a ninth, trading events such as price changes and stock-outsA period when a product is out of stock and cannot be sold, which caps sales and should stop advertising for that product until stock returns., if the log will also serve people explaining performance; mark it clearly so measurement changes can be filtered on their own.
Define the fields
Each entry needs enough to be found by date, understood without asking anyone, and traced to its evidence. Keep the field list short so that writing an entry takes a minute, and put the list in the header row of the log.
Field What to write Why it matters Date and time, with time zone When the change went live, not when it was planned Reports in different time zones put a change on different days End date For outages and temporary changes, when they ended Windows, not points, explain gaps Type One of the defined entry types Filtering and review System The property, container, site, account or dashboard changed Readers check only what touches their report What changed One or two plain sentences Understood without the ticket Expected effect Which metrics should move, in which direction, in which reports The test the monthly review applies Observed effect Filled in later: what actually moved Turns the log into evidence Link The container version, ticket, pull request or announcement The evidence behind the entry Owner Who made the change Who to ask Status Planned, live, reverted or backfilled Planned entries warn readers ahead of time Backfill the last year from versions and tickets
A log that starts today explains nothing in this year's comparisons. Backfill a year, source by source. The container's version list records every version with who created it and when, as the containers and workspaces guide describes; open each version with a name or description that mentions tags, triggers or consent and log the live date. Then search the ticket system for releases that touched forms, checkout, consent, domains or templates. Add the export link date, the dates of the tracking plan's versions, large campaign launches from the ad accounts, and outages from incident reports. Mark every backfilled row with that status and give a date range wherever the exact day is uncertain. Finally, plot the main conversion and session series by day for the year and look for steps with no entry beside them; each is a change still missing from the log.
Make logging part of every release
The log stays complete only if writing an entry is part of making a change, not a task for afterwards. Add "measurement change log updated" to the release checklist, the container publish checklist and the campaign launch checklist. Use the same text for the container version description and the log entry, so one piece of writing serves both. Log planned changes before they go live with the planned status, so that people reading reports that week know what is coming, then switch the status on the day. Changes to Consent ModeA Google tag feature that adjusts what tags collect according to the visitor's consent choices. settings and banner wording deserve the same treatment as tag changes, because they move every measured figure at once. A tracking plan change still goes through the plan's own change control; the log records only when it reached the data. Shipping the winner of an A/B test is a release like any other, and the tutorial on shipping a winning test variant logs it with the test's interval as the expected effect.
Connect the log to dashboards
A log nobody sees at the moment of reading a chart does little. The dashboard design guide notes that an events table blended in from a sheet the team keeps is the approach that scales, because the same entries appear on every page that shows the period. Connect the log as a data source, filter it to the systems and metrics each page reports, and show the relevant entries beside the time series as a short table or as labelled reference points. Add a line to each dashboard's notes pointing to the full log. If the team also marks dates inside analytics or ad tools, treat those marks as copies of log entries; the log remains the source.
Use the log when a metric moves
Make reading the log the first step of every investigation. When a metric steps, open the log for the days around the step and filter to the systems that feed the metric: often the answer is there, and the investigation becomes a check that the observed effect matches the expected one. The weekly checks in the data quality guide should do the same. Every analysis that compares periods should list the log entries that fall inside either period in its caveats, and a year-on-year comparison should check both years. When a movement has no entry, treat it as either a real change in the business or a missing entry, and find out which before reporting it.
Review it monthly
Once a month the owner reads the month's entries and does four things. Fill in the observed effect for each entry, comparing it with the expected effect, and open a ticket where they differ. Compare the container's version list and the release log with the entries, and chase any change that was made but not logged. Close outage windows that are still open. And add a short summary of the month's measurement changes to the monthly reporting pack, so that readers of the pack know which movements to discount. The review takes minutes when the log is kept well and is the early warning when it is not.
A year of entries, read (illustrative)
The example below shows a backfilled log explaining a year of conversion data that the team had stopped trusting.
Driftmallow Cushions explains three steps
Driftmallow Cushions is an invented online soft-furnishings shop. Its year-on-year report showed three steps in recorded purchases that nobody could explain, and the marketing team had started to discount the whole series. The new analyst backfilled the log from 63 container versions, 41 release tickets and the ad accounts, and three entries lined up with the steps.
| Date | Type | What changed | Effect on reported figures |
|---|---|---|---|
| 11–14 Feb (backfilled) | Consent | Banner redesign moved the reject option to the first layer | Measured sessions down 18% from the next day; recorded purchases down 15%; order system flat |
| 3 Jun 14:20 BST | Container release, version 37 | Purchase tag moved from the thank-you page view to the data layer purchase event | Recorded purchases down 7%, ending double counts from reloaded confirmation pages; analytics revenue now within 2% of the order system |
| 22 Sep to 6 Oct | Site release, then outage | Checkout moved to a payment subdomain missing from the cross-domain list | Purchases credited to the shop's own domain as a referral for two weeks; paid search key events down 31% in the window |
With the entries in place, the analyst rebuilt the year-on-year comparison: sessions compared from mid-February onwards only, purchases before June adjusted for the double counts, and the September window excluded from channel figures. The adjusted series showed purchases up 6% on the year rather than the 9% fall in the raw report. The consent entry also explained why ad platforms and analytics diverged from February, which the paid media agency had been asking about for months.
Keeping the log alive
Logs die in a familiar way: a burst of enthusiasm, a few weeks of good entries, a busy month in which nobody writes anything, and a quiet decision that the log is unreliable and not worth maintaining. Four habits prevent it.
- Make it cheap to write. A short field list, a dropdown for the type and the system, and a template line in the release checklist. An entry should take a minute.
- Make it visible when read. Entries that appear on dashboards and in analysis caveats get written, because people see their absence.
- Check it against the evidence. The monthly comparison with the version list and the tickets catches gaps before they become a year of them.
- Hand it over. When the analyst, the developer or the agency changes, the log and its owner are part of the handover, and the new owner reads the last year of entries in their first week.
Keep entries even when they turn out not to matter. A change that moved nothing is still evidence, and a reverted change still happened. Archive old years rather than deleting them, and keep the type list and Naming ConventionAn agreed pattern for naming campaigns, ads, events and tracking parameters so that reports can group them reliably. for systems stable, so that a filter written today still works next year.
Verify it worked
- Filter the log to one system, such as the container. Each row should show a date with time zone, a type, an expected effect and a link.
- Compare the container's version list with the log. Every version that touched tags, triggers or consent should have a matching entry.
- Check the backfilled rows. Each should carry the backfilled status, with a date range wherever the exact day is uncertain.
- Plot the main conversion and session series by day for the year. No step should sit without an entry beside it.
- Open one dashboard page. The relevant entries should appear beside the time series, with a note pointing to the full log.
If it goes wrong
- A step in the data has no entry. It is either a real business change or a missing entry. Ask the system owners, then log the answer.
- Entries stop appearing after a few weeks. Check that the release, container publish and campaign launch checklists still carry the "measurement change log updated" line.
- The dashboard shows no entries. Check that the log is connected as a data source and that the filter matches your system names.
- Observed effect differs from expected effect. Open a ticket and read the linked version to see what actually shipped.
- A change lands on different days in different reports. Add the missing time and time zone to the entry.
Next steps
- Read the data quality guide to bring the log into the weekly checks and start each investigation with it.
- Read the tracking outage SOP to see how each incident ends with a marked outage window.
- Read the dashboard design guide to place log entries beside the time series on every page.
Common mistakes
- One log per tool. Notes scattered across the container, the property, the dashboards and a wiki are never read together.
- Logging after the fact. Entries written weeks later have wrong dates and missing reasons.
- No expected effect. Without it, the monthly review cannot tell whether a change did what it was meant to.
- Dates without times or time zones. A change at night lands on different days in different reports.
- Recording measurement changes only. Without large campaign launches and outages, half the steps stay unexplained.
- Leaving it off the dashboards. A log that has to be looked up separately is looked up rarely.
- Starting from today. Without a backfill, this year's comparisons are still unexplained.
- No owner. Shared responsibility for a log means nobody checks whether it is complete.
The measurement planning hub places the log beside the measurement plan, the tracking plan and the data quality routine it supports.
