3 November 2026 · 9 min read

GA4 Erasure Requests: What You Can Delete, What Survives, and What Breaks

A client forwards an email: a customer has asked to be forgotten, and the client wants to know how to delete them from Google Analytics. GA4 has two deletion tools, they remove different things, and neither finishes inside the month the law gives you to reply.

Most teams meet this problem with the clock already running, and the first conversation goes badly because the platform is slower than the promise. Here is what each tool does, what survives it, and how to answer without overpromising.

Two tools, two different jobs

A data-deletion request removes the text that event parameters collected. A user deletion removes the events attached to one identifier. They are not interchangeable, and they live on different screens.

Both need Editor or above at property level. Analysts and viewers cannot open either screen, so a client on a read-only seat has to escalate.

What a data-deletion request actually removes

It erases the text collected by event parameters and replaces it with the literal string "(data deleted)". The event itself survives and still counts in your metrics. A purchase that carried an email address in a parameter comes back as a purchase with "(data deleted)" where the address used to sit. Revenue, transaction counts and user counts do not move.

If someone expects the numbers to drop after a deletion, they will not. The report showing "(data deleted)" is working as designed, not failing.

Deleting every parameter on every event also deletes the event name, because the event name is itself a parameter. That is the heaviest option in the list, and it is worth reading twice before selecting it.

The five request types

The form asks you to pick one scope:

  1. Delete all parameters on all events. The broadest option, and the only one that can remove auto-collected parameters.
  2. Delete all registered parameters on selected events. Up to 100 events from what you have collected.
  3. Delete selected parameters on all events. Up to 100 parameters from the full historical list.
  4. Delete selected registered parameters on selected events. Useful when one field is wrong on one event.
  5. Delete selected user properties. Up to 100 user properties, separate from event parameters.

After the scope you set a start and end date. The deletion runs on event dates, not on the date the data arrived, and it uses your property's time zone to interpret the range. There is also an optional filter: only delete parameter values that contain a given string. It behaves as a contains operator, ignores case, and refuses reserved values such as "(not set)".

The grace window, and the clock behind it

Submitting a request does not start a deletion. It starts a preview. For the first seven days, Editor users get an email and anyone with Editor can cancel from the request details page. During that period your reports and explorations already omit the data named in the request, so you see the effect before it becomes permanent.

From day eight the status changes to preview active, deletion in progress, and cancelling stops being an option. A property can hold a maximum of 12 requests in the grace or pending state at once.

Processing takes between 7 and 63 days depending on volume, and data has to be more than 12 days old before it can be deleted at all. Once the request completes, the deletion is not reversible. Editor users are emailed when it is submitted, when the deletion begins and when it finishes.

Read those numbers against a calendar. A request submitted today can take two months to complete.

Fields the request will not touch

Some values are protected by design. Analytics does not delete numeric parameters, text parameters derived from internal trusted identifiers, or the reserved values "", "(not set)" and "(data deleted)".

Auto-collected parameters are only available under the first scope. Even then, a documented exclusion list stays in place: geography fields such as city, country and region, device fields such as browser and platform, audience attributes such as age and gender, coordinates, and link and file fields.

Custom parameters are deletable when they were registered as custom dimensions. If the field you want gone is not in the parameter list, it was probably never registered. One more edge: a custom dimension fed by an auto-collected parameter does not take the auto-collected parameter with it when deleted.

Dimensions archived before July 2022 need a workaround, because deletion of archived parameters was not supported then. The documented route is to delete all parameters across all events, or all parameters for a set of events.

Deleting a user instead

When the request is about a person rather than a field, the User explorer is the tool. Open Explorations, pick the User explorer template, find the row for the identifier, select it and use the Delete user icon. The data stops displaying in explorations within 24 hours and is permanently removed within the following 63 days.

Each row is an Effective user ID: the user ID when you have one, otherwise the device ID, which is the client ID on web and the app instance ID in an app. That has a practical consequence. If your reporting identity blends user ID and device ID, a logged-out session attaches to the device ID and will not appear under the user you are trying to delete. The documented fix is to switch Reporting identity to By device only, open a fresh user explorer, find the pseudonymous ID and delete that as well. Analytics does not map the two identifiers for you, so you need your own way of linking them.

For repeatable work there is an API. A POST to the Admin API endpoint properties.submitUserDeletion takes one identifier per call: userId, clientId, appInstanceId, or userProvidedData carrying a single email address or phone number. Emails have to be lowercased, with periods before the @ removed for gmail and googlemail addresses and all spaces stripped. Phone numbers keep digits only with a plus prefix. The response returns a deletion request time, described as the moment before which all visitor data should be deleted. The scope is analytics.edit.

What the deletion breaks downstream

Three consequences are easy to miss until a client asks why a report changed shape.

Campaign attribution first. Once the deletion completes, historical campaign information is gone, so attribution credit for those users moves to whatever campaign data still exists, or to Direct. A marketing team that has been reporting channel performance from the same period will see the past shift under them.

Conversion reporting second. If the request names a field used in conversion reporting, all conversion reporting data for that date range is deleted and the reports show no data for it at all.

Consent mode third. Google's own guidance says to add seven calendar days to the end date of the request so the deletion also clears behavioural models trained on consented data. The worked example in the documentation moves a 1 January end date to 8 January.

Then the property hierarchy, which catches people on 360 accounts. A subproperty inherits deletions made at its source property, but a deletion at the subproperty stays there. Roll-ups go the other way: a deletion at a source property does not apply to the roll-up, so you submit a separate request inside it, and a deletion at the roll-up does not touch the sources.

The date range is not the collection date

A deletion request works on the dates the events happened. Conversion reporting, by contrast, propagates values through lookback windows and builds reports from interaction dates. Google's documentation is direct about the consequence: identifying the right range is your job, and if you find more data that should have gone, you file another request with that range.

For a customer who converted in March and asked to be forgotten in November, the March events are in scope and the November activity referencing them may fall outside the range you picked. Build the range from the identifier's whole history.

The copy in BigQuery

Anything you exported to BigQuery is yours. Google's export documentation states that when you export data to BigQuery you own that data and manage permissions through BigQuery access controls. So the Analytics interface will not clean up the warehouse copy, and a request that stops at the admin screen leaves a second copy live in a table anyone with dataset access can query.

If BigQuery Export is running, the deletion work has two halves, and the SQL half sits with whoever owns the dataset. Deleting rows from a partitioned table, or rebuilding the affected partitions, belongs in the runbook alongside the request form.

Where this collides with a UK erasure request

Under Article 17 of the UK GDPR an individual can ask for their personal data to be erased, and the ICO's guidance is specific about the timetable. You respond without undue delay and within one month, and the deadline runs to the same calendar date in the next month. A complex request or a batch from one person buys two further months, but only if you tell them inside the first month and explain why.

GA4 can take 63 days. That is more than two months, and it starts after a seven-day grace window you cannot shorten. So the honest position is that the Analytics side may complete after the statutory deadline, and the reply to the requester has to say what you did, when it takes effect, and what remains in progress.

The ICO also covers the part people skip. Erasure has to reach backup systems as well as live ones, and where data cannot be overwritten immediately it has to be put beyond use: not read, not processed, held only until the retention schedule replaces it. And you have to be clear with the individual about what will happen to their data when the request is fulfilled. A short written description of the deletion, the date range, the identifier used and the completion estimate satisfies that far better than a promise that the data is gone today.

A runbook that survives an audit

This is the sequence that keeps the timeline defensible:

  1. Log the request the day it arrives, with the identifier the requester gave you and the date you received it.
  2. Decide the scope. A field problem takes a data-deletion request. A person problem takes a user deletion, possibly two, one for the user ID and one for the device ID.
  3. Pull the identifier's history to set the date range, and extend the end date by seven days if consent mode modelling applies.
  4. Submit, then tell the requester what was requested and the documented processing window. Do not say the data is deleted.
  5. Handle the warehouse copy separately if BigQuery Export is on, and note who did it.
  6. Keep the completion email. It is the evidence that the request ran.

Erasure requests landing on your desk with no process behind them? Book a free tracking audit and we will map what your properties collect, where the copies live, and the deletion route for each one.

The part worth fixing before the next request

Every hard erasure case starts with a field that should never have been sent. Email addresses in event parameters, names in page titles, free text from a form pushed into a custom dimension: once they are in, removal is a request, a wait and a partial outcome, and the event still counts afterwards.

The cheap version of this is a short list of what your tags may collect, checked when a tag is added. The expensive version is the one in your inbox now.