Quick answer
Anonymization means processing data so people can no longer be identified by anyone using means reasonably likely to be used. Removing names is not enough, and a salted hash that stays stable is usually only pseudonymous. The test is whether a record can still be singled out, linked or inferred.
Formula: k = size of the smallest group of records sharing the same quasi-identifier values
What anonymization means
Anonymization is processing data so that the people in it can no longer be identified, by anyone, using any means reasonably likely to be used. The test is not whether the name column was removed; it is whether someone can still pick a person out, link their records, or learn something about them. If they can, the data is not anonymous, whatever it is called.
The EU's GDPR sets the standard in Recital 26: to decide whether a person is identifiable, account is taken of all the means reasonably likely to be used, by the controller or by anyone else, and data rendered anonymous in such a way that the person is no longer identifiable is outside the regulation. Data that has only been pseudonymised, where extra information could link it back to a person, remains personal data. This page explains the concepts; whether a particular dataset counts as anonymous is a legal judgement that depends on your data and context.
Anonymization, pseudonymization, masking and aggregation
These terms are used loosely, and the differences decide what you are allowed to do with the data.
| Technique | What it does | Can it be reversed or linked? | Likely status under GDPR |
|---|---|---|---|
| Removing the name column | Drops direct identifiers, keeps the rest | Often yes: other fields identify people | Usually still personal data |
| Pseudonymization | Replaces identifiers with a code, such as a keyed hash or a random id with a lookup table | Yes, by whoever holds the key, salt or table, and records stay linkable to each other | Still personal data |
| Masking | Hides part of a value, for example the last digits | Often partly | Depends on what remains |
| Aggregation | Reports counts or totals for groups, not rows | Not at the individual level, if groups are large enough | Can be anonymous, if the raw rows are gone |
| k-anonymity | Generalises fields so each person shares a profile with at least k others | Resists singling out; does not stop inference | Can contribute, rarely enough alone |
The Article 29 Working Party's opinion on anonymisation techniques (WP216) puts it directly: pseudonymisation is not a method of anonymisation. It merely reduces the linkability of a dataset with the original identity, and is a useful security measure. It also notes that if the controller keeps the original event-level data, a dataset handed out with identifiers removed or masked is still personal data; only aggregating to a level where individual events are no longer identifiable can make it anonymous.
The three tests: singling out, linkability, inference
WP216 describes an effective anonymisation as one that prevents anyone from doing three things:
| Test | Question | Example of failing it |
|---|---|---|
| Singling out | Can someone isolate one person's record? | A table with one row for the only visitor from a small town on one device |
| Linkability | Can two records be tied to the same person, within the data or across datasets? | The same visitor id on Monday and Tuesday; or an id that matches one in another dataset |
| Inference | Can a value be deduced about a person, even without finding their row? | Every visitor in a group converted, so anyone known to be in the group converted |
A dataset has to pass all three. Removing names addresses none of them reliably. Sweeney's study of US census data shows how little it takes: 87% of the US population (216 million of 248 million) had characteristics that likely made them unique on just 5-digit ZIP code, gender and date of birth, three fields that contain no name.
Why a salted hash is usually pseudonymous
A common shortcut is to replace an identifier with a hash and call the result anonymous. A hash is a one-way function, but one-way is not the same as unlinkable, for two reasons.
It can be replayed. WP216 notes that if the range of input values is known, they can be run through the hash function to derive the right value for a record, and that hash functions are designed to be fast, so brute force works. An IPv4 address has only 2<sup>32</sup> = 4,294,967,296 possible values, a small and fully enumerable set, so hashing IPs without a secret leaves them recoverable. WP216 adds that salting can reduce the likelihood of deriving the input but calculating the original value may still be feasible with reasonable means; a keyed hash, where a secret key is the additional input, is much harder to replay without the key.
It stays linkable. Whoever holds the key or salt can recompute the identifier, and the same input always gives the same output while the salt does not change. Sample data: take the documentation address 203.0.113.7 and one user-agent string, and compute an HMAC-SHA256 over them, keeping the first 12 characters.
| Salt | Monday id | Tuesday id | Same person linkable across days? |
|---|---|---|---|
| One salt kept for the site | 02bd63c5ae73 | 02bd63c5ae73 | Yes: identical ids |
| A new salt each day | 2fe13c063131 | 0744665c2ffc | No: the ids differ |
The first row is pseudonymous: the id is stable, so the person can be followed and counted as returning. The second row cannot be followed from one day to the next, but within a day the id still singles out a browser, and WP216 still lists a keyed hash with deletion of the key among pseudonymisation techniques that reduce linkability. The practical rule: ask who could recompute or link the identifier, and over what period. If anyone can, call it pseudonymous and handle it as personal data.
A k-anonymity check
k-anonymity groups people so that each shares the same values with at least k−1 others. Sample data: 1,000 visits grouped by country and device type.
| Group (country + device) | Visits |
|---|---|
| A | 412 |
| B | 288 |
| C | 150 |
| D | 96 |
| E | 41 |
| F | 13 |
| Total | 1,000 |
The value of k is the smallest group, here 13, so every visit shares its profile with at least 12 others, and 41 + 13 = 54 visits (5.4%) are in groups under 50. Now add exact browser version and screen size to the table. Suppose each of the 13 visits in group F had a different combination: group F splits into 13 groups of 1, and k falls from 13 to 1. Every extra field shrinks the groups, which is why releasing detailed fields next to a coarse "k = 13" claim is a common failure (missing quasi-identifiers).
k-anonymity also fails the inference test. If all 13 visits in group F bought something, then knowing someone belongs to group F is enough to know that they bought, without finding their row.
Common mistakes
- Removing names and stopping. Other fields identify people, as the ZIP, gender and date of birth figure shows.
- Hashing an identifier with no secret. If the input set is known, as with IP addresses or phone numbers, the hash can be reproduced.
- Treating a stable salted id as anonymous. If the same visitor gets the same id over time, the data is linkable.
- Keeping the raw data and releasing the cleaned copy. If the original event-level rows still exist, the released set stays personal data in WP216's reading.
- Treating truncation or masking as anonymization. They reduce detail; they do not by themselves rule out singling out or linkage.
- Checking only once. New data, new fields or new outside datasets can make a once-anonymous release identifiable again.
- Collecting more than the question needs. The cheapest anonymization is not holding the field at all; see data minimization.
- Using the word as a label. "Anonymous" in a privacy policy is a claim someone can check with the three tests.
Anonymization in TapCub
Take the salt table and look at how TapCub's own settings map onto it. TapCub names the choices in its documentation; this section describes what the collector does, not a legal conclusion.
- Know how the visitor id is made. The collector computes an HMAC over the visitor's address and user agent with a salt. The website script writes no cookie and no local storage (see cookieless analytics), and the address is used for the hash and for geography and then discarded; it is never written on the event row.
- Pick the salt behaviour. In Configuration › Settings › Privacy, the Balanced preset keeps a site-level salt, so a returning visitor keeps the same id (the first row of the table above). Strict rotates the salt each day, so the same browser is a different visitor tomorrow (the second row), and new vs returning, visitor profiles and retention are unavailable. If you need returning visitors, you need the stable id, and the data should be treated as pseudonymous.
- Check the collection tier. In Configuration › Privacy center, tier 1, named "anonymous aggregate", rotates the salt daily, keeps no persistent id and drops autocapture and identity operations. Tier 2, named "cookie-free pseudonym", separates visitors with an HMAC of address and user agent; it is the default. Where a region rule requires consent and none was given, the collector falls to tier 1 and still stores pageviews and custom events.
- Keep identifiers out of properties. With stripping on (the default), strings that look like an email address or a phone number in event properties are replaced by a placeholder.
- Plan for requests. Because the default stored data is pseudonymous, the privacy center also handles data-subject requests (access, erasure, restrict, portability).
Changing a preset affects only data collected afterwards; rows already stored are not rewritten. TapCub's documentation does not claim that any tier makes data legally anonymous; whether a tier meets that standard in your case is for your own legal review.
See also: What is data redaction? Methods and a chat transcript example.
Sources
- GDPR Recital 26, Not applicable to anonymous data: means reasonably likely to be used; pseudonymised data remains personal data.
- Article 29 Data Protection Working Party, Opinion 05/2014 on Anonymisation Techniques (WP216): singling out, linkability, inference; pseudonymisation is not anonymisation; hashing and salted hashing; k-anonymity.
- Latanya Sweeney, Simple Demographics Often Identify People Uniquely, Carnegie Mellon University Data Privacy Working Paper 3, 2000: 87% of the US population unique on ZIP, gender and date of birth.
Was this explanation clear?
We use your answers to fix and expand this page.
Couldn’t send. Please try again in a minute.
Thanks for the feedback. We’ll use it to improve this page.You’ve already sent feedback on this page. Thanks.

Data Engineer
Writes about event design, data quality, and collection workflows, with verifiable implementation methods for reliable analytics.