A shopper browses on her phone, pays in a store with a card and later emails support from a work address. Three systems hold three records. Identity resolution decides whether they belong to one person.

What is identity resolution?

Identity resolution is the work of deciding whether records belong to one person. Klaviyo says it finds that emails, phone numbers and device IDs belong to one person, then merges them into one profile (Klaviyo). Teams often store the links in an ID graph (Hightouch).

The idea is older than marketing software. Fellegi and Sunter posed the problem in 1969. They asked how to spot records in two files that stand for one person, object or event (Fellegi and Sunter, 1969). The wider field is entity resolution. Senzing calls identity resolution a subset of it. It is limited to people (Senzing).

Security teams use the same words in a narrower way. NIST defines the term as collecting information about an applicant to tell one person apart from the rest of the population a provider serves (NIST SP 800-63A). It is the first of three proofing steps. This entry covers the marketing sense: linking records to one person.

Which identifiers does identity resolution use?

Resolvers work from identifiers. Some are exact and tied to a login. Others are signals that point to a link. The table sorts the common ones.

IdentifierUsual matchWhere it comes fromWatch for
Member IDExactSign-up, login, scan at purchaseDuplicate accounts
Email addressExactSign-up, checkout, email clickShared inboxes, letter case
Phone numberExactSign-up, support callsDifferent formats
Payment card or walletExact once linkedLinked to the accountShared cards
App or device IDExact after loginApp installShared devices
Browser cookieSignal until loginWeb visitsBrowser limits on storage
IP addressSignalWeb logsShared networks
Name and addressFuzzyForms, mailTypos and nicknames

Safari limits cookies. WebKit says it blocks all third-party cookies by default. It also deletes script-written cookies and storage after seven days without user interaction with the site (WebKit). A login ID does not expire with the browser. Segment’s docs add that its resolver treats identifiers as case insensitive, so User_123 and user_123 land on one profile (Segment).

What is the difference between deterministic and probabilistic matching?

Deterministic matching links records on an exact ID. A shared email address or member ID is one. Sayers and colleagues say this method sorts records into two groups. The groups are matched and unmatched (Sayers et al., 2016).

Exact matching is strict. In the paper’s small example, linking two files by first and last name finds one matched record. A simple edit-distance score treats Steele and Steel as close. With a cutoff of 0.5, it adds two more links.

Probabilistic matching weighs evidence. Fellegi and Sunter set three outcomes for a pair of records: link, non-link and possible link, with limits on two kinds of error. Sayers explain two terms. The u-probability is the chance that two unmatched records agree on a field. The m-probability is an indicator of data quality.

Rarity counts. In the paper’s example, a chance match on the surname Smith has a u-probability of 0.42. A chance match on Sayers has 0.06. So two records that agree on a rare name are stronger evidence than two that agree on a common one. Two cutoffs then sort pairs into links, non-links and potential links. More links mean more clerical review.

What goes wrong in identity resolution?

Two errors matter. A false merge puts two people into one profile. A missed link leaves one person in two profiles. Fellegi and Sunter define both: linking a pair that is not a match, and failing to link a pair that is. Sayers list typing errors and missing data among the causes of missed links.

Tools add guardrails. Segment’s docs describe merge protection, which spots issues such as non-unique anonymous IDs. They also describe ID rules that force uniqueness on chosen IDs (Segment).

Measure the result. Pull a sample of pairs, have a person label each one and count both errors. Sayers cover how to assess the quality of a linkage method (Sayers et al., 2016).

For illustration, a team audits 200 pairs of records known to be one person and 200 pairs known to be two people. Rule A links on exact email. It links 150 of the true pairs and 4 of the false pairs. Of its 154 links, 150 are right, or 97.4%. It finds 75% of the true pairs. Rule B adds phone. It links 180 true pairs and 10 false pairs. Now 180 of 190 links are right, or 94.7%, and it finds 90% of the true pairs. Rule B finds more people and makes more false merges. The team picks by what each error costs.

What are the privacy limits of identity resolution?

Match keys are personal data in the hands of a company that can link them. GDPR Recital 30 names online IDs such as IP addresses and cookie IDs. It says their traces, joined with unique IDs and other data, can be used to build profiles of people and to identify them (GDPR Recital 30).

Hashing a key does not make it anonymous. Recital 26 covers data that has been pseudonymized. If extra data could tie it to a person, it still counts as data on an identifiable person (GDPR Recital 26).

The EU’s former Article 29 Working Party lists hash functions as one way to pseudonymize data. It calls this a useful security step. It adds that the step does not make data anonymous (Article 29 Working Party, 2014).

Collect only what the match needs. NIST asks for the minimum identity evidence needed to distinguish a unique identity. GDPR Article 5 points the same way: personal data must be limited to what the purpose needs (GDPR Article 5). This describes the text of the rules. It is not legal advice.

How do loyalty programs use identity resolution?

A loyalty program resolves identity at sign-up. Starbucks lets a member add multiple cards and link a credit card, debit card or mobile wallet to one personal account. It reserves the right to combine accounts that look like duplicates. Guest purchases earn no Stars (Starbucks Rewards Terms). Each rule keeps one person tied to one record.

A working setup follows five steps. Pick the exact keys, in order of trust. Make their formats match. Link on exact keys first. Add probable links for the rest, with a cutoff and a review queue. Audit a sample each quarter.

The member ID is the exact anchor. Probabilistic links reach the browsing that happens before a login. Our CDP guide says a CDP stitches the anonymous browser, the store purchase and the email click into one persistent profile per person. It adds that program activity is among the richest signals a CDP gets, since each event is tied to a verified identity and not a cookie.

Programs also give people a reason to identify. Our first-party data guide says members scan or log in to earn. So every purchase and engagement lands on a verified identity.

How GRAVTY handles identity resolution

Our telecom page describes one identity per member. It is resolved across email, phone, card and app. It stays consistent across brands and markets.

Our grocery page says identity resolution across email, phone, card and app attaches the shopper to the basket. The same member object shows up in every banner and channel. A banner is one store brand in the group.

Frequently asked questions

What is identity resolution?

Identity resolution is the work of deciding whether records belong to one person, then linking them into one profile. Klaviyo lists emails, phone numbers and device IDs as common IDs (Klaviyo). Teams often keep the result in an ID graph (Hightouch).

What is the difference between identity resolution and entity resolution?

Identity resolution is entity resolution for people. The wider field also covers firms, places and products (Senzing). Both use the same kinds of matching, and both work from records.

What is the difference between deterministic and probabilistic matching?

Deterministic matching uses an exact ID. Probabilistic matching weighs the odds. The first sorts records into matched and unmatched. The second adds a group of possible links for a person to review (Sayers et al., 2016).

Does hashing an email make it anonymous under GDPR?

Hashing alone does not make data anonymous. Recital 26 of GDPR says pseudonymized data stays personal if more data could tie it to a person. The Article 29 Working Party lists hash functions as one way to pseudonymize (Working Party, 2014). This is not legal advice.

How do loyalty programs resolve identity?

They start with the member ID. Starbucks lets a member add multiple cards and link payment methods to one personal account. It also reserves the right to combine duplicate accounts (Starbucks Rewards Terms). Probable links then cover the browsing that happens before login.