Identity resolution is the process of linking identifiers that appear across devices, channels and data sources, such as email addresses, phone numbers, cookie IDs, mobile advertising IDs and customer numbers, into a single profile of one person or household. In privacy in marketing, it is the capability that makes cross-channel personalization and measurement possible, and it is also one of the highest-risk data processing activities a marketing team can run.
The output is usually an identity graph: a data structure in which nodes are identifiers and edges record the evidence that they belong together. Because every link adds information about a person, identity resolution almost always processes personal data under laws such as the GDPR and US state privacy laws, even when individual identifiers are hashed.
How identity resolution links identifiers
Identity resolution works by collecting identifier pairs observed together and then clustering them into entities. When a customer logs in on a phone and a laptop with the same email, both device identifiers are linked to that email, and through it to each other.
Two matching methods dominate, and most commercial graphs combine them.
| Method | Evidence used | Accuracy | Privacy profile |
|---|---|---|---|
| Deterministic matching | Shared login, email, phone or customer ID | High precision, limited reach | Transparent if disclosed at collection |
| Probabilistic matching | IP address, location, device traits, timing | Lower precision, wider reach | Opaque and close to fingerprinting |
| Hybrid graphs | Deterministic core extended by models | Tunable by confidence threshold | Depends on how inferred links are used |
Probabilistic links carry a confidence score, and the operator chooses a threshold. A low threshold increases reach but merges different people, while a high threshold keeps profiles clean but leaves many identifiers unresolved. These false merges are both a data quality problem and a privacy problem, because one person may receive messages based on another person’s behavior.
Graphs also need ongoing maintenance. Devices change owners, cookies expire and shared household computers blur the line between people, so well-run systems apply edge decay: links that have not been observed again within a set period, often a few months, lose confidence or are removed. Without decay, a graph slowly accumulates stale and incorrect links that no longer describe anyone accurately.
Identifiers and identity graph types
The identifiers feeding a graph differ in stability, scope and legal sensitivity. The table below summarizes the most common ones.
| Identifier | Stability | Typical source |
|---|---|---|
| Hashed email | High, persists for years | Logins, newsletters, purchases |
| Phone number | High | Accounts, loyalty programs, SMS |
| First-party cookie ID | Weeks to months, browser limits apply | Website analytics and tags |
| Mobile advertising ID | Resettable, often unavailable | Apps after user permission |
| Postal address | High, shared within a household | Orders and offline records |
| Universal ID tokens | Rotated or encrypted per use | Industry frameworks such as UID2 |
Graphs are also classified by who owns them. A first-party identity graph is built by a brand from its own customer relationships, often inside a customer data platform. A third-party graph is assembled by a data provider from many sources and licensed to others, which raises far greater transparency and consent questions.
Identity resolution compared with related approaches
Identity resolution is often discussed next to other techniques that address the loss of third-party cookies. They differ in whether they need individual-level linking at all.
| Approach | Links individuals across sources | Main privacy concern |
|---|---|---|
| Identity resolution | Yes, by design | Profiling and loss of user control |
| Data clean room | Yes, but inside controlled queries | Output leakage if rules are weak |
| Contextual advertising | No | Minimal, based on page content |
| Modeled conversions | No, uses aggregate models | Low, depends on model inputs |
| Privacy-preserving attribution | No, uses aggregated reports | Low, designed with noise and limits |
The trend is toward keeping resolution inside first-party boundaries and sharing only aggregated or encrypted outputs, rather than sending raw identifiers to partners.
Legal and privacy requirements
Most identity resolution qualifies as profiling under the GDPR when it is used to analyze or predict personal preferences. It needs a valid legal basis, which for cross-context marketing is usually consent, plus clear notice that data from different sources will be combined.
- Hashing is not anonymization: a hashed email is still a stable identifier and remains personal data.
- Purpose limitation applies: data collected for order delivery cannot simply be repurposed for cross-site advertising.
- US state laws such as the California rules treat sharing for cross-context behavioral advertising as an activity users can opt out of.
- Data minimization favors linking only what a stated use case requires.
- Subject rights require that access and deletion requests reach every linked identifier, not only the one the user named.
The last point is easy to underestimate. A deletion request that removes an email but leaves linked device IDs in the graph does not fulfill the request, so the graph must support cascade deletion across all edges.
Practical design choices
Teams building or buying identity resolution face a few recurring decisions. The following decision table summarizes common situations.
| Situation | Recommended choice | Reason |
|---|---|---|
| Logged-in customer base | Deterministic first-party graph | Accurate and explainable to users |
| Need for partner matching | Clean room or encrypted tokens | Avoids sharing raw identifiers |
| Pressure for higher match rates | Raise thresholds before adding probabilistic links | Limits false merges |
| Consent not obtained | Keep data unlinked and use contextual methods | Linking would lack a legal basis |
Vendors often advertise match rates, but a high rate built on loose probabilistic links can mean many merged people. Asking for precision measured against a verified sample gives a more honest picture than reach alone.
Understanding it simply
A shopper may visit a store website on a phone, open its emails on a laptop and buy in a physical shop with a loyalty card. To the company, these look like three strangers. Identity resolution is the work of recognizing that they are one person.
It is like a librarian noticing that three borrowing slips with slightly different signatures belong to the same reader. That can make service smoother, but it also means the company knows much more about one person than any single slip revealed, which is why privacy rules apply.
Applied in practice
A mid-sized fashion retailer has an online shop, a mobile app and 40 physical stores. Its marketing team wants to stop sending discount emails for items a customer already bought in a store. The team builds a first-party deterministic graph that links online accounts, app logins and loyalty cards through verified email addresses, and it asks for consent to combine data at account creation.
Customers who decline remain unlinked and receive only general campaigns. Within one season, duplicate discount emails fall sharply, and the retailer avoids buying a third-party graph altogether. Its privacy team configures deletion requests to cascade through every linked identifier, and a quarterly review samples merged profiles to confirm that no two customers were joined by mistake.