What is identity resolution, and how does it work?

Identity resolution is the work of deciding whether separate records refer to the same real person or organisation, linking the ones that do, and keeping the evidence behind each decision. It compares the identifiers the records carry, such as names, dates of birth, addresses, email addresses, phone numbers, usernames and photos, and weighs each agreement by how rare the value is and whether it is independent of the others. Each pair of records ends up as a match, a non-match, or a possible match that a person reviews.

This piece covers where the method comes from, how it works, what each identifier is evidence of, and how to record a match so someone else can check it. Our guides to linking related accounts and running an OSINT investigation apply the same ideas to a platform's own data and to open sources.

Identity resolution grew out of record linkage, the work of assembling the records that belong to one person. A 2022 review by the Duke University statisticians Olivier Binette and Rebecca Steorts dates the first known paper on it to H. L. Dunn in 1946, who proposed binding each person's records together with the birth certificate number. For records with no such identifier, the review credits H. B. Newcombe and colleagues with the first automatic, computer-based method: their 1959 paper described linking 34,138 birth records from British Columbia to 114,471 marriage records. In the review's words, "No single piece of information was entirely reliable, but together they could be used for more accurate record linkage."

In 1969, Ivan Fellegi and Alan Sunter gave the idea a formal mathematical model in their paper A Theory for Record Linkage. As a US Census Bureau research report describes it, each pair of records gets a matching weight, comparing how likely its pattern of agreements and disagreements is among true matches with how likely it is among non-matches. Two thresholds then sort every pair into matches above the upper one, non-matches below the lower one, and possible matches in between, which the report says to "hold for clerical review" by a person. The thresholds come from the error rates the user will accept for false matches and false non-matches.

The steps are the same whether a person works by hand or software works across millions of records.

1. Standardise the records. Split names into parts, expand abbreviations, and write dates, phone numbers and addresses in one format. Normalise email addresses only where the provider does: Gmail delivers mail sent to any dotted version of a Gmail address to one inbox, but for Gmail used through work or school, "dots do change your address".
2. Decide which pairs to compare. Comparing every record with every other is rarely practical, so linkage compares pairs that share something, such as a surname or a date of birth, a step called blocking. The Duke review notes that blocking this way amounts to "a deterministic a priori judgment that these fields are error-free". A search on one spelling of a name does the same by hand: anyone recorded under another spelling never reaches the comparison.
3. Compare each field, allowing for near agreement. In a major census application described in the Census Bureau report, more than 25 percent of matches would not have been found by exact, character-by-character comparison.
4. Weigh each agreement by how rare the value is. In the Duke review's example, two records agreeing on the less common first name Xander are more likely to match than two agreeing only on John. Rarity is local too: the Census Bureau report notes that a surname such as Martinez is relatively rare in Minneapolis and more frequent in Los Angeles.
5. Decide: match, non-match, or possible match for a person to review, and write down why.

Deterministic matching links records by fixed rules. The simplest is exact matching, in which, as the Duke review puts it, "two record pairs are linked if they agree on all common attributes". The review notes that "the most commonly used record linkage methods are based on a series of deterministic rules", valued for being simple and easy to interpret.

Probabilistic matching weighs each agreement and disagreement and adds them up, as in the Fellegi and Sunter model, so strong evidence on several fields can outweigh a typo on one. The review cites studies in epidemiological research that "have shown consistent improvements of probabilistic methods over deterministic approaches". The two combine well: fixed rules for exclusions, such as two clearly different dates of birth, and weighing for the rest.

The terms overlap. The Census Bureau report lists entity resolution, object identification and fuzzy matching among the other names for record linkage, and notes that "An entity might be a business, a person, or some other type of unit". In this guide, and in the investigations work we do, identity resolution means the same task applied to people and organisations, where the question is who is behind a record.

Marketing uses the term differently. IAB Tech Lab, a non-profit standards consortium for digital advertising, describes an identity resolution service as one that matches multiple identifiers to provide "a unified view of an entity", using match keys such as hashed email addresses, hashed phone numbers, and names with postal addresses. That version links what advertisers and publishers know about each customer or prospect, while the investigator's version starts from one identifier and researches public sources to work out who is behind it.

Each identifier is evidence of something narrower than identity. Ask what an agreement shows, and how often it would happen between two different people.

A name on its own is weak evidence. In its 2020 Census name data, the Census Bureau counted 2,369,644 people with the last name Smith and 3,476,721 with the first name Michael, and found that just 69 first names covered a quarter of the population, as did 258 last names. The bureau publishes no counts of full names, but those figures show how quickly a common one collides. One person's name also varies: the Census Bureau report treats Bill and William as a difference that appears between true matches, and the Wolfsberg Group, an association of banks, gives Piotr Czajkowski, Pyotr Tchaikovsky and Piotr Chaykovskiy in its guidance on news screening as transcriptions of one name.

A date of birth narrows the field without settling it. The CDC's provisional count of US births in 2025 was 3,606,400, about 9,900 on an average day, so a matching date fits thousands of people. It works better as an exclusion. The Wolfsberg guidance lists differences in "key secondary identifier information", such as a date of birth, or a date of incorporation for a company, among the measures that may justify discounting a news alert as a false match, and treats profession and residence as "weaker" grounds.

An address shows that two records were connected to one place when each was written. People move: the Census Bureau's American Community Survey found that 11.8 percent of people had moved to a different residence in 2024. Households share addresses too, and the Census Bureau report gives the case of a husband and wife, two different people whose records can agree on "last name, age, house number, street name, telephone".

A personal email address is among the stronger identifiers, because one inbox usually belongs to one person and some providers never reissue an address: Google says a removed Gmail address "can't be used by anyone in the future." Families and small businesses share inboxes, though, and a work address belongs to the job as much as to the person.

A phone number is strong evidence when both records used it at about the same time. The Federal Communications Commission estimated in 2018 that approximately 35 million US numbers are disconnected and made available for reassignment each year, and Princeton researchers who studied number recycling note that "after a while, carriers reassign them to new subscribers." A number that verified one account in 2019 and another in 2026 may belong to two people, so treat the dates as part of the identifier.

A username is evidence in proportion to how unusual it is. In a 2011 study of usernames across online services, researchers at INRIA found that the chance two identical usernames belong to the same person depends strongly on how unlikely anyone else is to choose that string. In their data, 70 percent of usernames contained part of the person's real name, so a username agreement often repeats a name agreement, and most people used two or three different usernames, so a missing match proves little.

A photo shows that the same image appears in two places, which links whoever uploaded it. It does not link the person pictured, because images are copied: the FBI has warned of impersonators using "pictures stolen from open social media profiles of the impersonated individual." Treat resemblance between different photos as support for a match that rests on other evidence.

A common way to overstate a match is to count one fact several times. The Fellegi and Sunter weights add up the evidence field by field, assuming that agreement on one field says nothing about agreement on another once you know whether the pair truly matches. The Census Bureau report shows what goes wrong when that fails: agreement on a surname's phonetic code adds nothing once the surnames agree, and counting it anyway "erroneously increases the total agreement weight".

In an investigation, the same problem appears whenever several details come from one source. A profile's name, city and employer were typed by one person, so a second profile that copied them agrees three times on one claim. A registry filing and a reporter's news story are usually two sources, and agreement between them counts twice unless one was copied from the other. Agreement counts once per source, however many fields it fills. For each agreement, ask whether it would still hold if the others turned out to be a coincidence or a copy.

Records that agree too neatly deserve the same question, since a synthetic identity is built so that every piece of it checks out, as our guide to detecting synthetic identity fraud explains.

A conflict is a disagreement that two records about one person should not show. Many differences are ordinary: a nickname, a typo, a transliteration, an old address, a married name. Others are hard to explain if the records describe one person, such as dates of birth years apart, different middle names written out in full, or activity recorded after a death. Generational suffixes need care, because a father and son can share a name and some datasets drop suffixes. The Duke review shows two records of deaths in El Salvador that could describe one person, "but they could also be father and son".

Decide before you start which conflicts block a match and which only lower its weight. A hard conflict holds the pair as a non-match until something explains it, and a soft one goes in the record with the explanation you accepted. Our guide to adverse media screening applies the same thinking to whether a news story is about your person or a namesake.

A matching decision can go wrong in two ways. A false merge links records that belong to different people. A missed link leaves apart records that belong to one person. The Duke review describes Fellegi and Sunter's two error rates as the probability "of linking when the records do not match" and the probability "of not linking when the records do match", and moving the thresholds trades one against the other.

Which error costs more depends on the use. A duplicate in a mailing list costs a second letter. A false merge in a fraud case attaches one person's history to someone who did nothing, and a missed link lets a banned operator return under a new account. Set the bar for a match from that cost, before you look at the records.

Our public benchmark of researching individuals, RECON, is built around the same asymmetry. Agents start from a name and a few identifying details and must recover facts "without confusing people who share a name". A correct answer earns one point, a wrong answer loses one, and a missing answer earns zero, so a confident answer about the wrong person scores below an honest gap.

Matches are often treated as transitive. As the Duke review puts it, "In many cases, knowledge that records A and B are a match, and that B and C are a match, means that necessarily A and C are also a match." That holds when both links are right. When one is wrong, the chain joins two people who share nothing directly.

Researchers at the University of Connecticut who studied the clustering of linked records call this the chaining problem: when records are grouped whenever any two are similar enough, A and C can land in one cluster even though they are too different to match on their own. In their tests on 1,000,000 synthetic records of 100,000 people, methods prone to chaining failed to group all the records correctly for 1,981 people, against 241 for the method they designed to avoid it.

In an investigation, a chain often runs through a value that belongs to more than one person, such as a business inbox, a family phone or a brand's account. Before merging the ends of a chain, compare those two records directly and look for conflicts, and be wary of a cluster that keeps growing through one shared value.

Write each decision down so that someone who was not there can check it, or undo it. For every pair you link or decline to link, record:

• The two records, where each came from, and when.
• Each identifier compared, with its value in each record and after standardising.
• How common each agreeing value is, and how you know.
• When each record used the value, for anything that changes hands, such as a phone number.
• Which agreements depend on others, so that each counts once.
• Every conflict, with the explanation you accepted or the reason it blocked the match.
• The decision, who made it and when, and what evidence would change it.

If an identifier was missing from one record, say it was not compared, so that nobody later reads silence as agreement.

The case is a composite. Quellmore Ceramics and the people in it are invented, its domain uses the .example names reserved for documentation, and its phone number comes from the 555-0100 to 555-0199 range reserved for fictional use.

Before releasing a first large payout to a new ceramics seller, a marketplace's trust and safety team wants to confirm that the person named in the seller's payout details is the person running the shop. It has three records, two of them its own.

• The seller account, opened in March 2026: display name M. Teague, store name Quellmore Ceramics, contact email hello@quellmore.example, verified phone number +1 503 555 0147, shipping from Portland, Oregon.
• The payout details entered during the marketplace's own seller onboarding: holder name Marisol A. Teague, born in 1991, a Portland address, and the same phone number, verified the same week.
• The public social profile @quellmoreceramics: posting since 2021, a profile photo of a woman at a potter's wheel, a link to quellmore.example, and posts signed Dana.

The seller account and the payout details. The phone number was verified separately on each within days, which shows control of the same number at the same time. The names are compatible, and the city agrees, but the name could not carry a match alone: the 2020 Census counted 23,089 people with the last name Teague and 49,544 with the first name Marisol. With a strong, recent shared identifier and no conflict, the team links the two as one person.

The seller account and the social profile. The store and the profile link to each other and the handle repeats the store's name, but whoever runs the brand wrote all of that, so it counts as one fact: the profile belongs to the business. Looking for something written by someone else, the team finds a 2023 local newspaper feature about the studio. It names the two founders, Marisol Teague and Dana Quarrington, and runs the profile's photo with a caption naming the woman at the wheel as Dana.

The conflict. Chaining the two links would attach the profile and its photo to Marisol, but the posts are signed Dana and the caption names the woman in the photo as Dana. The team links the profile to Quellmore Ceramics and to Dana, and keeps it off Marisol's record. Only the payout details carry a date of birth, so the file says it was not compared.

The decision. The seller account and the payout details belong to Marisol A. Teague: the shared phone links them, and the seller's display name, M. Teague, points to her rather than to her co-founder. The payout goes ahead. The file records each link with its evidence, and two questions stay open: whether Dana has access to the seller account, and who answers hello@quellmore.example.

Each check has a known way around it.

• Copied identifiers. An impersonator who uses a real person's name and photo makes the records agree on purpose, and the matches then point at the victim. The dates on each record often expose it.
• Manufactured agreement. A synthetic identity agrees with itself because its maker kept it consistent.
• Shared values. A household, a small business or a shared inbox makes several people look like one, and a chain through the shared value merges them.
• Deliberate separation. A person who uses a different name, email address and handle on every service leaves nothing for an agreement to find, which produces missed links.
• Reassignment. A recycled phone number or a reused handle connects a new owner to an old owner's records.
• Copied errors. A wrong date or address copied from one database into several looks like corroboration. Trace each agreement back to where it was first recorded.

Most disagreements between records about one person have ordinary causes, and a fair review checks those first.

• People change their names and move house, and names are spelled differently across scripts and languages. Check the dates before treating two names or addresses as a conflict.
• Records are typed by people. In one region in the Census Bureau report, 25 percent of first names and 15 percent of last names did not agree character by character among pairs that were true matches.
• Datasets alter names. The Census Bureau's processing of 2020 Census names counts SMITHJR and SMITH III as SMITH, and found that PEÑA had been converted to PE?A, which the bureau corrected to PENA.
• A missing value is not a conflict. If one record has no date of birth, the match neither gains nor loses from it.
• Families, flatmates and business partners share phones, addresses and inboxes, so a shared value there shows a connection and leaves open which person a record belongs to.
• When the person is a customer or an applicant, asking them about a difference is often the quickest way to settle it.

Resolving an identity by hand means searching each identifier across registries, news, professional profiles and the web, and tracking which finding belongs to whom. Sixtyfour's agent starts from an identifier you already have, such as a name, an email address, a phone number, a username or a company, researches public sources, and links what it finds to the right person or company. Every finding comes with its source, and anything it could not establish is reported as open. A person on your team makes the decision. There is more on how investigators use it on our investigations page.

Identity resolution decides whether separate records belong to the same person by weighing each agreement on how rare it is and whether it is independent of the others, treating clear conflicts as blocks and holding uncertain pairs for a person to review. Check chains link by link, set the bar for a match from the cost of each kind of error, and write every decision down with its evidence so someone else can check it.