Quick answer
An identity graph stores identifiers such as visitor ids, device ids and login ids as nodes, and the rules that link them as edges. Its value depends on whether those edges can be explained: a login link can be audited, a probabilistic link can quietly merge different people.
What is an identity graph?
An identity graph is a data structure that stores every identifier you hold for people as nodes and every rule that says "these two identifiers belong to the same person" as edges. A cookie or visitor id, an app device id, a mini program OpenID, a hashed email and a login id are all nodes. A login on a device, an email typed into a checkout form, or a server event that carries both a device id and an account id are edges. A connected group of nodes is treated as one person.
That framing matters more than any vendor diagram. A list of identifiers tells you nothing. The graph's value sits in the edges, and every edge is a claim that someone, or some algorithm, made. If you can't say why an edge exists, you can't say whether your "people" count is right. This is why the thesis of this page is simple: an identity graph is only as useful as its linking rules are explainable. Deterministic edges (a login id) can be audited; probabilistic edges quietly merge different people. For analytics, prefer deterministic links.
Identity graphs are the data structure behind identity resolution, the process of deciding which records describe the same person. This page stays on the structure and the matching rules; the resolution process itself has its own entry.
Nodes and edges: what goes into the graph
| Node (identifier) | Where it comes from | How stable it is |
|---|---|---|
| Web visitor id or first-party cookie | Analytics script on your site | One browser on one device; resets when cookies are cleared or a salt rotates |
| App device id | Mobile SDK | One install; usually survives app updates, not reinstalls |
| Mini program OpenID | WeChat and similar platforms | One user inside one mini program |
| Login id (user ID) | Your own account system | Stable for as long as the account exists |
| Hashed email or phone | Forms, checkout, CRM | Stable, but people share family emails and change numbers |
| IP address, device model, time zone | Network and device signals | Weak; shared by households, offices and mobile carriers |
Edges come in two kinds:
- Deterministic edge. Both identifiers appeared together in one authenticated event: a user logged in on this device, or your server sent an event carrying both the device id and the account id. The edge is a fact you can point to.
- Probabilistic edge. A model scored two identifiers as likely to be the same person, based on shared IP address, similar device signals, overlapping times or location. The edge is a guess with a confidence score.
Deterministic vs probabilistic matching
| Question | Deterministic matching | Probabilistic matching |
|---|---|---|
| What creates the edge | The same login id, account id or verified email seen with both identifiers | A score from shared IP, device attributes, location and timing |
| Can you explain one merge | Yes: "device B logged in as u_42 on 5 Nov" | Only as a score, such as "0.83 similarity" |
| Typical error | Missed links (people who never log in) | Wrong links (two people on one Wi-Fi merged) |
| Coverage | Limited to people who identify themselves | Higher, including anonymous traffic |
| Audit trail | Each bind can be logged and undone | Hard to reproduce once the model or thresholds change |
| Privacy exposure | Uses identifiers your users gave you | Relies on device and network signals; often restricted (see below) |
| Good for | Product analytics, funnels, retention, support context | Reach estimates in advertising, where some error is tolerated |
The practical rule follows from the error column. A deterministic graph is incomplete but correct: it undercounts merges, so your person count is a little too high. A probabilistic graph is complete but unverifiable: it overcounts merges, and you can't tell which ones are wrong without auditing them one by one. For analytics that drive product decisions, a known undercount beats an unknown mix of errors.
Worked example: three devices and one login
Sample data: one shopper uses a laptop browser, an iPhone app and an Android tablet over a week, and a second browser shows up on the same home network.
| Identifier | Type | First seen | Edge to login u_42 |
|---|---|---|---|
| w-7f3 | Web visitor (laptop browser) | 2 Nov | Deterministic: logged in on 4 Nov |
| i-91c | App device (iPhone) | 3 Nov | Deterministic: logged in on 3 Nov |
| a-22d | App device (Android tablet) | 6 Nov | Deterministic: logged in on 8 Nov |
| w-c10 | Web visitor (phone browser) | 7 Nov | None. Probabilistic only: same home IP and city as w-7f3 |
Counting people:
- No graph: 4 identifiers = 4 "people".
- Deterministic graph: w-7f3, i-91c and a-22d connect through u_42, so they form 1 person; w-c10 stays anonymous. Total = 2 people.
- Probabilistic graph: a rule "same IP and same city within 24 hours" adds an edge from w-c10 to w-7f3. Total = 1 person.
Is w-c10 the same shopper on a phone, or a partner on the same Wi-Fi? The deterministic graph admits it doesn't know. The probabilistic graph has already decided, and every metric downstream inherits that decision.
Now scale it up. Sample data for one month:
- 10,000 device and visitor ids are seen.
- 2,400 of those ids logged in at least once, and they belong to 1,500 login ids.
- Deterministic person count = 10,000 − 2,400 + 1,500 = 9,100.
- A probabilistic rule proposes merging another 800 anonymous ids into existing people, which would give 9,100 − 800 = 8,300.
- An analyst audits 50 of the proposed merges by hand and finds 14 wrong: 14 ÷ 50 = 28%.
- Expected wrong merges = 800 × 0.28 = 224, or 224 ÷ 8,300 = 2.7% of the final people.
That 2.7% isn't spread evenly. Shared networks cluster in households, offices and campuses, so the wrong merges concentrate in exactly the segments you might want to compare. And unlike the deterministic undercount, nothing in the dashboard tells you which 224 people are fused.
What wrong merges cost
- Inflated per-person metrics. Two people fused into one doubles that "person's" sessions and orders, which pushes average order count and lifetime value up.
- Broken funnels and retention. A person who signed up on Tuesday inherits another person's visits from last month, so time-to-convert and returning-user rates shift.
- Wrong attribution. One merged person's purchase gets credited to the other person's ad click.
- Support and privacy mistakes. An agent sees someone else's browsing history in a support chat, or a deletion request removes the wrong records.
- Hard reversal. Once merged events have been rolled up into daily reports, splitting the person again means recomputing history.
A missed link costs less: the same person counts twice until they log in on the second device. That error shrinks as login coverage grows, and it never attaches one person's data to another.
Identity graphs in marketing
"Identity graph marketing" usually means a third-party graph: a data company links cookies, mobile advertising ids, hashed emails and household data across many brands so advertisers can target or measure the same person on different sites. That is a different object from the first-party graph an analytics tool builds from your own logins. Cross-company graphs rely on the techniques that browsers and platforms now limit, covered in cross-site tracking. Pulling first-party data into one profile per customer is the job of a customer data platform; a CDP's identity graph is usually the first-party, mostly deterministic kind.
Privacy boundaries
- Identifiers are personal data. Recital 30 of the GDPR notes that people may be associated with online identifiers from their devices, such as IP addresses and cookie identifiers, and that these can be combined to create profiles. A graph that joins them is processing personal data, with the obligations that come with it.
- Device graphs and fingerprints count as tracking on Apple platforms. Apple's user privacy guidance lists "device graph identifiers" and fingerprint ids among identifiers that need App Tracking Transparency permission when used to join your app's data with third-party data, and it forbids deriving device data to uniquely identify a device.
- Keep the graph inside your own data. Joining your own users' logins across your own website and apps is a different risk from joining them with other companies' data.
Identity graphs in TapCub
Take the worked example above. In TapCub the three devices join only when the shopper logs in:
- Send the login. Call
tapcub.login('u_42')after sign-in on the website, and the sameloginfrom the app SDKs. A mini program calls it after it has an OpenID; see identity on mini programs. Whether identification runs depends on the project's identity and consent settings in Configuration › Privacy center. - Watch the edges form. The first time a device binds to a login id, that session and the device's earlier anonymous events receive the user id. If the device later logs in as someone else, rows already stored stay with the old person, the device mapping moves, and a new session starts. Values such as
null,guest,anonymous, all zeros and-1are rejected, so a bad placeholder can't merge thousands of devices into one person. - Audit every merge. All-platform analytics › Users › Identity map logs each bind, backfill, account switch, split and rejection, and a card compares visitor, device and person counts so you can see how much duplication identity removed. Owners and admins can split a device that was bound wrongly. A person keeps up to 50 devices by default; past that, the oldest is unbound and the log records it.
- Read the person. The shopper appears in All-platform analytics › Users › People with a day-by-day timeline across web and app.
Every edge in this graph comes from an identifier you sent: login, zero-code login sources or server events. The cookie-free visitor id is a hash of address and browser data used to count visitors; it never creates a person on its own. TapCub does not offer probabilistic household or fingerprint matching, so w-c10 in the example stays anonymous until it logs in.
See also: What is cross-device tracking? How logins connect devices.
See also: What is entity resolution? Methods, matching and examples.
Sources
- Apple Developer, User privacy and data use.
- EUR-Lex, Regulation (EU) 2016/679 (GDPR), Recital 30.
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.

Technical Engineer
Focuses on tracking code, SDKs, and API integrations, with practical guidance on implementation checks and troubleshooting.